Insert Training at step 4 in WIZARD_STEPS (7→8), STEP_LABELS, and
STEP_ROUTES. Bump Identity→5, Inference→6, Search→7, Integrations→8 in
their respective saveStep calls. Cap resumeAt at 8. Await loadStatus()
before loadDbPairs() in FineTuneView onMounted so optedIn is set before
the early-exit guard runs.
Inserts a new optional Training Export step between Resume and Identity
in the setup wizard. Users can opt in to saving cover letters for
fine-tuning dataset export. Consent copy distinguishes local vs. cloud
storage. WIZARD_STEPS bumped to 7; router, and adjacent step
back/next navigation updated accordingly.
- PATCH /api/settings/fine-tune/opt-in — toggle training_export_opt_in in user.yaml
- GET /api/settings/fine-tune/db-pairs — list DB jobs with exclusion flags (403 without opt-in)
- PATCH /api/settings/fine-tune/db-pairs/{id}/exclude|include — per-job exclusion toggle
- GET /api/settings/fine-tune/export — NDJSON streaming download of all training pairs (DB + file)
- POST/GET /api/settings/fine-tune/cloud-request|cloud-status — Phase 2 stubs (501)
- finetune_status now includes opted_in field
- 6 new API tests; all 17 tests pass
Add excluded_from_training column to jobs table (migration 009 + _MIGRATIONS
entry for existing DBs). Add get_db_pairs(), get_training_pairs(), and
set_training_exclusion() helpers for the cover letter training export pipeline.
Add test_training_export.py with 8 tests covering all helpers (all passing).
Premium/ultra users with a custom_writing_model in their session are
routed to that model as the first cf-orch candidate; all other tiers
use the shared Qwen2.5-3B-Instruct base. complete_json() is unchanged
since fine-tuned writing models aren't trained for structured output.
Adds _request_tier and _request_writing_model ContextVars. Resolution
order: USER_WRITING_MODELS env var (Monday path) then Heimdall meta
(future path via peregrine#110).
Vue+FastAPI (api+web services) is the only frontend. The peregrine-cloud
Streamlit container was still running alongside the new stack and is now
removed. Port 8505 freed.
- app/cloud_session.py: CloudSessionFactory(product="peregrine") from
cf-core v0.16.0; get_session / require_tier FastAPI dependencies;
session_middleware_dep sets request-scoped user_id ContextVar
- app/llm.py: _request_user_id ContextVar + set/get helpers;
_allocate_orch_async includes user_id in payload when present so
premium users get their custom model path from cf-orch UserModelRegistry
- app/main.py: session_middleware_dep wired as global FastAPI dependency;
runs on every request, zero function-signature changes needed
Force-added to bypass resume_matcher/ gitignore (CF-specific patch files).
- Add _try_rerank() to job_ranker.py: after stack_score sort, rerank
top-50 candidates by (resume_text, description) cross-encoder relevance
- Add resume_text param to rank_jobs(); graceful no-op when empty
- Add reranker gap-reordering pass in rewrite_for_ats(): gaps sorted by
(jd_text, term) relevance before section grouping and LLM rewrite
- Both integrations fall back silently on ImportError or inference failure
- 13 new tests; CF_RERANKER_MOCK=1 for zero-weight test runs
- Closes#108
- Seed resumes table with a full UX designer base resume (Alex Rivera persona)
- Add ATS gap reports and optimized resumes for Spotify (job 1), Duolingo (job 2), NPR (job 3)
- Each gap report highlights role-specific keyword opportunities (audio UX, gamification, public media)
- Optimized resumes tailor the base resume framing to each company's emphasis
- Seed company_research for Asana (phone_screen), Notion (interviewing), Figma (hired)
- Includes company_brief, ceo_brief, talking_points, tech_brief, funding_brief,
competitors_brief, accessibility_brief for each
- Update demo/config/user.yaml career_summary to match UX designer persona
- Fixes mismatch between "software engineer" summary and UX/design job seeds
- Adds music and education mission preference notes
Six new screenshots (dashboard, card-stack review with approve/reject tint,
apply workspace, interviews kanban) plus an animated GIF of the swipe-review
flow. Adds demo link above the fold.
Replace the legacy Streamlit `app` service with a FastAPI `api` service
running dev_api:app on port 8601. The Vue SPA (nginx) proxies /api/ →
api:8601 internally so no host port is needed on the api container.
Move web service from port 8507 → 8504 to match the documented demo URL
(demo.circuitforge.tech/peregrine via Caddy → host:8504).
--color-primary in dark mode is a medium-light green (#6ab870); white on green
yields ~2.2:1 contrast (fails WCAG AA 4.5:1 minimum). Using --color-surface
(dark navy in dark mode, near-white in light mode) ensures the text always
contrasts strongly with the primary background regardless of theme.
Also tints banner background with 8% primary via color-mix() so it reads as
visually distinct from the page surface without being loud.
- JobCardStack: expose resetCard() to restore card after a blocked action
- JobReviewView: call resetCard() when approve/reject returns false; prevents
card going blank after demo guard blocks the action
- useApi: add 'demo-blocked' to ApiError union; return truthy error from the
403 interceptor so store callers bail early (no optimistic UI update)
- ApplyView: add HintChip to desktop split-pane layout (was mobile-only)
- HintChip: fix text color — --app-primary-light is near-white in light theme,
causing invisible text; switch to --color-text for cross-theme contrast
- vite.config.ts: support VITE_API_TARGET env var for dev proxy override
- migrations/006: add date_posted, hired_feedback columns and references_ table
(columns existed in live DB but were missing from migration history)
- DemoBanner: commit component and test (were untracked)
Adds location blocks for /peregrine/assets/ and /peregrine/ so the SPA
works correctly when accessed via a Caddy prefix that does not strip the
path (e.g. direct host access without reverse proxy stripping).
- CRITICAL: Remove X-CF-Tier header trust from _get_effective_tier; use
Heimdall in cloud mode and APP_TIER env var in single-tenant only
- HIGH: Add update_message_body helper + PUT /api/messages/{id} endpoint;
updateMessageBody store action; approveDraft now persists edits to DB
before calling approve so history always shows the final approved text
- Cleanup: Remove dead canDraftLlm ref, checkLlmAvailable function, and
v-else-if Enable LLM drafts link; show Draft reply button unconditionally
- MEDIUM: Cap GET /api/messages limit param with Query(ge=1, le=1000)
- Test: Update test_draft_without_llm_returns_402 to patch effective_tier
instead of sending X-CF-Tier header
- scripts/messaging.py: add logged_at param to create_message; use provided value or fall back to _now_utc()
- dev-api.py: add logged_at: Optional[str] = None to MessageCreateBody
- web/src/stores/messaging.ts: remove logged_at from Omit, add as optional intersection so callers can pass it through
- web/src/components/MessageLogModal.vue: pass logged_at in handleSubmit payload; move @keydown.esc from backdrop to modal-dialog (which holds focus); compute localNow fresh inside watch so it reflects actual open time
Allocations from peregrine cloud containers were showing pipeline=null
in cf-orch analytics. Adding CF_APP_NAME to both app and api service
blocks so LLMRouter passes it as the pipeline tag on each allocation.
Move POST /api/jobs/:id/survey/analyze off the FastAPI worker thread
by routing it through the LLM task queue (same pattern as cover_letter,
company_research, resume_optimize).
- Extract prompt builders + run_survey_analyze() to scripts/survey_assistant.py
- Add survey_analyze to LLM_TASK_TYPES (task_scheduler.py) with 2.5 GB VRAM budget
(text mode: phi3:mini; visual mode uses vision service's own VRAM pool)
- Add elif branch in task_runner._run_task; result stored as JSON in error col
- Replace sync endpoint body with submit_task(); add GET /survey/analyze/task poll
- Update survey.ts store: analyze() now fires task + polls at 3s interval;
silently attaches to existing in-flight task when is_new=false
- SurveyView button label shows task stage while polling
Fixes load-test spike: ~22 greenlets blocking on LLM inference at 100 concurrent
users, causing 90s poll timeouts on cover_letter and research tasks.
Documents the cf-orch allocation pattern for cf-text and cf-voice as
openai_compat backends with a cf_orch block. Products enable these when
CF_ORCH_URL is set; the router allocates via the broker and calls the
managed service directly. No catalog or leaf details here — those live
in cf-orch node profiles (The Orchard trunk/leaf split).
