fix: search prefs wizard data loss, resume sync link, docs + GUI help links

Bug fixes (filed as #125–#128): - Wizard step 7 read data.titles instead of data.search.titles — user-entered job titles and locations were silently dropped on every wizard run (#125) - GET /api/settings/search returned "titles" key but store expected "job_titles" — Settings → Search Prefs always showed empty even when data existed (#126) - remote_only preference not persisted during wizard setup (#127) - apply-to-profile didn't set default_resume_id in user.yaml, so future Resume Profile saves never synced back to the library entry (#128) Also: - Wizard step headings corrected (off-by-one after Training step was inserted) - Ollama host in wizard inference step now reads from saved wizard state - Resume upload during wizard now creates a library entry and sets it as default Docs: - New: docs/user-guide/daily-workflow.md — end-to-end daily usage guide - Updated: docs/user-guide/settings.md — rewritten for Vue SPA (was Streamlit) - mkdocs.yml nav: Daily Workflow added as first User Guide entry GUI help links: - web/src/composables/useDocsUrl.ts — shared docs base URL composable - Home: "Daily Workflow guide ↗" link in subtitle - Job Review: "? Docs" link in title row - Resume Library: "? Help" link in header - Settings → Resume Profile: "? Help" link in page header - Settings → Search Prefs: "? Help" link in page header
2026-06-15 16:52:56 -07:00 · 2026-06-15 16:52:56 -07:00 · 82c26074d8
commit 82c26074d8
parent f799aff4e0
26 changed files with 1301 additions and 249 deletions
--- a/.env.example
+++ b/.env.example
@ -2,10 +2,10 @@
 # Auto-generated by the setup wizard, or fill in manually.
 # NEVER commit .env to git.
-STREAMLIT_PORT=8502
+VUE_PORT=8506
 OLLAMA_PORT=11434
 VLLM_PORT=8000
-CF_TEXT_PORT=8006
+CF_TEXT_PORT=8008
 SEARXNG_PORT=8888
 VISION_PORT=8002
 VISION_MODEL=vikhyatk/moondream2
--- a/.gitignore
+++ b/.gitignore
@ -60,3 +60,4 @@ demo/seed_demo.py
 tests/e2e/results/demo/
 tests/e2e/results/cloud/
 tests/e2e/results/local/
 config/wizard-test/
--- a/README.md
+++ b/README.md
@ -70,7 +70,7 @@ cd peregrine
 ./manage.sh start
 ```
-Open **http://localhost:8502** — the setup wizard walks you through the rest.
+Open **http://localhost:8506** — the setup wizard walks you through the rest.
 > **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.
@ -78,10 +78,11 @@ Open **http://localhost:8502** — the setup wizard walks you through the rest.
 ### Inference profiles
 ```bash
-./manage.sh start                       # remote — no GPU; LLM calls go to Anthropic / OpenAI
+./manage.sh start                       # cpu — local Ollama on CPU (recommended default)
 ./manage.sh start --profile cpu         # local Ollama on CPU (or Metal via native Ollama on macOS)
 ./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 cf-orch     # no local LLM — route to CircuitForge GPU cluster
 ./manage.sh start --profile remote      # no local LLM — use cloud API keys
 ```
 ---
@ -109,7 +110,7 @@ Open **http://localhost:8502** — the setup wizard walks you through the rest.
 | **Voice guidelines** (custom writing style and tone) | Premium with LLM ¹ |
 | 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 |
+| Human-in-the-loop operator (CAPTCHAs, phone calls, wet signatures) | Premium |
 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.
--- a/config/resume_keywords.yaml
+++ b/config/resume_keywords.yaml
@ -1,17 +1,47 @@
 domains:
 - B2B SaaS
 - enterprise software
 - cybersecurity
 - security
 - compliance
 - post-sale lifecycle
 - SaaS metrics
 - web security
 - risk management
 - Fortune 500
 - enterprise accounts
 - consulting
 - CS advisory
 - startup
 keywords:
 - churn reduction
 - escalation management
 - cross-functional
 - product feedback loop
 - customer advocacy
 - NPS
 - net promoter score
 - QBR
 - quarterly business review
 - executive relationships
 - EBR
 - renewal
 - expansion
 - upsell
 - health score
 - time-to-value
 - TTV
 - onboarding
 - playbook
 - success plan
 - stakeholder management
 - executive sponsor
 - risk identification
 - at-risk accounts
 - forecasting
 - GRR
 - NRR
 - ARR
 skills:
 - Customer Success
 - Technical Account Management
@ -21,3 +51,19 @@ skills:
 - project management
 - onboarding
 - renewal management
 - executive communication
 - CS leadership
 - team building
 - cross-functional collaboration
 - customer segmentation
 - success planning
 - account management
 - risk management
 - Salesforce
 - Gainsight
 - ChurnZero
 - Zendesk
 - Jira
 - Notion
 - Slack
 - Looker
--- a/dev-api.py
+++ b/dev-api.py
@ -116,12 +116,35 @@ def _load_demo_seed(db_path: str, seed_file: str) -> None:
        con.close()
 def _load_data_env() -> None:
    """Load API keys written by the wizard into the running process.
    The wizard saves keys to <data_dir>/.env (next to staging.db).  The main
    _load_env() call targets the image-baked /app/.env, which is a different
    path.  This helper bridges the gap by force-overriding env vars that are
    unset or empty (compose injects empty strings for optional vars).
    """
    data_env = Path(DB_PATH).parent / ".env"
    if not data_env.exists():
        return
    for line in data_env.read_text().splitlines():
        line = line.strip()
        if not line or line.startswith("#") or "=" not in line:
            continue
        key, _, value = line.partition("=")
        key, value = key.strip(), value.strip()
        if value and not os.environ.get(key):
            os.environ[key] = value
@asynccontextmanager
 async def lifespan(app: FastAPI):
    """Load .env, run migrations, and (in demo mode) seed the demo DB."""
    # Load .env before any runtime env reads — safe because lifespan doesn't run
    # when dev_api is imported by tests (only when uvicorn actually starts).
    _load_env(PEREGRINE_ROOT / ".env")
    # Also load wizard-saved keys from the data directory (overrides empty compose vars).
    _load_data_env()
    from scripts.db_migrate import migrate_db
    migrate_db(Path(DB_PATH))
@ -1166,9 +1189,17 @@ def apply_resume_to_profile(resume_id: int):
    with open(resume_path, "w", encoding="utf-8") as f:
        yaml.dump(current_profile, f, allow_unicode=True, default_flow_style=False)
-    from scripts.db import update_resume_synced_at as _mark_synced
+    from scripts.db import update_resume_synced_at as _mark_synced, set_default_resume as _set_default
    _mark_synced(db_path, resume_id)
    # Establish this entry as the default so future Profile saves sync back to it
    _set_default(db_path, resume_id)
    _user_yaml = db_path.parent / "config" / "user.yaml"
    if _user_yaml.exists():
        _prof = yaml.safe_load(_user_yaml.read_text(encoding="utf-8")) or {}
        _prof["default_resume_id"] = resume_id
        _user_yaml.write_text(yaml.dump(_prof, default_flow_style=False, allow_unicode=True))
    return {
        "ok":             True,
        "backup_id":      backup["id"],
@ -3250,7 +3281,23 @@ async def upload_resume(file: UploadFile):
        resume_path.parent.mkdir(parents=True, exist_ok=True)
        with open(resume_path, "w") as f:
            yaml.dump(result, f, allow_unicode=True, default_flow_style=False)
        # Also add to resume library and mark as default
        import json as _json
        from scripts.db import create_resume as _create_r, set_default_resume as _set_default
        db_path = Path(_request_db.get() or DB_PATH)
        resume_name = Path(file.filename).stem or "Uploaded Resume"
        library_entry = _create_r(
            db_path,
            name=resume_name,
            text=raw_text,
            source="upload",
            struct_json=_json.dumps(result),
        )
        _set_default(db_path, library_entry["id"])
        result["exists"] = True
        result["library_id"] = library_entry["id"]
        return {"ok": True, "data": result}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
@ -3313,6 +3360,10 @@ def get_search_prefs():
                for b in boards
            ]
        # Normalize title key — wizard saved "titles", settings canonical is "job_titles"
        if "titles" in profile and "job_titles" not in profile:
            profile["job_titles"] = profile.pop("titles")
        return profile
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
@ -3817,6 +3868,26 @@ def save_deploy_config(payload: dict):
    return {"ok": True, "note": "Restart required to apply changes"}
 class OrchUrlPayload(BaseModel):
    orch_url: str = ""
@app.get("/api/settings/system/orch-url")
 def get_orch_url():
    """Return the saved Orchard coordinator URL."""
    cfg = _load_wizard_yaml()
    return {"orch_url": cfg.get("cf_orch_url", "")}
@app.post("/api/settings/system/orch-url")
 def save_orch_url(payload: OrchUrlPayload):
    """Persist the Orchard coordinator URL to user.yaml."""
    cfg = _load_wizard_yaml()
    cfg["cf_orch_url"] = payload.orch_url.strip()
    _save_wizard_yaml(cfg)
    return {"ok": True}
 # ── Settings: Fine-Tune ───────────────────────────────────────────────────────
 _TRAINING_JSONL = Path("/Library/Documents/JobSearch/training_data/cover_letters.jsonl")
@ -4350,6 +4421,7 @@ def wizard_status():
            "linkedin": cfg.get("linkedin", ""),
            "career_summary": cfg.get("career_summary", ""),
            "services": cfg.get("services", {}),
            "cf_orch_url": cfg.get("cf_orch_url", ""),
        },
    }
@ -4371,8 +4443,8 @@ def wizard_save_step(payload: WizardStepPayload):
    step = payload.step
    data = payload.data
-    if step < 1 or step > 7:
+    if step < 1 or step > 8:
-        raise HTTPException(status_code=400, detail="step must be 1–7")
+        raise HTTPException(status_code=400, detail="step must be 1–8")
    updates: dict = {"wizard_step": step}
@ -4398,13 +4470,16 @@ def wizard_save_step(payload: WizardStepPayload):
            with open(resume_path, "w") as f:
                yaml.dump(resume, f, allow_unicode=True, default_flow_style=False)
-    elif step == 4:
+    elif step in (4, 5):
        # Step 4 (legacy) or step 5 (current) — identity fields.
        # Step 4 was the original numbering before the training step was inserted
        # between resume and identity; both are accepted for backward compat.
        for field in ("name", "email", "phone", "linkedin", "career_summary"):
            if field in data:
                updates[field] = data[field]
-    elif step == 5:
+    elif step == 6:
-        # Write API keys to .env (never store in user.yaml)
+        # Step 6 — inference: API keys + optional Orchard coordinator URL.
        env_path = Path(_wizard_yaml_path()).parent.parent / ".env"
        env_lines = env_path.read_text().splitlines() if env_path.exists() else []
@ -4422,18 +4497,24 @@ def wizard_save_step(payload: WizardStepPayload):
            env_lines = _set_env_key(env_lines, "OPENAI_COMPAT_URL", data["openai_url"])
        if data.get("openai_key"):
            env_lines = _set_env_key(env_lines, "OPENAI_COMPAT_KEY", data["openai_key"])
-        if any(data.get(k) for k in ("anthropic_key", "openai_url", "openai_key")):
+        if data.get("orch_url"):
            env_lines = _set_env_key(env_lines, "GPU_SERVER_URL", data["orch_url"])
            updates["cf_orch_url"] = data["orch_url"]
        if any(data.get(k) for k in ("anthropic_key", "openai_url", "openai_key", "orch_url")):
            env_path.parent.mkdir(parents=True, exist_ok=True)
            env_path.write_text("\n".join(env_lines) + "\n")
        if "services" in data:
            updates["services"] = data["services"]
-    elif step == 6:
+    elif step == 7:
-        # Persist search preferences to search_profiles.yaml in canonical format:
+        # Step 7 — search preferences.
-        #   profiles: [{name, titles, locations, boards, ...}]
+        # Wizard sends { search: { titles, locations, remote_only } }; fall back to
-        titles = data.get("titles", [])
+        # top-level keys for direct API callers that omit the "search" wrapper.
-        locations = data.get("locations", [])
+        search = data.get("search", {})
        titles = search.get("titles", data.get("titles", data.get("job_titles", [])))
        locations = search.get("locations", data.get("locations", []))
        remote_only = search.get("remote_only", data.get("remote_only", False))
        search_path = _search_prefs_path()
        existing_search: dict = {}
        if search_path.exists():
@ -4450,14 +4531,15 @@ def wizard_save_step(payload: WizardStepPayload):
        if default_profile is None:
            default_profile = {"name": "default"}
            profiles_list.append(default_profile)
-        default_profile["titles"] = titles
+        default_profile["job_titles"] = titles
        default_profile["locations"] = locations
        default_profile["remote_only"] = remote_only
        existing_search["profiles"] = profiles_list
        search_path.parent.mkdir(parents=True, exist_ok=True)
        with open(search_path, "w") as f:
            yaml.dump(existing_search, f, allow_unicode=True, default_flow_style=False)
-    # Step 7 (integrations) has no extra side effects here — connections are
+    # Step 8 (integrations) has no extra side effects here — connections are
    # handled by the existing /api/settings/system/integrations/{id}/connect.
    try:
@ -4484,6 +4566,39 @@ def _fetch_cforch_nodes() -> list[dict]:
        return []
 def _probe_ollama() -> bool:
    """Return True if Ollama is reachable from inside the container."""
    candidates = [
        "http://host.docker.internal:11434/api/tags",
        "http://ollama:11434/api/tags",
    ]
    for url in candidates:
        try:
            r = requests.get(url, timeout=2)
            if r.status_code == 200:
                return True
        except Exception:
            pass
    return False
 def _probe_searxng() -> bool:
    """Return True if SearXNG is reachable from inside the container."""
    candidates = [
        "http://searxng:8080/",
        "http://host.docker.internal:8888/",
        "http://host.docker.internal:8080/",
    ]
    for url in candidates:
        try:
            r = requests.get(url, timeout=2)
            if r.status_code < 500:
                return True
        except Exception:
            pass
    return False
@app.get("/api/wizard/hardware")
 def wizard_hardware():
    """Detect local GPUs, suggest an inference profile, and report cf-orch nodes."""
@ -4502,35 +4617,71 @@ def wizard_hardware():
                "vram_free_mb": gpu["vram_free_mb"],
            })
    ollama_running = _probe_ollama()
    searxng_running = _probe_searxng()
    # If no GPU but Ollama is already running, default to cpu rather than remote
    if suggested == "cpu" and not gpus and not ollama_running:
        suggested = "remote"
    return {
        "gpus": gpus,
        "suggested_profile": suggested,
        "profiles": list(_WIZARD_PROFILES),
        "cf_orch_available": len(orch_nodes) > 0,
        "cf_orch_gpus": orch_summary,
        "ollama_running": ollama_running,
        "searxng_running": searxng_running,
    }
 def _container_safe_url(url: str) -> str:
    """Replace localhost/127.0.0.1 with host.docker.internal so tests reach the host."""
    import re as _re
    return _re.sub(r"(https?://)(?:localhost|127\.0\.0\.1)\b", r"\1host.docker.internal", url)
 class WizardInferenceTestPayload(BaseModel):
    profile: str = "remote"
    anthropic_key: str = ""
    openai_url: str = ""
    openai_key: str = ""
    orch_url: str = ""
    ollama_host: str = "localhost"
    ollama_port: int = 11434
@app.post("/api/wizard/inference/test")
 def wizard_test_inference(payload: WizardInferenceTestPayload):
-    """Test LLM or Ollama connectivity.
+    """Test LLM, Ollama, or Orchard coordinator connectivity.
-    Always returns {ok, message} — a connection failure is reported as a
+    Always returns {ok, message} — a connection failure is a soft warning so
-    soft warning (message), not an HTTP error, so the wizard can let the
+    the wizard lets the user continue past a temporarily-unreachable service.
    user continue past a temporarily-down Ollama instance.
    """
-    if payload.profile == "remote":
+    if payload.profile == "cf-orch":
        orch_url = _container_safe_url(payload.orch_url.rstrip("/")) if payload.orch_url else ""
        if not orch_url:
            return {"ok": False, "message": "Enter the Orchard coordinator URL first."}
        try:
            resp = requests.get(f"{orch_url}/api/nodes", timeout=5,
                                headers={"Accept": "application/json"})
            if resp.status_code == 200:
                nodes = resp.json().get("nodes", [])
                n = len(nodes)
                return {"ok": True, "message": f"Orchard reachable — {n} node(s) online."}
            return {"ok": False, "message": f"Orchard returned HTTP {resp.status_code}."}
        except Exception as exc:
            return {
                "ok": False,
                "message": (
                    f"Cannot reach Orchard at {payload.orch_url} — "
                    "check the URL and that the coordinator is running. "
                    f"({exc})"
                ),
            }
    elif payload.profile == "remote":
        try:
            # Temporarily inject key if provided (don't persist yet)
            env_override = {}
            if payload.anthropic_key:
                env_override["ANTHROPIC_API_KEY"] = payload.anthropic_key
@ -4554,15 +4705,16 @@ def wizard_test_inference(payload: WizardInferenceTestPayload):
                        os.environ[k] = v
        except Exception as exc:
            return {"ok": False, "message": f"LLM test failed: {exc}"}
    else:
-        # Local profile — ping Ollama
+        # Local profiles (cpu, single-gpu, dual-gpu) — ping Ollama
-        ollama_url = f"http://{payload.ollama_host}:{payload.ollama_port}"
+        host = payload.ollama_host or "localhost"
        ollama_url = _container_safe_url(f"http://{host}:{payload.ollama_port}")
        try:
            resp = requests.get(f"{ollama_url}/api/tags", timeout=5)
            ok = resp.status_code == 200
            message = "Ollama is running." if ok else f"Ollama returned HTTP {resp.status_code}."
        except Exception:
            # Soft-fail: user can skip and configure later
            return {
                "ok": False,
                "message": (
--- a/docs/getting-started/docker-profiles.md
+++ b/docs/getting-started/docker-profiles.md
@ -1,69 +1,129 @@
 # Docker Profiles
-Peregrine uses Docker Compose profiles to start only the services your hardware can support. Choose a profile with `make start PROFILE=<name>`.
+Peregrine uses Docker Compose profiles to start only the services your hardware supports. Choose a profile with `./manage.sh start --profile <name>`.
 `manage.sh` delegates to `make`, which auto-detects Docker vs Podman and applies the correct GPU overlay — `compose.gpu.yml` for Docker, `compose.podman-gpu.yml` for Podman (CDI-based). You do not need to specify the overlay manually.
 ---
 ## Profile Reference
 | Profile | Services started | Use case |
-|---------|----------------|----------|
+|---------|-----------------|----------|
-| `remote` | `app`, `searxng` | No GPU. LLM calls go to an external API (Anthropic, OpenAI-compatible). |
+| `cpu` | `web`, `api`, `ollama`, `searxng` | No GPU. Local models on CPU. Recommended default for new installs. |
-| `cpu` | `app`, `ollama`, `searxng` | No GPU. Runs local models on CPU — functional but slow. |
+| `single-gpu` | `web`, `api`, `ollama`, `vision`, `searxng` | One NVIDIA GPU. Covers cover letters, research, and vision. |
-| `single-gpu` | `app`, `ollama`, `vision`, `searxng` | One NVIDIA GPU. Covers cover letters, research, and vision (survey screenshots). |
+| `dual-gpu` | `web`, `api`, `ollama`, `vllm`, `vision`, `searxng` | Two NVIDIA GPUs. GPU split controlled by `DUAL_GPU_MODE`. |
-| `dual-gpu` | `app`, `ollama`, `vllm`, `vision`, `searxng` | Two NVIDIA GPUs. GPU 0 = Ollama (cover letters), GPU 1 = vLLM (research). |
+| `cf-orch` | `web`, `api`, `searxng` | No local LLM. Inference routed to CircuitForge GPU cluster. Requires Paid license. |
 | `remote` | `web`, `api`, `searxng` | No local LLM. Inference goes to cloud API keys (Anthropic, OpenAI-compatible). |
 | `memory` | (any + memory flag) | Enables RAM-optimised container limits for low-RAM machines. Combine with another profile. |
 ---
 ## Service Descriptions
-| Service | Image / Source | Port | Purpose |
+| Service | Image / Source | Host Port | Purpose |
-|---------|---------------|------|---------|
+|---------|---------------|-----------|---------|
-| `app` | `Dockerfile` (Streamlit) | 8501 | The main Peregrine UI |
+| `web` | `Dockerfile.web` (Nginx + Vue SPA) | `VUE_PORT` (default 8506) | Main UI — serves the Vue frontend and proxies `/api/` to `api` |
 | `api` | `Dockerfile` (FastAPI) | Internal only (proxied through `web`) | REST API — all backend logic |
 | `ollama` | `ollama/ollama` | 11434 | Local model inference — cover letters and general tasks |
-| `vllm` | `vllm/vllm-openai` | 8000 | High-throughput local inference — research tasks |
+| `vllm` | `vllm/vllm-openai` | 8000 | High-throughput inference — research tasks |
 | `vision` | `scripts/vision_service/` | 8002 | Moondream2 — survey screenshot analysis |
-| `searxng` | `searxng/searxng` | 8888 | Private meta-search engine — company research web scraping |
+| `searxng` | `searxng/searxng` | 8888 | Private meta-search — company research web scraping |
 The `web` container runs Nginx internally on port 80, mapped to `VUE_PORT` on the host. The Nginx config proxies `/api/` requests to `api:8601` — the FastAPI container is not exposed directly.
 ---
 ## Choosing a Profile
 ### remote
 Use `remote` if:
 - You have no NVIDIA GPU
 - You plan to use Anthropic Claude or another API-hosted model exclusively
 - You want the fastest startup (only two containers)
 You must configure at least one external LLM backend in **Settings → LLM Backends**.
 ### cpu
 Use `cpu` if:
- You have no GPU but want to run models locally (e.g. for privacy)
+- You have no GPU but want local inference (good for privacy)
 - Acceptable for light use; cover letter generation may take several minutes per request
-Pull a model after the container starts:
+Pull a model after starting:
 ```bash
-docker exec -it peregrine-ollama-1 ollama pull llama3.1:8b
+docker exec -it peregrine-ollama-1 ollama pull llama3.2:3b
 ```
 `llama3.2:3b` is the recommended CPU model — it runs on machines with 8 GB of system RAM.
 ### single-gpu
 Use `single-gpu` if:
 - You have one NVIDIA GPU with at least 8 GB VRAM
 - Recommended for most single-user installs
- The vision service (Moondream2) starts on the same GPU using 4-bit quantisation (~1.5 GB VRAM)
+
 The vision service (Moondream2) starts on the same GPU using 4-bit quantisation (~1.5 GB VRAM). Pull a model after starting:
 ```bash
 docker exec -it peregrine-ollama-1 ollama pull llama3.1:8b
 ```
 ### dual-gpu
 Use `dual-gpu` if:
 - You have two or more NVIDIA GPUs
- GPU 0 handles Ollama (cover letters, quick tasks)
+- Default: GPU 0 handles Ollama (cover letters), GPU 1 handles vLLM (research)
- GPU 1 handles vLLM (research, long-context tasks)
+
- The vision service shares GPU 0 with Ollama
+See [Dual-GPU Modes](#dual-gpu-modes) below to configure how the two GPUs are split.
 ### cf-orch
 Use `cf-orch` if:
 - You have access to a CircuitForge GPU cluster running the cf-orch coordinator
 - No local GPU required — inference is handled by the cluster
 - Requires a Paid or higher license
 Set `CF_ORCH_URL` in `.env` to your coordinator address:
 ```bash
 CF_ORCH_URL=http://10.1.10.71:7700
 ```
 The wizard hardware step lets you enter the URL interactively and verifies the connection before saving.
 ### remote
 Use `remote` if:
 - You have no local GPU and no cf-orch cluster
 - You are using Anthropic Claude, OpenAI, or another cloud API exclusively
 Configure at least one external LLM backend in **Settings → LLM Backends** after first login.
 ### memory (add-on)
 Use the `memory` add-on alongside any profile for machines with limited RAM:
 ```bash
 ./manage.sh start --profile single-gpu --profile memory
 ```
 This applies conservative container memory limits to prevent the OOM (out-of-memory) killer from terminating containers.
 ---
 ## Dual-GPU Modes
 When using `dual-gpu`, `DUAL_GPU_MODE` in `.env` controls how the second GPU is used:
 | Mode | GPU 0 | GPU 1 | Use case |
 |------|-------|-------|----------|
 | `mixed` (default) | Ollama | vLLM | Best overall: fast cover letters + high-throughput research |
 | `ollama` | Ollama | Ollama | Both GPUs run Ollama; no vLLM; useful if vLLM models are too large for one card |
 | `vllm` | vLLM | vLLM | Both GPUs run vLLM (tensor parallel); maximum research throughput |
 Set in `.env`:
 ```bash
 DUAL_GPU_MODE=mixed    # default
 # DUAL_GPU_MODE=ollama
 # DUAL_GPU_MODE=vllm
 ```
 The Makefile expands `dual-gpu` into `--profile dual-gpu-$(DUAL_GPU_MODE)` before passing it to `docker compose`. The `compose.gpu.yml` overlay defines the `dual-gpu-mixed`, `dual-gpu-ollama`, and `dual-gpu-vllm` profile variants.
 ---
@ -75,40 +135,69 @@ Use `dual-gpu` if:
 | 4–8 GB | `single-gpu` | Run smaller models (3B–8B parameters) |
 | 8–16 GB | `single-gpu` | Run 8B–13B models comfortably |
 | 16–24 GB | `single-gpu` | Run 13B–34B models |
-| 24 GB+ | `single-gpu` or `dual-gpu` | 70B models with quantisation |
+| 24 GB+ (one card) | `single-gpu` | 70B models with quantisation |
 | 16+ GB (two cards) | `dual-gpu` | Parallel cover letters + research |
 ---
 ## How preflight.py Works
-`make start` calls `scripts/preflight.py` before launching Docker. Preflight does the following:
+`./manage.sh start` calls `scripts/preflight.py` before launching Docker. Preflight does the following:
-1. **Port conflict detection** — checks whether `STREAMLIT_PORT`, `OLLAMA_PORT`, `VLLM_PORT`, `SEARXNG_PORT`, and `VISION_PORT` are already in use. Reports any conflicts and suggests alternatives.
+1. **Port conflict detection** — checks whether `VUE_PORT`, `OLLAMA_PORT`, `VLLM_PORT`, `SEARXNG_PORT`, and `VISION_PORT` are already in use. Reports any conflicts and suggests alternatives.
-2. **GPU enumeration** — queries `nvidia-smi` for GPU count and VRAM per card.
+2. **External service adoption** — if Ollama or SearXNG are already running on their configured ports (common when using native Ollama on macOS, or a shared SearXNG instance), preflight writes a `compose.override.yml` that stubs out the duplicate containers. The running process is adopted rather than replaced.
-3. **RAM check** — reads `/proc/meminfo` (Linux) or `vm_stat` (macOS) to determine available system RAM.
+3. **GPU enumeration** — queries `nvidia-smi` for GPU count and VRAM per card. On Apple Silicon Macs, falls back to `system_profiler SPDisplaysDataType` and returns unified memory as the VRAM figure.
-4. **KV cache offload** — if GPU VRAM is less than 10 GB, preflight calculates `CPU_OFFLOAD_GB` (the amount of KV cache to spill to system RAM) and writes it to `.env`. The vLLM container picks this up via `--cpu-offload-gb`.
+4. **RAM check** — reads `/proc/meminfo` (Linux) or `vm_stat` (macOS) for available system RAM.
-5. **Profile recommendation** — writes `RECOMMENDED_PROFILE` to `.env`. This is informational; `make start` uses the `PROFILE` variable you specify (defaulting to `remote`).
+5. **KV cache offload** — if GPU VRAM is less than 10 GB, preflight calculates `CPU_OFFLOAD_GB` and writes it to `.env`. The vLLM container picks this up via `--cpu-offload-gb` to overflow the KV cache to system RAM.
-You can run preflight independently:
+6. **Profile recommendation** — writes `RECOMMENDED_PROFILE` to `.env`. This is informational only; `./manage.sh start --profile <name>` uses the profile you specify.
 Run preflight independently at any time:
 ```bash
-make preflight
+./manage.sh preflight
 # or
-python scripts/preflight.py
+conda run -n cf python scripts/preflight.py
 ```
 ---
 ## Podman Support
 Podman is fully supported as a Docker drop-in. `install.sh` detects whether Podman or Docker is available, and `manage.sh`/`make` use it automatically.
 ### GPU setup for Podman (CDI)
 Podman uses the CDI (Container Device Interface) standard for GPU passthrough, rather than Docker's `--gpus all` flag. Generate the CDI spec once after driver installation:
 ```bash
 sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml
 ```
 Without this step, GPU profiles start but containers have no GPU access.
 ### Rootless Podman
 Rootless Podman is supported. If you encounter permission errors on the Docker socket, ensure `podman.socket` is running for your user:
 ```bash
 systemctl --user enable --now podman.socket
 ```
 The `make` layer auto-detects rootless Podman and uses `XDG_RUNTIME_DIR/podman/podman.sock` instead of `/var/run/docker.sock`.
 ---
 ## Customising Ports
-Edit `.env` before running `make start`:
+Edit `.env` before running `./manage.sh start`:
 ```bash
-STREAMLIT_PORT=8501
+VUE_PORT=8506          # main UI (Vue SPA)
 OLLAMA_PORT=11434
 VLLM_PORT=8000
 SEARXNG_PORT=8888
@ -116,3 +205,15 @@ VISION_PORT=8002
 ```
 All containers read from `.env` via the `env_file` directive in `compose.yml`.
 ---
 ## Wizard Test Instance
 A separate compose file is available for testing first-run and onboarding wizard flows without touching your main data:
 ```bash
 docker compose -f compose.wizard-test.yml --project-name peregrine-wizard up -d
 ```
 The wizard test instance runs on port **8507** with ephemeral storage — every `docker compose restart` wipes the database back to a clean slate. Uses the same images as the main instance but mounts a minimal LLM config so the wizard detection endpoints work correctly.
--- a/docs/getting-started/installation.md
+++ b/docs/getting-started/installation.md
@ -7,7 +7,7 @@ This page walks through a full Peregrine installation from scratch.
 ## Prerequisites
 - **Git** — to clone the repository
- **Internet connection** — `install.sh` downloads Docker and other dependencies
+- **Internet connection** — `install.sh` downloads Docker/Podman and other dependencies
 - **Operating system**: Ubuntu/Debian, Fedora/RHEL, Arch Linux, or macOS (with Docker Desktop)
 !!! warning "Windows"
@ -34,16 +34,28 @@ bash install.sh
 1. **Detects your platform** (Ubuntu/Debian, Fedora/RHEL, Arch, macOS)
 2. **Installs Git** if not already present
-3. **Installs Docker Engine** and the Docker Compose v2 plugin via the official Docker repositories
+3. **Installs Docker Engine** (or Podman if Docker is not available) via official repositories
 4. **Adds your user to the `docker` group** so you do not need `sudo` for docker commands (Linux only — log out and back in after this)
-5. **Detects NVIDIA GPUs** — if `nvidia-smi` is present and working, installs the NVIDIA Container Toolkit and configures Docker to use it
+5. **Detects NVIDIA GPUs** — if `nvidia-smi` is present and working, installs the NVIDIA Container Toolkit and configures Docker/Podman to use it
 6. **Creates `.env` from `.env.example`** — edit `.env` to customise ports and model storage paths before starting
 !!! note "macOS"
-    `install.sh` installs Docker Desktop via Homebrew (`brew install --cask docker`) then exits. Open Docker Desktop, start it, then re-run the script.
+    `install.sh` installs Docker Desktop via Homebrew (`brew install --cask docker`) then exits. Open Docker Desktop, start it, then re-run the script. Ollama can also run natively for Metal GPU-accelerated inference — see the macOS note in Step 4.
 !!! note "GPU requirement"
-    For GPU support, `nvidia-smi` must return output before you run `install.sh`. Install your NVIDIA driver first. The Container Toolkit installation will fail silently if the driver is not present.
+    For GPU support, `nvidia-smi` must return output before you run `install.sh`. Install your NVIDIA driver first.
 ---
 ## Step 2a — Podman users: GPU CDI setup
 If you prefer rootless Podman over Docker, `install.sh` detects it and manages.sh/make use it automatically. For GPU profiles to work with Podman you must generate a CDI spec first:
 ```bash
 sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml
 ```
 This needs to be done once after driver installation. Without it, GPU profiles will start but containers will not have GPU access. Docker users can skip this step — Docker uses `--gpus all` instead of CDI.
 ---
@ -52,15 +64,21 @@ bash install.sh
 The `.env` file controls ports and volume mount paths. The defaults work for most single-user installs:
 ```bash
-# Default ports
+# Main UI port
-STREAMLIT_PORT=8501
+VUE_PORT=8506
-OLLAMA_PORT=11434
+
-VLLM_PORT=8000
+# Model paths — use full absolute paths, not ~ (tilde does not expand inside containers)
-SEARXNG_PORT=8888
+DOCS_DIR=/home/yourname/Documents/JobSearch
-VISION_PORT=8002
+OLLAMA_MODELS_DIR=/home/yourname/models/ollama
 # Inference model defaults
 OLLAMA_DEFAULT_MODEL=llama3.2:3b
 # External API keys — only needed for the "remote" profile or BYOK unlock
 ANTHROPIC_API_KEY=
 ```
-Change `STREAMLIT_PORT` if 8501 is taken on your machine.
+Change `VUE_PORT` if 8506 is taken on your machine. See [Docker Profiles](docker-profiles.md) for a full port reference.
 ---
@ -69,21 +87,24 @@ Change `STREAMLIT_PORT` if 8501 is taken on your machine.
 Choose a profile based on your hardware:
 ```bash
-make start                        # remote — no GPU, use API-only LLMs
+./manage.sh start                        # cpu — local Ollama on CPU (recommended default)
-make start PROFILE=cpu            # cpu — local models on CPU (slow)
+./manage.sh start --profile single-gpu   # one NVIDIA GPU
-make start PROFILE=single-gpu     # single-gpu — one NVIDIA GPU
+./manage.sh start --profile dual-gpu     # two NVIDIA GPUs
-make start PROFILE=dual-gpu       # dual-gpu — GPU 0 = Ollama, GPU 1 = vLLM
+./manage.sh start --profile remote       # no local LLM — use cloud API keys only
 ```
-`make start` runs `preflight.py` first, which checks for port conflicts and writes GPU/RAM recommendations back to `.env`. Then it calls `docker compose --profile <PROFILE> up -d`.
+`manage.sh start` runs `preflight.py` first, which checks for port conflicts and writes GPU/RAM recommendations to `.env`. Then it calls `docker compose` (or `podman compose`) with the right compose file overlay for your hardware.
 !!! tip "macOS with native Ollama"
    If you installed Ollama natively via Homebrew for Metal GPU inference, start with `--profile cpu`. The container API on port 8506 connects to your host's Ollama at `localhost:11434` automatically.
 ---
 ## Step 5 — Open the UI
-Navigate to **http://localhost:8501** (or whatever `STREAMLIT_PORT` you set).
+Navigate to **http://localhost:8506** (or whatever `VUE_PORT` you set).
-The first-run wizard launches automatically. See [First-Run Wizard](first-run-wizard.md) for a step-by-step guide through all seven steps.
+The first-run wizard launches automatically. See [First-Run Wizard](first-run-wizard.md) for a step-by-step guide.
 ---
@ -96,7 +117,7 @@ The first-run wizard launches automatically. See [First-Run Wizard](first-run-wi
 | Fedora 39/40 | Yes | |
 | RHEL / Rocky / AlmaLinux | Yes | |
 | Arch Linux / Manjaro | Yes | |
-| macOS (Apple Silicon) | Yes | Docker Desktop required; no GPU support |
+| macOS (Apple Silicon) | Yes | Docker Desktop required; GPU via native Ollama (Metal) |
 | macOS (Intel) | Yes | Docker Desktop required; no GPU support |
 | Windows | No | Use WSL2 with Ubuntu |
@ -107,20 +128,23 @@ The first-run wizard launches automatically. See [First-Run Wizard](first-run-wi
 Only NVIDIA GPUs are supported. AMD ROCm is not currently supported.
 Requirements:
 - NVIDIA driver installed and `nvidia-smi` working before running `install.sh`
 - CUDA 12.x recommended (CUDA 11.x may work but is untested)
 - Minimum 8 GB VRAM for `single-gpu` profile with default models
- For `dual-gpu`: GPU 0 is assigned to Ollama, GPU 1 to vLLM
+- **Podman users:** CDI spec required — see Step 2a above
-If your GPU has less than 10 GB VRAM, `preflight.py` will calculate a `CPU_OFFLOAD_GB` value and write it to `.env`. The vLLM container picks this up via `--cpu-offload-gb` to overflow KV cache to system RAM.
+For `dual-gpu`, both cards must be NVIDIA. GPU 0 handles Ollama (cover letters, general tasks) and GPU 1 handles the research workload. The exact behaviour is controlled by `DUAL_GPU_MODE` — see [Docker Profiles](docker-profiles.md#dual-gpu-modes).
 If your GPU has less than 10 GB VRAM, `preflight.py` calculates a `CPU_OFFLOAD_GB` value and writes it to `.env`. The vLLM container picks this up via `--cpu-offload-gb` to overflow KV cache to system RAM.
 ---
 ## Stopping Peregrine
 ```bash
-make stop       # stop all containers
+./manage.sh stop       # stop all containers
-make restart    # stop then start again (runs preflight first)
+./manage.sh restart    # stop then start again (runs preflight first)
 ```
 ---
@ -128,7 +152,7 @@ make restart    # stop then start again (runs preflight first)
 ## Reinstalling / Clean State
 ```bash
-make clean      # removes containers, images, and data volumes (destructive)
+./manage.sh clean      # removes containers, images, and data volumes (destructive)
 ```
 You will be prompted to type `yes` to confirm.
--- a/docs/user-guide/daily-workflow.md
+++ b/docs/user-guide/daily-workflow.md
@ -0,0 +1,142 @@
 # Daily Workflow
 This page describes how Peregrine fits into a typical active job search. The core loop is short: find jobs, triage them, generate and send applications, track what happens next.
 ---
 ## The Core Loop
 ```
 Run Discovery → Review Jobs → Apply Workspace → Track in Interviews
 ```
 Each stage feeds the next. You can run the full loop in under ten minutes on a good day, or spend longer editing cover letters and doing interview prep when you need to.
 ---
 ## Starting Your Day
 ### 1. Run Discovery
 Open the **Home** page and click **Run Discovery**. Peregrine queries all your configured job boards simultaneously and stores results in the local database.
 - Discovery runs one search profile at a time. Each profile produces results per board, then moves to the next.
 - A summary at the end shows how many new jobs were found vs. already known.
 - Jobs you have already seen (by URL) are skipped automatically.
 If some jobs came back with short descriptions, click **Fill Missing Descriptions** to enrich them in the background while you work.
 See [Job Discovery](job-discovery.md) for search profile configuration and board details.
 ---
 ### 2. Review the Queue
 Navigate to **Job Review**. New jobs arrive with status `pending` and appear in the review queue.
 For each job you can:
 - **Approve** — sends it into the application pipeline
 - **Reject** — archives it out of the queue
 Sort by **Match Score** (high to low) to see the best keyword matches first. The match score compares the job description against your resume keywords — a rough signal, not a hard filter.
 Jobs with incoming email leads (a recruiter contacted you about this role) sort to the top automatically.
 See [Job Review](job-review.md) for sorting, keyword gaps, and bulk actions.
 ---
 ### 3. Write and Send Applications
 Navigate to **Apply Workspace**. All approved jobs appear here.
 For each job:
 1. Click **Generate Cover Letter** — runs as a background task using your resume and career summary.
 2. Read and edit the result. The generator uses your mission alignment notes when it detects company fit.
 3. Click **Export PDF** to save a formatted PDF to your documents directory.
 4. Apply externally (via the company site or board).
 5. Click **Mark Applied** to move the job into the Interviews kanban.
 See [Apply Workspace](apply-workspace.md) for cover letter configuration, PDF formatting, and ATS optimization.
 ---
 ### 4. Track Interviews
 The **Interviews** page is a kanban board. Jobs move through stages as your search progresses:
 ```
 applied → phone_screen → interviewing → offer → hired
 ```
 When a job moves to **phone_screen**, Peregrine automatically kicks off a company research brief in the background — a one-page summary of the company, recent news, leadership, and accessibility signals.
 Use **Interview Prep** to review talking points, practice Q&A, and get live reference cards during calls.
 See [Interviews](interviews.md) for stage transitions, research briefs, and prep tools.
 ---
 ## Managing Your Resume
 Peregrine has two resume views that work together:
 ### Resume Library (`/resumes`)
 An archive of every resume version — uploaded originals, AI-optimised variants, and auto-backups. The starred entry is your **active default**.
 - **Import** a PDF, DOCX, ODT, or plain text file to add a version to the library.
 - **★ Set as Default** marks the entry as the active resume used for cover letter generation and keyword matching.
 - **⇩ Apply to profile** pushes a library entry into the structured Resume Profile (see below), and links it so future profile edits sync back automatically.
 ### Resume Profile (`Settings → Resume Profile`)
 A structured editor for personal details, work experience, education, and skills. This is the data the cover letter generator reads directly.
 - When content was applied from the library, the view shows a sync status and date.
 - Saving the Resume Profile automatically updates the linked library entry — keeping them in sync without manual effort.
 - You can replace the current profile by uploading a new file directly from this view.
 **Recommended flow:** upload to the library → set as default → "Apply to profile" → edit in Resume Profile as needed. Your library stays current automatically.
 ---
 ## Keeping Search Preferences Fresh
 Go to **Settings → Search Prefs** to update what Peregrine searches for.
 Key fields:
 | Field | What it does |
 |-------|-------------|
 | Job Titles | The roles searched across all boards |
 | Locations | Geographic scope (leave blank for unrestricted) |
 | Remote only | Filter to remote positions only |
 | Exclude Keywords | Drop any job title containing these words before it enters the database |
 | Job Boards | Enable or disable specific sources |
 | Blocklists | Companies, industries, or locations to always skip |
 Click **Suggest** next to any field to get AI-generated suggestions based on your resume profile.
 Changes take effect on the next discovery run — no restart needed.
 ---
 ## Weekly Habits
 **Clean up the queue** — reject stale pending jobs at least once a week so the queue stays scannable.
 **Update your search prefs** — if you are getting too many mismatches, add more terms to Exclude Keywords. If the queue is thin, broaden Locations or add boards.
 **Check Interviews** — move any stalled jobs to the right stage so the kanban reflects reality. The research brief appears in Interview Prep once a job reaches `phone_screen`.
 **Tune your resume keywords** — go to **Settings → Skills** if you want to add or reweight keywords used for match scoring.
 ---
 ## Tips
 - **Match score is a triage signal, not a gate.** A score of 40 might be a perfect cultural fit that uses different terminology. Read the description.
 - **Cover letters improve with context.** The richer your career summary and mission alignment notes (Settings → My Profile), the more specific and accurate the generated letters.
 - **Company research auto-runs.** You do not need to request it manually — it starts the moment a job hits `phone_screen`.
 - **Everything is local.** Your database, resume, and application history live in `data/staging.db` and `data/config/`. Back them up like any other important file.
--- a/docs/user-guide/settings.md
+++ b/docs/user-guide/settings.md
@ -1,6 +1,8 @@
 # Settings
-The Settings page is accessible from the sidebar. It contains all configuration for Peregrine, organised into tabs.
+Access Settings from the sidebar. The page has a navigation panel on the left (desktop) or a chip bar at the top (mobile). Each section is described below.
 For an overview of how settings fit into your daily use, see [Daily Workflow](daily-workflow.md).
 ---
@ -10,143 +12,177 @@ Personal information used in cover letters, research briefs, and interview prep.
 | Field | Description |
 |-------|-------------|
-| Name | Your full name |
+| Full name | Your name as it appears in generated documents |
-| Email | Contact email address |
+| Email | Contact email |
 | Phone | Contact phone number |
-| LinkedIn | LinkedIn profile URL |
+| LinkedIn URL | Used in cover letter headers |
-| Career summary | 2–4 sentence professional summary |
+| Career summary | 2–4 sentences that anchor all LLM-generated content |
 | NDA companies | Companies you cannot mention in research briefs (previous employers under NDA) |
 | Docs directory | Where PDFs and exported documents are saved (default: `~/Documents/JobSearch`) |
 ### Mission Preferences
-Optional notes about industries you genuinely care about. When the cover letter generator detects alignment with one of these industries, it injects your note into paragraph 3 of the cover letter.
+Optional notes about industries you genuinely care about. When the cover letter generator detects alignment with one of these industries, it injects your note into the generated letter.
-| Field | Tag | Example |
+| Field | Tag |
-|-------|-----|---------|
+|-------|-----|
-| Music industry note | `music` | "I've played in bands for 15 years and care deeply about how artists get paid" |
+| Music industry note | `music` |
-| Animal welfare note | `animal_welfare` | "I volunteer at my local shelter every weekend" |
+| Animal welfare note | `animal_welfare` |
-| Education note | `education` | "I tutored underserved kids and care deeply about literacy" |
+| Education note | `education` |
 Leave a field blank to use a generic default when alignment is detected.
 ### Research Brief Preferences
-Controls optional sections in company research briefs. Both are for personal decision-making only and are never included in applications.
+Controls optional sections in company research briefs. Both are for personal decision-making only and never appear in applications.
-| Setting | Section added |
+| Setting | Section added to brief |
-|---------|--------------|
+|---------|----------------------|
-| Candidate accessibility focus | Disability inclusion and accessibility signals (ADA, ERGs, WCAG) |
+| Accessibility focus | Disability inclusion signals (ADA, ERGs, WCAG) |
-| Candidate LGBTQIA+ focus | LGBTQIA+ inclusion signals (ERGs, non-discrimination policies, culture) |
+| LGBTQIA+ focus | Inclusion signals (ERGs, non-discrimination policies) |
 ---
 ## Search
 Manage search profiles. Equivalent to editing `config/search_profiles.yaml` directly, but with a form UI.
 - Add, edit, and delete profiles
 - Configure titles, locations, boards, custom boards, exclude keywords, and mission tags
 - Changes are saved to `config/search_profiles.yaml`
 ---
 ## LLM Backends
 Configure which LLM backends Peregrine uses and in what order.
 | Setting | Description |
 |---------|-------------|
 | Enabled toggle | Whether a backend is considered in the fallback chain |
 | Base URL | API endpoint (for `openai_compat` backends) |
 | Model | Model name or `__auto__` (vLLM auto-detects the loaded model) |
 | API key | API key if required |
 | Test button | Sends a short ping to verify the backend is reachable |
 ### Fallback chains
 Three independent fallback chains are configured:
 | Chain | Used for |
 |-------|---------|
 | `fallback_order` | Cover letter generation and general tasks |
 | `research_fallback_order` | Company research briefs |
 | `vision_fallback_order` | Survey screenshot analysis |
 ---
 ## Notion
 Configure Notion integration credentials. Requires:
 - Notion integration token (from [notion.so/my-integrations](https://www.notion.so/my-integrations))
 - Database ID (from the Notion database URL)
 The field map controls which Notion properties correspond to which Peregrine fields. Edit `config/notion.yaml` directly for advanced field mapping.
 ---
 ## Services
 Connection settings for local services:
 | Service | Default host:port |
 |---------|-----------------|
 | Ollama | localhost:11434 |
 | vLLM | localhost:8000 |
 | SearXNG | localhost:8888 |
 Each service has SSL and SSL-verify toggles for reverse-proxy setups.
 ---
 ## Resume Profile
-Edit your parsed resume data (work experience, education, skills, certifications). This is the same data extracted during the first-run wizard Resume step.
+A structured editor for your work experience, education, skills, and personal details. This is the primary data source for cover letter generation.
-Changes here affect all future cover letter generations.
+### Resume vs. Library
 The Resume Profile is backed by a structured YAML file (`plain_text_resume.yaml`). The **Resume Library** (`/resumes`, accessible from the sidebar) is a versioned archive of full resume texts. They stay in sync automatically when you use the "Apply to profile" flow — see [Daily Workflow — Managing Your Resume](daily-workflow.md#managing-your-resume).
 ### Uploading a resume
 If no profile exists yet, you can:
 - **Upload & Parse** — upload a PDF, DOCX, or ODT. Peregrine extracts structured data automatically.
 - **Fill in Manually** — start from a blank form.
 - **Run Setup Wizard** — re-enter the first-run wizard (self-hosted only).
 ### Editing the profile
 When a resume exists, the full form is shown. Sections:
 - **Career Summary** — used in every cover letter and research brief
 - **Personal Information** — name, email, phone, LinkedIn; synced from My Profile
 - **Work Experience** — title, company, period, location, industry, responsibilities, skills
 - **Education** — institution, degree, field, dates
 - **Skills, Domains, Keywords** — tags used for keyword matching; click **Suggest** for AI recommendations
 - **Certifications and Achievements** — optional; included in cover letter context
 Click **Save** to write changes. If a default library entry is linked, it updates automatically.
 ---
-## Email
+## Search Prefs
-Configure IMAP email sync. See [Email Sync](email-sync.md) for full setup instructions.
+Manage what Peregrine searches for across all job boards. Changes take effect on the next discovery run — no restart needed.
 | Field | Description |
 |-------|-------------|
 | Remote preference | Remote only, on-site only, or both |
 | Job Titles | Roles searched on every board |
 | Locations | Geographic scope; leave blank for unrestricted |
 | Exclude Keywords | Drop any job title containing these words before it enters the database |
 | Job Boards | Enable or disable specific sources; boards marked "coming soon" are tracked in the backlog |
 | Custom Board URLs | Additional job board URLs to include |
 | Blocklists | Companies, industries, or locations to always skip |
 Click **Suggest** next to Job Titles, Locations, or Exclude Keywords to get AI-generated suggestions based on your resume.
 ---
-## Skills
+## Connections
-Manage your `config/resume_keywords.yaml` — the list of skills and keywords used for match scoring.
+API credentials and authentication for external services.
-Add or remove keywords. Higher-weighted keywords count more toward the match score.
+| Service | What it enables |
 |---------|----------------|
 | Notion | Sync approved/applied jobs to a Notion database |
 | Airtable | Alternative sync target |
 | Google Drive | Document export |
 | Slack / Discord | Status notifications |
 | Google Calendar / Apple Calendar | Interview scheduling (Paid) |
 See [Integrations](integrations.md) for per-service setup instructions.
 ---
-## Integrations
+## System
-Connection cards for all 13 integrations. See [Integrations](integrations.md) for per-service details.
+*Not available in cloud mode.*
 LLM backend configuration and service connection settings.
 ### LLM Backends
 | Setting | Description |
 |---------|-------------|
 | Enabled toggle | Whether a backend is considered in the fallback chain |
 | Base URL | API endpoint for OpenAI-compatible backends |
 | Model | Model name or `__auto__` (vLLM auto-detects the loaded model) |
 | API key | Required for hosted APIs |
 | Test button | Sends a ping to verify the backend is reachable |
 Three independent fallback chains:
 | Chain | Used for |
 |-------|---------|
 | Cover letter chain | Cover letter generation and general tasks |
 | Research chain | Company research briefs |
 | Vision chain | Survey screenshot analysis |
 ### Service Hosts and Ports
 Connection settings for Ollama, vLLM, and SearXNG. Each service has an SSL toggle and SSL-verify toggle for reverse-proxy setups.
 ---
 ## Fine-Tune
-**Tier: Premium**
+*Tier: Premium only.*
 Tools for fine-tuning a cover letter model on your personal writing style.
- Export cover letter training data as JSONL
+1. **Export Training Data** — produces a JSONL file from your saved cover letters
- Configure training parameters (rank, epochs, learning rate)
+2. **Configure training** — rank, epochs, learning rate
- Start a fine-tuning run (requires `ogma` conda environment with Unsloth)
+3. **Start fine-tune** — runs via the `ogma` conda environment with Unsloth
- Register the output model with Ollama
+4. **Register model** — adds the output to Ollama as `alex-cover-writer:latest`
 ---
 ## License
 View your current license key, tier, and entitlements. Paste a new key here if you are upgrading or replacing a key.
 ---
 ## Data
 *Not available in cloud mode.*
 Export or delete your local data.
 | Action | What it does |
 |--------|-------------|
 | Export | Downloads `staging.db` and config files as a zip |
 | Purge pending jobs | Deletes all jobs with status `pending` |
 | Purge rejected jobs | Deletes all jobs with status `rejected` |
 | Factory reset | Removes all data and config; returns to first-run wizard |
 ---
 ## Privacy
 Controls for data collection and diagnostic logging. All collection is opt-in.
 ---
 ## Developer
-Developer and debugging tools.
+Developer and debugging tools. Only visible when dev mode is enabled or a `dev_tier_override` is set.
 | Option | Description |
 |--------|-------------|
-| Reset wizard | Sets `wizard_complete: false` and `wizard_step: 0`; resumes at step 1 on next page load |
+| Reset wizard | Sets `wizard_complete: false`; wizard restarts on next page load |
-| Dev tier override | Set `dev_tier_override` to `paid` or `premium` to test tier-gated features locally |
+| Dev tier override | Set tier to `paid` or `premium` to test tier-gated features locally |
-| Clear stuck tasks | Manually sets any `running` or `queued` background tasks to `failed` (also runs on app startup) |
+| Clear stuck tasks | Manually fails any `running` or `queued` background tasks |
-| View raw config | Shows the current `config/user.yaml` contents |
+| View raw config | Shows current `user.yaml` contents |
--- a/mkdocs.yml
+++ b/mkdocs.yml
@ -52,6 +52,7 @@ nav:
    - First-Run Wizard: getting-started/first-run-wizard.md
    - Docker Profiles: getting-started/docker-profiles.md
  - User Guide:
    - Daily Workflow: user-guide/daily-workflow.md
    - Job Discovery: user-guide/job-discovery.md
    - Job Review: user-guide/job-review.md
    - Apply Workspace: user-guide/apply-workspace.md
--- a/scripts/resume_parser.py
+++ b/scripts/resume_parser.py
@ -19,6 +19,14 @@ from docx import Document
 log = logging.getLogger(__name__)
 # Browser print artifact patterns — lines injected when a PDF is printed from a browser
 # (print header "MM/DD/YY, H:MM AM/PM <title>" and print footer "file:///... N/N")
 _BROWSER_ARTIFACT_RE = re.compile(
    r"^file:///"                                                      # file:// URL footer
    r"|^\d{1,2}/\d{1,2}/\d{2,4},\s+\d{1,2}:\d{2}\s+[AP]M\b",       # MM/DD/YY, H:MM AM/PM header
    re.I,
 )
 # ── Section header detection ──────────────────────────────────────────────────
 _SECTION_NAMES = {
@ -27,6 +35,8 @@ _SECTION_NAMES = {
    "education":  re.compile(r"^(education|academic|qualifications|degrees?|educational background|academic background)\s*:?\s*$", re.I),
    "skills":     re.compile(r"^(skills?|technical skills?|core competencies|competencies|expertise|areas? of expertise|key skills?|proficiencies|tools? & technologies)\s*:?\s*$", re.I),
    "achievements": re.compile(r"^(achievements?|accomplishments?|awards?|honors?|certifications?|publications?|volunteer)\s*:?\s*$", re.I),
    "projects":     re.compile(r"^(projects?|independent development|independent projects?|side projects?|personal projects?|open.?source|portfolio)\s*:?\s*$", re.I),
    "references": re.compile(r"^references?\s*:?\s*$", re.I),
 }
 # Degrees — used to detect education lines
@ -163,6 +173,8 @@ def _split_sections(text: str) -> dict[str, list[str]]:
        stripped = line.strip()
        if not stripped:
            continue
        if _BROWSER_ARTIFACT_RE.match(stripped):
            continue
        matched = False
        for section, pattern in _SECTION_NAMES.items():
            # Match if the line IS a section header (short + matches pattern)
@ -232,10 +244,14 @@ def _parse_experience(lines: list[str]) -> list[dict]:
      (A) Title | Company          (B) Title | Company | Dates
          Dates                        • bullet
          • bullet
      (C) Title\tDates             (tab-separated, common in DOCX exports)
          Company | Location
          • bullet
    """
    entries: list[dict] = []
    current: dict | None = None
    prev_line = ""
    seen_bullets = False  # True once we've appended the first bullet to current
    for line in lines:
        date_match = _DATE_RANGE_RE.search(line)
@ -243,12 +259,13 @@ def _parse_experience(lines: list[str]) -> list[dict]:
            if current:
                entries.append(current)
            # Title/company extraction — three layouts:
-            #  (A) Title on prev_line, "Company | Location | Dates" on date line
+            #  (A) Title on prev_line (not a bullet), "Company | Location | Dates" on date line
            #  (B) "Title | Company" on prev_line, dates on date line (same_line empty)
            #  (C) "Title | Company | Dates" all on one line
            same_line = _DATE_RANGE_RE.sub("", line)
            # Remove residual punctuation-only fragments like "()" left after date removal
            same_line = re.sub(r"[()[\]{}\s]+$", "", same_line).strip(" –—|-•")
            # Only use prev_line as title if it isn't bullet text (cleared after bullets)
            if prev_line and same_line.strip():
                # Layout A: title = prev_line, company = first segment of same_line
                title   = prev_line.strip()
@ -268,8 +285,19 @@ def _parse_experience(lines: list[str]) -> list[dict]:
                "bullets":    [],
            }
            prev_line = ""
            seen_bullets = False
        elif current is not None:
            is_bullet = bool(re.match(r"^[•\-–—*◦▪▸►]\s*", line))
            # Layout C: company/location on the line immediately after the date line,
            # before any bullets. Short non-date line = company, not a next-job header.
            if (not is_bullet and not seen_bullets and not current["company"]
                    and not _DATE_RE.search(line) and len(line.strip()) < 80):
                co_part = re.split(r"\s{2,}|[|,]\s*", line.strip(), maxsplit=1)[0]
                current["company"] = co_part.strip()
                prev_line = ""
                continue
            looks_like_header = (
                not is_bullet
                and " | " in line
@ -282,7 +310,10 @@ def _parse_experience(lines: list[str]) -> list[dict]:
                clean = re.sub(r"^[•\-–—*◦▪▸►]\s*", "", line).strip()
                if clean:
                    current["bullets"].append(clean)
-                prev_line = line
+                    seen_bullets = True
                # Clear prev_line after non-header content so the next date match
                # doesn't mistake a bullet as a job title (Layout A false-positive).
                prev_line = ""
        else:
            prev_line = line
@ -294,38 +325,76 @@ def _parse_experience(lines: list[str]) -> list[dict]:
 # ── Education ─────────────────────────────────────────────────────────────────
 _INSTITUTION_RE = re.compile(r"\b(university|college|institute|school|academy)\b", re.I)
 def _parse_education(lines: list[str]) -> list[dict]:
    """Parse education entries.
    Primary path: degree keyword detected (B.S., Master, etc.)
    Fallback path: year range detected without a degree keyword — handles resumes
    with courses, programmes, or non-degree study (e.g. "San Jose State University  2005-2006").
    """
    entries: list[dict] = []
    current: dict | None = None
    prev_line = ""
    for line in lines:
-        if _DEGREE_RE.search(line):
+        has_degree = bool(_DEGREE_RE.search(line))
        date_range = _DATE_RANGE_RE.search(line)
        has_year   = bool(re.search(r"\b(19|20)\d{2}\b", line))
        if has_degree or (has_year and date_range):
            if current:
                entries.append(current)
-            current = {
+            current = {"institution": "", "degree": "", "field": "", "graduation_year": ""}
-                "institution":      "",
+
                "degree":           "",
                "field":            "",
                "graduation_year":  "",
            }
            year_m = re.search(r"\b(19|20)\d{2}\b", line)
            if year_m:
                current["graduation_year"] = year_m.group(0)
            if has_degree:
                degree_m = _DEGREE_RE.search(line)
                if degree_m:
                    current["degree"] = degree_m.group(0).upper()
                remainder = _DEGREE_RE.sub("", _DATE_RE.sub("", line))
                remainder = re.sub(r"\b(19|20)\d{2}\b", "", remainder)
                current["field"] = remainder.strip(" ,–—|•.")
-            # Layout A: institution was on the line before the degree line
+                if prev_line and not _DEGREE_RE.search(prev_line) and not _DATE_RE.search(prev_line):
            if prev_line and not _DEGREE_RE.search(prev_line):
                    current["institution"] = prev_line.strip(" ,–—|•")
-        elif current is not None and not current["institution"]:
+            else:
-            # Layout B: institution follows the degree line
+                # Fallback: year-range line without a degree keyword.
-            clean = line.strip(" ,–—|•")
+                # Two layouts:
                #   (A) PDF: "Graphic Design, 2005–2006" with institution on prev_line
                #   (B) DOCX: "San Jose State University\t2005-2006" — institution on same line
                same = _DATE_RANGE_RE.sub("", line)
                same = re.sub(r"\b(19|20)\d{2}\b", "", same).strip(" ,–—|•\t")
                prev_clean = prev_line.strip(" ,–—|•") if prev_line else ""
                if same and _INSTITUTION_RE.search(prev_clean):
                    # Layout A: institution on prev_line (e.g. "San Jose State University")
                    current["institution"] = prev_clean
                    current["field"] = same
                elif same:
                    # Layout B: institution embedded on same line as year
                    current["institution"] = same
                elif prev_clean:
                    current["institution"] = prev_clean
            prev_line = ""  # consumed; prevent leaking into the next entry
        elif current is not None:
            clean = line.strip(" ,–—|•\t")
            if clean:
                if not current["institution"]:
                    current["institution"] = clean
                elif not current["field"]:
                    current["field"] = clean
                    prev_line = ""  # field consumed — don't seed the next entry
                    continue
            prev_line = line.strip()
        else:
            prev_line = line.strip()
    if current:
@ -336,13 +405,39 @@ def _parse_education(lines: list[str]) -> list[dict]:
 # ── Skills ────────────────────────────────────────────────────────────────────
 def _split_skill_tokens(line: str) -> list[str]:
    """Split a skills line on delimiters, but not on commas inside parentheses.
    Splits on |, •, ·, tab first (always separators), then on comma only when
    paren depth is zero — so "CRM Ticketing (Jira, Salesforce)" stays intact.
    """
    tokens: list[str] = []
    for part in re.split(r"[|•·\t]+", line):
        depth, buf = 0, ""
        for ch in part:
            if ch == "(":
                depth += 1
                buf += ch
            elif ch == ")":
                depth -= 1
                buf += ch
            elif ch == "," and depth == 0:
                tokens.append(buf)
                buf = ""
            else:
                buf += ch
        tokens.append(buf)
    return tokens
 def _parse_skills(lines: list[str]) -> list[str]:
    skills: list[str] = []
    for line in lines:
-        # Split on common delimiters
+        for item in _split_skill_tokens(line):
-        for item in re.split(r"[,|•·/]+", line):
+            # Strip only bullet/dash markers and whitespace, NOT parentheses —
-            clean = item.strip(" -–—*◦▪▸►()")
+            # many skills contain parens, e.g. "C++ (Arduino / Embedded)"
-            if 1 < len(clean) <= 50:
+            clean = item.strip(" -–—*◦▪▸►")
            if 1 < len(clean) <= 60:
                skills.append(clean)
    return skills
--- a/web/src/composables/useDocsUrl.ts
+++ b/web/src/composables/useDocsUrl.ts
@ -0,0 +1,5 @@
 const DOCS_BASE = 'https://docs.circuitforge.tech/peregrine'
 export function useDocsUrl(path: string): string {
  return `${DOCS_BASE}/${path}`
 }
--- a/web/src/stores/wizard.ts
+++ b/web/src/stores/wizard.ts
@ -2,7 +2,7 @@ import { ref, computed } from 'vue'
 import { defineStore } from 'pinia'
 import { useApiFetch } from '../composables/useApi'
-export type WizardProfile = 'remote' | 'cpu' | 'single-gpu' | 'dual-gpu'
+export type WizardProfile = 'remote' | 'cpu' | 'single-gpu' | 'dual-gpu' | 'cf-orch'
 export type WizardTier = 'free' | 'paid' | 'premium'
 export interface WorkExperience {
@ -36,6 +36,7 @@ export interface WizardInferenceData {
  anthropicKey: string
  openaiUrl: string
  openaiKey: string
  orchUrl: string
  ollamaHost: string
  ollamaPort: number
  services: Record<string, string | number>
@ -90,7 +91,8 @@ export const useWizardStore = defineStore('wizard', () => {
    anthropicKey: '',
    openaiUrl: '',
    openaiKey: '',
-    ollamaHost: 'localhost',
+    orchUrl: '',
    ollamaHost: '',
    ollamaPort: 11434,
    services: {},
    confirmed: false,
@ -127,6 +129,7 @@ export const useWizardStore = defineStore('wizard', () => {
        wizard_step: number
        saved_data: {
          inference_profile?: string
          cf_orch_url?: string
          tier?: string
          name?: string
          email?: string
@ -143,6 +146,8 @@ export const useWizardStore = defineStore('wizard', () => {
      if (saved.inference_profile)
        hardware.value.selectedProfile = saved.inference_profile as WizardProfile
      if (saved.cf_orch_url)
        inference.value.orchUrl = saved.cf_orch_url as string
      if (saved.tier)
        tier.value = saved.tier as WizardTier
      if (saved.name) identity.value.name = saved.name
@ -222,6 +227,7 @@ export const useWizardStore = defineStore('wizard', () => {
      anthropic_key: inference.value.anthropicKey,
      openai_url: inference.value.openaiUrl,
      openai_key: inference.value.openaiKey,
      orch_url: inference.value.orchUrl,
      ollama_host: inference.value.ollamaHost,
      ollama_port: inference.value.ollamaPort,
    }
--- a/web/src/stores/wizard/aiInterview.ts
+++ b/web/src/stores/wizard/aiInterview.ts
@ -55,7 +55,20 @@ export const useAiInterviewStore = defineStore('aiInterview', () => {
    })
    loading.value = false
    if (err || !data) {
      if (err?.kind === 'http' && err.status === 402) {
        error.value = 'AI profile assistant requires a Paid plan or a BYOK API key.'
      } else if (err?.kind === 'http' && err.status === 503) {
        try {
          const body = JSON.parse(err.detail) as { detail?: { error?: string } }
          error.value = body.detail?.error === 'llm_error'
            ? 'No LLM backend configured — add an API key in Settings → System first.'
            : 'Could not reach the assistant. Please try again.'
        } catch {
          error.value = 'Could not reach the assistant. Please try again.'
        }
      } else {
        error.value = 'Could not reach the assistant. Please try again.'
      }
      return
    }
    messages.value = [...messages.value, { role: 'assistant', content: data.reply }]
--- a/web/src/views/HomeView.vue
+++ b/web/src/views/HomeView.vue
@ -12,7 +12,10 @@
          {{ greeting }}
          <span v-if="isMidnight" aria-label="Late night session">🌙</span>
        </h1>
-        <p class="home__subtitle">Discover → Review → Apply</p>
+        <p class="home__subtitle">
          Discover → Review → Apply
          <a href="https://docs.circuitforge.tech/peregrine/user-guide/daily-workflow/" target="_blank" rel="noopener" class="home__docs-link" aria-label="Daily Workflow documentation">Daily Workflow guide ↗</a>
        </p>
      </div>
    </header>
@ -600,7 +603,22 @@ onUnmounted(() => {
  font-size: var(--text-sm);
  text-transform: uppercase;
  letter-spacing: 0.05em;
  display: flex;
  align-items: center;
  gap: var(--space-3);
  flex-wrap: wrap;
 }
 .home__docs-link {
  font-size: 0.7rem;
  text-transform: none;
  letter-spacing: 0;
  color: var(--color-text-muted);
  text-decoration: none;
  border: 1px solid var(--color-border);
  border-radius: var(--radius-full);
  padding: 1px 7px;
 }
 .home__docs-link:hover { color: var(--color-primary); border-color: var(--color-primary); }
 .home__metrics {
  display: grid;
--- a/web/src/views/JobReviewView.vue
+++ b/web/src/views/JobReviewView.vue
@ -9,6 +9,7 @@
    <header class="review__header">
      <div class="review__title-row">
        <h1 class="review__title">Review Jobs</h1>
        <a href="https://docs.circuitforge.tech/peregrine/user-guide/job-review/" target="_blank" rel="noopener" class="review__docs-link" aria-label="Job Review documentation">? Docs</a>
        <button class="help-btn" :aria-expanded="showHelp" @click="showHelp = !showHelp">
          <span aria-hidden="true">?</span>
          <span class="sr-only">Keyboard shortcuts</span>
@ -429,6 +430,17 @@ onUnmounted(() => {
  flex: 1;
 }
 .review__docs-link {
  font-size: 0.75rem;
  color: var(--color-text-muted);
  border: 1px solid var(--color-border);
  border-radius: var(--radius-full);
  padding: 2px 8px;
  text-decoration: none;
  white-space: nowrap;
  flex-shrink: 0;
 }
 .review__docs-link:hover { color: var(--color-primary); border-color: var(--color-primary); }
 .help-btn {
  width: 32px;
  height: 32px;
--- a/web/src/views/ResumesView.vue
+++ b/web/src/views/ResumesView.vue
@ -2,6 +2,7 @@
  <div class="rv">
    <div class="rv__header">
      <h1 class="rv__title">Resume Library</h1>
      <a href="https://docs.circuitforge.tech/peregrine/user-guide/daily-workflow/#managing-your-resume" target="_blank" rel="noopener" class="rv__help-link" aria-label="Resume Library documentation">? Help</a>
      <label class="btn-generate rv__import-btn">
        <span aria-hidden="true">📥</span> Import
        <input type="file" accept=".txt,.pdf,.docx,.odt,.yaml,.yml"
@ -314,7 +315,10 @@ onBeforeRouteLeave(() => {
 <style scoped>
 .rv { display: flex; flex-direction: column; gap: var(--space-4, 1rem); padding: var(--space-5, 1.25rem); height: 100%; }
-.rv__header { display: flex; align-items: center; justify-content: space-between; }
+.rv__header { display: flex; align-items: center; gap: var(--space-3); }
 .rv__header .btn-generate { margin-left: auto; }
 .rv__help-link { font-size: 0.75rem; color: var(--color-text-muted); border: 1px solid var(--color-border); border-radius: var(--radius-full); padding: 2px 8px; text-decoration: none; white-space: nowrap; flex-shrink: 0; }
 .rv__help-link:hover { color: var(--color-primary); border-color: var(--color-primary); }
 .rv__title { font-size: var(--font-xl, 1.25rem); font-weight: 700; margin: 0; }
 .rv__file-input { display: none; }
 .rv__import-btn { cursor: pointer; }
--- a/web/src/views/settings/ResumeProfileView.vue
+++ b/web/src/views/settings/ResumeProfileView.vue
@ -1,6 +1,9 @@
 <template>
  <div class="resume-profile">
    <div class="page-header">
      <h2>Resume Profile</h2>
      <a href="https://docs.circuitforge.tech/peregrine/user-guide/settings/#resume-profile" target="_blank" rel="noopener" class="help-link" aria-label="Resume Profile documentation">? Help</a>
    </div>
    <!-- Load error banner -->
    <div v-if="loadError" class="error-banner">
@ -401,6 +404,10 @@ async function handleUpload() {
 <style scoped>
 .resume-profile { max-width: 720px; margin: 0 auto; padding: var(--space-4); }
 .page-header { display: flex; align-items: center; gap: var(--space-3); margin-bottom: var(--space-6); }
 .page-header h2 { margin-bottom: 0; }
 .help-link { font-size: 0.75rem; color: var(--color-text-muted); border: 1px solid var(--color-border); border-radius: var(--radius-full); padding: 2px 8px; text-decoration: none; white-space: nowrap; flex-shrink: 0; }
 .help-link:hover { color: var(--color-primary); border-color: var(--color-primary); }
 h2 { font-size: 1.4rem; font-weight: 600; margin-bottom: var(--space-6); }
 h3 { font-size: 1rem; font-weight: 600; margin-bottom: var(--space-3); }
 .form-section { margin-bottom: var(--space-8); padding-bottom: var(--space-6); border-bottom: 1px solid var(--color-border); }
--- a/web/src/views/settings/SearchPrefsView.vue
+++ b/web/src/views/settings/SearchPrefsView.vue
@ -1,6 +1,9 @@
 <template>
  <div class="search-prefs">
    <div class="page-header">
      <h2>Search Preferences</h2>
      <a :href="docsUrl" target="_blank" rel="noopener" class="help-link" aria-label="Search Preferences documentation">? Help</a>
    </div>
    <p v-if="store.loadError" class="error-banner">{{ store.loadError }}</p>
    <!-- Remote Preference -->
@ -154,8 +157,10 @@
 <script setup lang="ts">
 import { ref, onMounted } from 'vue'
 import { useSearchStore } from '../../stores/settings/search'
 import { useDocsUrl } from '../../composables/useDocsUrl'
 const store = useSearchStore()
 const docsUrl = useDocsUrl('user-guide/settings/#search-prefs')
 const remoteOptions = [
  { value: 'remote' as const, label: 'Remote only' },
@ -186,6 +191,10 @@ onMounted(() => store.load())
 <style scoped>
 .search-prefs { max-width: 720px; margin: 0 auto; padding: var(--space-4); }
 .page-header { display: flex; align-items: center; gap: var(--space-3); margin-bottom: var(--space-6); }
 .page-header h2 { margin-bottom: 0; }
 .help-link { font-size: 0.75rem; color: var(--color-text-muted); border: 1px solid var(--color-border); border-radius: var(--radius-full); padding: 2px 8px; text-decoration: none; white-space: nowrap; flex-shrink: 0; }
 .help-link:hover { color: var(--color-primary); border-color: var(--color-primary); }
 h2 { font-size: 1.4rem; font-weight: 600; margin-bottom: var(--space-6); }
 h3 { font-size: 1rem; font-weight: 600; margin-bottom: var(--space-3); }
 .form-section { margin-bottom: var(--space-8); padding-bottom: var(--space-6); border-bottom: 1px solid var(--color-border); }
--- a/web/src/views/settings/SystemSettingsView.vue
+++ b/web/src/views/settings/SystemSettingsView.vue
@ -136,6 +136,29 @@
      <p v-if="store.deployError" class="error-msg">{{ store.deployError }}</p>
    </section>
    <!-- Orchard coordinator -->
    <section class="form-section">
      <h3>Orchard Coordinator</h3>
      <p class="section-note">
        The Orchard is CircuitForge's distributed GPU cluster. Requires a Paid license or higher.
        Leave blank to disable Orchard routing.
      </p>
      <div class="field-row">
        <label>Coordinator URL</label>
        <input
          v-model="orchUrl"
          type="url"
          placeholder="https://orch.circuitforge.tech"
          class="field-input-wide"
        />
        <button @click="saveOrchUrl" :disabled="orchSaving" class="btn-save-inline">
          {{ orchSaving ? 'Saving…' : 'Save' }}
        </button>
      </div>
      <p v-if="orchError" class="error">{{ orchError }}</p>
      <p v-if="orchSaved" class="success">Saved.</p>
    </section>
    <!-- BYOK Modal -->
    <Teleport to="body">
      <div v-if="store.byokPending.length > 0" class="modal-overlay" @click.self="store.cancelByok()">
@ -250,12 +273,39 @@ async function saveCoverLetterModel() {
  setTimeout(() => { clmSaved.value = false }, 3000)
 }
 // ── Orchard coordinator URL ───────────────────────────────────────────────────
 const orchUrl    = ref('')
 const orchSaving = ref(false)
 const orchError  = ref<string | null>(null)
 const orchSaved  = ref(false)
 async function loadOrchUrl() {
  const { data } = await useApiFetch<{ orch_url: string }>('/api/settings/system/orch-url')
  if (data) orchUrl.value = data.orch_url ?? ''
 }
 async function saveOrchUrl() {
  orchSaving.value = true
  orchError.value  = null
  orchSaved.value  = false
  const { error } = await useApiFetch('/api/settings/system/orch-url', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({ orch_url: orchUrl.value }),
  })
  orchSaving.value = false
  if (error) { orchError.value = 'Failed to save.'; return }
  orchSaved.value = true
  setTimeout(() => { orchSaved.value = false }, 3000)
 }
 onMounted(async () => {
  await store.loadLlm()
  const tasks = [
    store.loadServices(),
    store.loadFilePaths(),
    store.loadDeployConfig(),
    loadOrchUrl(),
  ]
  if (config.isCloud && tierOrder.indexOf(tier.value) >= tierOrder.indexOf('paid')) {
    tasks.push(loadCoverLetterModel())
@ -328,6 +378,7 @@ h3 { font-size: 1rem; font-weight: 600; margin-bottom: var(--space-3); }
 .field-row { display: flex; flex-direction: column; gap: 4px; margin-bottom: 14px; }
 .field-row label { font-size: 0.82rem; color: var(--color-text-muted); }
 .field-row input { background: var(--color-surface-alt); border: 1px solid var(--color-border); border-radius: 6px; color: var(--color-text); padding: 7px 10px; font-size: 0.88rem; }
 .field-input-wide { width: 100%; max-width: 400px; }
 .field-hint { font-size: 0.72rem; color: var(--color-text-muted); margin-top: 3px; }
 .btn-secondary { padding: 9px 18px; background: transparent; border: 1px solid var(--color-border); border-radius: 7px; color: var(--color-text-muted); cursor: pointer; font-size: 0.88rem; }
 .btn-danger {
--- a/web/src/views/wizard/WizardHardwareStep.vue
+++ b/web/src/views/wizard/WizardHardwareStep.vue
@ -13,9 +13,28 @@
        {{ wizard.hardware.gpus.join(', ') }}
      </div>
      <div v-else class="step__info">
-        No local NVIDIA GPUs detected. CPU or cf-orch mode recommended.
+        No local NVIDIA GPUs detected. CPU or Orchard mode recommended.
      </div>
      <!-- Service status -->
      <div class="hw-services">
        <div class="hw-svc" :class="ollamaRunning ? 'hw-svc--up' : 'hw-svc--down'">
          <span class="hw-svc__dot" aria-hidden="true" />
          <span class="hw-svc__name">Ollama</span>
          <span class="hw-svc__status">{{ ollamaRunning ? 'running' : 'not detected' }}</span>
        </div>
        <div class="hw-svc" :class="searxngRunning ? 'hw-svc--up' : 'hw-svc--down'">
          <span class="hw-svc__dot" aria-hidden="true" />
          <span class="hw-svc__name">SearXNG</span>
          <span class="hw-svc__status">{{ searxngRunning ? 'running' : 'not detected' }}</span>
        </div>
      </div>
      <p v-if="!ollamaRunning" class="step__field-hint">
        Ollama not running — start it on the host before continuing, or choose Remote or Orchard mode.
        See <strong>Settings → Services</strong> after setup to manage services.
      </p>
      <div class="step__field">
        <label class="step__label" for="hw-profile">Inference profile</label>
        <select id="hw-profile" v-model="selectedProfile" class="step__select">
@ -23,7 +42,7 @@
          <option value="single-gpu">Single GPU — local Ollama + one GPU</option>
          <option value="dual-gpu">Dual GPU — local Ollama + two GPUs</option>
          <option value="cf-orch">
-            cf-orch — CircuitForge GPU cluster
+            Orchard — CircuitForge GPU cluster
            {{ orchAvailable ? `(${orchGpus.length} GPU(s) available)` : '(configure endpoint below)' }}
          </option>
          <option value="remote">Remote — use cloud API keys</option>
@ -49,7 +68,7 @@
        </div>
        <div class="step__field">
-          <label class="step__label" for="orch-url">cf-orch coordinator URL</label>
+          <label class="step__label" for="orch-url">Orchard coordinator URL</label>
          <input
            id="orch-url"
            v-model="orchUrl"
@ -58,14 +77,14 @@
            placeholder="http://10.1.10.71:7700"
          />
          <p class="step__field-hint">
-            The coordinator serves public inference endpoints for paid+ users.
+            The Orchard coordinator serves public inference endpoints for Paid+ users.
            Leave blank to use the default cluster URL from Settings.
          </p>
        </div>
        <div class="step__tier-note">
          <span aria-hidden="true">🔒</span>
-          cf-orch inference requires a <strong>Paid</strong> license or higher.
+          Orchard inference requires a <strong>Paid</strong> license or higher.
          You can select this profile now; it will activate once your license is verified.
        </div>
      </template>
@ -75,7 +94,7 @@
        class="step__warning"
      >
        ⚠️ No local GPUs detected — a GPU profile may not work. Choose CPU
-        or cf-orch if you have access to the cluster.
+        or Orchard if you have access to the cluster.
      </div>
    </template>
@ -107,6 +126,10 @@ const orchAvailable = ref(false)
 const orchGpus = ref<Array<{ node: string; name: string; vram_total_mb: number; vram_free_mb: number }>>([])
 const orchUrl = ref('')
 // local service probe results
 const ollamaRunning = ref(false)
 const searxngRunning = ref(false)
 onMounted(async () => {
  detecting.value = true
  const { data } = await useApiFetch<{
@ -115,6 +138,8 @@ onMounted(async () => {
    profiles: string[]
    cf_orch_available: boolean
    cf_orch_gpus: Array<{ node: string; name: string; vram_total_mb: number; vram_free_mb: number }>
    ollama_running: boolean
    searxng_running: boolean
  }>('/api/wizard/hardware')
  detecting.value = false
  if (!data) return
@ -128,6 +153,8 @@ onMounted(async () => {
  orchAvailable.value = data.cf_orch_available ?? false
  orchGpus.value = data.cf_orch_gpus ?? []
  ollamaRunning.value = data.ollama_running ?? false
  searxngRunning.value = data.searxng_running ?? false
 })
 async function next() {
@ -140,3 +167,40 @@ async function next() {
  if (ok) router.push('/setup/tier')
 }
 </script>
 <style scoped>
 .hw-services {
  display: flex;
  gap: var(--space-4);
  margin: var(--space-3) 0 var(--space-2);
  flex-wrap: wrap;
 }
 .hw-svc {
  display: flex;
  align-items: center;
  gap: var(--space-2);
  padding: var(--space-2) var(--space-3);
  border-radius: var(--radius-full);
  font-size: 0.8rem;
  font-weight: 500;
  border: 1px solid var(--color-border-light);
  background: var(--color-surface-alt);
 }
 .hw-svc--up  { border-color: color-mix(in srgb, var(--color-success) 40%, transparent); }
 .hw-svc--down { opacity: 0.65; }
 .hw-svc__dot {
  width: 8px;
  height: 8px;
  border-radius: var(--radius-full);
  flex-shrink: 0;
 }
 .hw-svc--up   .hw-svc__dot { background: var(--color-success); }
 .hw-svc--down .hw-svc__dot { background: var(--color-text-muted); }
 .hw-svc__name  { color: var(--color-text); }
 .hw-svc__status { color: var(--color-text-muted); }
 </style>
--- a/web/src/views/wizard/WizardIdentityStep.vue
+++ b/web/src/views/wizard/WizardIdentityStep.vue
@ -1,6 +1,6 @@
 <template>
  <div class="step">
-    <h2 class="step__heading">Step 4 — Your Identity</h2>
+    <h2 class="step__heading">Step 5 — Your Identity</h2>
    <p class="step__caption">
      Used in cover letters, research briefs, and interview prep. You can update
      this any time in Settings → My Profile.
--- a/web/src/views/wizard/WizardInferenceStep.vue
+++ b/web/src/views/wizard/WizardInferenceStep.vue
@ -1,6 +1,6 @@
 <template>
  <div class="step">
-    <h2 class="step__heading">Step 5 — Inference & API Keys</h2>
+    <h2 class="step__heading">Step 6 — Inference & API Keys</h2>
    <p class="step__caption">
      Configure how Peregrine generates AI content. You can adjust this any time
      in Settings → System.
@ -36,7 +36,35 @@
      </div>
    </template>
-    <!-- Local mode -->
+    <!-- Orchard mode -->
    <template v-else-if="isCfOrch">
      <div class="step__info">
        Orchard mode: Peregrine routes AI generation through the CircuitForge GPU cluster.
      </div>
      <div class="step__field">
        <label class="step__label" for="inf-orch-url">Orchard coordinator URL</label>
        <input id="inf-orch-url" v-model="form.orchUrl" type="url"
               class="step__input" placeholder="https://orch.circuitforge.tech" />
      </div>
      <div v-if="isPaid" class="step__check-row">
        <label class="step__checkbox-label">
          <input
            type="checkbox"
            class="step__checkbox"
            :checked="form.orchUrl === MANAGED_ORCH_URL"
            @change="onUseManagedOrchard"
          />
          <span>Use CircuitForge managed Orchard</span>
        </label>
        <span class="step__check-hint">
          Auto-fills your Paid+ cluster endpoint ({{ MANAGED_ORCH_URL }})
        </span>
      </div>
    </template>
    <!-- Local mode (CPU / single-gpu / dual-gpu) -->
    <template v-else>
      <div class="step__info">
        Local mode ({{ wizard.hardware.selectedProfile }}): Peregrine uses
@ -81,12 +109,19 @@
 import { reactive, ref, computed } from 'vue'
 import { useRouter } from 'vue-router'
 import { useWizardStore } from '../../stores/wizard'
 import { useAppConfigStore } from '../../stores/appConfig'
 import './wizard.css'
 const wizard = useWizardStore()
 const config = useAppConfigStore()
 const router = useRouter()
 const MANAGED_ORCH_URL = 'https://orch.circuitforge.tech'
 const isRemote = computed(() => wizard.hardware.selectedProfile === 'remote')
 const isCfOrch = computed(() => wizard.hardware.selectedProfile === 'cf-orch')
 const isPaid = computed(() => config.tier !== 'free')
 const showAdvanced = ref(false)
 const testing = ref(false)
 const testResult = ref<{ ok: boolean; message: string } | null>(null)
@ -95,19 +130,42 @@ const form = reactive({
  anthropicKey: wizard.inference.anthropicKey,
  openaiUrl: wizard.inference.openaiUrl,
  openaiKey: wizard.inference.openaiKey,
  orchUrl: wizard.inference.orchUrl,
 })
 const savedSvcs = wizard.inference.services as Record<string, string | number>
 const services = reactive([
-  { key: 'ollama', label: 'Ollama', host: 'ollama', port: 11434 },
+  {
-  { key: 'searxng', label: 'SearXNG', host: 'searxng', port: 8080 },
+    key: 'ollama',
    label: 'Ollama',
    host: (savedSvcs['ollama_host'] as string) || wizard.inference.ollamaHost || 'localhost',
    port: (savedSvcs['ollama_port'] as number) || wizard.inference.ollamaPort || 11434,
  },
  {
    key: 'searxng',
    label: 'SearXNG',
    host: (savedSvcs['searxng_host'] as string) || 'searxng',
    port: (savedSvcs['searxng_port'] as number) || 8080,
  },
 ])
 function onUseManagedOrchard(e: Event) {
  const checked = (e.target as HTMLInputElement).checked
  form.orchUrl = checked ? MANAGED_ORCH_URL : ''
 }
 async function runTest() {
  testing.value = true
  testResult.value = null
  wizard.inference.anthropicKey = form.anthropicKey
  wizard.inference.openaiUrl = form.openaiUrl
  wizard.inference.openaiKey = form.openaiKey
  wizard.inference.orchUrl = form.orchUrl
  const ollamaSvc = services.find(s => s.key === 'ollama')
  if (ollamaSvc) {
    wizard.inference.ollamaHost = ollamaSvc.host
    wizard.inference.ollamaPort = ollamaSvc.port
  }
  testResult.value = await wizard.testInference()
  testing.value = false
 }
@ -115,10 +173,10 @@ async function runTest() {
 function back() { router.push('/setup/identity') }
 async function next() {
  // Sync form back to store
  wizard.inference.anthropicKey = form.anthropicKey
  wizard.inference.openaiUrl = form.openaiUrl
  wizard.inference.openaiKey = form.openaiKey
  wizard.inference.orchUrl = form.orchUrl
  const svcMap: Record<string, string | number> = {}
  services.forEach(s => {
@ -131,6 +189,7 @@ async function next() {
    anthropic_key: form.anthropicKey,
    openai_url: form.openaiUrl,
    openai_key: form.openaiKey,
    orch_url: form.orchUrl,
    services: svcMap,
  })
  if (ok) router.push('/setup/search')
@ -166,4 +225,33 @@ async function next() {
 .svc-port {
  text-align: right;
 }
 .step__check-row {
  display: flex;
  flex-direction: column;
  gap: var(--space-1);
  margin-bottom: var(--space-4);
 }
 .step__checkbox-label {
  display: flex;
  align-items: center;
  gap: var(--space-2);
  cursor: pointer;
  font-size: 0.9rem;
  color: var(--color-text);
 }
 .step__checkbox {
  width: 1rem;
  height: 1rem;
  accent-color: var(--color-primary);
  flex-shrink: 0;
 }
 .step__check-hint {
  font-size: 0.8rem;
  color: var(--color-text-muted);
  padding-left: calc(1rem + var(--space-2));
 }
 </style>
--- a/web/src/views/wizard/WizardIntegrationsStep.vue
+++ b/web/src/views/wizard/WizardIntegrationsStep.vue
@ -1,6 +1,6 @@
 <template>
  <div class="step">
-    <h2 class="step__heading">Step 7 — Integrations</h2>
+    <h2 class="step__heading">Step 8 — Integrations</h2>
    <p class="step__caption">
      Optional. Connect external tools to supercharge your workflow.
      You can configure these any time in Settings → System.
@ -54,6 +54,7 @@ const wizard = useWizardStore()
 const config = useAppConfigStore()
 const router = useRouter()
 const isPaid = computed(() =>
  wizard.tier === 'paid' || wizard.tier === 'premium',
 )
@ -87,7 +88,12 @@ async function finish() {
  // Save integration selections (step 7) then mark wizard complete
  await wizard.saveStep(8, { integrations: [...checkedIds.value] })
  const ok = await wizard.complete()
-  if (ok) router.replace('/')
+  if (ok) {
    // Update store before navigating so the router guard sees wizard as complete
    // without waiting for a full config.load() round-trip.
    config.wizardComplete = true
    router.replace('/')
  }
 }
 </script>
--- a/web/src/views/wizard/WizardResumeStep.vue
+++ b/web/src/views/wizard/WizardResumeStep.vue
@ -2,7 +2,7 @@
  <div class="step">
    <h2 class="step__heading">Step 3 — Your Resume</h2>
    <p class="step__caption">
-      Upload a resume to auto-populate your profile, or build it manually.
+      Upload a resume to auto-populate your profile, build it manually, or let an AI guide you.
    </p>
    <!-- Tabs -->
@ -13,14 +13,31 @@
        class="resume-tab"
        :class="{ 'resume-tab--active': tab === 'upload' }"
        @click="tab = 'upload'"
-      >Upload File</button>
+      >
        <span class="resume-tab__icon" aria-hidden="true">📄</span>
        <span class="resume-tab__label">Upload File</span>
      </button>
      <button
        role="tab"
        :aria-selected="tab === 'manual'"
        class="resume-tab"
        :class="{ 'resume-tab--active': tab === 'manual' }"
        @click="tab = 'manual'"
-      >Build Manually</button>
+      >
        <span class="resume-tab__icon" aria-hidden="true">✏️</span>
        <span class="resume-tab__label">Build Manually</span>
      </button>
      <button
        role="tab"
        :aria-selected="tab === 'ai'"
        class="resume-tab resume-tab--ai"
        :class="{ 'resume-tab--active': tab === 'ai' }"
        @click="tab = 'ai'"
      >
        <span class="resume-tab__icon" aria-hidden="true">{{ hasAiAccess ? '✨' : '🔒' }}</span>
        <span class="resume-tab__label">AI Assistant</span>
        <span v-if="!hasAiAccess" class="resume-tab__badge">Paid</span>
      </button>
    </div>
    <!-- Upload tab -->
@ -106,6 +123,34 @@
      </button>
    </div>
    <!-- AI assistant tab -->
    <div v-if="tab === 'ai'" class="resume-ai">
      <div v-if="!hasAiAccess" class="ai-gate">
        <p class="ai-gate__icon" aria-hidden="true">🔒</p>
        <p class="ai-gate__heading">AI Assistant requires a Paid plan</p>
        <p class="ai-gate__body">
          Upgrade to Paid, or bring your own LLM key in
          <strong>Settings → LLM Backends</strong> to unlock the AI profile assistant for free.
        </p>
        <p class="ai-gate__body">
          In the meantime, use <button class="ai-gate__link" @click="tab = 'upload'">Upload File</button>
          or <button class="ai-gate__link" @click="tab = 'manual'">Build Manually</button>.
        </p>
      </div>
      <div v-else class="ai-embed">
        <p class="ai-embed__intro">
          The AI assistant will ask you a few questions to build your profile.
          Your answers are saved locally — nothing is sent anywhere without your approval.
        </p>
        <a href="/wizard/ai-profile" class="btn-primary ai-embed__cta">
          Open AI Assistant →
        </a>
        <p class="ai-embed__note">
          Opens in a focused view. Come back here to continue the wizard once you're done.
        </p>
      </div>
    </div>
    <div v-if="validationError" class="step__warning" style="margin-top: var(--space-4)">
      {{ validationError }}
    </div>
@ -120,17 +165,21 @@
 </template>
 <script setup lang="ts">
-import { ref } from 'vue'
+import { ref, computed } from 'vue'
 import { useRouter } from 'vue-router'
 import { useWizardStore } from '../../stores/wizard'
 import type { WorkExperience } from '../../stores/wizard'
 import { useApiFetch } from '../../composables/useApi'
 import { useAppConfigStore } from '../../stores/appConfig'
 import './wizard.css'
 const wizard = useWizardStore()
 const router = useRouter()
 const config = useAppConfigStore()
-const tab = ref<'upload' | 'manual'>(
+const hasAiAccess = computed(() => config.tier !== 'free' || config.byokUnlocked)
 const tab = ref<'upload' | 'manual' | 'ai'>(
  wizard.resume.experience.length > 0 ? 'manual' : 'upload',
 )
 const dragging = ref(false)
@ -223,30 +272,69 @@ async function next() {
 <style scoped>
 .resume-tabs {
  display: flex;
-  gap: 0;
+  gap: var(--space-2);
-  border-bottom: 2px solid var(--color-border-light);
+  border-bottom: 2px solid var(--color-border);
  margin-bottom: var(--space-6);
 }
 .resume-tab {
  display: flex;
  align-items: center;
  gap: var(--space-2);
  padding: var(--space-2) var(--space-5);
-  background: none;
+  background: var(--color-surface-alt);
-  border: none;
+  border: 1.5px solid var(--color-border);
  border-bottom: 2px solid transparent;
  border-radius: var(--radius-md) var(--radius-md) 0 0;
  margin-bottom: -2px;
  cursor: pointer;
  font-family: var(--font-body);
-  font-size: 0.9rem;
+  font-size: 0.875rem;
  color: var(--color-text-muted);
-  transition: color var(--transition), border-color var(--transition);
+  transition: color var(--transition), background var(--transition), border-color var(--transition);
 }
 .resume-tab:hover:not(.resume-tab--active) {
  background: var(--color-surface-raised);
  color: var(--color-text);
  border-color: var(--color-border);
 }
 .resume-tab--active {
  background: var(--color-surface);
  color: var(--color-primary);
-  border-bottom-color: var(--color-primary);
+  border-color: var(--color-border);
  border-bottom-color: var(--color-surface);
  font-weight: 600;
 }
 .resume-tab--ai.resume-tab--active {
  color: var(--color-accent);
  border-bottom-color: var(--color-surface);
 }
 .resume-tab__icon {
  font-size: 1rem;
  line-height: 1;
 }
 .resume-tab__label {
  /* explicit — keeps tab text from being an accessibility mystery */
 }
 .resume-tab__badge {
  font-size: 0.7rem;
  font-weight: 700;
  padding: 1px var(--space-2);
  background: color-mix(in srgb, var(--color-accent) 15%, transparent);
  color: var(--color-accent);
  border: 1px solid color-mix(in srgb, var(--color-accent) 30%, transparent);
  border-radius: var(--radius-full);
  margin-left: var(--space-1);
  text-transform: uppercase;
  letter-spacing: 0.04em;
 }
 .upload-zone {
  display: flex;
  flex-direction: column;
@ -310,4 +398,86 @@ async function next() {
  grid-template-columns: 1fr 1fr;
  gap: var(--space-4);
 }
 /* ── AI tab panels ──────────────────────────────────── */
 .resume-ai {
  min-height: 200px;
  display: flex;
  flex-direction: column;
  justify-content: center;
 }
 .ai-gate {
  text-align: center;
  padding: var(--space-8) var(--space-4);
  display: flex;
  flex-direction: column;
  align-items: center;
  gap: var(--space-3);
  background: var(--color-surface-alt);
  border: 1px dashed var(--color-border);
  border-radius: var(--radius-lg);
 }
 .ai-gate__icon {
  font-size: 2rem;
  margin: 0;
 }
 .ai-gate__heading {
  font-size: 1rem;
  font-weight: 600;
  color: var(--color-text);
  margin: 0;
 }
 .ai-gate__body {
  font-size: 0.875rem;
  color: var(--color-text-muted);
  line-height: 1.55;
  margin: 0;
  max-width: 380px;
 }
 .ai-gate__link {
  background: none;
  border: none;
  padding: 0;
  font-family: var(--font-body);
  font-size: inherit;
  color: var(--color-primary);
  cursor: pointer;
  text-decoration: underline;
  text-underline-offset: 2px;
 }
 .ai-embed {
  display: flex;
  flex-direction: column;
  align-items: flex-start;
  gap: var(--space-4);
  padding: var(--space-6);
  background: color-mix(in srgb, var(--color-accent) 6%, var(--color-surface));
  border: 1px solid color-mix(in srgb, var(--color-accent) 25%, transparent);
  border-radius: var(--radius-lg);
 }
 .ai-embed__intro {
  font-size: 0.9rem;
  color: var(--color-text);
  line-height: 1.6;
  margin: 0;
 }
 .ai-embed__cta {
  text-decoration: none;
  display: inline-flex;
  align-items: center;
 }
 .ai-embed__note {
  font-size: 0.8rem;
  color: var(--color-text-muted);
  margin: 0;
 }
 </style>
--- a/web/src/views/wizard/WizardSearchStep.vue
+++ b/web/src/views/wizard/WizardSearchStep.vue
@ -1,6 +1,6 @@
 <template>
  <div class="step">
-    <h2 class="step__heading">Step 6 — Search Preferences</h2>
+    <h2 class="step__heading">Step 7 — Search Preferences</h2>
    <p class="step__caption">
      Tell Peregrine what roles and markets to watch. You can add more profiles
      in Settings → Search later.