Enrich memory extractor with tool calls/results in transcript ...

- _EXTRACT_SYSTEM now explains 4 transcript entry types and instructs
  LLM to trust tool results over chat, return source/source_context
- _extract_facts builds tool_call_map, appends [Tool call] and
  [Tool result] lines with truncation (500/200 chars)
- Transcript capped at 12k chars (head+tail, drop middle)
- Parse source/source_context from LLM response; map confidence:
  tool_call/auto_discovery=95, user_explicit=90, default=70
- Add TODO comment about deferred semantic deduplication

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Document pgvector migration in memory system ...

- Add PostgreSQL vs SQLite comparison table

- Document migrate_pgvector.py usage

- Update storage section to mention asyncpg + aiosqlite

Add pgvector migration script for memory_facts ...

- ALTER TABLE memory_facts: embedding, source, confidence, expires_at, source_context

- CREATE INDEX: hnsw(embedding), expires, source+category

- Safe to run multiple times (IF NOT EXISTS)

- Reads DATABASE_URL from settings

Wire pgvector semantic search into memory system ...

- Add vector(768) column + HNSW index to memory_facts

- Add LLMBackend.embed() with Ollama + fallback implementation

- MemoryStore: cosine-distance search with ILIKE fallback

- New memory tool params: source, confidence, expires_days, source_context

- Update extractor, sqlite_store, deps wiring

- Add pgvector to requirements

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Slim eval rubric to 3 levels with one reference per axis ...

Five anchors per axis (10/30/50/75/100, even after the earlier shift)
were both redundant and amplified the model's snap-to-round-numbers
prior. Cut to three level descriptions per axis (weak / typical /
strong) with a single non-round reference score (53) on `typical`.
Re-state the scale as open-ended with no upper bound to make the
"future Navi may exceed past ceilings" intent explicit.

- rubric_v1.yaml: anchors → levels (5 → 3 per axis), reference score
  53 only on typical, scale framed as fully open-ended.
- judge.py: render_rubric_for_prompt walks the new `levels` shape and
  surfaces the reference score only when present.
- expert prompts (strict_critic, pragmatist, tech_lead): drop the
  example output blocks (their concrete numbers were misleading the
  judges), rewrite the scale paragraph for the new structure.
- schema.py: docstring no longer pins ">100" as the open-scale marker.

User intent: dynamics, not absolute scores. Weekly aggregates over
three averaged experts smooth individual snap-to-5 into continuous
trends; the rubric is a calibration aid, not a grading ceiling.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Fight rubric-anchor snapping in eval judges ...

Judges were clustering scores onto the rubric's round anchor values
(30, 50, 75, 100) instead of producing fine-grained continuous scores,
which made small differences between sessions invisible.

- rubric_v1.yaml: shift anchors off round numbers (33/51/77/102),
  reframe the scale as open-ended integers ≥ 0 with illustrative level
  descriptions, and tell judges explicitly not to round to anchors.
- expert prompts (strict_critic, pragmatist, tech_lead): mirror the
  scale framing and add an example output with deliberately non-round
  scores between anchors.
- judge.py: bump expert temperature 0.2 → 0.5 so the judges produce
  more varied, non-deterministic scores.

Old v1 evaluations in the DB are not comparable to new ones; user
intends to wipe and re-run from scratch, so versions are not bumped.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Stop image_view hallucinations on inline-attached images ...

The model was inventing fake paths/URLs (e.g. files.oaiusercontent.com,
/home/ubuntu/navi-1/input_file_0.png) and calling image_view on them
when the user attached an image directly in chat — the image was
already in the multimodal context, but the tool description and lack
of a signal pushed the model to "load" it anyway.

- websocket.py: when a user message has inline images, append a brief
  note that they are already in context.
- image_view.py: soften the description — keep proactive use for paths
  and URLs the model genuinely cannot see, but tell it inline images
  don't need this tool.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Rewrite eval_system.md as user guide; preserve original spec as eval_system_design.md ...

- docs/eval_system.md: replaced stale spec with current user guide
  covering UI tabs, CLI, rubric, experts, versioning, API
- docs/eval_system_design.md: preserved original design spec for
  reference

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Add eval system Phase 5 — debug UI ...

Self-contained SPA at /debug/eval (route already wired in 8e0eed6).
Single index.html in the existing debug/ style — vanilla JS, embedded
CSS, no framework, no build step. Four tabs:

- Sessions — filterable table (profile / status / limit), eval status
  pill, headline avg scores, click-through to detail
- Detail — session metadata + every stored eval run, axes laid out as
  axis × expert grids with inline averages, expert comments, button to
  re-evaluate this single session
- Stats — weekly per-axis means table, optional complexity-bucket split
- Run — form to trigger any scope (unevaluated / single / all), live
  status panel polling /eval/run/{id} every 2.5s, run history with
  click-to-attach

Hash routing: #detail/<session_id> deep-links to a session.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

changed llm

Add eval system Phase 4 — read endpoints and background runner ...

REST surface for the debug UI:
- GET /eval/sessions  — overview list with eval status / latest avg /
  feedback counts (single SQL: sessions ⨝ feedback ⨝ latest run)
- GET /eval/sessions/{id} — session detail with all evaluations
- GET /eval/stats — weekly per-axis means; optional complexity-bucket split
- POST /eval/run — fire-and-forget background eval, returns run_id
- GET /eval/run/{id}, GET /eval/runs — poll progress and history

Pulled the runner loop out of cli into runner.py so both the CLI and
the REST endpoint share the same loop. State for in-flight runs lives
in an in-memory registry (single-process, cleared on restart).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add eval system Phase 3 — judge runner end to end ...

Fills in the stubs from Phase 2:
- judge.render_session: full transcript with tool_call/tool_result folding,
  reactions inlined per assistant block, planning_logs appendix, no
  compression-summary substitution
- judge.run_expert: real LLM call, fence-tolerant JSON parse, single retry
  with corrective nudge on schema or parse error
- judge.evaluate_session: asyncio.gather across the three experts
- db.EvalDB: insert_evaluation_run (txn), list_evaluations,
  evaluated_session_ids, feedback_by_index helper
- cli `run` (filters: --session, --since, --limit, --re-evaluate-all,
  --dry-run, --model, --backend) and `show` (groups by eval_run_id, prints
  per-expert axes plus averaged scores)

Verified end-to-end against a real 10-message secretary session:
all three experts returned valid JSON first try; spread between strict
critic and the others surfaced as expected.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add eval system Phase 2 — rubric, expert prompts, judge skeleton ...

Drafts the v1 rubric (7 axes, anchors at 10/30/50/75/100, open scale),
three independent expert prompts (strict_critic / pragmatist / tech_lead)
that all return the same JSON shape, and the orchestration scaffolding:
schema.py (pydantic models), judge.py (rubric loader, score averaging,
fence-tolerant JSON parser, new_run_metadata), cli.py with argparse for
run / show / stats. Real LLM calls and transcript rendering land in
Phase 3 — the stubs raise NotImplementedError.

`python -m debug.eval` works as the entry point. Anchor `examples` are
left empty for now; user fills them with real session_ids later without
bumping rubric_version.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add eval system Phase 1 — message feedback signal ...

Spec at docs/eval_system.md describes the full LLM-as-judge plan;
this commit lands only the in-app feedback layer:
- debug/eval/ Python package with EvalDB (asyncpg) and FastAPI router
  exposing /eval/feedback (set / clear / list)
- message_feedback postgres table keyed by (session_id, message_index)
- thumbs up / down on each completed assistant block in the webclient,
  optimistic update with rollback on failure

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

changed llm & new ollama param

changed model

Strengthen todo progress discipline

Add structured planning review and adaptive depth

Tune reflect autonomy guidance

Improve compression and memory prompts

Remove tool-call-like examples from prompts

Improve prompt resilience and project context use

Fix profile prompt inconsistencies

Tune profile sampling configs

?

temp -

Add Android client documentation ...

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Open external links in browser on Android; add NaviAndroid UA token ...

- shouldOverrideUrlLoading: external URLs open via Intent, Navi server stays in WebView
- User-Agent gets "NaviAndroid/1.0" suffix for platform detection
- webclient/composables/usePlatform.js: exports isAndroid for future Android-specific behavior

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Collapse thinking and plan cards by default ...

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Replace LaTeX math symbols with Unicode before markdown rendering ...

Handles arrows, comparison operators, logic symbols and common math notation
that LLMs produce but marked.js doesn't render.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>