64 changed files with 538 additions and 3415 deletions
--- a/.env.example
+++ b/.env.example
@ -45,8 +45,7 @@ FORGEJO_API_URL=https://git.opensourcesolarpunk.com/api/v1
 # Set CF_LICENSE_KEY to authenticate with the hosted coordinator.
 # Leave both blank for local self-hosted cf-orch or bare-metal inference.
 CF_LICENSE_KEY=
-GPU_SERVER_URL=https://orch.circuitforge.tech
+CF_ORCH_URL=https://orch.circuitforge.tech
 # CF_ORCH_URL is also accepted as a backward-compat alias for GPU_SERVER_URL
 # cf-orch agent — GPU profiles only (single-gpu, dual-gpu-*)
 # The agent registers this node with the cf-orch coordinator and reports VRAM stats.
--- a/.forgejo/workflows/ci.yml
+++ b/.forgejo/workflows/ci.yml
@ -23,9 +23,6 @@ jobs:
          python-version: '3.12'
          cache: pip
      - name: Install system dependencies
        run: sudo apt-get update -q && sudo apt-get install -y libsqlcipher-dev
      - name: Install dependencies
        run: pip install -r requirements.txt
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@ -9,109 +9,6 @@ Format follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 ---
 ## [0.9.5] — 2026-05-08
 ### Fixed
 - **Theme: dark/explicit themes show correct page background** — `index.html` inline style
  set `html, body { background: #eaeff8 }` hardcoded. `body` paints on top of `html`, so
  even when `html { background: var(--color-surface) }` correctly resolved to `#16202e` in
  dark mode, the body's hardcoded light background covered it — producing dark cards on a
  light page. Fixed by: (1) removing body background from the inline style; (2) adding a
  tiny blocking `<script>` that reads `cf-theme` / `cf-hacker-mode` from localStorage and
  sets `data-theme` on `<html>` before first paint; (3) adding
  `html[data-theme="dark"|"solarized-dark"|"hacker"]` rules so FOUT prevention fires the
  right background immediately on load.
 ---
 ## [0.9.4] — 2026-05-08
 ### Added
 - **Messages view — expandable email timeline** — click any email item to lazy-load
  and read the full body inline (HTML stripped to plain text via `DOMParser`).
  Bodies are fetched on-demand via the new `GET /api/contacts/{id}` endpoint to avoid
  loading 50KB+ email bodies on every page view.
 - **Messages view — compose bar** — action buttons (Log call, Log note, Use template,
  Draft reply with LLM, Call via Osprey) moved from the always-visible header into a
  sticky bottom compose bar triggered by a ＋ New toggle. Reduces visual clutter when
  just reading the thread.
 - **Home view — "Skip review" checkbox** — when adding jobs by URL, a checkbox (default
  on) sends them directly to the Apply queue, bypassing Job Review.
 - **ContactsView — sync status** — shows last completed sync time and a spinner when
  an email sync is running.
 - **imap_sync: Indeed alert parser** — `parse_indeed_alert()` extracts job title,
  company, location, salary, and canonical URL from Indeed Job Alert digest emails.
 - **scrape_url: Oracle HCM support** — Playwright-based scraper for Oracle HCM
  CandidateExperience portals (React SPAs requiring JS execution).
 - **manage.sh** — compose engine auto-detection (docker compose / podman compose /
  podman-compose), `build` command, and cloud/demo stack shortcuts.
 - **theme.css** — `--color-overlay` token for modal/dialog backdrops.
 ### Fixed
 - **Messages view layout** — changed `height: 100%` to `height: 100dvh` with a mobile
  override for the 56px tab bar. `height: 100%` was resolving to "shrink-wrap" because
  `.app-main` has no explicit height; compose bar is now correctly pinned to the bottom.
 - **Accessibility: danger button contrast** — `btn--danger` used `color: white` on
  `--app-accent` (Talon Orange), yielding 2.8:1 contrast (fails WCAG AA 4.5:1 for
  normal text). Fixed to `color: var(--app-accent-text)` (dark navy, 5.5:1).
 - **Accessibility: warning badge contrast** — `tab-badge` in Job Review used `color: white`
  on `--color-warning` (amber). Same fix applied.
 - **Theme: Interviews signal banners** — hardcoded `rgba(245,158,11,…)` / `rgba(39,174,…)`
  / `rgba(192,57,…)` replaced with `color-mix()` against `--color-warning/success/error`.
 - **Theme: Interviews signal count** — `color: #e67e22` hardcode replaced with
  `var(--app-accent)`.
 - **Theme: References academic tag chip** — `color: #7c3aed` hardcode replaced with
  `var(--status-synced)`; background uses `color-mix()` with the same token.
 - **Theme: Interviews signal-move button** — `color: #fff` on `--color-primary` fails
  in dark mode (light green bg); fixed to `var(--color-text-inverse)`.
 - **Modal backdrops** — `rgba(0,0,0,0.5)` replaced with `var(--color-overlay)` for
  theme consistency.
 ---
 ## [0.9.3] — 2026-05-05
 ### Added
 - **Editable resume review** — proposed summary and experience bullets in the review modal
  are now editable text areas. Edits flow through `apply_review_decisions()` and override
  the LLM output in the final resume struct. Preview textarea in Apply Workspace is also
  editable, with manual changes preserved through the approve step via `preview_text_override`.
 ### Fixed
 - **Double bullets in resume optimizer** — `_section_text_for_prompt` now strips existing
  bullet characters before prefixing with `•`, and `_reparse_experience_bullets` uses a
  greedy strip regex so `• •` patterns can no longer survive parsing.
 - **Asterisk markup in summary** — added `_clean_summary_markup()` to strip LLM-generated
  markdown bullet chars (`*`, `-`, etc.) from career summary output; injected no-markdown
  rule into the LLM prompt's CRITICAL RULES list.
 - **Light theme dark CSS bleed** — `peregrine.css` media dark override now scoped to
  `:root:not([data-theme])` (auto mode only) instead of `:root:not([data-theme="hacker"])`.
  Fixes dark navy `--app-primary-light`/`--app-accent-light` bleeding into light themes
  (light, solarized-light, colorblind) on dark-OS machines.
 ---
 ## [0.9.2] — 2026-05-02
 ### Added
 - **Cover letter training export** (#111) — opt-in consent gate (`training_export_opt_in`
  in `user.yaml`, default off) lets users export applied-job cover letters as Alpaca-format
  JSONL for local fine-tuning. Per-job exclude/restore curation in Settings → Fine-Tune.
  Streaming JSONL download merges DB pairs with any previously uploaded file pairs.
  Cloud fine-tune Phase 2 stub (501) reserved for cf-orch integration.
 - **WizardTrainingStep** — new onboarding consent step inserted between Resume and Identity;
  skippable, opt-in default off, cloud-aware privacy copy.
 - **a11y:** confirmed-state toggle (no optimistic DOM divergence), visible Premium tier gate
  with upgrade link, `aria-live` region on pairs list, cloud-aware consent copy.
 ---
 ## [0.9.0] — 2026-04-20
 ### Added
--- a/README.md
+++ b/README.md
@ -1,143 +1,213 @@
-<div align="center">
+# Peregrine
  <img src="web/public/peregrine.svg" alt="Peregrine" width="120" />
-  <h1>Peregrine</h1>
+> **Primary development** happens at [git.opensourcesolarpunk.com](https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine) — GitHub and Codeberg are push mirrors. Issues and PRs are welcome on either platform.
-  <p><strong>Job search pipeline — by <a href="https://circuitforge.tech">Circuit Forge LLC</a></strong></p>
+[![License: BSL 1.1](https://img.shields.io/badge/License-BSL_1.1-blue.svg)](./LICENSE-BSL)
-
+[![CI](https://github.com/CircuitForge/peregrine/actions/workflows/ci.yml/badge.svg)](https://github.com/CircuitForge/peregrine/actions/workflows/ci.yml)
  <p><em>AI for the tasks the system made hard on purpose.</em></p>
  [![License: MIT / BSL 1.1](https://img.shields.io/badge/License-MIT%20%2F%20BSL%201.1-blue.svg)](#license)
  [![CI](https://github.com/CircuitForgeLLC/peregrine/actions/workflows/ci.yml/badge.svg)](https://github.com/CircuitForgeLLC/peregrine/actions/workflows/ci.yml)
 [![Docs](https://img.shields.io/badge/docs-docs.circuitforge.tech-orange)](https://docs.circuitforge.tech/peregrine/)
  [![Version](https://img.shields.io/badge/version-0.9.0-green)](https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine/releases)
-  <p>
+**Job search pipeline — by [Circuit Forge LLC](https://circuitforge.tech)**
    <a href="https://demo.circuitforge.tech/peregrine"><strong>Live Demo</strong></a> —
    no account required, nothing saved &nbsp;|&nbsp;
    <a href="https://docs.circuitforge.tech/peregrine/">Docs</a> &nbsp;|&nbsp;
    <a href="https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine/issues">Issues</a>
  </p>
-  <blockquote>
+> *"Tools for the jobs that the system made hard on purpose."*
-    <strong>Primary development</strong> happens at
+
-    <a href="https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine">git.opensourcesolarpunk.com/Circuit-Forge/peregrine</a>.
+**[Try the live demo](https://demo.circuitforge.tech/peregrine)** — no account required, nothing saved.
    GitHub and Codeberg are push mirrors. Issues and PRs are welcome on any platform.
  </blockquote>
 </div>
 ---
 ![Job review — swipe right to approve, left to skip](docs/screenshots/02-review-swipe.gif)
 <table>
 <tr>
-<td><img src="docs/screenshots/01-dashboard.png" alt="Dashboard with pipeline stats and discovery controls"/></td>
+<td><img src="docs/screenshots/01-dashboard.png" alt="Dashboard with pipeline stats"/></td>
-<td><img src="docs/screenshots/02-review.png" alt="Job review — approve, skip, or reject with keyboard shortcuts"/></td>
+<td><img src="docs/screenshots/04-interviews.png" alt="Interview kanban with recruiter emails attached"/></td>
 </tr>
 <tr>
-<td><img src="docs/screenshots/03-apply.png" alt="Apply workspace with LLM-drafted cover letter"/></td>
+<td><img src="docs/screenshots/03-apply.png" alt="Apply workspace with AI cover letter draft"/></td>
-<td><img src="docs/screenshots/04-interviews.png" alt="Interview kanban with company research and recruiter emails"/></td>
+<td><img src="docs/screenshots/02-review.png" alt="Job review card with match score and ghost-post detection"/></td>
 </tr>
 </table>
 ---
-## Why Peregrine?
+Job search is a second job nobody hired you for.
-Job search is a second job nobody hired you for. ATS (applicant tracking system) filters designed to reject. Boards that show the same listing eight times. Cover letter number forty-seven for a role that might already be filled. Hours of prep for a phone screen that lasts twelve minutes.
+ATS filters designed to reject. Job boards that show the same listing eight times. Cover letter number forty-seven for a role that might already be filled. Hours of prep for a phone screen that lasts twelve minutes.
- **Handles the full pipeline.** Discover, filter, match, draft, track — one tool, one database, no duct tape.
+Peregrine handles the pipeline — discovery, matching, tracking, drafting, and prep — so you can spend your time doing the work you actually want to be doing.
- **LLM is optional and local-first.** Discovery and tracking work with no LLM at all. When you do configure one, it runs on your hardware by default. Cloud inference is a fallback, not the default path.
+
- **Ghost-post detection baked in.** Listings that have been open too long or look like sourcing traps get flagged before you spend time on them.
+**LLM support is optional.** The full discovery and tracking pipeline works without one. When you do configure a backend, the LLM drafts the parts that are genuinely miserable — cover letters, company research briefs, interview prep sheets — and waits for your approval before anything goes anywhere.
- **Human approval at every step.** LLM drafts cover letters and research briefs; you approve before anything goes anywhere. Peregrine never submits an application on your behalf.
+
- **Privacy · Safety · Accessibility** are architectural constraints, not aspirational copy. No PII (personally identifiable information) logging, no behavioral profiling, no dark patterns.
+### What Peregrine does not do
 Peregrine does **not** submit job applications for you. You still have to go to each employer's site and click apply yourself.
 This is intentional. Automated mass-applying is a bad experience for everyone — it's also a trust violation with employers who took the time to post a real role. Peregrine is a preparation and organization tool, not a bot.
 What it *does* cover is everything before and after that click: finding the jobs, matching them against your resume, generating cover letters and prep materials, and once you've applied — tracking where you stand, classifying the emails that come back, and surfacing company research when an interview lands on your calendar. The submit button is yours. The rest of the grind is ours.
 > **Exception:** [AIHawk](https://github.com/nicolomantini/LinkedIn-Easy-Apply) is a separate, optional tool that handles LinkedIn Easy Apply automation. Peregrine integrates with it for AIHawk-compatible profiles, but it is not part of Peregrine's core pipeline.
 ---
 ## Quick Start
-One-line install:
+**1. Clone and install dependencies** (Docker, NVIDIA toolkit if needed):
 ```bash
 bash <(curl -fsSL https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine/raw/branch/main/install.sh)
 ```
 Or clone and run manually:
 ```bash
 git clone https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine
 cd peregrine
 ./manage.sh setup
 ./manage.sh start
 ```
-Open **http://localhost:8502** — the setup wizard walks you through the rest.
+**2. Start Peregrine:**
 > **macOS / Apple Silicon:** install Ollama natively via Homebrew before starting for Metal GPU-accelerated inference. `install.sh` handles this automatically.
 > **Windows:** use WSL2 with Ubuntu.
 ### Inference profiles
 ```bash
-./manage.sh start                       # remote — no GPU; LLM calls go to Anthropic / OpenAI
+./manage.sh start                          # remote profile (API-only, no GPU)
-./manage.sh start --profile cpu         # local Ollama on CPU (or Metal via native Ollama on macOS)
+./manage.sh start --profile cpu            # local Ollama (CPU, or Metal GPU on Apple Silicon — see below)
-./manage.sh start --profile single-gpu  # Ollama + vision on GPU 0 (NVIDIA only)
+./manage.sh start --profile single-gpu    # Ollama + Vision on GPU 0  (NVIDIA only)
-./manage.sh start --profile dual-gpu    # Ollama + vLLM on two NVIDIA GPUs
+./manage.sh start --profile dual-gpu      # Ollama + Vision + vLLM (GPU 0 + 1)  (NVIDIA only)
 ```
 Or use `make` directly:
 ```bash
 make start                        # remote profile
 make start PROFILE=single-gpu
 ```
 **3.** Open http://localhost:8501 — the setup wizard guides you through the rest.
 > **macOS / Apple Silicon:** Docker Desktop must be running. For Metal GPU-accelerated inference, install Ollama natively before starting — `install.sh` will prompt you to do this. See [Apple Silicon GPU](#apple-silicon-gpu) below.
 > **Windows:** Not supported — use WSL2 with Ubuntu.
 ### Installing to `/opt` or other system directories
 If you clone into a root-owned directory (e.g. `sudo git clone ... /opt/peregrine`), two things need fixing:
 **1. Git ownership warning** (`fatal: detected dubious ownership`) — `./manage.sh setup` fixes this automatically. If you need git to work *before* running setup:
 ```bash
 git config --global --add safe.directory /opt/peregrine
 ```
 **2. Preflight write access** — preflight writes `.env` and `compose.override.yml` into the repo directory. Fix ownership once:
 ```bash
 sudo chown -R $USER:$USER /opt/peregrine
 ```
 After that, run everything without `sudo`.
 ### Podman
 Podman is rootless by default — **no `sudo` needed.** `./manage.sh setup` will configure `podman-compose` if it isn't already present.
 ### Docker
 After `./manage.sh setup`, log out and back in for docker group membership to take effect. Until then, prefix commands with `sudo`. After re-login, `sudo` is no longer required.
 ---
 ## Inference Profiles
 | Profile | Services started | Use case |
 |---------|-----------------|----------|
 | `remote` | app + searxng | No GPU; LLM calls go to Anthropic / OpenAI |
 | `cpu` | app + ollama + searxng | No GPU; local models on CPU. On Apple Silicon, use with native Ollama for Metal acceleration — see below. |
 | `single-gpu` | app + ollama + vision + searxng | One **NVIDIA** GPU: cover letters, research, vision |
 | `dual-gpu` | app + ollama + vllm + vision + searxng | Two **NVIDIA** GPUs: GPU 0 = Ollama, GPU 1 = vLLM |
 ### Apple Silicon GPU
 Docker Desktop on macOS runs in a Linux VM — it cannot access the Apple GPU. Metal-accelerated inference requires Ollama to run **natively** on the host.
 `install.sh` handles this automatically: it offers to install Ollama via Homebrew, starts it as a background service, and explains what happens next. If Ollama is running on port 11434 when you start Peregrine, preflight detects it, stubs out the Docker Ollama container, and routes inference through the native process — which uses Metal automatically.
 To do it manually:
 ```bash
 brew install ollama
 brew services start ollama          # starts at login, uses Metal GPU
 ./manage.sh start --profile cpu     # preflight adopts native Ollama; Docker container is skipped
 ```
 The `cpu` profile label is a slight misnomer in this context — Ollama will be running on the GPU. `single-gpu` and `dual-gpu` profiles are NVIDIA-specific and not applicable on Mac.
 ---
 ## First-Run Wizard
 On first launch the setup wizard walks through seven steps:
 1. **Hardware** — detects NVIDIA GPUs (Linux) or Apple Silicon GPU (macOS) and recommends a profile
 2. **Tier** — choose free, paid, or premium (or use `dev_tier_override` for local testing)
 3. **Identity** — name, email, phone, LinkedIn, career summary
 4. **Resume** — upload a PDF/DOCX for LLM parsing, or use the guided form builder
 5. **Inference** — configure LLM backends and API keys
 6. **Search** — job titles, locations, boards, keywords, blocklist
 7. **Integrations** — optional cloud storage, calendar, and notification services
 Wizard state is saved after each step — a crash or browser close resumes where you left off.
 Re-enter the wizard any time via **Settings → Developer → Reset wizard**.
 ---
 ## Features
 | Feature | Tier |
 |---------|------|
-| Job discovery — LinkedIn, Indeed, Glassdoor, Adzuna, The Ladders | Free |
+| Job discovery (JobSpy + custom boards) | Free |
-| Ghost-post detection | Free |
+| Resume keyword matching & gap analysis | Free |
-| Resume keyword matching and gap analysis | Free |
+| Document storage sync (Google Drive, Dropbox, OneDrive, MEGA, Nextcloud) | Free |
 | Document storage sync (Google Drive, Dropbox, OneDrive, Nextcloud) | Free |
 | Webhook notifications (Discord, Home Assistant) | Free |
 | Vue 3 SPA — full UI with onboarding wizard, job board, apply workspace, interview kanban | Free |
 | **Cover letter generation** | Free with LLM¹ |
 | **Company research briefs** | Free with LLM¹ |
-| **Interview prep and practice Q&A** | Free with LLM ¹ |
+| **Interview prep & practice Q&A** | Free with LLM¹ |
 | **Survey assistant** (culture-fit Q&A, screenshot analysis) | Free with LLM¹ |
 | **Wizard helpers** (career summary, bullet expansion, skill suggestions, job title suggestions, mission notes) | Free with LLM¹ |
 | Managed cloud LLM (no API key needed) | Paid |
-| Email sync and auto-classification | Paid |
+| Email sync & auto-classification | Paid |
 | LLM-powered keyword blocklist | Paid |
 | Job tracking integrations (Notion, Airtable, Google Sheets) | Paid |
 | Calendar sync (Google, Apple) | Paid |
 | Slack notifications | Paid |
 | CircuitForge shared cover-letter model | Paid |
-| **Voice guidelines** (custom writing style and tone) | Premium with LLM ¹ |
+| Vue 3 SPA — full UI with onboarding wizard, job board, apply workspace, sort/filter, research modal, draft cover letter | Free |
-| Cover letter model fine-tuning — your writing, your model | Premium |
+| **Voice guidelines** (custom writing style & tone) | Premium with LLM¹ ² |
 | Cover letter model fine-tuning (your writing, your model) | Premium |
 | Multi-user support | Premium |
 | Human-in-the-loop operator (CAPTCHAs, phone calls, wet signatures) | Ultra |
-¹ **BYOK (bring your own key) unlock:** configure any LLM backend — a local [Ollama](https://ollama.com) or vLLM instance, or your own API key (Anthropic, OpenAI-compatible) — and all "Free with LLM" and "Premium with LLM" features unlock at no charge.
+¹ **BYOK (bring your own key/backend) unlock:** configure any LLM backend — a local [Ollama](https://ollama.com) or vLLM instance,
 or your own API key (Anthropic, OpenAI-compatible) — and all features marked **Free with LLM** or **Premium with LLM**
 unlock at no charge. The paid tier earns its price by providing managed cloud inference so you
 don't need a key at all, plus integrations and email sync.
 ² **Voice guidelines** requires Premium tier without a configured LLM backend. With BYOK, it unlocks at any tier.
 ---
-## What Peregrine does not do
+## Email Sync
-Peregrine does **not** submit job applications for you. You still click apply on the employer's site.
+Monitors your inbox for job-related emails and automatically updates job stages (interview requests, rejections, survey links, offers).
-This is intentional. Automated mass-applying is a bad experience for everyone and a trust violation with employers who posted a real role. The submit button is yours. The rest of the grind is ours.
+Configure in **Settings → Email**. Requires IMAP access and, for Gmail, an App Password.
 ---
-## Stack
+## Integrations
-| Layer | Technology |
+Connect external services in **Settings → Integrations**:
-|-------|-----------|
+
-| Frontend | Vue 3 SPA (Vite) |
+- **Job tracking:** Notion, Airtable, Google Sheets
-| Backend | FastAPI + Python |
+- **Document storage:** Google Drive, Dropbox, OneDrive, MEGA, Nextcloud
-| Database | SQLite (local, per-user) |
+- **Calendar:** Google Calendar, Apple Calendar (CalDAV)
-| Job scraping | [JobSpy](https://github.com/Bunsly/JobSpy) + custom board scrapers |
+- **Notifications:** Slack, Discord (webhook), Home Assistant
 | LLM inference | Ollama, vLLM, Anthropic, OpenAI-compatible — configurable fallback chain |
 | Vision | moondream2 (survey screenshot analysis) |
 | Container | Docker / Podman |
 ---
-## manage.sh reference
+## CLI Reference (`manage.sh`)
 `manage.sh` is the single entry point for all common operations — no need to remember Make targets or Docker commands.
 ```
 ./manage.sh setup               Install Docker/Podman + NVIDIA toolkit
@ -146,38 +216,31 @@ This is intentional. Automated mass-applying is a bad experience for everyone an
 ./manage.sh restart             Restart all services
 ./manage.sh status              Show running containers
 ./manage.sh logs [service]      Tail logs (default: app)
-./manage.sh update              Pull latest images and rebuild app container
+./manage.sh update              Pull latest images + rebuild app container
 ./manage.sh preflight           Check ports + resources; write .env
 ./manage.sh test                Run test suite
-./manage.sh prepare-training    Scan docs for cover letters — outputs training JSONL
+./manage.sh prepare-training    Scan docs for cover letters → training JSONL
-./manage.sh finetune            Run LoRA fine-tune (requires single-gpu profile or higher)
+./manage.sh finetune            Run LoRA fine-tune (needs --profile single-gpu+)
 ./manage.sh open                Open the web UI in your browser
 ./manage.sh clean               Remove containers, images, volumes (asks to confirm)
 ```
 ---
-## Documentation
+## Developer Docs
-Full docs at **[docs.circuitforge.tech/peregrine](https://docs.circuitforge.tech/peregrine)**
+Full documentation at: https://docs.circuitforge.tech/peregrine
-Bug reports and feature requests: [Forgejo issues](https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine/issues)
+- [Installation guide](https://docs.circuitforge.tech/peregrine/getting-started/installation/)
-
+- [Adding a custom job board scraper](https://docs.circuitforge.tech/peregrine/developer-guide/adding-scrapers/)
---
+- [Adding an integration](https://docs.circuitforge.tech/peregrine/developer-guide/adding-integrations/)
-
+- [Contributing](https://docs.circuitforge.tech/peregrine/developer-guide/contributing/)
 ## Contributing
 Contributions are welcome. The discovery pipeline — scrapers, board integrations, matching logic — is MIT-licensed. Fork it, extend it, send PRs. AI features are BSL 1.1. See the [contributing guide](https://docs.circuitforge.tech/peregrine/developer-guide/contributing/) for conventions.
 ---
 ## License
-Peregrine uses a split license:
+Core discovery pipeline: [MIT](LICENSE-MIT)
-
+LLM features (cover letter generation, company research, interview prep, UI): [BSL 1.1](LICENSE-BSL)
 | Component | License |
 |-----------|---------|
 | Discovery pipeline — scrapers, matching, tracking | [MIT](LICENSE-MIT) |
 | LLM features — cover letter generation, company research, interview prep, survey assistant, fine-tuning | [BSL 1.1](LICENSE-BSL) — free for personal non-commercial self-hosting; commercial use or SaaS re-hosting requires a paid license; converts to MIT after four years |
 Fine-tuned model weights are proprietary and per-user — not redistributable.
 © 2026 Circuit Forge LLC
--- a/compose.cloud.yml
+++ b/compose.cloud.yml
@ -6,15 +6,46 @@
 # Caddy injects the Directus session cookie as X-CF-Session header before forwarding.
 # cloud_session.py resolves user_id → per-user db_path at session init.
 #
 # Services: api (FastAPI :8601), web (Vue :8508), searxng (internal)
 # Streamlit app service removed — Vue+FastAPI is the only frontend (peregrine#104).
 #
 # Usage:
 #   docker compose -f compose.cloud.yml --project-name peregrine-cloud up -d
 #   docker compose -f compose.cloud.yml --project-name peregrine-cloud down
-#   docker compose -f compose.cloud.yml --project-name peregrine-cloud logs api -f
+#   docker compose -f compose.cloud.yml --project-name peregrine-cloud logs app -f
 services:
  app:
    build:
      context: ..
      dockerfile: peregrine/Dockerfile.cfcore
    container_name: peregrine-cloud
    ports:
      - "8505:8501"
    volumes:
      - /devl/menagerie-data:/devl/menagerie-data  # per-user data trees
      - ./config/llm.cloud.yaml:/app/config/llm.yaml:ro  # cloud-safe backends only (no claude_code/copilot/anthropic)
    environment:
      - CLOUD_MODE=true
      - CLOUD_DATA_ROOT=/devl/menagerie-data
      - DIRECTUS_JWT_SECRET=${DIRECTUS_JWT_SECRET}
      - CF_SERVER_SECRET=${CF_SERVER_SECRET}
      - PLATFORM_DB_URL=${PLATFORM_DB_URL}
      - HEIMDALL_URL=${HEIMDALL_URL:-http://cf-license:8000}
      - HEIMDALL_ADMIN_TOKEN=${HEIMDALL_ADMIN_TOKEN}
      - STAGING_DB=/devl/menagerie-data/cloud-default.db  # fallback only — never used
      - DOCS_DIR=/tmp/cloud-docs
      - STREAMLIT_SERVER_BASE_URL_PATH=peregrine
      - PYTHONUNBUFFERED=1
      - PEREGRINE_CADDY_PROXY=1
      - CF_ORCH_URL=http://host.docker.internal:7700
      - CF_APP_NAME=peregrine
      - DEMO_MODE=false
      - FORGEJO_API_TOKEN=${FORGEJO_API_TOKEN:-}
    depends_on:
      searxng:
        condition: service_healthy
    extra_hosts:
      - "host.docker.internal:host-gateway"
    restart: unless-stopped
  api:
    build:
      context: ..
@ -37,8 +68,7 @@ services:
      - HEIMDALL_ADMIN_TOKEN=${HEIMDALL_ADMIN_TOKEN}
      - PYTHONUNBUFFERED=1
      - FORGEJO_API_TOKEN=${FORGEJO_API_TOKEN:-}
-      - GPU_SERVER_URL=${GPU_SERVER_URL:-http://host.docker.internal:7700}
+      - CF_ORCH_URL=http://host.docker.internal:7700
      - CF_ORCH_URL=${CF_ORCH_URL:-${GPU_SERVER_URL:-http://host.docker.internal:7700}}
      - CF_APP_NAME=peregrine
    extra_hosts:
      - "host.docker.internal:host-gateway"
--- a/compose.test-cfcore.yml
+++ b/compose.test-cfcore.yml
@ -29,8 +29,7 @@ services:
      - STAGING_DB=/devl/job-seeker/staging.db
      - PYTHONUNBUFFERED=1
      - STREAMLIT_SERVER_BASE_URL_PATH=
-      - GPU_SERVER_URL=${GPU_SERVER_URL:-http://host.docker.internal:7700}
+      - CF_ORCH_URL=http://host.docker.internal:7700
      - CF_ORCH_URL=${CF_ORCH_URL:-${GPU_SERVER_URL:-http://host.docker.internal:7700}}
    extra_hosts:
      - "host.docker.internal:host-gateway"
    restart: "no"
--- a/compose.yml
+++ b/compose.yml
@ -1,7 +1,48 @@
 # compose.yml — Peregrine by Circuit Forge LLC
-# Streamlit (app service) removed — Vue+FastAPI is the only frontend (#104)
+# Profiles: remote | cpu | single-gpu | dual-gpu-ollama
 services:
  app:
    build:
      context: ..
      dockerfile: peregrine/Dockerfile.cfcore
    command: >
      bash -c "streamlit run app/app.py
      --server.port=8501
      --server.headless=true
      --server.fileWatcherType=none
      2>&1 | tee /app/data/.streamlit.log"
    ports:
      - "${STREAMLIT_PORT:-8501}:8501"
    volumes:
      - ./config:/app/config
      - ./data:/app/data
      - ${DOCS_DIR:-~/Documents/JobSearch}:/docs
      - /var/run/docker.sock:/var/run/docker.sock
      - /usr/bin/docker:/usr/bin/docker:ro
    environment:
      - STAGING_DB=/app/data/staging.db
      - DOCS_DIR=/docs
      - ANTHROPIC_API_KEY=${ANTHROPIC_API_KEY:-}
      - OPENAI_COMPAT_URL=${OPENAI_COMPAT_URL:-}
      - OPENAI_COMPAT_KEY=${OPENAI_COMPAT_KEY:-}
      - PEREGRINE_GPU_COUNT=${PEREGRINE_GPU_COUNT:-0}
      - PEREGRINE_GPU_NAMES=${PEREGRINE_GPU_NAMES:-}
      - RECOMMENDED_PROFILE=${RECOMMENDED_PROFILE:-remote}
      - STREAMLIT_SERVER_BASE_URL_PATH=${STREAMLIT_BASE_URL_PATH:-}
      - FORGEJO_API_TOKEN=${FORGEJO_API_TOKEN:-}
      - FORGEJO_REPO=${FORGEJO_REPO:-}
      - FORGEJO_API_URL=${FORGEJO_API_URL:-}
      - PYTHONUNBUFFERED=1
      - PYTHONLOGGING=WARNING
      - PEREGRINE_CADDY_PROXY=1
    depends_on:
      searxng:
        condition: service_healthy
    extra_hosts:
      - "host.docker.internal:host-gateway"
    restart: unless-stopped
  api:
    build:
      context: ..
@ -20,9 +61,7 @@ services:
      - OPENAI_COMPAT_KEY=${OPENAI_COMPAT_KEY:-}
      - PEREGRINE_GPU_COUNT=${PEREGRINE_GPU_COUNT:-0}
      - PEREGRINE_GPU_NAMES=${PEREGRINE_GPU_NAMES:-}
-      - GPU_SERVER_URL=${GPU_SERVER_URL:-${CF_ORCH_URL:-http://host.docker.internal:7700}}
+      - CF_ORCH_URL=${CF_ORCH_URL:-http://host.docker.internal:7700}
      - CF_ORCH_URL=${CF_ORCH_URL:-${GPU_SERVER_URL:-http://host.docker.internal:7700}}
      - CF_APP_NAME=peregrine
      - PYTHONUNBUFFERED=1
    extra_hosts:
      - "host.docker.internal:host-gateway"
--- a/config/llm.yaml.example
+++ b/config/llm.yaml.example
@ -46,61 +46,11 @@ backends:
    type: vision_service
    supports_images: true
-  # ── cf-orch task-routed backends (preferred for GPU inference) ────────────
+  # ── cf-orch trunk services ─────────────────────────────────────────────────
-  # Use these when GPU_SERVER_URL is configured. The coordinator resolves
+  # These backends allocate via cf-orch rather than connecting to a static URL.
-  # product+task → model_id → node via assignments.yaml; no model IDs needed here.
+  # cf-orch starts the service on-demand and returns its URL; the router then
-  # Set enabled: true once GPU_SERVER_URL is configured.
+  # calls it directly using the openai_compat path.
-  cf_cover_letter:
+  # Set CF_ORCH_URL (env) or url below; leave enabled: false if cf-orch is
    type: openai_compat
    enabled: false
    base_url: http://localhost:8008/v1   # fallback when cf-orch is unavailable
    model: __auto__
    api_key: any
    supports_images: false
    cf_orch:
      product: peregrine
      task: cover_letter
      ttl_s: 3600
  cf_ats_rewrite:
    type: openai_compat
    enabled: false
    base_url: http://localhost:8008/v1
    model: __auto__
    api_key: any
    supports_images: false
    cf_orch:
      product: peregrine
      task: ats_rewrite
      ttl_s: 3600
  cf_job_research:
    type: openai_compat
    enabled: false
    base_url: http://localhost:8008/v1
    model: __auto__
    api_key: any
    supports_images: false
    cf_orch:
      product: peregrine
      task: job_research
      ttl_s: 3600
  cf_interview_prep:
    type: openai_compat
    enabled: false
    base_url: http://localhost:8008/v1
    model: __auto__
    api_key: any
    supports_images: false
    cf_orch:
      product: peregrine
      task: interview_prep
      ttl_s: 3600
  # ── cf-orch trunk services (service-based, legacy) ─────────────────────────
  # Generic service allocation — use the task-routed backends above when possible.
  # Set GPU_SERVER_URL (env) or url below; leave enabled: false if cf-orch is
  # not deployed in your environment.
  cf_text:
    type: openai_compat
--- a/dev-api.py
+++ b/dev-api.py
@ -48,21 +48,6 @@ _CLOUD_DATA_ROOT  = Path(os.environ.get("CLOUD_DATA_ROOT", "/devl/menagerie-data
 _DIRECTUS_SECRET  = os.environ.get("DIRECTUS_JWT_SECRET", "")
 IS_DEMO: bool = os.environ.get("DEMO_MODE", "").lower() in ("1", "true", "yes")
 # Resolve GPU inference server URL.
 # Priority: GPU_SERVER_URL → CF_ORCH_URL (backward compat) → cloud default when licensed.
 # Result is written back to CF_ORCH_URL so all downstream callers need no changes.
 _GPU_SERVER_URL: str | None = (
    os.environ.get("GPU_SERVER_URL")
    or os.environ.get("CF_ORCH_URL")
    or (
        "https://orch.circuitforge.tech"
        if os.environ.get("CF_LICENSE_KEY")
        else None
    )
 )
 if _GPU_SERVER_URL:
    os.environ["CF_ORCH_URL"] = _GPU_SERVER_URL
 # Per-request DB path — set by cloud_session_middleware; falls back to DB_PATH
 _request_db: ContextVar[str | None] = ContextVar("_request_db", default=None)
@ -129,38 +114,6 @@ app.include_router(_feedback_router, prefix="/api/feedback")
 _log = logging.getLogger("peregrine.session")
 # ── Structured auth logging ───────────────────────────────────────────────────
 # Writes one JSON line per request to /devl/peregrine-logs/auth.log when in
 # cloud mode. Rotates at 10 MB, keeps 5 files. Also logs to stdout in dev.
 _AUTH_LOG_DIR = Path(os.environ.get("PEREGRINE_LOG_DIR", "/devl/peregrine-logs"))
 class _JsonFormatter(logging.Formatter):
    def format(self, record: logging.LogRecord) -> str:
        payload = {
            "ts": datetime.now(timezone.utc).isoformat(),
            "level": record.levelname,
            "logger": record.name,
            "msg": record.getMessage(),
        }
        if hasattr(record, "auth_event"):
            payload.update(record.auth_event)
        return json.dumps(payload)
 def _setup_auth_logging() -> None:
    from logging.handlers import RotatingFileHandler
    _AUTH_LOG_DIR.mkdir(parents=True, exist_ok=True)
    handler = RotatingFileHandler(
        _AUTH_LOG_DIR / "auth.log", maxBytes=10 * 1024 * 1024, backupCount=5
    )
    handler.setFormatter(_JsonFormatter())
    handler.setLevel(logging.INFO)
    _log.addHandler(handler)
    _log.setLevel(logging.DEBUG)
 _setup_auth_logging()
 _seen_users: set[str] = set()  # track first-access events within this process lifetime
 def _demo_guard() -> None:
    """Raise 403 if running in demo mode. Call at the top of any write endpoint."""
@ -205,16 +158,6 @@ def _resolve_cf_user_id(cookie_str: str) -> str | None:
    return None
 def _auth_log(event: str, **kwargs) -> None:
    """Emit a structured INFO log line to the auth logger."""
    record = logging.LogRecord(
        name="peregrine.session", level=logging.INFO,
        pathname="", lineno=0, msg=event, args=(), exc_info=None,
    )
    record.auth_event = {"event": event, **kwargs}
    _log.handle(record)
@app.middleware("http")
 async def cloud_session_middleware(request: Request, call_next):
    """In cloud mode, resolve per-user staging.db from the X-CF-Session header."""
@ -222,36 +165,16 @@ async def cloud_session_middleware(request: Request, call_next):
        cookie_header = request.headers.get("X-CF-Session", "")
        user_id = _resolve_cf_user_id(cookie_header)
        if user_id:
            first_access = user_id not in _seen_users
            if first_access:
                _seen_users.add(user_id)
            user_db = str(_CLOUD_DATA_ROOT / user_id / "peregrine" / "staging.db")
            if user_db not in _migrated_db_paths:
                from scripts.db_migrate import migrate_db
                migrate_db(Path(user_db))
                _migrated_db_paths.add(user_db)
            _auth_log(
                "session_resolved",
                user_id=user_id,
                method=request.method,
                path=request.url.path,
                first_access=first_access,
            )
            token = _request_db.set(user_db)
            try:
                return await call_next(request)
            finally:
                _request_db.reset(token)
        else:
            # Only log failures on non-trivial paths (skip health checks / static assets)
            if request.url.path.startswith("/api/"):
                _auth_log(
                    "session_failed",
                    method=request.method,
                    path=request.url.path,
                    reason="no_user_id",
                    has_cookie=bool(cookie_header),
                )
    return await call_next(request)
@ -651,51 +574,6 @@ def resume_optimizer_task_status(job_id: int):
    return {"status": row["status"], "stage": row["stage"], "message": row["error"]}
 def _capture_review_corrections(
    db_path: Path,
    job_id: int,
    draft: dict,
    decisions: dict,
 ) -> None:
    """Persist (proposed, accepted) pairs when the user edits LLM output in the review UI.
    Only saves corrections where accepted=True AND the user actually modified the
    proposed text (proposed != accepted).  Rejections carry no training signal.
    """
    from scripts.db import save_resume_correction as _save_correction
    sections = {s["section"]: s for s in (draft.get("sections") or [])}
    # ── Summary correction ────────────────────────────────────────────────────
    summary_dec = decisions.get("summary", {})
    if summary_dec.get("accepted", True):
        edited_text = summary_dec.get("edited_text")
        proposed_summary = sections.get("summary", {}).get("proposed", "")
        if edited_text is not None and edited_text.strip() != proposed_summary.strip():
            _save_correction(db_path, job_id, "summary", proposed_summary, edited_text.strip())
    # ── Experience bullet corrections ─────────────────────────────────────────
    exp_sec = sections.get("experience", {})
    entry_diffs = {
        f"{e['title']}|{e['company']}": e
        for e in (exp_sec.get("entries") or [])
    }
    for entry_dec in (decisions.get("experience", {}).get("accepted_entries") or []):
        if not entry_dec.get("accepted", True):
            continue
        edited_bullets = entry_dec.get("edited_bullets")
        if edited_bullets is None:
            continue
        key = f"{entry_dec.get('title', '')}|{entry_dec.get('company', '')}"
        diff = entry_diffs.get(key)
        if diff is None:
            continue
        proposed_bullets = diff.get("proposed_bullets") or []
        cleaned = [b for b in edited_bullets if b.strip()]
        if cleaned != proposed_bullets:
            _save_correction(db_path, job_id, f"experience:{key}", proposed_bullets, cleaned)
@app.get("/api/jobs/{job_id}/resume_optimizer/review")
 def get_resume_review(job_id: int):
    """Return the pending review draft for this job (populated when task is awaiting_review)."""
@ -752,10 +630,6 @@ def preview_resume_review(job_id: int, body: ResumeReviewBody):
    # Step 1: apply section-level decisions
    struct = apply_review_decisions(draft, body.decisions)
    # Step 1b: capture (proposed, accepted) correction pairs for Avocet fine-tuning.
    # Only fires when accepted=True and the user actually edited the LLM output.
    _capture_review_corrections(db_path, job_id, draft, body.decisions)
    # Step 2: inject gap framing for rejected skills (adjacent / learning)
    framings = [f.model_dump() for f in body.gap_framings if f.mode in ("adjacent", "learning")]
    if framings:
@ -777,19 +651,6 @@ def preview_resume_review(job_id: int, body: ResumeReviewBody):
    return {"preview_text": preview_text, "preview_struct": struct}
@app.get("/api/resume_optimizer/corrections")
 def list_resume_corrections(job_id: int | None = None, limit: int = 200):
    """Return resume review correction pairs for Avocet import.
    Each record is a (proposed, accepted) pair from the review UI where the
    user edited the LLM output before accepting.  These are SFT (supervised
    fine-tuning) candidates that flow through Avocet for human review.
    """
    from scripts.db import get_resume_corrections as _get_corrections
    db_path = Path(_request_db.get() or DB_PATH)
    return {"corrections": _get_corrections(db_path, limit=limit, job_id=job_id)}
@app.post("/api/jobs/{job_id}/resume_optimizer/approve")
 def approve_resume(job_id: int, body: dict):
    """Save the user-approved assembled resume struct and mark the task complete.
@ -806,8 +667,7 @@ def approve_resume(job_id: int, body: dict):
        raise HTTPException(400, "preview_struct is required")
    from scripts.resume_optimizer import render_resume_text
-    override = (body.get("preview_text_override") or "").strip()
+    final_text = render_resume_text(struct)
    final_text = override if override else render_resume_text(struct)
    # Persist plain text + struct (struct enables YAML export later)
    _finalize(db_path=db_path, job_id=job_id, final_text=final_text)
@ -1816,16 +1676,6 @@ def list_contacts(job_id: Optional[int] = None, direction: Optional[str] = None,
    return {"total": total, "contacts": [dict(r) for r in rows]}
@app.get("/api/contacts/{contact_id}")
 def get_contact(contact_id: int):
    db = _get_db()
    row = db.execute("SELECT * FROM job_contacts WHERE id = ?", (contact_id,)).fetchone()
    db.close()
    if not row:
        raise HTTPException(status_code=404, detail="Contact not found")
    return dict(row)
 # ── References ─────────────────────────────────────────────────────────────────
 class ReferencePayload(BaseModel):
@ -2261,7 +2111,6 @@ def bulk_purge_jobs(body: BulkPurgeBody):
 class AddJobsBody(BaseModel):
    urls: List[str]
    skip_review: bool = True
@app.post("/api/jobs/add", status_code=202)
@ -2273,7 +2122,6 @@ def add_jobs_by_url(body: AddJobsBody):
        from scripts.task_runner import submit_task
        db_path = _db_path()
        existing = get_existing_urls(db_path)
        status = "approved" if body.skip_review else "pending"
        queued = 0
        for raw_url in body.urls:
            url = canonicalize_url(raw_url.strip())
@ -2283,7 +2131,6 @@ def add_jobs_by_url(body: AddJobsBody):
                "title": "Importing...", "company": "", "url": url,
                "source": "manual", "location": "", "description": "",
                "date_found": _dt.now().isoformat()[:10],
                "status": status,
            })
            if job_id:
                submit_task(db_path, "scrape_url", job_id)
@ -3734,13 +3581,8 @@ def finetune_status():
            db_count = task.get("result_count", 0) or 0
            pairs_count = max(pairs_count, db_count)
        status = task.get("status", "idle") if task else "idle"
        try:
            from scripts.user_profile import UserProfile
            _opted_in = UserProfile(Path(_user_yaml_path())).training_export_opt_in
        except Exception:
            _opted_in = False
        # Stub quota for self-hosted; cloud overrides via its own middleware
-        return {"status": status, "pairs_count": pairs_count, "quota_remaining": None, "opted_in": _opted_in}
+        return {"status": status, "pairs_count": pairs_count, "quota_remaining": None}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
@ -3821,117 +3663,6 @@ def finetune_local_status():
        return {"model_ready": False}
 # ── Settings: Fine-Tune — Training Export ─────────────────────────────────────
 class TrainingOptInBody(BaseModel):
    enabled: bool
 def _training_opt_in_required() -> None:
    """Raise 403 if training_export_opt_in is not enabled in user profile."""
    try:
        from scripts.user_profile import UserProfile
        profile = UserProfile(Path(_user_yaml_path()))
        if not profile.training_export_opt_in:
            raise HTTPException(
                status_code=403,
                detail="Training export is not enabled. Enable it in Settings → Fine-Tune.",
            )
    except FileNotFoundError:
        raise HTTPException(
            status_code=403,
            detail="Training export is not enabled. Enable it in Settings → Fine-Tune.",
        )
@app.patch("/api/settings/fine-tune/opt-in")
 def set_training_opt_in(body: TrainingOptInBody):
    try:
        from scripts.user_profile import UserProfile
        profile = UserProfile(Path(_user_yaml_path()))
        profile.training_export_opt_in = body.enabled
        profile.save()
        return {"ok": True, "enabled": profile.training_export_opt_in}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
@app.get("/api/settings/fine-tune/db-pairs")
 def list_db_pairs():
    _training_opt_in_required()
    try:
        from scripts.db import get_db_pairs
        db_path = Path(_request_db.get() or DB_PATH)
        pairs = get_db_pairs(db_path)
        excluded_count = sum(1 for p in pairs if p["excluded"])
        return {
            "pairs": pairs,
            "total": len(pairs),
            "excluded_count": excluded_count,
        }
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
@app.patch("/api/settings/fine-tune/db-pairs/{job_id}/exclude")
 def exclude_db_pair(job_id: int):
    _training_opt_in_required()
    try:
        from scripts.db import set_training_exclusion
        set_training_exclusion(Path(_request_db.get() or DB_PATH), job_id, excluded=True)
        return {"ok": True, "job_id": job_id}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
@app.patch("/api/settings/fine-tune/db-pairs/{job_id}/include")
 def include_db_pair(job_id: int):
    _training_opt_in_required()
    try:
        from scripts.db import set_training_exclusion
        set_training_exclusion(Path(_request_db.get() or DB_PATH), job_id, excluded=False)
        return {"ok": True, "job_id": job_id}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
@app.get("/api/settings/fine-tune/export")
 def export_training_jsonl():
    _training_opt_in_required()
    import json as _json
    from fastapi.responses import StreamingResponse
    from scripts.db import get_training_pairs
    db_path = Path(_request_db.get() or DB_PATH)
    db_pairs = get_training_pairs(db_path)
    file_pairs = _load_training_pairs()
    def _generate():
        for pair in db_pairs:
            yield _json.dumps(pair, ensure_ascii=False) + "\n"
        for pair in file_pairs:
            record = dict(pair)
            record.setdefault("source", "file")
            yield _json.dumps(record, ensure_ascii=False) + "\n"
    return StreamingResponse(
        _generate(),
        media_type="application/x-ndjson",
        headers={"Content-Disposition": 'attachment; filename="peregrine_training_pairs.jsonl"'},
    )
 # Phase 2 stubs — reserved, not yet implemented
@app.post("/api/settings/fine-tune/cloud-request")
 def cloud_finetune_request():
    raise HTTPException(status_code=501, detail="Cloud fine-tune is not yet available.")
@app.get("/api/settings/fine-tune/cloud-status")
 def cloud_finetune_status():
    raise HTTPException(status_code=501, detail="Cloud fine-tune is not yet available.")
 # ── Settings: License ─────────────────────────────────────────────────────────
 # _config_dir() / _license_path() / _tokens_path() are per-request (see helpers above)
--- a/docs/developer-guide/contributing.md
+++ b/docs/developer-guide/contributing.md
@ -11,7 +11,7 @@ Thank you for your interest in contributing to Peregrine. This guide covers the
 ## Fork and Clone
 ```bash
-git clone https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine
+git clone https://git.circuitforge.io/circuitforge/peregrine
 cd peregrine
 ```
--- a/docs/index.md
+++ b/docs/index.md
@ -12,7 +12,7 @@ Peregrine automates the full job search lifecycle: discovery, matching, cover le
 ```bash
 # 1. Clone and install dependencies
-git clone https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine
+git clone https://git.circuitforge.io/circuitforge/peregrine
 cd peregrine
 bash install.sh
@ -31,23 +31,20 @@ The first-run wizard guides you through hardware detection, tier selection, iden
 ## Feature Overview
-| Feature | Free | Paid† | Premium |
+| Feature | Free | Paid | Premium |
-|---------|------|-------|---------|
+|---------|------|------|---------|
 | Job discovery (JobSpy + custom boards) | Yes | Yes | Yes |
 | Resume keyword matching | Yes | Yes | Yes |
-| Cover letter generation | BYOK‡ | Yes | Yes |
+| Cover letter generation | - | Yes | Yes |
-| Company research briefs | BYOK‡ | Yes | Yes |
+| Company research briefs | - | Yes | Yes |
-| Interview prep & practice Q&A | BYOK‡ | Yes | Yes |
+| Interview prep & practice Q&A | - | Yes | Yes |
 | Email sync & auto-classification | - | Yes | Yes |
-| Survey assistant (culture-fit Q&A) | BYOK‡ | Yes | Yes |
+| Survey assistant (culture-fit Q&A) | - | Yes | Yes |
 | Integration connectors (Notion, Airtable, etc.) | Partial | Yes | Yes |
 | Calendar sync (Google, Apple) | - | Yes | Yes |
 | Cover letter model fine-tuning | - | - | Yes |
 | Multi-user support | - | - | Yes |
 † **Paid** gives access to CircuitForge's hosted inference — no API key required.  
 ‡ **BYOK** — configure any LLM backend in `config/llm.yaml` (local Ollama/vLLM or an API key) and these features unlock at no charge, regardless of tier.
 See [Tier System](reference/tier-system.md) for the full feature gate table.
 ---
@ -63,8 +60,8 @@ See [Tier System](reference/tier-system.md) for the full feature gate table.
 ## License
-Core discovery pipeline: [MIT](https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine/src/branch/main/LICENSE-MIT)
+Core discovery pipeline: [MIT](https://git.circuitforge.io/circuitforge/peregrine/src/branch/main/LICENSE-MIT)
-AI features (cover letter generation, company research, interview prep, UI): [BSL 1.1](https://git.opensourcesolarpunk.com/Circuit-Forge/peregrine/src/branch/main/LICENSE-BSL)
+AI features (cover letter generation, company research, interview prep, UI): [BSL 1.1](https://git.circuitforge.io/circuitforge/peregrine/src/branch/main/LICENSE-BSL)
 © 2026 Circuit Forge LLC
--- a/docs/screenshots/01-dashboard.png
+++ b/docs/screenshots/01-dashboard.png
--- a/docs/screenshots/02-review.png
+++ b/docs/screenshots/02-review.png
--- a/docs/screenshots/03-apply.png
+++ b/docs/screenshots/03-apply.png
--- a/docs/screenshots/04-interviews.png
+++ b/docs/screenshots/04-interviews.png
--- a/manage.sh
+++ b/manage.sh
@ -15,11 +15,6 @@ cd "$SCRIPT_DIR"
 PROFILE="${PROFILE:-remote}"
 # ── Compose engine detection ──────────────────────────────────────────────────
 COMPOSE="$(command -v docker >/dev/null 2>&1 && docker compose version >/dev/null 2>&1 \
    && echo "docker compose" \
    || (command -v podman >/dev/null 2>&1 && echo "podman compose" || echo "podman-compose"))"
 # ── Usage ────────────────────────────────────────────────────────────────────
 usage() {
    echo ""
@ -33,10 +28,9 @@ usage() {
    echo -e "    ${GREEN}start${NC}               Start Peregrine (preflight → up)"
    echo -e "    ${GREEN}stop${NC}                Stop all services"
    echo -e "    ${GREEN}restart${NC}             Restart all services"
    echo -e "    ${GREEN}build [service]${NC}     Rebuild image(s) without restarting (default: api web)"
    echo -e "    ${GREEN}status${NC}              Show running containers"
-    echo -e "    ${GREEN}logs [service]${NC}      Tail logs (default: api)"
+    echo -e "    ${GREEN}logs [service]${NC}      Tail logs (default: app)"
-    echo -e "    ${GREEN}update${NC}              Pull latest images + rebuild"
+    echo -e "    ${GREEN}update${NC}              Pull latest images + rebuild app"
    echo -e "    ${GREEN}preflight${NC}           Check ports + resources; write .env"
    echo -e "    ${GREEN}models${NC}              Check ollama models in config; pull any missing"
    echo -e "    ${GREEN}test${NC}                Run test suite"
@ -47,12 +41,6 @@ usage() {
    echo -e "    ${GREEN}clean${NC}               Remove containers, images, volumes (DESTRUCTIVE)"
    echo -e "    ${GREEN}open${NC}                Open the web UI in your browser"
    echo ""
    echo -e "  Cloud / demo commands:"
    echo -e "    ${GREEN}cloud-start${NC}         Start the cloud stack (peregrine-cloud)"
    echo -e "    ${GREEN}cloud-restart${NC}       Rebuild + restart the cloud stack"
    echo -e "    ${GREEN}demo-start${NC}          Start the demo stack (peregrine-demo)"
    echo -e "    ${GREEN}demo-restart${NC}        Rebuild + restart the demo stack"
    echo ""
    echo "  Profiles (set via --profile or PROFILE env var):"
    echo "    remote       API-only, no local inference (default)"
    echo "    cpu          Local Ollama inference on CPU"
@ -82,7 +70,7 @@ while [[ $# -gt 0 ]]; do
    esac
 done
-SERVICE="${1:-api}"   # used by `logs` command
+SERVICE="${1:-app}"   # used by `logs` command
 # ── Dependency guard ──────────────────────────────────────────────────────────
 # Commands that delegate to make; others (status, logs, update, open, setup) run fine without it.
@ -113,7 +101,7 @@ case "$CMD" in
    start)
        info "Starting Peregrine (PROFILE=${PROFILE})..."
        make start PROFILE="$PROFILE"
-        PORT="$(grep -m1 '^VUE_PORT=' .env 2>/dev/null | cut -d= -f2 || echo 8506)"
+        PORT="$(grep -m1 '^STREAMLIT_PORT=' .env 2>/dev/null | cut -d= -f2 || echo 8501)"
        success "Peregrine is up → http://localhost:${PORT}"
        ;;
@ -126,30 +114,33 @@ case "$CMD" in
    restart)
        info "Restarting (PROFILE=${PROFILE})..."
        make restart PROFILE="$PROFILE"
-        PORT="$(grep -m1 '^VUE_PORT=' .env 2>/dev/null | cut -d= -f2 || echo 8506)"
+        PORT="$(grep -m1 '^STREAMLIT_PORT=' .env 2>/dev/null | cut -d= -f2 || echo 8501)"
        success "Peregrine restarted → http://localhost:${PORT}"
        ;;
    status)
        # Auto-detect compose engine same way Makefile does
        COMPOSE="$(command -v docker >/dev/null 2>&1 && docker compose version >/dev/null 2>&1 \
            && echo "docker compose" \
            || (command -v podman >/dev/null 2>&1 && echo "podman compose" || echo "podman-compose"))"
        $COMPOSE ps
        ;;
    logs)
        COMPOSE="$(command -v docker >/dev/null 2>&1 && docker compose version >/dev/null 2>&1 \
            && echo "docker compose" \
            || (command -v podman >/dev/null 2>&1 && echo "podman compose" || echo "podman-compose"))"
        info "Tailing logs for: ${SERVICE}"
        $COMPOSE logs -f "$SERVICE"
        ;;
    build)
        BUILD_SVC="$([[ "${SERVICE}" == "api" ]] && echo "api web" || echo "${SERVICE}")"
        info "Building ${BUILD_SVC}..."
        $COMPOSE build $BUILD_SVC
        success "Build complete. Run './manage.sh restart' to apply."
        ;;
    update)
-        info "Pulling latest images and rebuilding..."
+        info "Pulling latest images and rebuilding app..."
        COMPOSE="$(command -v docker >/dev/null 2>&1 && docker compose version >/dev/null 2>&1 \
            && echo "docker compose" \
            || (command -v podman >/dev/null 2>&1 && echo "podman compose" || echo "podman-compose"))"
        $COMPOSE pull searxng ollama 2>/dev/null || true
-        $COMPOSE build api web
+        $COMPOSE build app web
        success "Update complete. Run './manage.sh restart' to apply."
        ;;
@ -176,7 +167,7 @@ case "$CMD" in
        ;;
    open)
-        PORT="$(grep -m1 '^VUE_PORT=' .env 2>/dev/null | cut -d= -f2 || echo 8506)"
+        PORT="$(grep -m1 '^STREAMLIT_PORT=' .env 2>/dev/null | cut -d= -f2 || echo 8501)"
        URL="http://localhost:${PORT}"
        info "Opening ${URL}"
        if command -v xdg-open &>/dev/null; then
@ -206,32 +197,6 @@ case "$CMD" in
            -v "${@:3}"
        ;;
    cloud-start)
        info "Starting cloud stack (peregrine-cloud)..."
        $COMPOSE -f compose.cloud.yml --project-name peregrine-cloud up -d
        success "Cloud stack up → http://localhost:8508"
        ;;
    cloud-restart)
        info "Rebuilding + restarting cloud stack (peregrine-cloud)..."
        $COMPOSE -f compose.cloud.yml --project-name peregrine-cloud build api web
        $COMPOSE -f compose.cloud.yml --project-name peregrine-cloud up -d
        success "Cloud stack restarted → http://localhost:8508"
        ;;
    demo-start)
        info "Starting demo stack (peregrine-demo)..."
        $COMPOSE -f compose.demo.yml --project-name peregrine-demo up -d
        success "Demo stack up → http://localhost:8504"
        ;;
    demo-restart)
        info "Rebuilding + restarting demo stack (peregrine-demo)..."
        $COMPOSE -f compose.demo.yml --project-name peregrine-demo build api web
        $COMPOSE -f compose.demo.yml --project-name peregrine-demo up -d
        success "Demo stack restarted → http://localhost:8504"
        ;;
    help|--help|-h)
        usage
        ;;
--- a/migrations/009_training_export.sql
+++ b/migrations/009_training_export.sql
@ -1 +0,0 @@
 ALTER TABLE jobs ADD COLUMN excluded_from_training INTEGER DEFAULT 0;
--- a/resume_matcher/apps/backend/app/cloud_session.py
+++ b/resume_matcher/apps/backend/app/cloud_session.py
@ -1,89 +0,0 @@
 """
 Peregrine cloud session — thin wrapper around cf_core.cloud_session.
 Sets request-scoped ContextVars with the authenticated user_id, tier, and
 custom writing model so that _allocate_orch_async in llm.py can forward them
 to cf-orch without any service function signature changes.
 Usage — add to main.py once:
    from app.cloud_session import session_middleware_dep
    app = FastAPI(..., dependencies=[Depends(session_middleware_dep)])
 From that point, any route (and every service/llm function it calls)
 has access to the current user context via llm.get_request_*() helpers.
 Writing model resolution order (first match wins):
  1. USER_WRITING_MODELS env var  — JSON dict mapping Directus UUID → model name
     e.g. USER_WRITING_MODELS={"5b99ca9f-...": "meghan-letter-writer:latest"}
     Use this for Monday; no Heimdall changes required.
  2. session.meta["custom_writing_model"]  — returned by Heimdall resolve endpoint
     once Heimdall is updated to expose user_preferences fields.
 """
 from __future__ import annotations
 import json
 import logging
 import os
 from fastapi import Depends, Request, Response
 from circuitforge_core.cloud_session import CloudSessionFactory, CloudUser, detect_byok
 log = logging.getLogger(__name__)
 __all__ = ["CloudUser", "get_session", "require_tier", "session_middleware_dep"]
 # JSON dict mapping Directus user UUID → custom writing model name.
 # Used until Heimdall's resolve endpoint exposes user_preferences.
 def _load_user_writing_models() -> dict[str, str]:
    raw = os.environ.get("USER_WRITING_MODELS", "").strip()
    if not raw:
        return {}
    try:
        return json.loads(raw)
    except json.JSONDecodeError:
        log.warning("USER_WRITING_MODELS is not valid JSON — ignoring")
        return {}
 _USER_WRITING_MODELS: dict[str, str] = _load_user_writing_models()
 _factory = CloudSessionFactory(
    product="peregrine",
    byok_detector=detect_byok,
 )
 get_session = _factory.dependency()
 require_tier = _factory.require_tier
 def session_middleware_dep(request: Request, response: Response) -> None:
    """Global FastAPI dependency — resolves the session and sets request-scoped
    ContextVars so llm._allocate_orch_async can forward them to cf-orch.
    Sets:
      - user_id: real cloud UUID, or None for local/anon sessions
      - tier: the resolved tier string (free/paid/premium/ultra/local)
      - writing_model: custom fine-tuned model from Heimdall meta, or None
    Add as a global dependency in main.py:
        app = FastAPI(..., dependencies=[Depends(session_middleware_dep)])
    """
    from app.llm import set_request_tier, set_request_user_id, set_request_writing_model
    session = _factory.resolve(request, response)
    user_id = session.user_id
    # Only forward real cloud UUIDs — local/dev/anon sessions use the shared catalog
    if user_id in (None, "local", "local-dev") or (user_id or "").startswith("anon-"):
        user_id = None
    set_request_user_id(user_id)
    set_request_tier(session.tier)
    # Resolution order: env-var map (Monday path) → Heimdall meta (future path)
    writing_model = (
        _USER_WRITING_MODELS.get(session.user_id)
        or session.meta.get("custom_writing_model")
    )
    set_request_writing_model(writing_model)
--- a/resume_matcher/apps/backend/app/llm.py
+++ b/resume_matcher/apps/backend/app/llm.py
@ -1,843 +0,0 @@
 """LiteLLM wrapper for multi-provider AI support."""
 import json
 import logging
 import os
 import re
 from contextlib import asynccontextmanager
 from contextvars import ContextVar
 from dataclasses import dataclass
 from typing import Any
 import httpx
 import litellm
 from pydantic import BaseModel
 from app.config import settings
 # LLM timeout configuration (seconds) - base values
 LLM_TIMEOUT_HEALTH_CHECK = 30
 LLM_TIMEOUT_COMPLETION = 120
 LLM_TIMEOUT_JSON = 180  # JSON completions may take longer
 # LLM-004: OpenRouter JSON-capable models (explicit allowlist)
 OPENROUTER_JSON_CAPABLE_MODELS = {
    # Anthropic models
    "anthropic/claude-3-opus",
    "anthropic/claude-3-sonnet",
    "anthropic/claude-3-haiku",
    "anthropic/claude-3.5-sonnet",
    "anthropic/claude-3.5-haiku",
    "anthropic/claude-haiku-4-5-20251001",
    "anthropic/claude-sonnet-4-20250514",
    "anthropic/claude-opus-4-20250514",
    # OpenAI models
    "openai/gpt-4-turbo",
    "openai/gpt-4",
    "openai/gpt-4o",
    "openai/gpt-4o-mini",
    "openai/gpt-3.5-turbo",
    "openai/gpt-5-nano-2025-08-07",
    # Google models
    "google/gemini-pro",
    "google/gemini-1.5-pro",
    "google/gemini-1.5-flash",
    "google/gemini-2.0-flash",
    "google/gemini-3-flash-preview",
    # DeepSeek models
    "deepseek/deepseek-chat",
    "deepseek/deepseek-reasoner",
    # Mistral models
    "mistralai/mistral-large",
    "mistralai/mistral-medium",
 }
 # JSON-010: JSON extraction safety limits
 MAX_JSON_EXTRACTION_RECURSION = 10
 MAX_JSON_CONTENT_SIZE = 1024 * 1024  # 1MB
 # Request-scoped user_id — set once by session_middleware_dep, read inside _allocate_orch_async.
 # ContextVar is safe for concurrent async requests: each request task gets its own copy.
 _request_user_id: ContextVar[str | None] = ContextVar("request_user_id", default=None)
 _request_tier: ContextVar[str | None] = ContextVar("request_tier", default=None)
 # Custom writing model for premium/ultra users — populated from Heimdall license key meta.
 # Set to None for all other tiers; complete() falls back to the shared base model.
 _request_writing_model: ContextVar[str | None] = ContextVar("request_writing_model", default=None)
 _PREMIUM_TIERS: frozenset[str] = frozenset({"premium", "ultra"})
 def set_request_user_id(user_id: str | None) -> None:
    _request_user_id.set(user_id)
 def get_request_user_id() -> str | None:
    return _request_user_id.get()
 def set_request_tier(tier: str | None) -> None:
    _request_tier.set(tier)
 def get_request_tier() -> str | None:
    return _request_tier.get()
 def set_request_writing_model(model: str | None) -> None:
    _request_writing_model.set(model)
 def get_request_writing_model() -> str | None:
    return _request_writing_model.get()
 class LLMConfig(BaseModel):
    """LLM configuration model."""
    provider: str
    model: str
    api_key: str
    api_base: str | None = None
@dataclass
 class _OrchAllocation:
    allocation_id: str
    url: str
    service: str
@asynccontextmanager
 async def _allocate_orch_async(
    coordinator_url: str,
    service: str,
    model_candidates: list[str],
    ttl_s: float,
    caller: str,
 ):
    """Async context manager that allocates a cf-orch service and releases on exit."""
    async with httpx.AsyncClient(timeout=120.0) as client:
        payload: dict[str, Any] = {
            "model_candidates": model_candidates,
            "ttl_s": ttl_s,
            "caller": caller,
        }
        uid = get_request_user_id()
        if uid:
            payload["user_id"] = uid
        resp = await client.post(
            f"{coordinator_url.rstrip('/')}/api/services/{service}/allocate",
            json=payload,
        )
        if not resp.is_success:
            raise RuntimeError(
                f"cf-orch allocation failed for {service!r}: "
                f"HTTP {resp.status_code} — {resp.text[:200]}"
            )
        data = resp.json()
        alloc = _OrchAllocation(
            allocation_id=data["allocation_id"],
            url=data["url"],
            service=service,
        )
        try:
            yield alloc
        finally:
            try:
                await client.delete(
                    f"{coordinator_url.rstrip('/')}/api/services/{service}/allocations/{alloc.allocation_id}",
                    timeout=10.0,
                )
            except Exception as exc:
                logging.debug("cf-orch release failed (non-fatal): %s", exc)
 def _normalize_api_base(provider: str, api_base: str | None) -> str | None:
    """Normalize api_base for LiteLLM provider-specific expectations.
    When using proxies/aggregators, users often paste a base URL that already
    includes a version segment (e.g., `/v1`). Some LiteLLM provider handlers
    append those segments internally, which can lead to duplicated paths like
    `/v1/v1/...` and cause 404s.
    """
    if not api_base:
        return None
    base = api_base.strip()
    if not base:
        return None
    base = base.rstrip("/")
    # Anthropic handler appends '/v1/messages'. If base already ends with '/v1',
    # strip it to avoid '/v1/v1/messages'.
    if provider == "anthropic" and base.endswith("/v1"):
        base = base[: -len("/v1")].rstrip("/")
    # Gemini handler appends '/v1/models/...'. If base already ends with '/v1',
    # strip it to avoid '/v1/v1/models/...'.
    if provider == "gemini" and base.endswith("/v1"):
        base = base[: -len("/v1")].rstrip("/")
    return base or None
 def _extract_text_parts(value: Any, depth: int = 0, max_depth: int = 10) -> list[str]:
    """Recursively extract text segments from nested response structures.
    Handles strings, lists, dicts with 'text'/'content'/'value' keys, and objects
    with text/content attributes. Limits recursion depth to avoid cycles.
    Args:
        value: Input value that may contain text in strings, lists, dicts, or objects.
        depth: Current recursion depth.
        max_depth: Maximum recursion depth before returning no content.
    Returns:
        A list of extracted text segments.
    """
    if depth >= max_depth:
        return []
    if value is None:
        return []
    if isinstance(value, str):
        return [value]
    if isinstance(value, list):
        parts: list[str] = []
        next_depth = depth + 1
        for item in value:
            parts.extend(_extract_text_parts(item, next_depth, max_depth))
        return parts
    if isinstance(value, dict):
        next_depth = depth + 1
        if "text" in value:
            return _extract_text_parts(value.get("text"), next_depth, max_depth)
        if "content" in value:
            return _extract_text_parts(value.get("content"), next_depth, max_depth)
        if "value" in value:
            return _extract_text_parts(value.get("value"), next_depth, max_depth)
        return []
    next_depth = depth + 1
    if hasattr(value, "text"):
        return _extract_text_parts(getattr(value, "text"), next_depth, max_depth)
    if hasattr(value, "content"):
        return _extract_text_parts(getattr(value, "content"), next_depth, max_depth)
    return []
 def _join_text_parts(parts: list[str]) -> str | None:
    """Join text parts with newlines, filtering empty strings.
    Args:
        parts: Candidate text segments.
    Returns:
        Joined string or None if the result is empty.
    """
    joined = "\n".join(part for part in parts if part).strip()
    return joined or None
 def _extract_message_text(message: Any) -> str | None:
    """Extract plain text from a LiteLLM message object across providers."""
    content: Any = None
    if hasattr(message, "content"):
        content = message.content
    elif isinstance(message, dict):
        content = message.get("content")
    return _join_text_parts(_extract_text_parts(content))
 def _extract_choice_text(choice: Any) -> str | None:
    """Extract plain text from a LiteLLM choice object.
    Tries message.content first, then choice.text, then choice.delta. Handles both
    object attributes and dict keys.
    Args:
        choice: LiteLLM choice object or dict.
    Returns:
        Extracted text or None if no content is found.
    """
    message: Any = None
    if hasattr(choice, "message"):
        message = choice.message
    elif isinstance(choice, dict):
        message = choice.get("message")
    content = _extract_message_text(message)
    if content:
        return content
    if hasattr(choice, "text"):
        content = _join_text_parts(_extract_text_parts(getattr(choice, "text")))
        if content:
            return content
    if isinstance(choice, dict) and "text" in choice:
        content = _join_text_parts(_extract_text_parts(choice.get("text")))
        if content:
            return content
    if hasattr(choice, "delta"):
        content = _join_text_parts(_extract_text_parts(getattr(choice, "delta")))
        if content:
            return content
    if isinstance(choice, dict) and "delta" in choice:
        content = _join_text_parts(_extract_text_parts(choice.get("delta")))
        if content:
            return content
    return None
 def _to_code_block(content: str | None, language: str = "text") -> str:
    """Wrap content in a markdown code block for client display."""
    text = (content or "").strip()
    if not text:
        text = "<empty>"
    return f"```{language}\n{text}\n```"
 def _load_stored_config() -> dict:
    """Load config from config.json file."""
    config_path = settings.config_path
    if config_path.exists():
        try:
            return json.loads(config_path.read_text())
        except (json.JSONDecodeError, OSError):
            return {}
    return {}
 def get_llm_config() -> LLMConfig:
    """Get current LLM configuration.
    Priority: config.json file > environment variables/settings
    """
    stored = _load_stored_config()
    return LLMConfig(
        provider=stored.get("provider", settings.llm_provider),
        model=stored.get("model", settings.llm_model),
        api_key=stored.get("api_key", settings.llm_api_key),
        api_base=stored.get("api_base", settings.llm_api_base),
    )
 def get_model_name(config: LLMConfig) -> str:
    """Convert provider/model to LiteLLM format.
    For most providers, adds the provider prefix if not already present.
    For OpenRouter, always adds 'openrouter/' prefix since OpenRouter models
    use nested prefixes like 'openrouter/anthropic/claude-3.5-sonnet'.
    """
    provider_prefixes = {
        "openai": "",  # OpenAI models don't need prefix
        "anthropic": "anthropic/",
        "openrouter": "openrouter/",
        "gemini": "gemini/",
        "deepseek": "deepseek/",
        "ollama": "ollama/",
    }
    prefix = provider_prefixes.get(config.provider, "")
    # OpenRouter is special: always add openrouter/ prefix unless already present
    # OpenRouter models use nested format: openrouter/anthropic/claude-3.5-sonnet
    if config.provider == "openrouter":
        if config.model.startswith("openrouter/"):
            return config.model
        return f"openrouter/{config.model}"
    # For other providers, don't add prefix if model already has a known prefix
    known_prefixes = ["openrouter/", "anthropic/", "gemini/", "deepseek/", "ollama/"]
    if any(config.model.startswith(p) for p in known_prefixes):
        return config.model
    # Add provider prefix for models that need it
    return f"{prefix}{config.model}" if prefix else config.model
 def _supports_temperature(provider: str, model: str) -> bool:
    """Return whether passing `temperature` is supported for this model/provider combo.
    Some models (e.g., OpenAI gpt-5 family) reject temperature values other than 1,
    and LiteLLM may error when temperature is passed.
    """
    _ = provider
    model_lower = model.lower()
    if "gpt-5" in model_lower:
        return False
    return True
 def _get_reasoning_effort(provider: str, model: str) -> str | None:
    """Return a default reasoning_effort for models that require it.
    Some OpenAI gpt-5 models may return empty message.content unless a supported
    `reasoning_effort` is explicitly set. This keeps downstream JSON parsing reliable.
    """
    _ = provider
    model_lower = model.lower()
    if "gpt-5" in model_lower:
        return "minimal"
    return None
 async def check_llm_health(
    config: LLMConfig | None = None,
    *,
    include_details: bool = False,
    test_prompt: str | None = None,
 ) -> dict[str, Any]:
    """Check if the LLM provider is accessible and working."""
    if config is None:
        config = get_llm_config()
    # Check if API key is configured (except for Ollama)
    if config.provider != "ollama" and not config.api_key:
        return {
            "healthy": False,
            "provider": config.provider,
            "model": config.model,
            "error_code": "api_key_missing",
        }
    model_name = get_model_name(config)
    prompt = test_prompt or "Hi"
    try:
        # Make a minimal test call with timeout
        # Pass API key directly to avoid race conditions with global os.environ
        kwargs: dict[str, Any] = {
            "model": model_name,
            "messages": [{"role": "user", "content": prompt}],
            "max_tokens": 16,
            "api_key": config.api_key,
            "api_base": _normalize_api_base(config.provider, config.api_base),
            "timeout": LLM_TIMEOUT_HEALTH_CHECK,
        }
        reasoning_effort = _get_reasoning_effort(config.provider, model_name)
        if reasoning_effort:
            kwargs["reasoning_effort"] = reasoning_effort
        response = await litellm.acompletion(**kwargs)
        content = _extract_choice_text(response.choices[0])
        if not content:
            # LLM-003: Empty response should mark health check as unhealthy
            logging.warning(
                "LLM health check returned empty content",
                extra={"provider": config.provider, "model": config.model},
            )
            result: dict[str, Any] = {
                "healthy": False,  # Fixed: empty content means unhealthy
                "provider": config.provider,
                "model": config.model,
                "response_model": response.model if response else None,
                "error_code": "empty_content",  # Changed from warning_code
                "message": "LLM returned empty response",
            }
            if include_details:
                result["test_prompt"] = _to_code_block(prompt)
                result["model_output"] = _to_code_block(None)
            return result
        result = {
            "healthy": True,
            "provider": config.provider,
            "model": config.model,
            "response_model": response.model if response else None,
        }
        if include_details:
            result["test_prompt"] = _to_code_block(prompt)
            result["model_output"] = _to_code_block(content)
        return result
    except Exception as e:
        # Log full exception details server-side, but do not expose them to clients
        logging.exception(
            "LLM health check failed",
            extra={"provider": config.provider, "model": config.model},
        )
        # Provide a minimal, actionable client-facing hint without leaking secrets.
        error_code = "health_check_failed"
        message = str(e)
        if "404" in message and "/v1/v1/" in message:
            error_code = "duplicate_v1_path"
        elif "404" in message:
            error_code = "not_found_404"
        elif "<!doctype html" in message.lower() or "<html" in message.lower():
            error_code = "html_response"
        result = {
            "healthy": False,
            "provider": config.provider,
            "model": config.model,
            "error_code": error_code,
        }
        if include_details:
            result["test_prompt"] = _to_code_block(prompt)
            result["model_output"] = _to_code_block(None)
            result["error_detail"] = _to_code_block(message)
        return result
 async def complete(
    prompt: str,
    system_prompt: str | None = None,
    config: LLMConfig | None = None,
    max_tokens: int = 4096,
    temperature: float = 0.7,
 ) -> str:
    """Make a completion request to the LLM."""
    if config is None:
        cf_orch_url = os.environ.get("CF_ORCH_URL", "").strip()
        if cf_orch_url:
            try:
                # Premium/ultra users get their personal fine-tuned writing model as the
                # first candidate; the base model is the fallback so cf-orch can
                # degrade gracefully if the personal model isn't loaded yet.
                tier = get_request_tier()
                writing_model = get_request_writing_model()
                model_candidates: list[str] = (
                    [writing_model, "Qwen2.5-3B-Instruct"]
                    if writing_model and tier in _PREMIUM_TIERS
                    else ["Qwen2.5-3B-Instruct"]
                )
                async with _allocate_orch_async(
                    cf_orch_url,
                    "vllm",
                    model_candidates=model_candidates,
                    ttl_s=300.0,
                    caller="peregrine-resume-matcher",
                ) as alloc:
                    orch_config = LLMConfig(
                        provider="openai",
                        model="__auto__",
                        api_key="any",
                        api_base=alloc.url.rstrip("/") + "/v1",
                    )
                    return await complete(prompt, system_prompt, orch_config, max_tokens, temperature)
            except Exception as exc:
                logging.warning("cf-orch allocation failed, falling back to default config: %s", exc)
        config = get_llm_config()
    model_name = get_model_name(config)
    messages = []
    if system_prompt:
        messages.append({"role": "system", "content": system_prompt})
    messages.append({"role": "user", "content": prompt})
    try:
        # Pass API key directly to avoid race conditions with global os.environ
        kwargs: dict[str, Any] = {
            "model": model_name,
            "messages": messages,
            "max_tokens": max_tokens,
            "api_key": config.api_key,
            "api_base": _normalize_api_base(config.provider, config.api_base),
            "timeout": LLM_TIMEOUT_COMPLETION,
        }
        if _supports_temperature(config.provider, model_name):
            kwargs["temperature"] = temperature
        reasoning_effort = _get_reasoning_effort(config.provider, model_name)
        if reasoning_effort:
            kwargs["reasoning_effort"] = reasoning_effort
        response = await litellm.acompletion(**kwargs)
        content = _extract_choice_text(response.choices[0])
        if not content:
            raise ValueError("Empty response from LLM")
        return content
    except Exception as e:
        # Log the actual error server-side for debugging
        logging.error(f"LLM completion failed: {e}", extra={"model": model_name})
        raise ValueError(
            "LLM completion failed. Please check your API configuration and try again."
        ) from e
 def _supports_json_mode(provider: str, model: str) -> bool:
    """Check if the model supports JSON mode."""
    # Models that support response_format={"type": "json_object"}
    json_mode_providers = ["openai", "anthropic", "gemini", "deepseek"]
    if provider in json_mode_providers:
        return True
    # LLM-004: OpenRouter models - use explicit allowlist instead of substring matching
    if provider == "openrouter":
        return model in OPENROUTER_JSON_CAPABLE_MODELS
    return False
 def _appears_truncated(data: dict) -> bool:
    """LLM-001: Check if JSON data appears to be truncated.
    Detects suspicious patterns indicating incomplete responses.
    """
    if not isinstance(data, dict):
        return False
    # Check for empty arrays that should typically have content
    suspicious_empty_arrays = ["workExperience", "education", "skills"]
    for key in suspicious_empty_arrays:
        if key in data and data[key] == []:
            # Log warning - these are rarely empty in real resumes
            logging.warning(
                "Possible truncation detected: '%s' is empty",
                key,
            )
            return True
    # Check for missing critical sections
    required_top_level = ["personalInfo"]
    for key in required_top_level:
        if key not in data:
            logging.warning(
                "Possible truncation detected: missing required section '%s'",
                key,
            )
            return True
    return False
 def _get_retry_temperature(attempt: int, base_temp: float = 0.1) -> float:
    """LLM-002: Get temperature for retry attempt - increases with each retry.
    Higher temperature on retries gives the model more variation to produce
    different (hopefully valid) output.
    """
    temperatures = [base_temp, 0.3, 0.5, 0.7]
    return temperatures[min(attempt, len(temperatures) - 1)]
 def _calculate_timeout(
    operation: str,
    max_tokens: int = 4096,
    provider: str = "openai",
 ) -> int:
    """LLM-005: Calculate adaptive timeout based on operation and parameters."""
    base_timeouts = {
        "health_check": LLM_TIMEOUT_HEALTH_CHECK,
        "completion": LLM_TIMEOUT_COMPLETION,
        "json": LLM_TIMEOUT_JSON,
    }
    base = base_timeouts.get(operation, LLM_TIMEOUT_COMPLETION)
    # Scale by token count (relative to 4096 baseline)
    token_factor = max(1.0, max_tokens / 4096)
    # Provider-specific latency adjustments
    provider_factors = {
        "openai": 1.0,
        "anthropic": 1.2,
        "openrouter": 1.5,  # More variable latency
        "ollama": 2.0,  # Local models can be slower
    }
    provider_factor = provider_factors.get(provider, 1.0)
    return int(base * token_factor * provider_factor)
 def _extract_json(content: str, _depth: int = 0) -> str:
    """Extract JSON from LLM response, handling various formats.
    LLM-001: Improved to detect and reject likely truncated JSON.
    LLM-007: Improved error messages for debugging.
    JSON-010: Added recursion depth and size limits.
    """
    # JSON-010: Safety limits
    if _depth > MAX_JSON_EXTRACTION_RECURSION:
        raise ValueError(f"JSON extraction exceeded max recursion depth: {_depth}")
    if len(content) > MAX_JSON_CONTENT_SIZE:
        raise ValueError(f"Content too large for JSON extraction: {len(content)} bytes")
    original = content
    # Remove markdown code blocks
    if "```json" in content:
        content = content.split("```json")[1].split("```")[0]
    elif "```" in content:
        parts = content.split("```")
        if len(parts) >= 2:
            content = parts[1]
            # Remove language identifier if present (e.g., "json\n{...")
            if content.startswith(("json", "JSON")):
                content = content[4:]
    content = content.strip()
    # If content starts with {, find the matching }
    if content.startswith("{"):
        depth = 0
        end_idx = -1
        in_string = False
        escape_next = False
        for i, char in enumerate(content):
            if escape_next:
                escape_next = False
                continue
            if char == "\\":
                escape_next = True
                continue
            if char == '"' and not escape_next:
                in_string = not in_string
                continue
            if in_string:
                continue
            if char == "{":
                depth += 1
            elif char == "}":
                depth -= 1
                if depth == 0:
                    end_idx = i
                    break
        # LLM-001: Check for unbalanced braces - loop ended without depth reaching 0
        if end_idx == -1 and depth != 0:
            logging.warning(
                "JSON extraction found unbalanced braces (depth=%d), possible truncation",
                depth,
            )
        if end_idx != -1:
            return content[: end_idx + 1]
    # Try to find JSON object in the content (only if not already at start)
    start_idx = content.find("{")
    if start_idx > 0:
        # Only recurse if { is found after position 0 to avoid infinite recursion
        return _extract_json(content[start_idx:], _depth + 1)
    # LLM-007: Log unrecognized format for debugging
    logging.error(
        "Could not extract JSON from response format. Content preview: %s",
        content[:200] if content else "<empty>",
    )
    raise ValueError(f"No JSON found in response: {original[:200]}")
 async def complete_json(
    prompt: str,
    system_prompt: str | None = None,
    config: LLMConfig | None = None,
    max_tokens: int = 4096,
    retries: int = 2,
 ) -> dict[str, Any]:
    """Make a completion request expecting JSON response.
    Uses JSON mode when available, with retry logic for reliability.
    """
    if config is None:
        cf_orch_url = os.environ.get("CF_ORCH_URL", "").strip()
        if cf_orch_url:
            try:
                async with _allocate_orch_async(
                    cf_orch_url,
                    "vllm",
                    model_candidates=["Qwen2.5-3B-Instruct"],
                    ttl_s=300.0,
                    caller="peregrine-resume-matcher",
                ) as alloc:
                    orch_config = LLMConfig(
                        provider="openai",
                        model="__auto__",
                        api_key="any",
                        api_base=alloc.url.rstrip("/") + "/v1",
                    )
                    return await complete_json(prompt, system_prompt, orch_config, max_tokens, retries)
            except Exception as exc:
                logging.warning("cf-orch allocation failed, falling back to default config: %s", exc)
        config = get_llm_config()
    model_name = get_model_name(config)
    # Build messages
    json_system = (
        system_prompt or ""
    ) + "\n\nYou must respond with valid JSON only. No explanations, no markdown."
    messages = [
        {"role": "system", "content": json_system},
        {"role": "user", "content": prompt},
    ]
    # Check if we can use JSON mode
    use_json_mode = _supports_json_mode(config.provider, config.model)
    last_error = None
    for attempt in range(retries + 1):
        try:
            # Build request kwargs
            # Pass API key directly to avoid race conditions with global os.environ
            kwargs: dict[str, Any] = {
                "model": model_name,
                "messages": messages,
                "max_tokens": max_tokens,
                "api_key": config.api_key,
                "api_base": _normalize_api_base(config.provider, config.api_base),
                "timeout": _calculate_timeout("json", max_tokens, config.provider),
            }
            if _supports_temperature(config.provider, model_name):
                # LLM-002: Increase temperature on retry for variation
                kwargs["temperature"] = _get_retry_temperature(attempt)
            reasoning_effort = _get_reasoning_effort(config.provider, model_name)
            if reasoning_effort:
                kwargs["reasoning_effort"] = reasoning_effort
            # Add JSON mode if supported
            if use_json_mode:
                kwargs["response_format"] = {"type": "json_object"}
            response = await litellm.acompletion(**kwargs)
            content = _extract_choice_text(response.choices[0])
            if not content:
                raise ValueError("Empty response from LLM")
            logging.debug(f"LLM response (attempt {attempt + 1}): {content[:300]}")
            # Extract and parse JSON
            json_str = _extract_json(content)
            result = json.loads(json_str)
            # LLM-001: Check if parsed result appears truncated
            if isinstance(result, dict) and _appears_truncated(result):
                logging.warning(
                    "Parsed JSON appears truncated, but proceeding with result"
                )
            return result
        except json.JSONDecodeError as e:
            last_error = e
            logging.warning(f"JSON parse failed (attempt {attempt + 1}): {e}")
            if attempt < retries:
                # Add hint to prompt for retry
                messages[-1]["content"] = (
                    prompt
                    + "\n\nIMPORTANT: Output ONLY a valid JSON object. Start with { and end with }."
                )
                continue
            raise ValueError(f"Failed to parse JSON after {retries + 1} attempts: {e}")
        except Exception as e:
            last_error = e
            logging.warning(f"LLM call failed (attempt {attempt + 1}): {e}")
            if attempt < retries:
                continue
            raise
    raise ValueError(f"Failed after {retries + 1} attempts: {last_error}")
--- a/resume_matcher/apps/backend/app/main.py
+++ b/resume_matcher/apps/backend/app/main.py
@ -1,88 +0,0 @@
 """FastAPI application entry point."""
 import asyncio
 import logging
 import sys
 from contextlib import asynccontextmanager
 from fastapi import Depends, FastAPI
 # Fix for Windows: Use ProactorEventLoop for subprocess support (Playwright)
 if sys.platform == "win32":
    asyncio.set_event_loop_policy(asyncio.WindowsProactorEventLoopPolicy())
 logger = logging.getLogger(__name__)
 from fastapi.middleware.cors import CORSMiddleware
 from app import __version__
 from app.cloud_session import session_middleware_dep
 from app.config import settings
 from app.database import db
 from app.pdf import close_pdf_renderer, init_pdf_renderer
 from app.routers import config_router, enrichment_router, health_router, jobs_router, resumes_router
@asynccontextmanager
 async def lifespan(app: FastAPI):
    """Application lifespan manager."""
    # Startup
    settings.data_dir.mkdir(parents=True, exist_ok=True)
    # PDF renderer uses lazy initialization - will initialize on first use
    # await init_pdf_renderer()
    yield
    # Shutdown - wrap each cleanup in try-except to ensure all resources are released
    try:
        await close_pdf_renderer()
    except Exception as e:
        logger.error(f"Error closing PDF renderer: {e}")
    try:
        db.close()
    except Exception as e:
        logger.error(f"Error closing database: {e}")
 app = FastAPI(
    title="Resume Matcher API",
    description="AI-powered resume tailoring for job descriptions",
    version=__version__,
    lifespan=lifespan,
    dependencies=[Depends(session_middleware_dep)],
 )
 # CORS middleware - origins configurable via CORS_ORIGINS env var
 app.add_middleware(
    CORSMiddleware,
    allow_origins=settings.cors_origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
 )
 # Include routers
 app.include_router(health_router, prefix="/api/v1")
 app.include_router(config_router, prefix="/api/v1")
 app.include_router(resumes_router, prefix="/api/v1")
 app.include_router(jobs_router, prefix="/api/v1")
 app.include_router(enrichment_router, prefix="/api/v1")
@app.get("/")
 async def root():
    """Root endpoint."""
    return {
        "name": "Resume Matcher API",
        "version": __version__,
        "docs": "/docs",
    }
 if __name__ == "__main__":
    import uvicorn
    uvicorn.run(
        "app.main:app",
        host=settings.host,
        port=settings.port,
        reload=True,
    )
--- a/scripts/db.py
+++ b/scripts/db.py
@ -171,7 +171,6 @@ _MIGRATIONS = [
    ("ats_gap_report",     "TEXT"),   # JSON gap report (free tier)
    ("date_posted",        "TEXT"),   # Original posting date from job board (shadow listing detection)
    ("hired_feedback",     "TEXT"),   # JSON: optional post-hire "what helped" response
    ("excluded_from_training", "INTEGER DEFAULT 0"),  # opt-out of training export
 ]
@ -234,11 +233,10 @@ def insert_job(db_path: Path = DEFAULT_DB, job: dict = None) -> Optional[int]:
        return None
    conn = sqlite3.connect(db_path)
    try:
        status = job.get("status", "pending")
        cursor = conn.execute(
            """INSERT INTO jobs
-               (title, company, url, source, location, is_remote, salary, description, date_found, date_posted, status)
+               (title, company, url, source, location, is_remote, salary, description, date_found, date_posted)
-               VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+               VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
            (
                job.get("title", ""),
                job.get("company", ""),
@ -250,7 +248,6 @@ def insert_job(db_path: Path = DEFAULT_DB, job: dict = None) -> Optional[int]:
                job.get("description", ""),
                job.get("date_found", ""),
                job.get("date_posted", "") or "",
                status,
            ),
        )
        conn.commit()
@ -1145,99 +1142,3 @@ def set_job_resume(db_path: Path = DEFAULT_DB, job_id: int = 0, resume_id: int =
        conn.commit()
    finally:
        conn.close()
 # ── Training export helpers ───────────────────────────────────────────────────
 def _strip_greeting(text: str) -> str:
    """Remove 'Dear X,' greeting line from cover letter text."""
    lines = text.splitlines()
    for i, line in enumerate(lines):
        stripped_line = line.strip()
        if stripped_line.lower().startswith("dear ") and stripped_line.endswith((",", ":")):
            rest = lines[i + 1:]
            while rest and not rest[0].strip():
                rest = rest[1:]
            result = "\n".join(rest).strip()
            return result if result else text.strip()
    return text.strip()
 def get_db_pairs(db_path: Path) -> list[dict]:
    """Return curation metadata for ALL qualifying jobs (included and excluded).
    Used by the curation UI. Includes excluded=True rows so users can restore them.
    """
    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    try:
        rows = conn.execute(
            "SELECT id, title, company, description, status, "
            "  excluded_from_training "
            "FROM jobs "
            "WHERE status IN ('applied','phone_screen','interviewing','offer','hired') "
            "  AND cover_letter IS NOT NULL AND cover_letter != '' "
            "ORDER BY applied_at DESC",
        ).fetchall()
    finally:
        conn.close()
    return [
        {
            "job_id":        row["id"],
            "title":         row["title"] or "",
            "company":       row["company"] or "",
            "status":        row["status"],
            "instruction":   (
                f"Write a cover letter for the {row['title'] or 'unknown'} "
                f"position at {row['company'] or 'unknown'}."
            ),
            "input_preview": (row["description"] or "")[:200],
            "excluded":      bool(row["excluded_from_training"]),
        }
        for row in rows
    ]
 def get_training_pairs(db_path: Path) -> list[dict]:
    """Return Alpaca-format training pairs for non-excluded qualifying jobs.
    Used by the JSONL export endpoint.
    """
    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    try:
        rows = conn.execute(
            "SELECT id, title, company, description, cover_letter "
            "FROM jobs "
            "WHERE status IN ('applied','phone_screen','interviewing','offer','hired') "
            "  AND cover_letter IS NOT NULL AND cover_letter != '' "
            "  AND excluded_from_training = 0 "
            "ORDER BY applied_at DESC",
        ).fetchall()
    finally:
        conn.close()
    return [
        {
            "instruction": (
                f"Write a cover letter for the {row['title'] or 'unknown'} "
                f"position at {row['company'] or 'unknown'}."
            ),
            "input":   row["description"] or "",
            "output":  _strip_greeting(row["cover_letter"]),
            "source":  "db",
            "job_id":  row["id"],
        }
        for row in rows
    ]
 def set_training_exclusion(db_path: Path, job_id: int, excluded: bool) -> None:
    """Set excluded_from_training flag on a job."""
    conn = sqlite3.connect(db_path)
    try:
        conn.execute(
            "UPDATE jobs SET excluded_from_training = ? WHERE id = ?",
            (1 if excluded else 0, job_id),
        )
        conn.commit()
    finally:
        conn.close()
--- a/scripts/imap_sync.py
+++ b/scripts/imap_sync.py
@ -392,7 +392,6 @@ def _has_todo_keyword(subject: str) -> bool:
 _LINKEDIN_ALERT_SENDER = "jobalerts-noreply@linkedin.com"
 _INDEED_ALERT_SENDER  = "jobalerts@indeed.com"
 # Social-proof / nav lines to skip when parsing alert blocks
 _ALERT_SKIP_PHRASES = {
@ -448,75 +447,6 @@ def parse_linkedin_alert(body: str) -> list[dict]:
    return jobs
 def parse_indeed_alert(body: str) -> list[dict]:
    """
    Parse the HTML body of an Indeed Job Alert email.
    Returns a list of dicts: {title, company, location, salary, url}.
    URL is canonicalised to https://www.indeed.com/viewjob?jk=<id>
    (tracking parameters stripped).
    """
    try:
        from bs4 import BeautifulSoup as _BS
    except ImportError:
        return []
    jobs: list[dict] = []
    soup = _BS(body, "html.parser")
    # Each job card is an <a> wrapping a job title — Indeed uses several layouts
    # across their email templates.  We try two strategies:
    #
    # Strategy A (2023+ layout): <td> blocks containing an <a> with /viewjob?jk=
    # Strategy B (older layout): <tr class="job"> blocks
    #
    # Both extract the canonical jk= key from the href.
    seen_jks: set[str] = set()
    for anchor in soup.find_all("a", href=True):
        href: str = anchor["href"]
        jk_m = re.search(r"[?&]jk=([a-z0-9]+)", href, re.IGNORECASE)
        if not jk_m:
            continue
        jk = jk_m.group(1)
        if jk in seen_jks:
            continue
        seen_jks.add(jk)
        title = anchor.get_text(separator=" ", strip=True)
        if not title or len(title) < 3:
            continue
        # Walk up to find the container cell/row and extract company + location
        container = anchor.find_parent(["td", "tr", "div"])
        company = location = salary = ""
        if container:
            text_lines = [
                t.strip() for t in container.get_text(separator="\n").splitlines()
                if t.strip() and t.strip().lower() != title.lower()
            ]
            if text_lines:
                company = text_lines[0]
            if len(text_lines) > 1:
                location = text_lines[1]
            # salary line often contains "$" or "/yr"
            for line in text_lines[2:]:
                if "$" in line or "/yr" in line.lower() or "/hour" in line.lower():
                    salary = line
                    break
        jobs.append({
            "title":    title,
            "company":  company,
            "location": location,
            "salary":   salary,
            "url":      f"https://www.indeed.com/viewjob?jk={jk}",
        })
    return jobs
 def _scan_todo_label(conn: imaplib.IMAP4, cfg: dict, db_path: Path,
                     active_jobs: list[dict],
                     known_message_ids: set) -> int:
@ -628,29 +558,20 @@ def _scan_unmatched_leads(conn: imaplib.IMAP4, cfg: dict,
        if mid in known_message_ids:
            continue
-        # ── Job alert digests — parse each card deterministically ───────
+        # ── LinkedIn Job Alert digest — parse each card individually ──────
-        from_lower = parsed["from_addr"].lower()
+        if _LINKEDIN_ALERT_SENDER in parsed["from_addr"].lower():
-        alert_cards: list[dict] = []
+            cards = parse_linkedin_alert(parsed["body"])
-        alert_source = ""
+            for card in cards:
        if _LINKEDIN_ALERT_SENDER in from_lower:
            alert_cards = parse_linkedin_alert(parsed["body"])
            alert_source = "linkedin"
        elif _INDEED_ALERT_SENDER in from_lower:
            alert_cards = parse_indeed_alert(parsed["body"])
            alert_source = "indeed"
        if alert_cards:
            for card in alert_cards:
                if card["url"] in existing_urls:
                    continue
                job_id = insert_job(db_path, {
                    "title": card["title"],
                    "company": card["company"],
                    "url": card["url"],
-                    "source":     alert_source,
+                    "source": "linkedin",
-                    "location":   card.get("location", ""),
+                    "location": card["location"],
                    "is_remote": 0,
-                    "salary":     card.get("salary", ""),
+                    "salary": "",
                    "description": "",
                    "date_found": datetime.now().isoformat()[:10],
                })
@ -659,7 +580,7 @@ def _scan_unmatched_leads(conn: imaplib.IMAP4, cfg: dict,
                    submit_task(db_path, "scrape_url", job_id)
                    existing_urls.add(card["url"])
                    new_leads += 1
-                    print(f"[imap] {alert_source} alert → {card['company']} — {card['title']}")
+                    print(f"[imap] LinkedIn alert → {card['company']} — {card['title']}")
            known_message_ids.add(mid)
            continue  # skip normal LLM extraction path
--- a/scripts/job_ranker.py
+++ b/scripts/job_ranker.py
@ -12,53 +12,10 @@ Usage:
 """
 from __future__ import annotations
 import logging
 import math
 import re
 from datetime import datetime, timezone
 _log = logging.getLogger(__name__)
 # Max jobs passed to the reranker (avoids excessive inference time on large stacks)
 _RERANK_POOL = 50
 def _try_rerank(resume_text: str, jobs: list[dict]) -> list[dict]:
    """Rerank jobs by cross-encoder relevance to resume text.
    Returns jobs sorted best-first by the reranker. Falls back silently to the
    input order if the reranker package is unavailable or inference fails.
    """
    if not jobs:
        return jobs
    try:
        from circuitforge_core.reranker import rerank
    except ImportError:
        return jobs
    try:
        descriptions = [j.get("description") or j.get("title", "") for j in jobs]
        results = rerank(resume_text, descriptions, top_n=len(jobs))
        # Map ranked candidates back to job dicts, handling duplicate descriptions
        idx_queue: dict[str, list[int]] = {}
        for i, d in enumerate(descriptions):
            idx_queue.setdefault(d, []).append(i)
        reranked: list[dict] = []
        used: set[int] = set()
        for r in results:
            for idx in idx_queue.get(r.candidate, []):
                if idx not in used:
                    reranked.append(jobs[idx])
                    used.add(idx)
                    break
        # Safety: append anything the reranker didn't return
        for i, j in enumerate(jobs):
            if i not in used:
                reranked.append(j)
        return reranked
    except Exception:
        _log.warning("Reranker pass failed; using stack_score order.", exc_info=True)
        return jobs
 # ── TUNING ─────────────────────────────────────────────────────────────────────
 # Adjust these constants to change how jobs are ranked.
@ -332,7 +289,6 @@ def rank_jobs(
    user_level: int = 3,
    limit: int = 10,
    min_score: float = 20.0,
    resume_text: str = "",
 ) -> list[dict]:
    """Score and rank pending jobs; return top-N above min_score.
@ -343,10 +299,6 @@ def rank_jobs(
        user_level:       Seniority level 1–7 (use seniority_from_experience()).
        limit:            Stack size; pass 0 to return all qualifying jobs.
        min_score:        Minimum stack_score to include (0–100).
        resume_text:      Plain-text resume for cross-encoder reranking pass.
                          When provided, the top-_RERANK_POOL candidates are
                          reranked by (resume, description) relevance before
                          the limit is applied. Graceful no-op when empty.
    Returns:
        Sorted list (best first) with 'stack_score' key added to each dict.
@ -358,10 +310,4 @@ def rank_jobs(
            scored.append({**job, "stack_score": s})
    scored.sort(key=lambda j: j["stack_score"], reverse=True)
    if resume_text and scored:
        pool = scored[:_RERANK_POOL]
        pool = _try_rerank(resume_text, pool)
        scored = pool + scored[_RERANK_POOL:]
    return scored[:limit] if limit > 0 else scored
--- a/scripts/resume_optimizer.py
+++ b/scripts/resume_optimizer.py
@ -233,22 +233,6 @@ def rewrite_for_ats(
    from scripts.llm_router import LLMRouter
    router = LLMRouter()
    # Rerank gaps by JD relevance so the most impactful terms are injected first.
    # Falls back silently to the incoming priority ordering on any error.
    jd_text = job.get("description", "")
    if jd_text and prioritized_gaps:
        try:
            from circuitforge_core.reranker import rerank as _rerank
            terms = [g["term"] for g in prioritized_gaps]
            results = _rerank(jd_text, terms, top_n=len(terms))
            term_rank = {r.candidate: r.rank for r in results}
            prioritized_gaps = sorted(
                prioritized_gaps,
                key=lambda g: term_rank.get(g["term"], len(prioritized_gaps)),
            )
        except Exception:
            pass  # keep original priority ordering
    # Group gaps by target section
    by_section: dict[str, list[str]] = {}
    for gap in prioritized_gaps:
@ -278,8 +262,7 @@ def rewrite_for_ats(
            f"3. Only rephrase existing content — replace vague verbs/nouns with the "
            f"   ATS-preferred equivalents listed above.\n"
            f"4. Keep the same number of bullet points in experience entries.\n"
-            f"5. Do NOT use markdown formatting — no **, __, or * for bullets.\n"
+            f"5. Return ONLY the rewritten section content, no labels or explanation."
            f"6. Return ONLY the rewritten section content, no labels or explanation."
            f"{voice_note}\n\n"
            f"Original {section} section:\n{original_content}"
        )
@ -306,8 +289,7 @@ def _section_text_for_prompt(resume: dict[str, Any], section: str) -> str:
        for exp in resume.get("experience", []):
            lines.append(f"{exp['title']} at {exp['company']} ({exp['start_date']}–{exp['end_date']})")
            for b in exp.get("bullets", []):
-                clean_b = re.sub(r"^[•\-–—*◦▪▸►\s]+", "", b).strip()
+                lines.append(f"  • {b}")
                lines.append(f"  • {clean_b}")
        return "\n".join(lines) if lines else "(empty)"
    return "(unsupported section)"
@ -316,7 +298,7 @@ def _apply_section_rewrite(resume: dict[str, Any], section: str, rewritten: str)
    """Return a new resume dict with the given section replaced by rewritten text."""
    updated = dict(resume)
    if section == "summary":
-        updated["career_summary"] = _clean_summary_markup(rewritten)
+        updated["career_summary"] = rewritten
    elif section == "skills":
        # LLM returns comma-separated or newline-separated skills
        skills = [s.strip() for s in re.split(r"[,\n•·]+", rewritten) if s.strip()]
@ -328,19 +310,6 @@ def _apply_section_rewrite(resume: dict[str, Any], section: str, rewritten: str)
    return updated
 def _clean_summary_markup(text: str) -> str:
    """Strip markdown/plain-text bullet markers from career summary lines.
    LLMs sometimes format summary content with '* item' or '• item' markdown.
    This converts those lines to unmarked text so the summary renders cleanly.
    """
    lines = []
    for line in text.splitlines():
        cleaned = re.sub(r"^[•*\-–—◦▪▸►]\s+", "", line.lstrip())
        lines.append(cleaned)
    return "\n".join(lines).strip()
 def _reparse_experience_bullets(
    original_entries: list[dict],
    rewritten_text: str,
@ -370,9 +339,9 @@ def _reparse_experience_bullets(
            chunk = remaining
        bullets = [
-            re.sub(r"^([•\-–—*◦▪▸►]\s*)+", "", line.strip()).strip()
+            re.sub(r"^[•\-–—*◦▪▸►]\s*", "", line).strip()
            for line in chunk.splitlines()
-            if re.match(r"^\s*[•\-–—*◦▪▸►]", line)
+            if re.match(r"^[•\-–—*◦▪▸►]\s*", line.strip())
        ]
        new_entry = dict(entry)
        new_entry["bullets"] = bullets if bullets else entry["bullets"]
@ -547,37 +516,27 @@ def apply_review_decisions(
            struct["skills"] = sorted(original_kept | approved_additions)
            break
-    # ── Summary: accept/reject + optional user-edited text ─────────────────
+    # ── Summary: accept proposed or revert to original ──────────────────────
-    summary_dec = decisions.get("summary", {})
+    if not decisions.get("summary", {}).get("accepted", True):
    if not summary_dec.get("accepted", True):
        for sec in sections:
            if sec["section"] == "summary":
                struct["career_summary"] = sec.get("original", struct.get("career_summary", ""))
                break
    else:
        edited_text = summary_dec.get("edited_text")
        if edited_text is not None:
            struct["career_summary"] = edited_text.strip()
-    # ── Experience: per-entry accept/reject + optional user-edited bullets ──
+    # ── Experience: per-entry accept/reject ─────────────────────────────────
-    exp_entry_map: dict[str, dict] = {
+    exp_decisions: dict[str, bool] = {
-        f"{ed.get('title', '')}|{ed.get('company', '')}": ed
+        f"{ed.get('title', '')}|{ed.get('company', '')}": ed.get("accepted", True)
        for ed in (decisions.get("experience", {}).get("accepted_entries") or [])
    }
    for sec in sections:
        if sec["section"] == "experience":
            for entry_diff in (sec.get("entries") or []):
                key = f"{entry_diff['title']}|{entry_diff['company']}"
-                entry_dec = exp_entry_map.get(key, {})
+                if not exp_decisions.get(key, True):
                accepted = entry_dec.get("accepted", True)
                edited_bullets = entry_dec.get("edited_bullets")
                    for exp_entry in (struct.get("experience") or []):
                        if (exp_entry.get("title") == entry_diff["title"] and
                                exp_entry.get("company") == entry_diff["company"]):
                        if not accepted:
                            exp_entry["bullets"] = entry_diff["original_bullets"]
                        elif edited_bullets is not None:
                            exp_entry["bullets"] = [b for b in edited_bullets if b.strip()]
            break
    return struct
--- a/scripts/scrape_url.py
+++ b/scripts/scrape_url.py
@ -57,7 +57,7 @@ _TIMEOUT = 12
 def _detect_board(url: str) -> str:
-    """Return 'linkedin', 'indeed', 'glassdoor', 'jobgether', 'oracle_hcm', or 'generic'."""
+    """Return 'linkedin', 'indeed', 'glassdoor', or 'generic'."""
    url_lower = url.lower()
    if "linkedin.com" in url_lower:
        return "linkedin"
@ -67,8 +67,6 @@ def _detect_board(url: str) -> str:
        return "glassdoor"
    if "jobgether.com" in url_lower:
        return "jobgether"
    if "oraclecloud.com" in url_lower and "hcmui" in url_lower:
        return "oracle_hcm"
    return "generic"
@ -203,70 +201,6 @@ def _scrape_jobgether(url: str) -> dict:
        return {"company": company, "source": "jobgether"} if company else {}
 def _scrape_oracle_hcm(url: str) -> dict:
    """Scrape an Oracle HCM CandidateExperience job page via Playwright.
    Oracle HCM portals are React SPAs that require JS execution. The prospect
    token in the URL path grants public access — no auth needed.
    """
    try:
        from playwright.sync_api import sync_playwright
    except ImportError:
        print("[scrape_url] Oracle HCM: Playwright not installed, falling back to generic")
        return _scrape_generic(url)
    try:
        with sync_playwright() as p:
            browser = p.chromium.launch(headless=True)
            try:
                ctx = browser.new_context(user_agent=_HEADERS["User-Agent"])
                page = ctx.new_page()
                page.goto(url, timeout=30_000)
                page.wait_for_load_state("networkidle", timeout=20_000)
                result = page.evaluate("""() => {
                    const sel = (s) => document.querySelector(s)?.textContent?.trim() || '';
                    const selInner = (s) => document.querySelector(s)?.innerText?.trim() || '';
                    // Title: try known HCM selectors then fall back to first h1
                    const title = sel('[class*="requisition-title"]')
                        || sel('[class*="JobTitle"]')
                        || sel('.job-title')
                        || sel('h1');
                    // Company: page header logo alt text, meta, or site-name span
                    const companyMeta = document.querySelector('meta[property="og:site_name"]')
                        ?.getAttribute('content') || '';
                    const company = sel('[class*="company-name"]')
                        || sel('[class*="siteName"]')
                        || sel('[class*="site-name"]')
                        || companyMeta;
                    // Location: job detail list items
                    const location = sel('[class*="job-location"]')
                        || sel('[data-testid*="location"]')
                        || sel('[class*="location"]');
                    // Description: main content div
                    const description = selInner('[class*="job-description"]')
                        || selInner('[class*="requisition-description"]')
                        || selInner('[class*="JobDescription"]')
                        || selInner('main article')
                        || selInner('main');
                    return { title, company, location, description };
                }""")
            finally:
                browser.close()
        result["source"] = "oracle_hcm"
        return {k: v for k, v in result.items() if v}
    except Exception as exc:
        print(f"[scrape_url] Oracle HCM Playwright error for {url}: {exc}")
        return {}
 def _parse_json_ld_or_og(html: str) -> dict:
    """Extract job fields from JSON-LD structured data, then og: meta tags."""
    soup = BeautifulSoup(html, "html.parser")
@ -344,8 +278,6 @@ def scrape_job_url(db_path: Path = DEFAULT_DB, job_id: int = None) -> dict:
            fields = _scrape_glassdoor(url)
        elif board == "jobgether":
            fields = _scrape_jobgether(url)
        elif board == "oracle_hcm":
            fields = _scrape_oracle_hcm(url)
        else:
            fields = _scrape_generic(url)
    except requests.RequestException as exc:
--- a/scripts/user_profile.py
+++ b/scripts/user_profile.py
@ -29,7 +29,6 @@ _DEFAULTS = {
    "tier": "free",
    "dev_tier_override": None,
    "wizard_complete": False,
    "training_export_opt_in": False,
    "wizard_step": 0,
    "dismissed_banners": [],
    "ui_preference": "streamlit",
@ -78,7 +77,6 @@ class UserProfile:
        self.tier: str = data.get("tier", "free")
        self.dev_tier_override: str | None = data.get("dev_tier_override") or None
        self.wizard_complete: bool = bool(data.get("wizard_complete", False))
        self.training_export_opt_in: bool = bool(data.get("training_export_opt_in", False))
        self.wizard_step: int = int(data.get("wizard_step", 0))
        self.dismissed_banners: list[str] = list(data.get("dismissed_banners", []))
        raw_pref = data.get("ui_preference", "streamlit")
@ -106,7 +104,6 @@ class UserProfile:
            "tier": self.tier,
            "dev_tier_override": self.dev_tier_override,
            "wizard_complete": self.wizard_complete,
            "training_export_opt_in": self.training_export_opt_in,
            "wizard_step": self.wizard_step,
            "dismissed_banners": self.dismissed_banners,
            "ui_preference": self.ui_preference,
--- a/tests/test_imap_sync.py
+++ b/tests/test_imap_sync.py
@ -203,73 +203,6 @@ def test_parse_linkedin_alert_empty_body():
    assert parse_linkedin_alert("No jobs here.") == []
 # ── Indeed alert parser ───────────────────────────────────────────────────────
 _INDEED_ALERT_HTML = """
 <html><body>
 <table>
  <tr>
    <td>
      <a href="https://www.indeed.com/viewjob?jk=abc123def456&utm_source=jobseeker_email">
        Senior Python Engineer
      </a>
      <br/>Acme Corp<br/>San Francisco, CA<br/>$130,000 - $160,000 a year
    </td>
  </tr>
  <tr>
    <td>
      <a href="https://www.indeed.com/viewjob?jk=999zzzqqq111&trk=email_alert">
        Staff Backend Engineer
      </a>
      <br/>Widgets Inc<br/>Remote
    </td>
  </tr>
  <tr>
    <td>
      <a href="https://www.indeed.com/rc/clk?jk=abc123def456&pos=0">Duplicate link</a>
    </td>
  </tr>
 </table>
 </body></html>
 """
 def test_parse_indeed_alert_extracts_jobs():
    from scripts.imap_sync import parse_indeed_alert
    jobs = parse_indeed_alert(_INDEED_ALERT_HTML)
    assert len(jobs) == 2
    assert jobs[0]["title"] == "Senior Python Engineer"
    assert jobs[0]["url"] == "https://www.indeed.com/viewjob?jk=abc123def456"
    assert jobs[1]["title"] == "Staff Backend Engineer"
    assert jobs[1]["url"] == "https://www.indeed.com/viewjob?jk=999zzzqqq111"
 def test_parse_indeed_alert_strips_tracking_params():
    from scripts.imap_sync import parse_indeed_alert
    jobs = parse_indeed_alert(_INDEED_ALERT_HTML)
    for job in jobs:
        assert "utm_source" not in job["url"]
        assert "trk=" not in job["url"]
 def test_parse_indeed_alert_deduplicates_jk():
    from scripts.imap_sync import parse_indeed_alert
    jobs = parse_indeed_alert(_INDEED_ALERT_HTML)
    urls = [j["url"] for j in jobs]
    assert len(urls) == len(set(urls))
 def test_parse_indeed_alert_empty_body():
    from scripts.imap_sync import parse_indeed_alert
    assert parse_indeed_alert("") == []
    assert parse_indeed_alert("<html><body>No jobs here</body></html>") == []
 def test_parse_indeed_alert_extracts_salary():
    from scripts.imap_sync import parse_indeed_alert
    jobs = parse_indeed_alert(_INDEED_ALERT_HTML)
    assert "$130,000" in jobs[0]["salary"]
 # ── _scan_unmatched_leads integration ─────────────────────────────────────────
 _ALERT_BODY = """\
--- a/tests/test_reranker_integration.py
+++ b/tests/test_reranker_integration.py
@ -1,234 +0,0 @@
 # tests/test_reranker_integration.py
 """Tests for reranker integration in job_ranker and resume_optimizer.
 Set CF_RERANKER_MOCK=1 to avoid loading real model weights during tests.
 """
 import os
 import pytest
 from unittest.mock import patch
 os.environ["CF_RERANKER_MOCK"] = "1"
 # ── Fixtures ──────────────────────────────────────────────────────────────────
 RESUME_TEXT = "Experienced Python engineer with Django REST and data pipeline experience."
 SAMPLE_JOBS = [
    {
        "id": 1,
        "title": "Python Developer",
        "company": "Acme",
        "description": "Python Django REST APIs data engineering pipelines",
        "date_found": "2026-04-01",
        "match_score": 70,
        "salary": "120000",
        "is_remote": 1,
        "location": "Remote",
    },
    {
        "id": 2,
        "title": "Data Analyst",
        "company": "Beta",
        "description": "SQL Excel Tableau business intelligence reporting dashboards",
        "date_found": "2026-04-02",
        "match_score": 60,
        "salary": "90000",
        "is_remote": 1,
        "location": "Remote",
    },
    {
        "id": 3,
        "title": "Frontend Engineer",
        "company": "Gamma",
        "description": "React TypeScript CSS frontend web UI component library",
        "date_found": "2026-04-01",
        "match_score": 50,
        "salary": "110000",
        "is_remote": 1,
        "location": "Remote",
    },
 ]
 SAMPLE_RESUME = {
    "name": "Alex Rivera",
    "email": "alex@example.com",
    "phone": "555-1234",
    "career_summary": "Experienced Python engineer with Django and REST API experience.",
    "skills": ["Python", "Django", "REST APIs"],
    "experience": [
        {
            "title": "Software Engineer",
            "company": "Acme Corp",
            "start_date": "2021",
            "end_date": "present",
            "bullets": ["Built REST APIs.", "Managed data pipelines."],
        }
    ],
    "education": [
        {
            "degree": "B.S.",
            "field": "Computer Science",
            "institution": "State University",
            "graduation_year": "2018",
        }
    ],
    "achievements": [],
 }
 PRIORITIZED_GAPS = [
    {"term": "Kubernetes", "section": "skills", "priority": 1, "rationale": "JD requires K8s"},
    {"term": "REST APIs", "section": "experience", "priority": 2, "rationale": "JD mentions REST"},
    {"term": "CI/CD", "section": "skills", "priority": 3, "rationale": "DevOps signal"},
 ]
 SAMPLE_JOB = {
    "title": "Senior Python Engineer",
    "company": "Acme",
    "description": "Python REST API developer with Kubernetes CI/CD and data pipeline experience.",
 }
 # ── _try_rerank ───────────────────────────────────────────────────────────────
 def test_try_rerank_empty_list():
    from scripts.job_ranker import _try_rerank
    assert _try_rerank("query", []) == []
 def test_try_rerank_returns_all_jobs():
    from scripts.job_ranker import _try_rerank
    result = _try_rerank(RESUME_TEXT, SAMPLE_JOBS)
    assert len(result) == len(SAMPLE_JOBS)
    assert {j["id"] for j in result} == {j["id"] for j in SAMPLE_JOBS}
 def test_try_rerank_preserves_job_fields():
    from scripts.job_ranker import _try_rerank
    result = _try_rerank(RESUME_TEXT, SAMPLE_JOBS)
    for job in result:
        assert "stack_score" not in job or True  # stack_score may or may not be present
        assert "id" in job
        assert "title" in job
 def test_try_rerank_graceful_fallback_on_import_error():
    """_try_rerank falls back to input order when reranker is unavailable."""
    import builtins
    real_import = builtins.__import__
    def mock_import(name, *args, **kwargs):
        if name == "circuitforge_core.reranker":
            raise ImportError("not installed")
        return real_import(name, *args, **kwargs)
    with patch("builtins.__import__", side_effect=mock_import):
        from scripts.job_ranker import _try_rerank as _tr
        result = _tr(RESUME_TEXT, SAMPLE_JOBS)
    assert result == SAMPLE_JOBS
 # ── rank_jobs ─────────────────────────────────────────────────────────────────
 def test_rank_jobs_no_resume_text():
    """rank_jobs without resume_text works normally; reranker not invoked."""
    from scripts.job_ranker import rank_jobs
    results = rank_jobs(SAMPLE_JOBS, ["Python Developer"], limit=10, min_score=0)
    assert len(results) >= 1
    assert all("stack_score" in j for j in results)
 def test_rank_jobs_with_resume_text_returns_same_job_ids():
    """rank_jobs with resume_text returns the same job set (possibly reordered)."""
    from scripts.job_ranker import rank_jobs
    without = rank_jobs(SAMPLE_JOBS, ["Python Developer"], limit=10, min_score=0)
    with_rr = rank_jobs(
        SAMPLE_JOBS, ["Python Developer"], limit=10, min_score=0, resume_text=RESUME_TEXT
    )
    assert {j["id"] for j in without} == {j["id"] for j in with_rr}
 def test_rank_jobs_with_resume_text_preserves_stack_score():
    """stack_score is set on all returned jobs regardless of reranker pass."""
    from scripts.job_ranker import rank_jobs
    results = rank_jobs(
        SAMPLE_JOBS, ["Python Developer"], limit=10, min_score=0, resume_text=RESUME_TEXT
    )
    assert all("stack_score" in j for j in results)
 def test_rank_jobs_limit_respected_after_rerank():
    """limit parameter is still respected when reranker is active."""
    from scripts.job_ranker import rank_jobs
    results = rank_jobs(
        SAMPLE_JOBS, ["Python Developer"], limit=2, min_score=0, resume_text=RESUME_TEXT
    )
    assert len(results) <= 2
 def test_rank_jobs_empty_description_falls_back_to_title():
    """Jobs with no description use title as the reranker candidate string."""
    from scripts.job_ranker import rank_jobs
    jobs_no_desc = [{**j, "description": None} for j in SAMPLE_JOBS]
    results = rank_jobs(
        jobs_no_desc, ["Python Developer"], limit=10, min_score=0, resume_text=RESUME_TEXT
    )
    assert {j["id"] for j in results} == {j["id"] for j in SAMPLE_JOBS}
 # ── rewrite_for_ats (reranker gap reordering) ─────────────────────────────────
 def test_rewrite_for_ats_runs_with_reranker(monkeypatch):
    """rewrite_for_ats completes without error when reranker is active."""
    from scripts.resume_optimizer import rewrite_for_ats
    with patch("scripts.llm_router.LLMRouter") as MockRouter:
        mock_bullets = '["Built REST APIs and Kubernetes pipelines."]'
        MockRouter.return_value.complete.return_value = mock_bullets
        result = rewrite_for_ats(SAMPLE_RESUME, PRIORITIZED_GAPS, SAMPLE_JOB)
    assert isinstance(result, dict)
    assert "skills" in result
    assert "experience" in result
 def test_rewrite_for_ats_preserves_unrewritten_sections(monkeypatch):
    """Sections with no gaps are passed through unchanged."""
    from scripts.resume_optimizer import rewrite_for_ats
    # Only 'skills' gaps — education should be untouched
    skills_only_gaps = [g for g in PRIORITIZED_GAPS if g["section"] == "skills"]
    with patch("scripts.llm_router.LLMRouter") as MockRouter:
        MockRouter.return_value.complete.return_value = '["AWS certified engineer."]'
        result = rewrite_for_ats(SAMPLE_RESUME, skills_only_gaps, SAMPLE_JOB)
    assert result["education"] == SAMPLE_RESUME["education"]
 def test_rewrite_for_ats_reranker_fallback_on_error(monkeypatch):
    """rewrite_for_ats completes even if reranker raises an exception."""
    from scripts.resume_optimizer import rewrite_for_ats
    from circuitforge_core.reranker import reset_reranker
    # Patch rerank to raise so we test the fallback path
    with patch("circuitforge_core.reranker.rerank", side_effect=RuntimeError("boom")):
        with patch("scripts.llm_router.LLMRouter") as MockRouter:
            MockRouter.return_value.complete.return_value = '["Built pipelines."]'
            result = rewrite_for_ats(SAMPLE_RESUME, PRIORITIZED_GAPS, SAMPLE_JOB)
    assert isinstance(result, dict)
    reset_reranker()
 def test_rewrite_for_ats_empty_gaps_no_reranker_call():
    """rewrite_for_ats with empty gaps returns original resume unchanged."""
    from scripts.resume_optimizer import rewrite_for_ats
    with patch("scripts.llm_router.LLMRouter") as MockRouter:
        result = rewrite_for_ats(SAMPLE_RESUME, [], SAMPLE_JOB)
        MockRouter.return_value.complete.assert_not_called()
    assert result["skills"] == SAMPLE_RESUME["skills"]
--- a/tests/test_training_export.py
+++ b/tests/test_training_export.py
@ -1,224 +0,0 @@
 """Tests for cover letter training export helpers."""
 import json
 import sqlite3
 import pytest
 from pathlib import Path
 def _make_db(tmp_path: Path) -> Path:
    from scripts.db import init_db
    db = tmp_path / "test.db"
    init_db(db)
    # excluded_from_training column is added by _migrate_db via _MIGRATIONS — no manual ALTER needed
    return db
 def _insert_job(db: Path, *, title="Engineer", company="Acme", status="applied",
                cover_letter="Dear Hiring Manager,\n\nI am excited.", description="Build stuff.",
                excluded=0) -> int:
    conn = sqlite3.connect(db)
    cur = conn.execute(
        "INSERT INTO jobs (title, company, url, source, location, is_remote, salary, "
        "description, date_found, status, cover_letter, excluded_from_training) "
        "VALUES (?,?,?,?,?,?,?,?,?,?,?,?)",
        (title, company, f"https://example.com/{title}", "test", "Remote", 1, "",
         description, "2026-01-01", status, cover_letter, excluded),
    )
    conn.commit()
    job_id = cur.lastrowid
    conn.close()
    return job_id
 def test_get_training_pairs_returns_applied_jobs(tmp_path):
    from scripts.db import get_training_pairs
    db = _make_db(tmp_path)
    _insert_job(db, title="Engineer", company="Acme", status="applied")
    pairs = get_training_pairs(db)
    assert len(pairs) == 1
    assert pairs[0]["source"] == "db"
    assert pairs[0]["instruction"] == "Write a cover letter for the Engineer position at Acme."
    assert "job_id" in pairs[0]
 def test_get_training_pairs_strips_greeting(tmp_path):
    from scripts.db import get_training_pairs
    db = _make_db(tmp_path)
    _insert_job(db, cover_letter="Dear Hiring Manager,\n\nI am excited to apply.\n\nSincerely, Me")
    pairs = get_training_pairs(db)
    assert not pairs[0]["output"].startswith("Dear")
    assert "I am excited" in pairs[0]["output"]
 def test_get_training_pairs_excludes_non_applied(tmp_path):
    from scripts.db import get_training_pairs
    db = _make_db(tmp_path)
    _insert_job(db, title="PendingJob", status="pending")
    _insert_job(db, title="ApprovedJob", status="approved")
    pairs = get_training_pairs(db)
    assert len(pairs) == 0
 def test_get_training_pairs_excludes_opted_out(tmp_path):
    from scripts.db import get_training_pairs
    db = _make_db(tmp_path)
    _insert_job(db, excluded=1)
    pairs = get_training_pairs(db)
    assert len(pairs) == 0
 def test_get_training_pairs_null_description_gives_empty_input(tmp_path):
    from scripts.db import get_training_pairs
    db = _make_db(tmp_path)
    conn = sqlite3.connect(db)
    conn.execute(
        "INSERT INTO jobs (title, company, url, source, location, is_remote, salary, "
        "date_found, status, cover_letter, excluded_from_training) "
        "VALUES (?,?,?,?,?,?,?,?,?,?,?)",
        ("Dev", "Corp", "https://x.com/1", "test", "Remote", 1, "",
         "2026-01-01", "applied", "Great letter body", 0),
    )
    conn.commit()
    conn.close()
    pairs = get_training_pairs(db)
    assert pairs[0]["input"] == ""
 def test_get_db_pairs_includes_excluded_with_flag(tmp_path):
    from scripts.db import get_db_pairs
    db = _make_db(tmp_path)
    _insert_job(db, excluded=0)
    _insert_job(db, title="Other", excluded=1)
    pairs = get_db_pairs(db)
    assert len(pairs) == 2
    excluded = [p for p in pairs if p["excluded"]]
    included = [p for p in pairs if not p["excluded"]]
    assert len(excluded) == 1
    assert len(included) == 1
 def test_set_training_exclusion_excludes(tmp_path):
    from scripts.db import get_training_pairs, set_training_exclusion
    db = _make_db(tmp_path)
    job_id = _insert_job(db)
    assert len(get_training_pairs(db)) == 1
    set_training_exclusion(db, job_id, excluded=True)
    assert len(get_training_pairs(db)) == 0
 def test_set_training_exclusion_restores(tmp_path):
    from scripts.db import get_training_pairs, set_training_exclusion
    db = _make_db(tmp_path)
    job_id = _insert_job(db, excluded=1)
    assert len(get_training_pairs(db)) == 0
    set_training_exclusion(db, job_id, excluded=False)
    assert len(get_training_pairs(db)) == 1
 def test_strip_greeting_returns_original_when_no_body(tmp_path):
    from scripts.db import _strip_greeting
    # A letter that is only a salutation with no body should return the original text
    result = _strip_greeting("Dear Hiring Manager,")
    assert result == "Dear Hiring Manager,"
 def test_user_profile_training_opt_in_defaults_false(tmp_path):
    from scripts.user_profile import UserProfile
    yaml_path = tmp_path / "user.yaml"
    yaml_path.write_text("name: Test\nemail: test@example.com\n")
    profile = UserProfile(yaml_path)
    assert profile.training_export_opt_in is False
 def test_user_profile_training_opt_in_roundtrip(tmp_path):
    from scripts.user_profile import UserProfile
    yaml_path = tmp_path / "user.yaml"
    yaml_path.write_text("name: Test\nemail: test@example.com\n")
    profile = UserProfile(yaml_path)
    profile.training_export_opt_in = True
    profile.save()
    reloaded = UserProfile(yaml_path)
    assert reloaded.training_export_opt_in is True
 # ── API tests ─────────────────────────────────────────────────────────────────
@pytest.fixture()
 def api_client(tmp_path, monkeypatch):
    """TestClient with a fresh DB and user.yaml for training export endpoints."""
    import yaml
    from fastapi.testclient import TestClient
    db = _make_db(tmp_path)
    yaml_path = tmp_path / "config" / "user.yaml"
    yaml_path.parent.mkdir(parents=True)
    yaml_path.write_text(yaml.dump({"name": "Test", "email": "t@t.com"}))
    monkeypatch.setenv("STAGING_DB", str(db))
    monkeypatch.setattr("dev_api.DB_PATH", str(db))
    monkeypatch.setattr("dev_api._user_yaml_path", lambda: str(yaml_path))
    from dev_api import app
    return TestClient(app), db, yaml_path
 def test_opt_in_toggle(api_client):
    client, db, yaml_path = api_client
    resp = client.patch("/api/settings/fine-tune/opt-in", json={"enabled": True})
    assert resp.status_code == 200
    assert resp.json()["enabled"] is True
    import yaml as _yaml
    data = _yaml.safe_load(yaml_path.read_text())
    assert data["training_export_opt_in"] is True
 def test_db_pairs_blocked_without_opt_in(api_client):
    client, db, yaml_path = api_client
    resp = client.get("/api/settings/fine-tune/db-pairs")
    assert resp.status_code == 403
 def test_db_pairs_returns_jobs_when_opted_in(api_client):
    client, db, yaml_path = api_client
    _insert_job(db, title="Engineer", company="Acme")
    client.patch("/api/settings/fine-tune/opt-in", json={"enabled": True})
    resp = client.get("/api/settings/fine-tune/db-pairs")
    assert resp.status_code == 200
    data = resp.json()
    assert data["total"] >= 1
    assert data["pairs"][0]["title"] == "Engineer"
 def test_exclude_and_restore(api_client):
    client, db, yaml_path = api_client
    job_id = _insert_job(db)
    client.patch("/api/settings/fine-tune/opt-in", json={"enabled": True})
    resp = client.patch(f"/api/settings/fine-tune/db-pairs/{job_id}/exclude")
    assert resp.status_code == 200
    pairs = client.get("/api/settings/fine-tune/db-pairs").json()["pairs"]
    assert any(p["job_id"] == job_id and p["excluded"] for p in pairs)
    client.patch(f"/api/settings/fine-tune/db-pairs/{job_id}/include")
    pairs = client.get("/api/settings/fine-tune/db-pairs").json()["pairs"]
    assert any(p["job_id"] == job_id and not p["excluded"] for p in pairs)
 def test_export_jsonl_blocked_without_opt_in(api_client):
    client, db, yaml_path = api_client
    resp = client.get("/api/settings/fine-tune/export")
    assert resp.status_code == 403
 def test_export_jsonl_streams_valid_records(api_client):
    client, db, yaml_path = api_client
    _insert_job(db, cover_letter="Dear Sir,\n\nGreat role body.", description="Build things.")
    client.patch("/api/settings/fine-tune/opt-in", json={"enabled": True})
    resp = client.get("/api/settings/fine-tune/export")
    assert resp.status_code == 200
    assert "attachment" in resp.headers.get("content-disposition", "")
    lines = [l for l in resp.text.strip().splitlines() if l]
    assert len(lines) >= 1
    record = json.loads(lines[0])
    assert "instruction" in record
    assert "input" in record
    assert "output" in record
    assert record["source"] == "db"
--- a/web/index.html
+++ b/web/index.html
@ -5,20 +5,11 @@
    <link rel="icon" type="image/svg+xml" href="/favicon.svg" />
    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
    <title>Peregrine — Job Search Assistant</title>
-    <!-- Apply stored theme before first paint — prevents FOUT flash on dark/hacker themes.
+    <!-- Inline background prevents blank flash before CSS bundle loads -->
-         Mirrors the logic in useTheme.initTheme(). Must run before the <style> below. -->
+    <!-- Matches --color-surface light / dark from theme.css -->
    <script>try{if(localStorage.getItem('cf-hacker-mode')==='true'){document.documentElement.dataset.theme='hacker';}else{var t=localStorage.getItem('cf-theme');if(t&&t!=='auto')document.documentElement.dataset.theme=t;}}catch(e){}</script>
    <!-- FOUT prevention: background only on html (body is transparent). Gotcha #14.
         body paints on top of html — a hardcoded body background covers html's CSS-
         variable-driven color even when it resolves correctly. Keep background off body.
         Covers auto mode (media query) and all explicit theme choices. -->
    <style>
-      html, body { margin: 0; min-height: 100vh; }
+      html, body { margin: 0; background: #eaeff8; min-height: 100vh; }
-      html { background: #eaeff8; }
+      @media (prefers-color-scheme: dark) { html, body { background: #16202e; } }
      @media (prefers-color-scheme: dark) { html:not([data-theme]) { background: #16202e; } }
      html[data-theme="dark"]          { background: #16202e; }
      html[data-theme="solarized-dark"] { background: #002b36; }
      html[data-theme="hacker"]        { background: #0a0c0a; }
    </style>
    <!-- Plausible analytics: cookie-free, GDPR-compliant, self-hosted.
         Skips localhost/127.0.0.1. Reports to hostname + circuitforge.tech rollup. -->
--- a/web/src/App.vue
+++ b/web/src/App.vue
@ -71,10 +71,6 @@ html {
 }
 body {
  /* Gotcha #14: do NOT set background on body. body paints on top of html —
     a hardcoded body background will cover html's CSS-variable-driven color
     even when html { background: var(--color-surface) } resolves correctly.
     FOUT prevention lives in index.html on html only, not body. */
  min-height: 100dvh;   /* dynamic viewport — mobile chrome-aware. Gotcha #13. */
  overflow-x: hidden;
 }
@ -139,8 +135,8 @@ body {
  bottom: calc(72px + env(safe-area-inset-bottom));
  left: 50%;
  transform: translateX(-50%);
-  background: var(--color-surface-raised, #f5f7fc);
+  background: var(--color-surface-raised, #2a3650);
-  color: var(--color-text, #1a2338);
+  color: var(--color-text, #eaeff8);
  padding: 10px 20px;
  border-radius: var(--radius-md, 8px);
  font-size: 0.9rem;
--- a/web/src/assets/peregrine.css
+++ b/web/src/assets/peregrine.css
@ -58,9 +58,6 @@ body {
  --score-low:      var(--color-error);     /* < 30% */
  --score-none:     var(--color-text-muted);
  /* ── Hover overlay ── */
  --color-hover: rgba(0, 0, 0, 0.06);  /* subtle darkening on light surfaces */
  /* ── Motion tokens ── */
  --swipe-exit:         300ms;
  --swipe-spring:       400ms cubic-bezier(0.34, 1.56, 0.64, 1);
@ -82,7 +79,7 @@ body {
 /* ── Dark mode ─────────────────────────────────────── */
 /* Covers both: OS-level dark preference AND explicit dark theme selection in UI */
@media (prefers-color-scheme: dark) {
-  :root:not([data-theme]) {
+  :root:not([data-theme="hacker"]) {
    --app-primary:       #68A8D8;   /* Falcon Blue (dark) — 6.54:1 on #16202e ✅ AA */
    --app-primary-hover: #7BBDE6;
    --app-primary-light: #0D1F35;
@ -94,8 +91,6 @@ body {
    --score-mid-high:   #5ba3d9;  /* lighter blue for dark bg */
    --color-hover: rgba(255, 255, 255, 0.07);  /* subtle lightening on dark surfaces */
    --status-synced:    #9b8fea;
    --status-survey:    #b08fea;
    --status-phone:     #4ec9be;
@ -117,8 +112,6 @@ body {
  --score-mid-high:   #5ba3d9;
  --color-hover: rgba(255, 255, 255, 0.07);
  --status-synced:    #9b8fea;
  --status-survey:    #b08fea;
  --status-phone:     #4ec9be;
--- a/web/src/assets/theme.css
+++ b/web/src/assets/theme.css
@ -63,9 +63,6 @@
  --shadow-md: 0 4px 12px rgba(26, 35, 56, 0.1),  0 2px 4px rgba(26, 35, 56, 0.06);
  --shadow-lg: 0 10px 30px rgba(26, 35, 56, 0.12), 0 4px 8px rgba(26, 35, 56, 0.06);
  /* Overlay — modal/dialog scrim */
  --color-overlay: rgba(0, 0, 0, 0.5);
  /* Transitions */
  --transition: 200ms ease;
  --transition-slow: 400ms ease;
--- a/web/src/components/AppNav.vue
+++ b/web/src/components/AppNav.vue
@ -59,6 +59,9 @@
        <Cog6ToothIcon class="sidebar__icon" aria-hidden="true" />
        <span class="sidebar__label">Settings</span>
      </RouterLink>
      <button class="sidebar__classic-btn" @click="switchToClassic" title="Switch to Classic (Streamlit) UI">
        ⚡ Classic
      </button>
    </div>
  </nav>
@ -131,6 +134,23 @@ function exitHackerMode() {
  restoreTheme()
 }
 const _apiBase = import.meta.env.BASE_URL.replace(/\/$/, '')
 async function switchToClassic() {
  // Persist preference via API so Streamlit reads streamlit from user.yaml
  // and won't re-set the cookie back to vue (avoids the ?prgn_switch rerun cycle)
  try {
    await fetch(_apiBase + '/api/settings/ui-preference', {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({ preference: 'streamlit' }),
    })
  } catch { /* non-fatal — cookie below is enough for immediate redirect */ }
  document.cookie = 'prgn_ui=streamlit; path=/; SameSite=Lax'
  // Navigate to root (no query params) — Caddy routes to Streamlit based on cookie
  window.location.href = window.location.origin + '/'
 }
 const navLinks = computed(() => [
  { to: '/',           icon: HomeIcon,                   label: 'Home' },
  { to: '/review',     icon: ClipboardDocumentListIcon,  label: 'Job Review' },
@ -301,6 +321,29 @@ const mobileLinks = [
  margin: 0;
 }
 .sidebar__classic-btn {
  display: flex;
  align-items: center;
  width: 100%;
  padding: var(--space-2) var(--space-3);
  margin-top: var(--space-1);
  background: none;
  border: none;
  border-radius: var(--radius-md);
  color: var(--color-text-muted);
  font-size: var(--text-xs);
  font-weight: 500;
  cursor: pointer;
  opacity: 0.6;
  transition: opacity 150ms, background 150ms;
  white-space: nowrap;
 }
 .sidebar__classic-btn:hover {
  opacity: 1;
  background: var(--color-surface-alt);
 }
 /* ── Theme picker ───────────────────────────────────── */
 .sidebar__theme {
  padding: var(--space-2) var(--space-3);
--- a/web/src/components/HintChip.vue
+++ b/web/src/components/HintChip.vue
@ -32,7 +32,7 @@ function dismiss(): void {
  display: flex;
  align-items: flex-start;
  gap: var(--space-2, 8px);
-  background: var(--color-surface, #eaeff8);
+  background: var(--color-surface, #0d1829);
  border: 1px solid var(--app-primary, #2B6CB0);
  border-radius: var(--radius-md, 8px);
  padding: var(--space-2, 8px) var(--space-3, 12px);
@ -59,5 +59,5 @@ function dismiss(): void {
  line-height: 1;
 }
-.hint-chip__dismiss:hover { color: var(--color-text, #1a2338); }
+.hint-chip__dismiss:hover { color: var(--color-text, #eaeff8); }
 </style>
--- a/web/src/components/InterviewCard.vue
+++ b/web/src/components/InterviewCard.vue
@ -151,7 +151,7 @@ async function reclassifySignal(sig: StageSignal, newLabel: StageSignal['stage_s
 }
 const scoreClass = computed(() => {
-  const s = props.job.match_score ?? 0
+  const s = (props.job.match_score ?? 0) * 100
  if (s >= 85) return 'score--high'
  if (s >= 65) return 'score--mid'
  return 'score--low'
@ -159,7 +159,7 @@ const scoreClass = computed(() => {
 const scoreLabel = computed(() =>
  props.job.match_score != null
-    ? `${Math.round(props.job.match_score)}%`
+    ? `${Math.round(props.job.match_score * 100)}%`
    : '—'
 )
@ -588,7 +588,7 @@ async function saveFeedback() {
  background: var(--color-hover);
 }
 .btn-chip-active {
-  background: var(--app-primary-light);
+  background: var(--color-primary-muted, #e8f0ff);
  color: var(--color-primary); border-color: var(--color-primary);
  font-weight: 600;
 }
--- a/web/src/components/ResumeLibraryCard.vue
+++ b/web/src/components/ResumeLibraryCard.vue
@ -106,7 +106,7 @@ onMounted(load)
 }
 .rlc__title {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  font-weight: 600;
  margin: 0;
  display: flex;
@ -128,7 +128,7 @@ onMounted(load)
 .rlc__name {
  font-weight: 500;
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
 }
 .rlc__meta {
@ -143,7 +143,7 @@ onMounted(load)
 }
 .rlc__empty {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--color-text-muted, #64748b);
 }
@ -153,7 +153,7 @@ onMounted(load)
 }
 .rlc__loading {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--color-text-muted, #64748b);
 }
@ -183,7 +183,7 @@ onMounted(load)
  padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem);
  cursor: pointer;
  border-radius: var(--radius-sm, 0.25rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
 }
 .rlc__picker-item:hover,
--- a/web/src/components/ResumeOptimizerPanel.vue
+++ b/web/src/components/ResumeOptimizerPanel.vue
@ -112,15 +112,16 @@
            <span class="rop__preview-badge">Preview — not yet saved</span>
          </div>
          <textarea
-            v-model="previewText"
+            :value="previewText"
-            class="rop__textarea"
+            class="rop__textarea rop__textarea--preview"
-            aria-label="Resume preview — editable before approving"
+            aria-label="Resume preview text"
-            spellcheck="true"
+            spellcheck="false"
            readonly
          />
          <p class="rop__preview-hint">
-            Review and edit the assembled resume above. Click
+            Review the assembled resume above. If it looks right, click
-            <strong>Approve &amp; Save</strong> to lock it in, or go back to adjust
+            <strong>Approve &amp; Save</strong> to lock it in. You can also go back and adjust
-            your section-level decisions.
+            your review decisions.
          </p>
          <div class="rop__save-to-library">
            <label class="rop__save-toggle">
@ -491,10 +492,7 @@ async function approveResume() {
  if (!previewStruct.value) return
  approvingResume.value = true
-  const body: Record<string, unknown> = {
+  const body: Record<string, unknown> = { preview_struct: previewStruct.value }
    preview_struct: previewStruct.value,
    preview_text_override: previewText.value,
  }
  if (saveToLibrary.value) {
    body.save_to_library = true
    body.resume_name = savedResumeName.value.trim() || `Optimized for job ${props.jobId}`
@ -578,7 +576,7 @@ onUnmounted(stopPolling)
 }
 .rop__tier-note {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-text-muted, #64748b);
  background: var(--app-surface-alt, #f8fafc);
  border: 1px solid var(--app-border, #e2e8f0);
@ -603,13 +601,13 @@ onUnmounted(stopPolling)
 .rop__hint,
 .rop__empty {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-text-muted, #64748b);
  margin: 0;
 }
 .rop__error {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-danger, #dc2626);
  margin: 0;
 }
@ -618,7 +616,7 @@ onUnmounted(stopPolling)
  display: flex;
  align-items: center;
  gap: var(--space-2, 0.5rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-text-muted, #64748b);
 }
@ -643,7 +641,7 @@ onUnmounted(stopPolling)
  border-radius: var(--radius-sm, 0.25rem);
  border-left: 3px solid transparent;
  background: var(--app-surface-alt, #f8fafc);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  cursor: pointer;
  user-select: none;
 }
@ -706,7 +704,7 @@ onUnmounted(stopPolling)
 }
 .rop__wordcount {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-text-muted, #64748b);
 }
@ -727,7 +725,7 @@ onUnmounted(stopPolling)
  background: color-mix(in srgb, var(--app-danger, #dc2626) 8%, transparent);
  border: 1px solid color-mix(in srgb, var(--app-danger, #dc2626) 30%, transparent);
  border-radius: var(--radius-md, 0.5rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-danger, #dc2626);
 }
@ -736,7 +734,7 @@ onUnmounted(stopPolling)
  min-height: 20rem;
  padding: var(--space-3, 0.75rem);
  font-family: var(--font-mono, monospace);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  line-height: 1.6;
  border: 1px solid var(--app-border, #e2e8f0);
  border-radius: var(--radius-md, 0.5rem);
@ -762,7 +760,7 @@ onUnmounted(stopPolling)
  color: #fff;
  border: none;
  border-radius: var(--radius-md, 0.5rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  font-weight: 500;
  cursor: pointer;
  transition: background 0.15s;
@ -781,7 +779,7 @@ onUnmounted(stopPolling)
  color: var(--app-text, #1e293b);
  border: 1px solid var(--app-border, #e2e8f0);
  border-radius: var(--radius-md, 0.5rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  font-weight: 500;
  cursor: pointer;
  transition: background 0.15s;
@ -799,7 +797,7 @@ onUnmounted(stopPolling)
 }
 .rop__review-intro {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-text-muted, #64748b);
  margin: 0;
  padding: var(--space-3, 0.75rem) var(--space-4, 1rem);
@ -819,7 +817,7 @@ onUnmounted(stopPolling)
 }
 .rop__review-section-title {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  font-weight: 600;
  text-transform: uppercase;
  letter-spacing: 0.05em;
@ -852,7 +850,7 @@ onUnmounted(stopPolling)
  gap: var(--space-1, 0.25rem);
  padding: 0.3em 0.75em;
  border-radius: var(--radius-full, 9999px);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  border: 1.5px solid var(--app-border, #e2e8f0);
  background: var(--app-surface, #fff);
  cursor: pointer;
@ -887,7 +885,7 @@ onUnmounted(stopPolling)
  gap: var(--space-1, 0.25rem);
  padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem);
  border-radius: var(--radius-sm, 0.25rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
 }
 .rop__diff-col--original {
@ -936,7 +934,7 @@ onUnmounted(stopPolling)
 }
 .rop__exp-company {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-text-muted, #64748b);
 }
@ -945,7 +943,7 @@ onUnmounted(stopPolling)
  display: inline-flex;
  align-items: center;
  gap: var(--space-1, 0.25rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  cursor: pointer;
  color: var(--app-text, #1e293b);
 }
@ -975,7 +973,7 @@ onUnmounted(stopPolling)
  background: none;
  border: none;
  color: var(--app-accent, #6366f1);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  cursor: pointer;
  padding: 0;
 }
@ -994,7 +992,7 @@ onUnmounted(stopPolling)
  background: var(--app-surface-alt, #f8fafc);
  border: 1px solid var(--app-border, #e2e8f0);
  border-radius: var(--radius-sm, 0.25rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
 }
 .rop__history-date {
@ -1060,7 +1058,7 @@ onUnmounted(stopPolling)
  display: inline-flex;
  align-items: center;
  gap: var(--space-1, 0.25rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  cursor: pointer;
  color: var(--app-text, #1e293b);
 }
@ -1068,7 +1066,7 @@ onUnmounted(stopPolling)
 .rop__framing-context {
  width: 100%;
  padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  font-family: inherit;
  line-height: 1.5;
  border: 1px solid var(--app-border, #e2e8f0);
@ -1101,9 +1099,13 @@ onUnmounted(stopPolling)
  border-radius: var(--radius-full, 9999px);
 }
 .rop__textarea--preview {
  background: color-mix(in srgb, var(--app-accent, #6366f1) 3%, var(--app-surface, #fff));
  cursor: default;
 }
 .rop__preview-hint {
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-text-muted, #64748b);
  margin: 0;
 }
@ -1132,7 +1134,7 @@ onUnmounted(stopPolling)
  color: var(--app-text-muted, #64748b);
  border: 1px solid var(--app-border, #e2e8f0);
  border-radius: var(--radius-md, 0.5rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  cursor: pointer;
  transition: background 0.15s;
 }
@ -1163,7 +1165,7 @@ onUnmounted(stopPolling)
  display: inline-flex;
  align-items: center;
  gap: var(--space-2, 0.5rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  color: var(--app-text, #1e293b);
  cursor: pointer;
  user-select: none;
@ -1171,7 +1173,7 @@ onUnmounted(stopPolling)
 .rop__resume-name-input {
  padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem);
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  font-family: inherit;
  border: 1px solid var(--app-border, #e2e8f0);
  border-radius: var(--radius-sm, 0.25rem);
--- a/web/src/components/ResumeReviewModal.vue
+++ b/web/src/components/ResumeReviewModal.vue
@ -63,9 +63,7 @@
            <SummaryPage
              :section="summarySection!"
              :accepted="summaryAccepted"
              :edited-proposed="summaryEdited"
              @update:accepted="summaryAccepted = $event"
              @update:editedProposed="summaryEdited = $event"
            />
          </template>
@ -74,9 +72,7 @@
            <ExperiencePage
              :entry="currentEntry!"
              :accepted="expAccepted[currentPage.entryKey!] ?? true"
              :edited-bullets="expEdited[currentPage.entryKey!] ?? currentEntry!.proposed_bullets"
              @update:accepted="expAccepted[currentPage.entryKey!] = $event"
              @update:editedBullets="expEdited[currentPage.entryKey!] = $event"
            />
          </template>
@ -259,17 +255,11 @@ function goTo(idx: number) {
 const approvedSkills = ref<Set<string>>(new Set(skillsSection.value?.added ?? []))
 const skillFramings = ref<Map<string, GapFraming>>(new Map())
 const summaryAccepted = ref(true)
 const summaryEdited = ref<string>(summarySection.value?.proposed ?? '')
 const expAccepted = ref<Record<string, boolean>>(
  Object.fromEntries(
    (expSection.value?.entries ?? []).map(e => [`${e.title}|${e.company}`, true])
  )
 )
 const expEdited = ref<Record<string, string[]>>(
  Object.fromEntries(
    (expSection.value?.entries ?? []).map(e => [`${e.title}|${e.company}`, [...e.proposed_bullets]])
  )
 )
 function toggleSkill(skill: string) {
  interactedPages.value = new Set([...interactedPages.value, 'skills'])
@ -332,22 +322,15 @@ function emitSubmit() {
    decisions.skills = { approved_additions: [...approvedSkills.value] }
  }
  if (summarySection.value) {
-    decisions.summary = {
+    decisions.summary = { accepted: summaryAccepted.value }
      accepted: summaryAccepted.value,
      edited_text: summaryEdited.value,
    }
  }
  if (expSection.value) {
    decisions.experience = {
-      accepted_entries: expSection.value.entries.map(e => {
+      accepted_entries: expSection.value.entries.map(e => ({
        const key = `${e.title}|${e.company}`
        return {
        title: e.title,
        company: e.company,
-          accepted: expAccepted.value[key] ?? true,
+        accepted: expAccepted.value[`${e.title}|${e.company}`] ?? true,
-          edited_bullets: expEdited.value[key] ?? e.proposed_bullets,
+      })),
        }
      }),
    }
  }
--- a/web/src/components/ResumeSyncConfirmModal.vue
+++ b/web/src/components/ResumeSyncConfirmModal.vue
@ -131,7 +131,7 @@ defineEmits<{
  background: var(--color-error, #dc2626);
  color: #fff; border: none;
  border-radius: var(--radius-md); cursor: pointer;
-  font-size: var(--text-sm); font-weight: 600;
+  font-size: var(--font-sm); font-weight: 600;
 }
 .btn-danger:hover { filter: brightness(1.1); }
 .btn-secondary {
@ -140,7 +140,7 @@ defineEmits<{
  color: var(--color-text);
  border: 1px solid var(--color-border);
  border-radius: var(--radius-md); cursor: pointer;
-  font-size: var(--text-sm);
+  font-size: var(--font-sm);
 }
 .btn-secondary:hover { background: var(--color-surface-alt); }
 </style>
--- a/web/src/components/resume-review/ConfirmPage.vue
+++ b/web/src/components/resume-review/ConfirmPage.vue
@ -46,11 +46,11 @@ const emit = defineEmits<{
 <style scoped>
 .rp-confirm { display: flex; flex-direction: column; gap: var(--space-4, 1rem); }
 .rp__heading { font-size: var(--font-lg, 1.125rem); font-weight: 600; margin: 0; color: var(--color-text, #1a2338); }
-.rp__hint { font-size: var(--text-sm); color: var(--color-text-muted, #4a5c7a); margin: 0; }
+.rp__hint { font-size: var(--font-sm, 0.875rem); color: var(--color-text-muted, #4a5c7a); margin: 0; }
 .rp-confirm__list { list-style: none; padding: 0; margin: 0; display: flex; flex-direction: column; gap: var(--space-2, 0.5rem); }
-.rp-confirm__item { display: flex; align-items: center; gap: var(--space-3, 0.75rem); padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem); background: var(--color-surface-alt, #dde4f0); border-radius: var(--radius-sm, 0.25rem); font-size: var(--text-sm); }
+.rp-confirm__item { display: flex; align-items: center; gap: var(--space-3, 0.75rem); padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem); background: var(--color-surface-alt, #dde4f0); border-radius: var(--radius-sm, 0.25rem); font-size: var(--font-sm, 0.875rem); }
 .rp-confirm__status { margin-left: auto; font-size: var(--font-xs, 0.75rem); color: var(--color-text-muted, #4a5c7a); text-transform: capitalize; }
-.rp__error { color: var(--color-error, #c0392b); font-size: var(--text-sm); margin: 0; }
+.rp__error { color: var(--color-error, #c0392b); font-size: var(--font-sm, 0.875rem); margin: 0; }
 .rp-confirm__actions { display: flex; gap: var(--space-3, 0.75rem); flex-wrap: wrap; }
 .tab__dot { width: 8px; height: 8px; border-radius: 50%; flex-shrink: 0; background: var(--tab-color, #94a3b8); }
 .tab__dot--unvisited   { --tab-color: var(--color-text-muted, #94a3b8); }
@ -64,7 +64,7 @@ const emit = defineEmits<{
  padding: var(--space-3, 0.75rem) var(--space-4, 1rem);
  background: var(--color-accent, #c4732a); color: #fff;
  border: none; border-radius: var(--radius-md, 0.5rem);
-  font-size: var(--text-sm); font-weight: 600; cursor: pointer;
+  font-size: var(--font-sm, 0.875rem); font-weight: 600; cursor: pointer;
 }
 .btn-generate:disabled { opacity: 0.6; cursor: not-allowed; }
 .btn-secondary {
@ -72,6 +72,6 @@ const emit = defineEmits<{
  padding: var(--space-3, 0.75rem) var(--space-4, 1rem);
  background: var(--color-surface-alt, #dde4f0); color: var(--color-text, #1a2338);
  border: 1px solid var(--color-border, #a8b8d0); border-radius: var(--radius-md, 0.5rem);
-  font-size: var(--text-sm); font-weight: 600; cursor: pointer;
+  font-size: var(--font-sm, 0.875rem); font-weight: 600; cursor: pointer;
 }
 </style>
--- a/web/src/components/resume-review/ExperiencePage.vue
+++ b/web/src/components/resume-review/ExperiencePage.vue
@ -9,25 +9,11 @@
          <li v-for="b in entry.original_bullets" :key="b">{{ b }}</li>
        </ul>
      </div>
-      <div class="rp__diff-col rp__diff-col--editable">
+      <div class="rp__diff-col">
-        <span class="rp__diff-label">Proposed — edit below</span>
+        <span class="rp__diff-label">Proposed</span>
-        <div class="rp__bullet-edit-list" role="list" :aria-label="`Edit proposed bullets for ${entry.title}`">
+        <ul class="rp__bullet-list">
-          <div
+          <li v-for="b in entry.proposed_bullets" :key="b">{{ b }}</li>
-            v-for="(bullet, idx) in editedBullets"
+        </ul>
            :key="idx"
            class="rp__bullet-edit-row"
            role="listitem"
          >
            <textarea
              class="rp__bullet-textarea"
              :value="bullet"
              :aria-label="`Bullet ${idx + 1}`"
              rows="2"
              spellcheck="true"
              @input="updateBullet(idx, ($event.target as HTMLTextAreaElement).value)"
            />
          </div>
        </div>
      </div>
    </div>
    <label class="rp__accept-toggle">
@ -42,7 +28,7 @@
 </template>
 <script setup lang="ts">
-const props = defineProps<{
+defineProps<{
  entry: {
    title: string
    company: string
@ -50,46 +36,21 @@ const props = defineProps<{
    proposed_bullets: string[]
  }
  accepted: boolean
  editedBullets: string[]
 }>()
 const emit = defineEmits<{
  'update:accepted': [v: boolean]
  'update:editedBullets': [v: string[]]
 }>()
 function updateBullet(idx: number, value: string) {
  const next = props.editedBullets.map((b, i) => (i === idx ? value : b))
  emit('update:editedBullets', next)
 }
 </script>
 <style scoped>
 .rp-exp { display: flex; flex-direction: column; gap: var(--space-4, 1rem); }
 .rp__heading { font-size: var(--font-lg, 1.125rem); font-weight: 600; margin: 0; color: var(--color-text, #1a2338); }
-.rp__company { font-size: var(--text-sm); color: var(--color-text-muted, #4a5c7a); margin: 0; }
+.rp__company { font-size: var(--font-sm, 0.875rem); color: var(--color-text-muted, #4a5c7a); margin: 0; }
 .rp__diff-pair { display: grid; grid-template-columns: 1fr 1fr; gap: var(--space-4, 1rem); }
@media (max-width: 600px) { .rp__diff-pair { grid-template-columns: 1fr; } }
 .rp__diff-col { display: flex; flex-direction: column; gap: var(--space-2, 0.5rem); }
 .rp__diff-col--editable { gap: var(--space-2, 0.5rem); }
 .rp__diff-label { font-size: var(--font-xs, 0.75rem); font-weight: 600; text-transform: uppercase; letter-spacing: 0.05em; color: var(--color-text-muted, #4a5c7a); }
-.rp__bullet-list { margin: 0; padding-left: var(--space-4, 1rem); font-size: var(--text-sm); line-height: 1.6; background: var(--color-surface-alt, #dde4f0); border-radius: var(--radius-sm, 0.25rem); padding: var(--space-3, 0.75rem) var(--space-3, 0.75rem) var(--space-3, 0.75rem) var(--space-6, 1.5rem); }
+.rp__bullet-list { margin: 0; padding-left: var(--space-4, 1rem); font-size: var(--font-sm, 0.875rem); line-height: 1.6; background: var(--color-surface-alt, #dde4f0); border-radius: var(--radius-sm, 0.25rem); padding: var(--space-3, 0.75rem) var(--space-3, 0.75rem) var(--space-3, 0.75rem) var(--space-6, 1.5rem); }
-.rp__bullet-edit-list { display: flex; flex-direction: column; gap: var(--space-2, 0.5rem); }
+.rp__accept-toggle { display: inline-flex; align-items: center; gap: var(--space-2, 0.5rem); cursor: pointer; font-size: var(--font-sm, 0.875rem); }
 .rp__bullet-edit-row { display: flex; align-items: flex-start; gap: var(--space-1, 0.25rem); }
 .rp__bullet-textarea {
  flex: 1;
  font-size: var(--text-sm);
  line-height: 1.5;
  padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem);
  background: var(--color-surface, #eaeff8);
  border: 1.5px solid var(--color-accent, #c4732a);
  border-radius: var(--radius-sm, 0.25rem);
  color: var(--color-text, #1a2338);
  resize: vertical;
  width: 100%;
  box-sizing: border-box;
  font-family: inherit;
 }
 .rp__bullet-textarea:focus { outline: 2px solid var(--color-accent, #c4732a); outline-offset: 2px; }
 .rp__accept-toggle { display: inline-flex; align-items: center; gap: var(--space-2, 0.5rem); cursor: pointer; font-size: var(--text-sm); }
 </style>
--- a/web/src/components/resume-review/SkillsPage.vue
+++ b/web/src/components/resume-review/SkillsPage.vue
@ -58,7 +58,7 @@ const emit = defineEmits<{
 <style scoped>
 .rp-skills { display: flex; flex-direction: column; gap: var(--space-4, 1rem); }
 .rp__heading { font-size: var(--font-lg, 1.125rem); font-weight: 600; margin: 0; color: var(--color-text, #1a2338); }
-.rp__hint { font-size: var(--text-sm); color: var(--color-text-muted, #4a5c7a); margin: 0; }
+.rp__hint { font-size: var(--font-sm, 0.875rem); color: var(--color-text-muted, #4a5c7a); margin: 0; }
 .rp__skill-list { display: flex; flex-direction: column; gap: var(--space-3, 0.75rem); }
 .rp__skill-group { display: flex; flex-direction: column; gap: var(--space-2, 0.5rem); }
 .rp__skill-chip {
@ -66,13 +66,13 @@ const emit = defineEmits<{
  padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem);
  border: 1px solid var(--color-border, #a8b8d0);
  border-radius: var(--radius-md, 0.5rem);
-  cursor: pointer; font-size: var(--text-sm);
+  cursor: pointer; font-size: var(--font-sm, 0.875rem);
  background: var(--color-surface-raised, #f5f7fc);
  transition: background var(--transition, 200ms ease);
 }
 .rp__skill-chip--approved { background: var(--color-primary-light, #e8f2e7); border-color: var(--color-primary, #2d5a27); }
 .rp__framing { padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem); display: flex; flex-direction: column; gap: var(--space-2, 0.5rem); background: var(--color-surface-alt, #dde4f0); border-radius: var(--radius-sm, 0.25rem); }
 .rp__framing-label { font-size: var(--font-xs, 0.75rem); font-weight: 600; color: var(--color-text-muted, #4a5c7a); }
-.rp__framing-context { border: 1px solid var(--color-border, #a8b8d0); border-radius: var(--radius-sm, 0.25rem); padding: var(--space-2, 0.5rem); font-size: var(--text-sm); resize: vertical; }
+.rp__framing-context { border: 1px solid var(--color-border, #a8b8d0); border-radius: var(--radius-sm, 0.25rem); padding: var(--space-2, 0.5rem); font-size: var(--font-sm, 0.875rem); resize: vertical; }
-.rp__removed { font-size: var(--text-sm); color: var(--color-text-muted, #4a5c7a); font-style: italic; }
+.rp__removed { font-size: var(--font-sm, 0.875rem); color: var(--color-text-muted, #4a5c7a); font-style: italic; }
 </style>
--- a/web/src/components/resume-review/SummaryPage.vue
+++ b/web/src/components/resume-review/SummaryPage.vue
@ -6,15 +6,9 @@
        <span class="rp__diff-label" aria-label="Original">Original</span>
        <p class="rp__diff-text">{{ section.original || '(empty)' }}</p>
      </div>
-      <div class="rp__diff-col rp__diff-col--editable">
+      <div class="rp__diff-col">
-        <span class="rp__diff-label" aria-label="Proposed — editable">Proposed</span>
+        <span class="rp__diff-label" aria-label="Proposed">Proposed</span>
-        <textarea
+        <p class="rp__diff-text">{{ section.proposed }}</p>
          class="rp__edit-textarea"
          :value="editedProposed"
          :aria-label="`Edit proposed summary`"
          spellcheck="true"
          @input="emit('update:editedProposed', ($event.target as HTMLTextAreaElement).value)"
        />
      </div>
    </div>
    <label class="rp__accept-toggle">
@ -34,12 +28,10 @@ import type { TextDiff } from '../ResumeReviewModal.vue'
 defineProps<{
  section: TextDiff
  accepted: boolean
  editedProposed: string
 }>()
 const emit = defineEmits<{
  'update:accepted': [v: boolean]
  'update:editedProposed': [v: string]
 }>()
 </script>
@ -49,23 +41,7 @@ const emit = defineEmits<{
 .rp__diff-pair { display: grid; grid-template-columns: 1fr 1fr; gap: var(--space-4, 1rem); }
@media (max-width: 600px) { .rp__diff-pair { grid-template-columns: 1fr; } }
 .rp__diff-col { display: flex; flex-direction: column; gap: var(--space-2, 0.5rem); }
 .rp__diff-col--editable { gap: var(--space-2, 0.5rem); }
 .rp__diff-label { font-size: var(--font-xs, 0.75rem); font-weight: 600; text-transform: uppercase; letter-spacing: 0.05em; color: var(--color-text-muted, #4a5c7a); }
-.rp__diff-text { font-size: var(--text-sm); line-height: 1.6; padding: var(--space-3, 0.75rem); background: var(--color-surface-alt, #dde4f0); border-radius: var(--radius-sm, 0.25rem); margin: 0; }
+.rp__diff-text { font-size: var(--font-sm, 0.875rem); line-height: 1.6; padding: var(--space-3, 0.75rem); background: var(--color-surface-alt, #dde4f0); border-radius: var(--radius-sm, 0.25rem); margin: 0; }
-.rp__edit-textarea {
+.rp__accept-toggle { display: inline-flex; align-items: center; gap: var(--space-2, 0.5rem); cursor: pointer; font-size: var(--font-sm, 0.875rem); }
  font-size: var(--text-sm);
  line-height: 1.6;
  padding: var(--space-3, 0.75rem);
  background: var(--color-surface, #eaeff8);
  border: 1.5px solid var(--color-accent, #c4732a);
  border-radius: var(--radius-sm, 0.25rem);
  color: var(--color-text, #1a2338);
  resize: vertical;
  min-height: 7rem;
  width: 100%;
  box-sizing: border-box;
  font-family: inherit;
 }
 .rp__edit-textarea:focus { outline: 2px solid var(--color-accent, #c4732a); outline-offset: 2px; }
 .rp__accept-toggle { display: inline-flex; align-items: center; gap: var(--space-2, 0.5rem); cursor: pointer; font-size: var(--text-sm); }
 </style>
--- a/web/src/router/index.ts
+++ b/web/src/router/index.ts
@ -46,7 +46,6 @@ export const router = createRouter({
        { path: 'hardware',   component: () => import('../views/wizard/WizardHardwareStep.vue') },
        { path: 'tier',       component: () => import('../views/wizard/WizardTierStep.vue') },
        { path: 'resume',     component: () => import('../views/wizard/WizardResumeStep.vue') },
        { path: 'training',   component: () => import('../views/wizard/WizardTrainingStep.vue') },
        { path: 'identity',   component: () => import('../views/wizard/WizardIdentityStep.vue') },
        { path: 'inference',  component: () => import('../views/wizard/WizardInferenceStep.vue') },
        { path: 'search',     component: () => import('../views/wizard/WizardSearchStep.vue') },
--- a/web/src/stores/prep.test.ts
+++ b/web/src/stores/prep.test.ts
@ -27,7 +27,6 @@ describe('usePrepStore', () => {
        tech_brief: null, funding_brief: null, red_flags: null, accessibility_brief: null,
        generated_at: '2026-03-20T12:00:00' }, error: null })  // research
      .mockResolvedValueOnce({ data: [], error: null })          // contacts
      .mockResolvedValueOnce({ data: [], error: null })          // qa
      .mockResolvedValueOnce({ data: { status: 'none', stage: null, message: null }, error: null }) // task
      .mockResolvedValueOnce({ data: { id: 1, title: 'Engineer', company: 'Acme', url: null,
        description: 'Build things.', cover_letter: null, match_score: 80,
@ -51,7 +50,6 @@ describe('usePrepStore', () => {
        tech_brief: null, funding_brief: null, red_flags: null, accessibility_brief: null,
        generated_at: null }, error: null })
      .mockResolvedValueOnce({ data: [], error: null })
      .mockResolvedValueOnce({ data: [], error: null })          // qa
      .mockResolvedValueOnce({ data: { status: 'none', stage: null, message: null }, error: null })
      .mockResolvedValueOnce({ data: { id: 1, title: 'Old Job', company: 'OldCo', url: null,
        description: null, cover_letter: null, match_score: null, keyword_gaps: null }, error: null })
@ -64,7 +62,6 @@ describe('usePrepStore', () => {
    mockApiFetch
      .mockResolvedValueOnce({ data: null, error: { kind: 'http', status: 404, detail: '' } })  // 404 → null
      .mockResolvedValueOnce({ data: [], error: null })
      .mockResolvedValueOnce({ data: [], error: null })          // qa
      .mockResolvedValueOnce({ data: { status: 'none', stage: null, message: null }, error: null })
      .mockResolvedValueOnce({ data: { id: 2, title: 'New Job', company: 'NewCo', url: null,
        description: null, cover_letter: null, match_score: null, keyword_gaps: null }, error: null })
@ -105,7 +102,6 @@ describe('usePrepStore', () => {
        tech_brief: null, funding_brief: null, red_flags: null, accessibility_brief: null,
        generated_at: null }, error: null })
      .mockResolvedValueOnce({ data: [], error: null })
      .mockResolvedValueOnce({ data: [], error: null })          // qa
      .mockResolvedValueOnce({ data: { status: 'none', stage: null, message: null }, error: null })
      .mockResolvedValueOnce({ data: { id: 1, title: 'Eng', company: 'Acme', url: null,
        description: null, cover_letter: null, match_score: null, keyword_gaps: null }, error: null })
@ -116,12 +112,11 @@ describe('usePrepStore', () => {
    // Mock first poll → completed
    mockApiFetch
      .mockResolvedValueOnce({ data: { status: 'completed', stage: null, message: null }, error: null })
-      // re-fetch on completed: research, contacts, qa, task, fullJob
+      // re-fetch on completed: research, contacts, task, fullJob
      .mockResolvedValueOnce({ data: { company_brief: 'Updated!', ceo_brief: null, talking_points: null,
        tech_brief: null, funding_brief: null, red_flags: null, accessibility_brief: null,
        generated_at: '2026-03-20T13:00:00' }, error: null })
      .mockResolvedValueOnce({ data: [], error: null })
      .mockResolvedValueOnce({ data: [], error: null })          // qa
      .mockResolvedValueOnce({ data: { status: 'completed', stage: null, message: null }, error: null })
      .mockResolvedValueOnce({ data: { id: 1, title: 'Eng', company: 'Acme', url: null,
        description: 'Now with content', cover_letter: null, match_score: null, keyword_gaps: null }, error: null })
@ -139,7 +134,6 @@ describe('usePrepStore', () => {
        tech_brief: null, funding_brief: null, red_flags: null, accessibility_brief: null,
        generated_at: null }, error: null })
      .mockResolvedValueOnce({ data: [], error: null })
      .mockResolvedValueOnce({ data: [], error: null })          // qa
      .mockResolvedValueOnce({ data: { status: 'none', stage: null, message: null }, error: null })
      .mockResolvedValueOnce({ data: { id: 1, title: 'Eng', company: 'Acme', url: null,
        description: null, cover_letter: null, match_score: null, keyword_gaps: null }, error: null })
@ -168,7 +162,6 @@ describe('usePrepStore', () => {
        tech_brief: null, funding_brief: null, red_flags: null, accessibility_brief: null,
        generated_at: '2026-03-20T12:00:00' }, error: null })  // research OK
      .mockResolvedValueOnce({ data: null, error: { kind: 'http', status: 500, detail: 'DB error' } }) // contacts fail
      .mockResolvedValueOnce({ data: [], error: null })          // qa
      .mockResolvedValueOnce({ data: { status: 'none', stage: null, message: null }, error: null })    // task OK
      .mockResolvedValueOnce({ data: { id: 1, title: 'Engineer', company: 'Acme', url: null,
        description: 'Build things.', cover_letter: null, match_score: 80,
--- a/web/src/stores/settings/fineTune.test.ts
+++ b/web/src/stores/settings/fineTune.test.ts
@ -1,11 +1,8 @@
 import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest'
 import { setActivePinia, createPinia } from 'pinia'
 import { useFineTuneStore } from './fineTune'
 import type { DbPair } from './fineTune'
 vi.mock('../../composables/useApi', () => ({ useApiFetch: vi.fn() }))
 vi.mock('../appConfig', () => ({ useAppConfigStore: vi.fn(() => ({ isDemo: false })) }))
 vi.mock('../../composables/useToast', () => ({ showToast: vi.fn() }))
 import { useApiFetch } from '../../composables/useApi'
 const mockFetch = vi.mocked(useApiFetch)
@ -39,47 +36,4 @@ describe('useFineTuneStore', () => {
    expect(mockFetch).toHaveBeenCalledWith('/api/settings/fine-tune/status')
    store.stopPolling()
  })
  it('toggleOptIn updates optedIn state', async () => {
    mockFetch.mockResolvedValue({ data: { ok: true, enabled: true }, error: null })
    const store = useFineTuneStore()
    await store.toggleOptIn(true)
    expect(store.optedIn).toBe(true)
  })
  it('loadDbPairs no-ops when not opted in', async () => {
    const store = useFineTuneStore()
    store.optedIn = false
    await store.loadDbPairs()
    expect(store.dbPairs).toEqual([])
    expect(mockFetch).not.toHaveBeenCalledWith('/api/settings/fine-tune/db-pairs')
  })
  it('loadDbPairs fetches when opted in', async () => {
    const pairs: DbPair[] = [{ job_id: 1, title: 'Eng', company: 'Acme', status: 'applied', instruction: 'Write...', input_preview: 'Build', excluded: false }]
    mockFetch.mockResolvedValue({ data: { pairs, total: 1, excluded_count: 0 }, error: null })
    const store = useFineTuneStore()
    store.optedIn = true
    await store.loadDbPairs()
    expect(store.dbPairs).toHaveLength(1)
  })
  it('excludeDbPair marks pair excluded and increments count', async () => {
    mockFetch.mockResolvedValue({ data: { ok: true }, error: null })
    const store = useFineTuneStore()
    store.dbPairs = [{ job_id: 1, title: 'Eng', company: 'Acme', status: 'applied', instruction: 'Write...', input_preview: 'Build', excluded: false }]
    await store.excludeDbPair(1)
    expect(store.dbPairs[0].excluded).toBe(true)
    expect(store.dbExcludedCount).toBe(1)
  })
  it('includeDbPair marks pair included and decrements excludedCount', async () => {
    mockFetch.mockResolvedValue({ data: { ok: true }, error: null })
    const store = useFineTuneStore()
    store.dbPairs = [{ job_id: 1, title: 'Eng', company: 'Acme', status: 'applied', instruction: 'Write...', input_preview: 'Build', excluded: true }]
    store.dbExcludedCount = 1
    await store.includeDbPair(1)
    expect(store.dbPairs[0].excluded).toBe(false)
    expect(store.dbExcludedCount).toBe(0)
  })
 })
--- a/web/src/stores/settings/fineTune.ts
+++ b/web/src/stores/settings/fineTune.ts
@ -10,16 +10,6 @@ export interface TrainingPair {
  source_file: string
 }
 export interface DbPair {
  job_id: number
  title: string
  company: string
  status: string
  instruction: string
  input_preview: string
  excluded: boolean
 }
 export const useFineTuneStore = defineStore('settings/fineTune', () => {
  const step = ref(1)
  const inFlightJob = ref(false)
@ -32,11 +22,6 @@ export const useFineTuneStore = defineStore('settings/fineTune', () => {
  const pairsLoading = ref(false)
  let _pollTimer: ReturnType<typeof setInterval> | null = null
  const optedIn = ref(false)
  const dbPairs = ref<DbPair[]>([])
  const dbPairsLoading = ref(false)
  const dbExcludedCount = ref(0)
  function resetStep() { step.value = 1 }
  async function loadStatus() {
@ -46,7 +31,6 @@ export const useFineTuneStore = defineStore('settings/fineTune', () => {
    pairsCount.value = data.pairs_count ?? 0
    quotaRemaining.value = data.quota_remaining ?? null
    inFlightJob.value = ['queued', 'running'].includes(data.status)
    optedIn.value = (data as any).opted_in ?? false
  }
  function startPolling() {
@ -84,60 +68,6 @@ export const useFineTuneStore = defineStore('settings/fineTune', () => {
    }
  }
  async function toggleOptIn(enabled: boolean) {
    const { data } = await useApiFetch<{ ok: boolean; enabled: boolean }>(
      '/api/settings/fine-tune/opt-in',
      { method: 'PATCH', body: JSON.stringify({ enabled }), headers: { 'Content-Type': 'application/json' } },
    )
    if (data) optedIn.value = data.enabled
  }
  async function loadDbPairs() {
    if (!optedIn.value) { dbPairs.value = []; return }
    dbPairsLoading.value = true
    const { data } = await useApiFetch<{ pairs: DbPair[]; total: number; excluded_count: number }>(
      '/api/settings/fine-tune/db-pairs',
    )
    dbPairsLoading.value = false
    if (data) {
      dbPairs.value = data.pairs
      dbExcludedCount.value = data.excluded_count
    }
  }
  async function excludeDbPair(jobId: number) {
    const { data } = await useApiFetch<{ ok: boolean }>(
      `/api/settings/fine-tune/db-pairs/${jobId}/exclude`,
      { method: 'PATCH' },
    )
    if (data?.ok) {
      dbPairs.value = dbPairs.value.map(p =>
        p.job_id === jobId ? { ...p, excluded: true } : p,
      )
      dbExcludedCount.value += 1
    }
  }
  async function includeDbPair(jobId: number) {
    const { data } = await useApiFetch<{ ok: boolean }>(
      `/api/settings/fine-tune/db-pairs/${jobId}/include`,
      { method: 'PATCH' },
    )
    if (data?.ok) {
      dbPairs.value = dbPairs.value.map(p =>
        p.job_id === jobId ? { ...p, excluded: false } : p,
      )
      dbExcludedCount.value = Math.max(0, dbExcludedCount.value - 1)
    }
  }
  function downloadExport() {
    const a = document.createElement('a')
    a.href = '/api/settings/fine-tune/export'
    a.download = 'peregrine_training_pairs.jsonl'
    a.click()
  }
  return {
    step,
    inFlightJob,
@ -155,14 +85,5 @@ export const useFineTuneStore = defineStore('settings/fineTune', () => {
    submitJob,
    loadPairs,
    deletePair,
    optedIn,
    dbPairs,
    dbPairsLoading,
    dbExcludedCount,
    toggleOptIn,
    loadDbPairs,
    excludeDbPair,
    includeDbPair,
    downloadExport,
  }
 })
--- a/web/src/stores/survey.test.ts
+++ b/web/src/stores/survey.test.ts
@ -54,20 +54,14 @@ describe('useSurveyStore', () => {
  })
  it('analyze stores result including mode and rawInput', async () => {
    vi.useFakeTimers()
    const mockApiFetch = vi.mocked(useApiFetch)
    // POST → task accepted
    mockApiFetch.mockResolvedValueOnce({ data: { task_id: 7, is_new: true }, error: null })
    // Poll → completed with result
    mockApiFetch.mockResolvedValueOnce({
-      data: { status: 'completed', stage: null, message: null,
+      data: { output: '1. B — reason', source: 'text_paste' },
              result: { output: '1. B — reason', source: 'text_paste' } },
      error: null,
    })
    const store = useSurveyStore()
    await store.analyze(1, { text: 'Q1: test', mode: 'quick' })
    await vi.advanceTimersByTimeAsync(3000)
    expect(store.analysis).not.toBeNull()
    expect(store.analysis!.output).toBe('1. B — reason')
@ -75,7 +69,6 @@ describe('useSurveyStore', () => {
    expect(store.analysis!.mode).toBe('quick')
    expect(store.analysis!.rawInput).toBe('Q1: test')
    expect(store.loading).toBe(false)
    vi.useRealTimers()
  })
  it('analyze sets error on failure', async () => {
--- a/web/src/stores/wizard.ts
+++ b/web/src/stores/wizard.ts
@ -43,14 +43,13 @@ export interface WizardInferenceData {
  testMessage: string
 }
-// Total mandatory steps (integrations step 8 is optional/skip-able)
+// Total mandatory steps (integrations step 7 is optional/skip-able)
-export const WIZARD_STEPS = 8
+export const WIZARD_STEPS = 6
-export const STEP_LABELS = ['Hardware', 'Tier', 'Resume', 'Training', 'Identity', 'Inference', 'Search', 'Integrations']
+export const STEP_LABELS = ['Hardware', 'Tier', 'Resume', 'Identity', 'Inference', 'Search', 'Integrations']
 export const STEP_ROUTES = [
  '/setup/hardware',
  '/setup/tier',
  '/setup/resume',
  '/setup/training',
  '/setup/identity',
  '/setup/inference',
  '/setup/search',
@ -164,7 +163,7 @@ export const useWizardStore = defineStore('wizard', () => {
      }
      // Resume at next step after last completed
-      const resumeAt = Math.max(1, Math.min(data.wizard_step + 1, 8))
+      const resumeAt = Math.max(1, Math.min(data.wizard_step + 1, 7))
      currentStep.value = resumeAt
      return routeForStep(resumeAt)
    } finally {
--- a/web/src/views/ContactsView.vue
+++ b/web/src/views/ContactsView.vue
@ -1,5 +1,5 @@
 <script setup lang="ts">
-import { ref, onMounted } from 'vue'
+import { ref, computed, onMounted } from 'vue'
 import { useApiFetch } from '../composables/useApi'
 import HintChip from '../components/HintChip.vue'
 import { useAppConfigStore } from '../stores/appConfig'
@ -26,8 +26,6 @@ const error       = ref<string | null>(null)
 const search     = ref('')
 const direction  = ref<'all' | 'inbound' | 'outbound'>('all')
 const searchInput = ref('')
 const syncing     = ref(false)
 const syncStatus  = ref<{ status: string; last_completed_at: string | null } | null>(null)
 let debounceTimer: ReturnType<typeof setTimeout> | null = null
 async function fetchContacts() {
@ -78,45 +76,9 @@ const signalLabel: Record<string, string> = {
  rejected:            '✖ Rejected',
  positive_response:   '✅ Positive',
  survey_received:     '📋 Survey',
  event_rescheduled:   '🔄 Rescheduled',
  neutral:             '— Neutral',
 }
-async function fetchSyncStatus() {
+onMounted(fetchContacts)
  const { data } = await useApiFetch<{ status: string; last_completed_at: string | null }>(
    '/api/email/sync/status'
  )
  if (data) syncStatus.value = data
 }
 async function triggerSync() {
  syncing.value = true
  await useApiFetch('/api/tasks/email-sync', { method: 'POST' })
  // Poll until the task finishes or we give up after 60 s
  const deadline = Date.now() + 60_000
  const poll = setInterval(async () => {
    await fetchSyncStatus()
    if (syncStatus.value?.status === 'completed' || Date.now() > deadline) {
      clearInterval(poll)
      syncing.value = false
      fetchContacts()
    }
  }, 2000)
 }
 function formatSyncTime(iso: string | null): string {
  if (!iso) return 'never'
  const d = new Date(iso)
  const diff = Date.now() - d.getTime()
  if (diff < 60_000)  return 'just now'
  if (diff < 3_600_000) return `${Math.floor(diff / 60_000)}m ago`
  if (diff < 86_400_000) return `${Math.floor(diff / 3_600_000)}h ago`
  return d.toLocaleDateString([], { month: 'short', day: 'numeric' })
 }
 onMounted(async () => {
  await Promise.all([fetchContacts(), fetchSyncStatus()])
 })
 </script>
 <template>
@ -129,20 +91,6 @@ onMounted(async () => {
    <header class="contacts-header">
      <h1 class="contacts-title">Contacts</h1>
      <span class="contacts-count" v-if="total > 0">{{ total }} total</span>
      <div class="contacts-sync">
        <span v-if="syncStatus" class="sync-last">
          Last sync: {{ formatSyncTime(syncStatus.last_completed_at) }}
        </span>
        <button
          class="btn-sync"
          :disabled="syncing"
          @click="triggerSync"
          :aria-label="syncing ? 'Email sync running' : 'Sync email now'"
        >
          <span :class="['sync-icon', { 'sync-icon--spinning': syncing }]">↻</span>
          {{ syncing ? 'Syncing…' : 'Sync email' }}
        </button>
      </div>
    </header>
    <div class="contacts-toolbar">
@ -167,16 +115,8 @@ onMounted(async () => {
    <div v-if="loading" class="contacts-empty">Loading…</div>
    <div v-else-if="error" class="contacts-empty contacts-empty--error">{{ error }}</div>
    <div v-else-if="contacts.length === 0 && !search" class="contacts-empty contacts-empty--setup">
      <p>No contacts yet.</p>
      <p class="contacts-empty-hint">
        Connect your inbox in
        <a href="/settings?tab=connections" class="setup-link">Settings → Connections</a>
        then hit <strong>Sync email</strong> to import recruiter emails automatically.
      </p>
    </div>
    <div v-else-if="contacts.length === 0" class="contacts-empty">
-      No contacts found for that search.
+      No contacts found{{ search ? ' for that search' : '' }}.
    </div>
    <div v-else class="contacts-table-wrap">
@ -399,69 +339,4 @@ onMounted(async () => {
 .text-muted {
  color: var(--color-text-muted);
 }
 .contacts-sync {
  display: flex;
  align-items: center;
  gap: var(--space-3);
  margin-left: auto;
 }
 .sync-last {
  font-size: var(--text-xs);
  color: var(--color-text-muted);
 }
 .btn-sync {
  display: flex;
  align-items: center;
  gap: 6px;
  padding: 6px 14px;
  background: transparent;
  border: 1px solid var(--color-border);
  border-radius: 7px;
  color: var(--color-text-muted);
  font-size: var(--text-sm);
  cursor: pointer;
  white-space: nowrap;
 }
 .btn-sync:hover:not(:disabled) {
  border-color: var(--app-primary);
  color: var(--app-primary);
 }
 .btn-sync:disabled {
  opacity: 0.6;
  cursor: not-allowed;
 }
 .sync-icon {
  font-size: 1rem;
  line-height: 1;
  display: inline-block;
 }
 .sync-icon--spinning {
  animation: spin 1s linear infinite;
 }
@keyframes spin {
  to { transform: rotate(360deg); }
 }
 .contacts-empty--setup {
  padding: var(--space-10) var(--space-4);
 }
 .contacts-empty-hint {
  margin-top: var(--space-2);
  font-size: var(--text-sm);
  color: var(--color-text-muted);
 }
 .setup-link {
  color: var(--app-primary);
  text-decoration: underline;
 }
 </style>
--- a/web/src/views/HomeView.vue
+++ b/web/src/views/HomeView.vue
@ -159,10 +159,6 @@
              rows="4"
              aria-label="Job URLs to add"
            />
            <label class="add-jobs__skip-review">
              <input type="checkbox" v-model="skipReview" />
              Skip review — add directly to Apply queue
            </label>
            <button
              class="action-btn action-btn--primary"
              :disabled="!urlInput.trim()"
@ -443,14 +439,13 @@ const runEnrich       = () => runTask('enrich', '/api/tasks/enrich')
 const addTab   = ref<'url' | 'csv'>('url')
 const urlInput = ref('')
 const skipReview = ref(true)
 async function addByUrl() {
  const urls = urlInput.value.split('\n').map(u => u.trim()).filter(Boolean)
  await useApiFetch('/api/jobs/add', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({ urls, skip_review: skipReview.value }),
+    body: JSON.stringify({ urls }),
  })
  urlInput.value = ''
  store.refresh()
@ -796,16 +791,6 @@ onUnmounted(() => {
 .add-jobs__textarea:focus { outline: 2px solid var(--app-primary); outline-offset: 1px; }
 .add-jobs__skip-review {
  display: flex;
  align-items: center;
  gap: var(--space-2);
  font-size: var(--text-sm);
  color: var(--color-text-muted);
  cursor: pointer;
  user-select: none;
 }
 /* ── Danger Zone ──────────────────────────────────────── */
 .danger-zone {
--- a/web/src/views/InterviewsView.vue
+++ b/web/src/views/InterviewsView.vue
@ -682,7 +682,7 @@ function formatRejectionDate(job: PipelineJob): string {
  padding: 1px 8px; font-size: 0.75em; font-weight: 700; margin-left: var(--space-1);
  color: var(--color-text-muted);
 }
-.pre-list-signal-count { margin-left: auto; font-size: 0.75em; font-weight: 700; color: var(--app-accent); }
+.pre-list-signal-count { margin-left: auto; font-size: 0.75em; font-weight: 700; color: #e67e22; }
 /* Collapsible pre-list body */
 .pre-list-body {
@ -713,15 +713,15 @@ function formatRejectionDate(job: PipelineJob): string {
  border-top: 1px solid transparent;
  display: flex; flex-direction: column; gap: 4px;
 }
-.pre-signal-banner[data-color="amber"] { background: color-mix(in srgb, var(--color-warning) 8%, var(--color-surface)); border-top-color: color-mix(in srgb, var(--color-warning) 40%, transparent); }
+.pre-signal-banner[data-color="amber"] { background: rgba(245,158,11,0.08); border-top-color: rgba(245,158,11,0.4); }
-.pre-signal-banner[data-color="green"] { background: color-mix(in srgb, var(--color-success) 8%, var(--color-surface)); border-top-color: color-mix(in srgb, var(--color-success) 40%, transparent); }
+.pre-signal-banner[data-color="green"] { background: rgba(39,174,96,0.08);  border-top-color: rgba(39,174,96,0.4);  }
-.pre-signal-banner[data-color="red"]   { background: color-mix(in srgb, var(--color-error)   8%, var(--color-surface)); border-top-color: color-mix(in srgb, var(--color-error)   40%, transparent); }
+.pre-signal-banner[data-color="red"]   { background: rgba(192,57,43,0.08);  border-top-color: rgba(192,57,43,0.4);  }
 .signal-label   { font-size: 0.82em; }
 .signal-subject { font-size: 0.78em; color: var(--color-text-muted); }
 .signal-actions { display: flex; gap: 6px; align-items: center; }
 .btn-signal-move {
-  background: var(--color-primary); color: var(--color-text-inverse);
+  background: var(--color-primary); color: #fff;
  border: none; border-radius: 4px; padding: 2px 8px; font-size: 0.78em; cursor: pointer;
 }
 .btn-signal-dismiss {
@ -767,7 +767,7 @@ function formatRejectionDate(job: PipelineJob): string {
  background: var(--color-hover);
 }
 .btn-chip-active {
-  background: var(--app-primary-light);
+  background: var(--color-primary-muted, #e8f0ff);
  color: var(--color-primary); border-color: var(--color-primary);
  font-weight: 600;
 }
--- a/web/src/views/JobReviewView.vue
+++ b/web/src/views/JobReviewView.vue
@ -496,7 +496,7 @@ onUnmounted(() => {
 .tab-badge {
  background: var(--color-warning);
-  color: var(--app-accent-text);
+  color: white;
  font-size: 0.65rem;
  font-weight: 700;
  border-radius: 999px;
--- a/web/src/views/MessagingView.vue
+++ b/web/src/views/MessagingView.vue
@ -34,8 +34,35 @@
      </div>
      <template v-else>
        <!-- Action bar -->
        <div class="action-bar" role="toolbar" aria-label="Message actions">
          <button class="btn btn--ghost" @click="openLogModal('call_note')">Log call</button>
          <button class="btn btn--ghost" @click="openLogModal('in_person')">Log note</button>
          <button class="btn btn--ghost" @click="openTemplateModal('apply')">Use template</button>
          <button
            class="btn btn--primary"
            :disabled="store.loading"
            @click="requestDraft"
          >
            {{ store.loading ? 'Drafting…' : 'Draft reply with LLM' }}
          </button>
          <!-- Osprey (Phase 2 stub) — aria-disabled, never hidden -->
          <button
            class="btn btn--osprey"
            aria-disabled="true"
            :title="ospreyTitle"
            @mouseenter="handleOspreyHover"
            @focus="handleOspreyHover"
          >
            📞 Call via Osprey
          </button>
        </div>
        <!-- Draft pending announcement (screen reader) -->
-        <div aria-live="polite" aria-atomic="true" class="sr-only">{{ draftAnnouncement }}</div>
+        <div aria-live="polite" aria-atomic="true" class="sr-only">
          {{ draftAnnouncement }}
        </div>
        <!-- Error banner -->
        <p v-if="store.error" class="thread-error" role="alert">{{ store.error }}</p>
@ -49,15 +76,9 @@
            v-for="item in timeline"
            :key="item._key"
            class="timeline__item"
-            :class="[
+            :class="[`timeline__item--${item.type}`, item.approved_at === null && item.type === 'draft' ? 'timeline__item--draft-pending' : '']"
              `timeline__item--${item.type}`,
              item.approved_at === null && item.type === 'draft' ? 'timeline__item--draft-pending' : '',
              item.type !== 'draft' ? 'timeline__item--expandable' : '',
              expandedKeys.has(item._key) ? 'timeline__item--open' : '',
            ]"
            role="listitem"
            :aria-label="`${typeLabel(item.type)}, ${item.direction || ''}, ${item.logged_at}`"
            @click="item.type !== 'draft' && toggleExpand(item)"
          >
            <span class="timeline__icon" aria-hidden="true">{{ typeIcon(item.type) }}</span>
            <div class="timeline__content">
@ -68,29 +89,19 @@
                <span
                  v-if="item.type === 'draft' && item.approved_at === null"
                  class="timeline__badge timeline__badge--pending"
-                >Pending approval</span>
+                >
                  Pending approval
                </span>
                <span
                  v-if="item.type === 'draft' && item.approved_at !== null"
                  class="timeline__badge timeline__badge--approved"
-                >Approved</span>
+                >
-                <span v-if="item.type !== 'draft'" class="timeline__expand-hint" aria-hidden="true">
+                  Approved
                  {{ expandedKeys.has(item._key) ? '▲' : '▼' }}
                </span>
              </div>
              <p v-if="item.subject" class="timeline__subject">{{ item.subject }}</p>
-              <!-- Expandable body for non-draft items -->
+              <!-- Draft body is editable before approval -->
              <template v-if="item.type !== 'draft' && expandedKeys.has(item._key)">
                <div class="timeline__body-wrap" @click.stop>
                  <div v-if="bodyCache[item.id] === null" class="timeline__body-loading">
                    Loading…
                  </div>
                  <pre v-else-if="bodyCache[item.id]" class="timeline__body">{{ bodyCache[item.id] }}</pre>
                  <p v-else class="timeline__body-empty">No body content.</p>
                </div>
              </template>
              <!-- Draft: editable textarea + actions -->
              <template v-if="item.type === 'draft' && item.approved_at === null">
                <textarea
                  :ref="el => setDraftRef(item.id, el)"
@ -108,50 +119,25 @@
                    v-if="item.to_addr"
                    :href="`mailto:${item.to_addr}?subject=${encodeURIComponent(item.subject ?? '')}&body=${encodeURIComponent(item.body ?? '')}`"
                    class="btn btn--ghost btn--sm"
-                    target="_blank" rel="noopener"
+                    target="_blank"
-                  >Open in email client</a>
+                    rel="noopener"
                  >
                    Open in email client
                  </a>
                  <button class="btn btn--ghost btn--sm btn--danger" @click="confirmDelete(item.id)">
                    Discard
                  </button>
                </div>
              </template>
              <template v-else>
                <p class="timeline__body">{{ item.body }}</p>
              </template>
            </div>
          </li>
          <li v-if="timeline.length === 0" class="timeline__empty">
            No messages logged yet for this job.
          </li>
        </ul>
        <!-- Compose bar (sticky footer) -->
        <div class="compose-bar" role="toolbar" aria-label="Compose actions">
          <div v-if="composing" class="compose-bar__actions">
            <button class="btn btn--ghost btn--sm" @click="triggerAction(() => openLogModal('call_note'))">Log call</button>
            <button class="btn btn--ghost btn--sm" @click="triggerAction(() => openLogModal('in_person'))">Log note</button>
            <button class="btn btn--ghost btn--sm" @click="triggerAction(() => openTemplateModal('apply'))">Use template</button>
            <button
              class="btn btn--primary btn--sm"
              :disabled="store.loading"
              @click="triggerAction(requestDraft)"
            >
              <span v-if="store.loading" class="btn__spinner" aria-hidden="true"></span>
              {{ store.loading ? 'Drafting…' : 'Draft reply with LLM' }}
            </button>
            <button
              class="btn btn--osprey btn--sm"
              aria-disabled="true"
              :title="ospreyTitle"
              @mouseenter="handleOspreyHover"
              @focus="handleOspreyHover"
            >📞 Call via Osprey</button>
          </div>
          <button
            class="btn compose-bar__toggle"
            :class="composing ? 'btn--ghost' : 'btn--primary'"
            @click="composing = !composing"
            :aria-expanded="composing"
            aria-controls="compose-actions"
          >{{ composing ? '✕ Close' : '＋ New' }}</button>
        </div>
      </template>
    </main>
@ -244,8 +230,8 @@ const jobContacts = ref<JobContact[]>([])
 watch(selectedJobId, async (id) => {
  if (id === null) { jobContacts.value = []; return }
-  const { data } = await useApiFetch<{ total: number; contacts: JobContact[] }>(`/api/contacts?job_id=${id}`)
+  const { data } = await useApiFetch<JobContact[]>(`/api/contacts?job_id=${id}`)
-  jobContacts.value = data?.contacts ?? []
+  jobContacts.value = data ?? []
 })
 const timeline = computed<TimelineItem[]>(() => {
@ -276,31 +262,6 @@ const timeline = computed<TimelineItem[]>(() => {
  )
 })
 // ── Body expansion ────────────────────────────────────────────────────────
 const expandedKeys = ref(new Set<string>())
 const bodyCache = ref<Record<number, string | null>>({})  // null = still loading
 async function toggleExpand(item: TimelineItem) {
  const key = item._key
  const next = new Set(expandedKeys.value)
  if (next.has(key)) { next.delete(key); expandedKeys.value = next; return }
  next.add(key)
  expandedKeys.value = next
  if (key.startsWith('jc-') && !(item.id in bodyCache.value)) {
    bodyCache.value = { ...bodyCache.value, [item.id]: null }
    const { data } = await useApiFetch<{ body: string | null }>(`/api/contacts/${item.id}`)
    const raw = data?.body ?? ''
    const text = raw.trimStart().startsWith('<')
      ? (new DOMParser().parseFromString(raw, 'text/html').body.textContent ?? '').trim()
      : raw.trim()
    bodyCache.value = { ...bodyCache.value, [item.id]: text }
  }
 }
 // ── Compose bar ────────────────────────────────────────────────────────────
 const composing = ref(false)
 function triggerAction(fn: () => void) { composing.value = false; fn() }
 // ── Draft body edits (local, before approve) ──────────────────────────────
 const draftBodyEdits = ref<Record<number, string>>({})
@ -454,15 +415,8 @@ onUnmounted(() => {
 <style scoped>
 .messaging-layout {
  display: flex;
-  height: 100dvh;
+  height: 100%;
  min-height: 0;
  overflow: hidden;
 }
@media (max-width: 1023px) {
  .messaging-layout {
    height: calc(100dvh - 56px - env(safe-area-inset-bottom, 0px));
  }
 }
 /* ── Left panel ─────────────────────── */
@ -511,6 +465,11 @@ onUnmounted(() => {
  flex: 1; display: flex; align-items: center; justify-content: center;
  color: var(--color-text-muted);
 }
 .action-bar {
  display: flex; flex-wrap: wrap; gap: var(--space-2); align-items: center;
  padding: var(--space-3) var(--space-4);
  border-bottom: 1px solid var(--color-border-light);
 }
 .btn--osprey {
  opacity: 0.5; cursor: not-allowed;
  background: none; border: 1px dashed var(--color-border);
@ -518,21 +477,6 @@ onUnmounted(() => {
  color: var(--color-text-muted); font-size: var(--text-sm);
  padding: var(--space-2) var(--space-3); min-height: 36px;
 }
 /* Compose bar */
 .compose-bar {
  flex-shrink: 0;
  display: flex; flex-direction: column; align-items: flex-end;
  gap: var(--space-2);
  padding: var(--space-3) var(--space-4);
  border-top: 1px solid var(--color-border-light);
  background: var(--color-surface);
 }
 .compose-bar__actions {
  display: flex; flex-wrap: wrap; gap: var(--space-2); align-items: center;
  width: 100%; justify-content: flex-start;
 }
 .compose-bar__toggle { align-self: flex-end; min-width: 90px; justify-content: center; }
 .thread-error {
  margin: var(--space-2) var(--space-4);
  color: var(--app-accent); font-size: var(--text-sm);
@ -563,27 +507,10 @@ onUnmounted(() => {
  font-size: var(--text-xs); font-weight: 700;
  padding: 1px 6px; border-radius: var(--radius-full);
 }
-.timeline__badge--pending { background: var(--color-accent-light); color: var(--color-accent); }
+.timeline__badge--pending { background: #fef3c7; color: #d97706; }
-.timeline__badge--approved { background: var(--color-primary-light); color: var(--color-primary); }
+.timeline__badge--approved { background: #d1fae5; color: #065f46; }
 .timeline__subject { font-size: var(--text-sm); font-weight: 500; margin: 0; }
-.timeline__expand-hint {
+.timeline__body { font-size: var(--text-sm); white-space: pre-wrap; margin: 0; color: var(--color-text); }
  font-size: var(--text-xs); color: var(--color-text-muted); margin-left: auto;
  transition: transform 150ms ease;
 }
 .timeline__item--expandable { cursor: pointer; }
 .timeline__item--expandable:hover { border-color: var(--app-primary); }
 .timeline__body-wrap {
  margin-top: var(--space-2);
  border-top: 1px solid var(--color-border-light);
  padding-top: var(--space-2);
 }
 .timeline__body {
  font-size: var(--text-sm); white-space: pre-wrap; margin: 0;
  color: var(--color-text); max-height: 280px; overflow-y: auto;
  font-family: var(--font-body);
 }
 .timeline__body-loading { font-size: var(--text-xs); color: var(--color-text-muted); }
 .timeline__body-empty { font-size: var(--text-xs); color: var(--color-text-muted); margin: 0; }
 .timeline__draft-body {
  width: 100%; font-size: var(--text-sm); font-family: var(--font-body);
  padding: var(--space-2); border: 1px solid var(--color-border);
@ -595,45 +522,20 @@ onUnmounted(() => {
 .timeline__empty { color: var(--color-text-muted); font-size: var(--text-sm); padding: var(--space-2); }
 /* Buttons */
-.btn {
+.btn { padding: var(--space-2) var(--space-3); border-radius: var(--radius-md); font-size: var(--text-sm); font-weight: 500; cursor: pointer; min-height: 36px; }
  padding: var(--space-2) var(--space-3);
  border-radius: var(--radius-md);
  font-size: var(--text-sm);
  font-weight: 500;
  cursor: pointer;
  min-height: 36px;
  display: inline-flex;
  align-items: center;
  gap: 6px;
  transition: background 120ms ease, border-color 120ms ease, opacity 120ms ease, transform 80ms ease;
 }
 .btn:active:not(:disabled) { transform: translateY(1px); }
 .btn:focus-visible { outline: 2px solid var(--app-primary); outline-offset: 2px; }
 .btn--sm { padding: var(--space-1) var(--space-3); min-height: 30px; font-size: var(--text-xs); }
 .btn--primary { background: var(--app-primary); color: var(--color-surface); border: none; }
-.btn--primary:hover:not(:disabled) { opacity: 0.88; }
+.btn--primary:hover:not(:disabled) { opacity: 0.9; }
 .btn--primary:disabled { opacity: 0.5; cursor: not-allowed; }
 .btn--ghost { background: none; border: 1px solid var(--color-border); color: var(--color-text); }
-.btn--ghost:hover:not(:disabled) { background: var(--color-surface-alt); border-color: var(--app-primary); color: var(--app-primary); }
+.btn--ghost:hover { background: var(--color-surface-alt); }
-.btn--danger { background: var(--app-accent); color: var(--app-accent-text); border: none; }
+.btn--danger { background: var(--app-accent); color: white; border: none; }
-.btn--danger:hover:not(:disabled) { opacity: 0.88; }
+.btn--danger:hover { opacity: 0.9; }
 /* Spinner inside buttons */
 .btn__spinner {
  width: 13px;
  height: 13px;
  border: 2px solid rgba(255,255,255,0.35);
  border-top-color: white;
  border-radius: 50%;
  animation: btn-spin 0.65s linear infinite;
  flex-shrink: 0;
 }
@keyframes btn-spin { to { transform: rotate(360deg); } }
 /* Modals (delete confirm) */
 .modal-backdrop {
  position: fixed; inset: 0;
-  background: var(--color-overlay);
+  background: rgba(0,0,0,0.5);
  display: flex; align-items: center; justify-content: center;
  z-index: 200;
 }
--- a/web/src/views/ReferencesView.vue
+++ b/web/src/views/ReferencesView.vue
@ -442,7 +442,7 @@ onMounted(fetchRefs)
 .tag-chip--technical  { background: var(--app-primary-light);     color: var(--app-primary); }
 .tag-chip--managerial { background: rgba(39, 174, 96, 0.12);      color: var(--color-success); }
 .tag-chip--character  { background: rgba(212, 137, 26, 0.12);     color: var(--score-mid); }
-.tag-chip--academic   { background: color-mix(in srgb, var(--status-synced) 12%, var(--color-surface)); color: var(--status-synced); }
+.tag-chip--academic   { background: rgba(103, 58, 183, 0.12);     color: #7c3aed; }
 .ref-card__actions {
  display: flex;
--- a/web/src/views/ResumesView.vue
+++ b/web/src/views/ResumesView.vue
@ -332,7 +332,7 @@ onBeforeRouteLeave(() => {
 .rv__item-star { color: var(--color-warning, #f59e0b); font-size: 1rem; flex-shrink: 0; margin-top: 2px; }
 .rv__item-info { display: flex; flex-direction: column; gap: 2px; }
-.rv__item-name { font-weight: 500; font-size: var(--text-sm); }
+.rv__item-name { font-weight: 500; font-size: var(--font-sm, 0.875rem); }
 .rv__item-meta { font-size: var(--font-xs, 0.75rem); color: var(--color-text-muted, #64748b); }
 .rv__item-source { font-size: var(--font-xs, 0.75rem); color: var(--color-accent, #6366f1); }
@ -340,7 +340,7 @@ onBeforeRouteLeave(() => {
 .rv__preview-header { display: flex; align-items: flex-start; justify-content: space-between; flex-wrap: wrap; gap: var(--space-3, 0.75rem); }
 .rv__preview-meta { display: flex; align-items: center; gap: var(--space-2, 0.5rem); flex-wrap: wrap; }
 .rv__preview-name { font-size: var(--font-lg, 1.125rem); font-weight: 600; margin: 0; }
-.rv__preview-words { font-size: var(--text-sm); color: var(--color-text-muted, #64748b); }
+.rv__preview-words { font-size: var(--font-sm, 0.875rem); color: var(--color-text-muted, #64748b); }
 .rv__default-badge {
  font-size: var(--font-xs, 0.75rem); font-weight: 600;
  background: var(--color-success, #16a34a); color: #fff;
@ -352,7 +352,7 @@ onBeforeRouteLeave(() => {
  border: 1px solid var(--color-error, #dc2626);
  border-radius: var(--radius-md, 0.5rem);
  padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem);
-  cursor: pointer; font-size: var(--text-sm);
+  cursor: pointer; font-size: var(--font-sm, 0.875rem);
 }
 .rv__delete-btn:disabled { opacity: 0.4; cursor: not-allowed; }
@ -364,13 +364,13 @@ onBeforeRouteLeave(() => {
 .rv__textarea {
  flex: 1; min-height: 400px; padding: var(--space-3, 0.75rem);
  border: 1px solid var(--color-border, #e2e8f0); border-radius: var(--radius-md, 0.5rem);
-  font-family: monospace; font-size: var(--text-sm); resize: vertical;
+  font-family: monospace; font-size: var(--font-sm, 0.875rem); resize: vertical;
  background: var(--color-surface-alt, #f8fafc);
  color: var(--color-text);
 }
 .rv__textarea:not([readonly]) { background: var(--color-surface); }
 .rv__edit-actions { display: flex; gap: var(--space-2, 0.5rem); }
-.rv__error { color: var(--color-error, #dc2626); font-size: var(--text-sm); }
+.rv__error { color: var(--color-error, #dc2626); font-size: var(--font-sm, 0.875rem); }
 .rv__download-menu { position: relative; }
 .rv__download-dropdown {
@ -382,11 +382,11 @@ onBeforeRouteLeave(() => {
 .rv__download-dropdown button {
  width: 100%; text-align: left; background: none; border: none;
  padding: var(--space-2, 0.5rem) var(--space-3, 0.75rem);
-  cursor: pointer; font-size: var(--text-sm); border-radius: var(--radius-sm, 0.25rem);
+  cursor: pointer; font-size: var(--font-sm, 0.875rem); border-radius: var(--radius-sm, 0.25rem);
 }
 .rv__download-dropdown button:hover { background: var(--color-surface-alt, #f8fafc); }
-.rv__loading, .rv__empty { color: var(--color-text-muted, #64748b); font-size: var(--text-sm); }
+.rv__loading, .rv__empty { color: var(--color-text-muted, #64748b); font-size: var(--font-sm, 0.875rem); }
 /* Button styles — defined locally since no global button sheet exists yet */
 .btn-secondary {
@ -396,7 +396,7 @@ onBeforeRouteLeave(() => {
  border-radius: var(--radius-md, 0.5rem);
  color: var(--color-text-muted);
  cursor: pointer;
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  white-space: nowrap;
 }
 .btn-secondary:hover:not(:disabled) {
@ -412,7 +412,7 @@ onBeforeRouteLeave(() => {
  border: none;
  border-radius: var(--radius-md, 0.5rem);
  cursor: pointer;
-  font-size: var(--text-sm);
+  font-size: var(--font-sm, 0.875rem);
  font-weight: 600;
  white-space: nowrap;
  display: inline-flex;
--- a/web/src/views/settings/FineTuneView.vue
+++ b/web/src/views/settings/FineTuneView.vue
@ -3,31 +3,16 @@ import { ref, onMounted, onUnmounted } from 'vue'
 import { storeToRefs } from 'pinia'
 import { useFineTuneStore } from '../../stores/settings/fineTune'
 import { useAppConfigStore } from '../../stores/appConfig'
 import { showToast } from '../../composables/useToast'
 const store = useFineTuneStore()
 const config = useAppConfigStore()
-const { step, inFlightJob, jobStatus, pairsCount, quotaRemaining, pairs, pairsLoading,
+const { step, inFlightJob, jobStatus, pairsCount, quotaRemaining, pairs, pairsLoading } = storeToRefs(store)
        optedIn, dbPairs, dbPairsLoading, dbExcludedCount } = storeToRefs(store)
 const fileInput = ref<HTMLInputElement | null>(null)
 const selectedFiles = ref<File[]>([])
 const uploadResult = ref<{ file_count: number } | null>(null)
 const extractError = ref<string | null>(null)
 const modelReady = ref<boolean | null>(null)
 const toggling = ref(false)
 const toggleSaved = ref(false)
 async function handleOptInChange(e: Event) {
  const enabled = (e.target as HTMLInputElement).checked
  toggling.value = true
  toggleSaved.value = false
  await store.toggleOptIn(enabled)
  await store.loadDbPairs()
  toggling.value = false
  toggleSaved.value = true
  setTimeout(() => { toggleSaved.value = false }, 2000)
 }
 async function handleUpload() {
  if (!selectedFiles.value.length) return
@ -60,9 +45,7 @@ async function checkLocalModel() {
 onMounted(async () => {
  store.startPolling()
  await store.loadStatus()
  await store.loadPairs()
  await store.loadDbPairs()
  if (store.step === 3 && !config.isCloud) await checkLocalModel()
 })
 onUnmounted(() => { store.stopPolling(); store.resetStep() })
@ -72,115 +55,6 @@ onUnmounted(() => { store.stopPolling(); store.resetStep() })
  <div class="fine-tune-view">
    <h2>Fine-Tune Model</h2>
    <!-- Training Export: consent toggle (always visible) -->
    <section class="form-section training-export-consent">
      <h3>Training Export</h3>
      <p class="section-note">
        When enabled, your applied-job cover letters are available as a local dataset file
        for fine-tuning a language model to your writing style.
      </p>
      <label class="toggle-label" :class="{ 'toggle-saving': toggling }">
        <input
          type="checkbox"
          :checked="optedIn"
          :disabled="toggling"
          @change="handleOptInChange"
          aria-describedby="opt-in-desc"
        />
        Include cover letters in training export
        <span v-if="toggling" class="toggle-status" aria-live="polite">Saving…</span>
        <span v-else-if="toggleSaved" class="toggle-status" aria-live="polite">Saved</span>
      </label>
      <p class="section-note" id="opt-in-desc">
        <template v-if="!config.isCloud">
          Your cover letters stay on your device unless you explicitly request cloud fine-tuning.
        </template>
        <template v-else>
          Your cover letters are stored on your CircuitForge account and are not shared with any
          third party unless you explicitly request cloud fine-tuning.
        </template>
        <span v-if="!optedIn" class="opt-out-receipt">
          Training export is off — cover letters remain local only.
          You can change this in Settings at any time.
        </span>
      </p>
    </section>
    <!-- From Applied Jobs: curation list (only when opted in) -->
    <section v-if="optedIn" class="form-section">
      <h3>From Applied Jobs</h3>
      <div class="db-pairs-header">
        <span class="pairs-count">
          {{
            dbPairs.filter(p => !p.excluded).length === 1
              ? '1 pair available'
              : `${dbPairs.filter(p => !p.excluded).length} pairs available`
          }}
          <span
            v-if="dbExcludedCount > 0"
            class="excluded-badge"
            :title="`${dbExcludedCount} pair(s) excluded — use Restore to re-include`"
          >{{ dbExcludedCount }} excluded</span>
        </span>
        <div class="db-pairs-actions">
          <button
            class="btn-secondary"
            :disabled="dbPairs.filter(p => !p.excluded).length === 0"
            @click="store.downloadExport()"
          >
            Download JSONL <span aria-hidden="true">↓</span>
          </button>
          <div class="cloud-finetune-wrap">
            <button
              class="btn-secondary"
              :disabled="config.tier !== 'premium' || dbPairs.filter(p => !p.excluded).length === 0"
              @click="config.tier === 'premium' && showToast('Cloud fine-tuning coming soon')"
            >
              Request Cloud Fine-Tune
            </button>
            <p v-if="config.tier !== 'premium'" class="tier-gate-note">
              Available on Premium.
              <a href="/settings?tab=license" class="upgrade-link">Upgrade your plan →</a>
            </p>
          </div>
        </div>
      </div>
      <p class="section-note download-advisory">
        The downloaded file contains your cover letters in plain text (JSONL format).
        Store it in a secure location.
      </p>
      <div aria-live="polite" aria-atomic="false" aria-label="Applied jobs training pairs">
        <div v-if="dbPairsLoading" class="pairs-loading">Loading…</div>
        <ul v-else-if="dbPairs.length > 0" class="pairs-items db-pairs-items">
          <li
            v-for="pair in dbPairs"
            :key="pair.job_id"
            class="pair-item"
            :class="{ 'pair-excluded': pair.excluded }"
          >
            <div class="pair-info">
              <span class="pair-instruction">{{ pair.title }} · {{ pair.company }}</span>
              <span class="pair-source">{{ pair.status }}</span>
            </div>
            <button
              v-if="!pair.excluded"
              class="pair-delete"
              title="Exclude from training export"
              @click="store.excludeDbPair(pair.job_id)"
            >Exclude</button>
            <button
              v-else
              class="pair-restore"
              title="Restore to training export"
              @click="store.includeDbPair(pair.job_id)"
            >Restore</button>
          </li>
        </ul>
        <p v-else class="section-note">No applied jobs with cover letters found.</p>
      </div>
    </section>
    <!-- Wizard steps indicator -->
    <div class="wizard-steps">
      <span :class="['step', step >= 1 ? 'active' : '']">1. Upload</span>
@ -315,21 +189,4 @@ onUnmounted(() => { store.stopPolling(); store.resetStep() })
 .pair-source { font-size: 0.75rem; color: var(--color-text-muted); }
 .pair-delete { flex-shrink: 0; background: none; border: none; color: var(--color-error); cursor: pointer; font-size: 0.9rem; padding: 2px 4px; border-radius: var(--radius-sm); transition: background 150ms; }
 .pair-delete:hover { background: var(--color-error); color: #fff; }
 .training-export-consent { border: 1px solid var(--color-border-light); border-radius: var(--radius-md); padding: var(--space-4, 1rem); margin-bottom: var(--space-6, 1.5rem); }
 .toggle-label { display: flex; align-items: center; gap: var(--space-2, 0.5rem); font-size: 0.9rem; font-weight: 500; cursor: pointer; flex-wrap: wrap; }
 .toggle-label.toggle-saving { opacity: 0.7; }
 .toggle-label input[type="checkbox"] { width: 16px; height: 16px; accent-color: var(--color-primary); cursor: pointer; flex-shrink: 0; }
 .toggle-status { font-size: 0.8rem; color: var(--color-text-muted); margin-left: var(--space-1, 0.25rem); }
 .opt-out-receipt { display: block; margin-top: var(--space-1, 0.25rem); color: var(--color-text-muted); font-size: 0.8rem; }
 .db-pairs-header { display: flex; align-items: flex-start; justify-content: space-between; flex-wrap: wrap; gap: var(--space-3, 0.75rem); margin-bottom: var(--space-4, 1rem); }
 .db-pairs-actions { display: flex; align-items: flex-start; gap: var(--space-2, 0.5rem); flex-wrap: wrap; }
 .cloud-finetune-wrap { display: flex; flex-direction: column; gap: var(--space-1, 0.25rem); }
 .tier-gate-note { font-size: 0.8rem; color: var(--color-text-muted); margin: 0; }
 .upgrade-link { color: var(--color-primary); text-decoration: underline; }
 .excluded-badge { margin-left: var(--space-2, 0.5rem); background: var(--color-warning-bg, #fef3c7); color: var(--color-warning-fg, #92400e); font-size: 0.75rem; padding: 1px 6px; border-radius: var(--radius-full, 9999px); }
 .db-pairs-items { max-height: 320px; }
 .pair-excluded { opacity: 0.5; }
 .pair-restore { flex-shrink: 0; background: none; border: 1px solid var(--color-border); color: var(--color-text-muted); cursor: pointer; font-size: 0.8rem; padding: 2px 8px; border-radius: var(--radius-sm); }
 .pair-restore:hover { background: var(--color-surface-alt); }
 .download-advisory { margin-top: var(--space-2, 0.5rem); font-style: italic; }
 </style>
--- a/web/src/views/wizard/WizardIdentityStep.vue
+++ b/web/src/views/wizard/WizardIdentityStep.vue
@ -74,7 +74,7 @@ const form = reactive({
  careerSummary: wizard.identity.careerSummary,
 })
-function back() { router.push('/setup/training') }
+function back() { router.push('/setup/resume') }
 async function next() {
  validationError.value = ''
@ -92,7 +92,7 @@ async function next() {
  }
  wizard.identity = { ...form }
-  const ok = await wizard.saveStep(5, {
+  const ok = await wizard.saveStep(4, {
    name: form.name,
    email: form.email,
    phone: form.phone,
--- a/web/src/views/wizard/WizardInferenceStep.vue
+++ b/web/src/views/wizard/WizardInferenceStep.vue
@ -127,7 +127,7 @@ async function next() {
  })
  wizard.inference.services = svcMap
-  const ok = await wizard.saveStep(6, {
+  const ok = await wizard.saveStep(5, {
    anthropic_key: form.anthropicKey,
    openai_url: form.openaiUrl,
    openai_key: form.openaiKey,
--- a/web/src/views/wizard/WizardIntegrationsStep.vue
+++ b/web/src/views/wizard/WizardIntegrationsStep.vue
@ -85,7 +85,7 @@ function back() { router.push('/setup/search') }
 async function finish() {
  // Save integration selections (step 7) then mark wizard complete
-  await wizard.saveStep(8, { integrations: [...checkedIds.value] })
+  await wizard.saveStep(7, { integrations: [...checkedIds.value] })
  const ok = await wizard.complete()
  if (ok) router.replace('/')
 }
--- a/web/src/views/wizard/WizardResumeStep.vue
+++ b/web/src/views/wizard/WizardResumeStep.vue
@ -216,7 +216,7 @@ async function next() {
    experience: wizard.resume.experience,
    ...(wizard.resume.parsedData ?? {}),
  }})
-  if (ok) router.push('/setup/training')
+  if (ok) router.push('/setup/identity')
 }
 </script>
--- a/web/src/views/wizard/WizardSearchStep.vue
+++ b/web/src/views/wizard/WizardSearchStep.vue
@ -143,7 +143,7 @@ async function next() {
  wizard.search.titles = [...form.titles]
  wizard.search.locations = [...form.locations]
-  const ok = await wizard.saveStep(7, {
+  const ok = await wizard.saveStep(6, {
    search: {
      titles: form.titles,
      locations: form.locations,
--- a/web/src/views/wizard/WizardTrainingStep.vue
+++ b/web/src/views/wizard/WizardTrainingStep.vue
@ -1,74 +0,0 @@
 <script setup lang="ts">
 import { ref } from 'vue'
 import { useRouter } from 'vue-router'
 import { useAppConfigStore } from '../../stores/appConfig'
 const router = useRouter()
 const config = useAppConfigStore()
 const optIn = ref(false)
 const saving = ref(false)
 async function next() {
  saving.value = true
  try {
    if (optIn.value) {
      await fetch('/api/settings/fine-tune/opt-in', {
        method: 'PATCH',
        headers: { 'Content-Type': 'application/json' },
        body: JSON.stringify({ enabled: true }),
      })
    }
    router.push('/setup/identity')
  } finally {
    saving.value = false
  }
 }
 function back() { router.push('/setup/resume') }
 </script>
 <template>
  <div class="wizard-step">
    <h2 class="step-title">
      Training Export
      <span class="optional-badge">Optional</span>
    </h2>
    <p class="step-body">
      Would you like to save your cover letters for training export?
      This lets you build a personal dataset to fine-tune a language model to your writing style.
    </p>
    <p v-if="!config.isCloud" class="step-body-note">
      Your data stays on your device unless you explicitly request cloud fine-tuning.
    </p>
    <p v-else class="step-body-note">
      Your cover letters are stored on your CircuitForge account.
      They are not shared with any third party unless you request cloud fine-tuning.
    </p>
    <label class="opt-in-label">
      <input type="checkbox" v-model="optIn" />
      Yes, include my cover letters in training export
    </label>
    <div class="step-actions">
      <button class="btn-ghost" @click="back">
        <span aria-hidden="true">←</span> Back
      </button>
      <button class="btn-primary" :disabled="saving" :aria-busy="saving" @click="next">
        {{ saving ? 'Saving…' : 'Continue' }}
        <span v-if="!saving" aria-hidden="true">→</span>
      </button>
    </div>
  </div>
 </template>
 <style scoped>
 .wizard-step { display: flex; flex-direction: column; gap: var(--space-5, 1.25rem); }
 .step-title { font-family: var(--font-display); font-size: 1.25rem; font-weight: 700; display: flex; align-items: center; gap: var(--space-2, 0.5rem); }
 .optional-badge { font-family: var(--font-body); font-size: 0.75rem; font-weight: 500; background: var(--color-surface-alt); color: var(--color-text-muted); padding: 2px 8px; border-radius: var(--radius-full, 9999px); }
 .step-body { font-size: 0.9rem; color: var(--color-text); line-height: 1.6; }
 .step-body-note { font-size: 0.85rem; color: var(--color-text-muted); line-height: 1.5; margin-top: calc(-1 * var(--space-3, 0.75rem)); }
 .opt-in-label { display: flex; align-items: flex-start; gap: var(--space-2, 0.5rem); font-size: 0.9rem; cursor: pointer; }
 .opt-in-label input[type="checkbox"] { margin-top: 2px; width: 16px; height: 16px; accent-color: var(--color-primary); flex-shrink: 0; cursor: pointer; }
 .step-actions { display: flex; gap: var(--space-3, 0.75rem); justify-content: flex-end; padding-top: var(--space-4, 1rem); border-top: 1px solid var(--color-border-light); }
 </style>
		`@ -1 +0,0 @@`
			`ALTER TABLE jobs ADD COLUMN excluded_from_training INTEGER DEFAULT 0;`