Fresh redeploy hardening: lint baseline, test floor + CI, #236, #180

[overcuriousity]

Implements ROADMAP Tier 1 and Tier 2 — the work needed to redeploy fresh and run in production. Spec and plan are committed under docs/superpowers/.

Deliberately out of scope: #248/#247 (memory confidence, episodic/semantic split), the eval harness and published benchmark, FTS5 search, session export/import, Phase 3 god-file splits, and the mechanical refactors #204–#210.

What landed

Area	Result
Lint baseline	581 findings → 0; 48 files reformatted; 4 rule ignores documented with rationale
CI	New .github/workflows/ci.yml: ruff check + ruff format --check + pytest, all blocking, on push and PR
Tests	29 → 124 (95 new)
Dead deps	asyncpg and aiosqlite removed
#236	Nightly skill-documentation condensation removed
#180	Zero-config local embedding fallback

Commit-by-commit

The two style: commits are mechanical and land first so every later diff stays readable. Review them separately from the rest.

1. style: ruff auto-fixes + formatter (no --unsafe-fixes)
2. style: remaining findings + documented ignores
3. test: convergence validators + credential redaction, plus the CI workflow
4. test: tool-call rescue parsers + NL translator helpers
5. chore: dead deps + stale docs
6. fix: Remove skill documentation condensation from dreaming cycle #236
7. feat: local ONNX embedding provider
8. feat: Local embedding fallback for zero-config memory resilience #180 wiring

Test coverage added

Only synchronous, IO-free surfaces — no mocks, no pytest-asyncio, no network.

• All 8 Convergence Protocol stage-2 validators, each with true-positive and false-positive cases. The false positives are the point: an over-firing validator blocks legitimate agent actions.
• All 6 tool-call rescue parser paths plus rescue_tool_calls end to end, including the case that ordinary prose must never be rescued into a phantom tool call.
• Credential redaction, config secret extraction, NL detection, JSON control-character repair.

The async orchestrators (run_convergence_protocol, process_tool_call) remain untested — they need a stubbed backend pool, which is follow-up work.

#236 — remove skill condensation

prompt_assembler injects only the skills table of contents, so condensing documentation reduced no prompt size while lossily LLM-rewriting every skill nightly. Losses compound across cycles, and because skill_store embeds summary + documentation, condensation also flattens the embedding signal.

Note the ordering hazard: DREAM_SKILLS_CONDENSATION_PROMPT.txt was listed in prompt_loader.REQUIRED_FILES, where a missing file is a startup failure. Deleting the template without that registry edit would brick startup. Verified with prompt_loader.validate_all().

#180 — zero-config local embeddings

A fresh deployment with no embeddings endpoint previously failed at startup.

• New infra/local_embeddings.py: lazily loaded ONNX all-MiniLM-L6-v2 (384-dim), cached in data/.embedding_cache/, fetched by URL rather than adding a huggingface_hub dependency.
• Wired at the single choke point llm_utils.embed(), so all 15 _embed() call sites benefit from one edit.
• onnxruntime + tokenizers ship as an optional extra (uv sync --extra local-embeddings), keeping the required dependency count flat.
• Scope is zero-config only. A configured endpoint stays authoritative and is never silently replaced — writing 384-dim vectors into a 1536-dim store would corrupt the index.
• Dimension guard: LocalVectorBackend now records the provider and dimension that built the store and refuses to open under a mismatched provider, instead of failing as an opaque numpy shape error mid-query. Pre-existing stores adopt their dimension from a stored vector on first open.

Verified end to end: with no endpoint configured, embed() downloads the model and returns 384-dim unit-norm vectors.

Reviewer notes

• sqlalchemy was kept deliberately. It is not imported directly; workers/scheduler_thread.py:28 uses APScheduler's optional SQLAlchemy jobstore, which APScheduler does not install itself. Removing it breaks scheduling at runtime, not at import.
• Scoped --unsafe-fixes use. Commit 2 applies ruff's own fixes for a narrow, behaviour-preserving rule selection (SIM102/103/108/117/118, C401, C408, F401, B905) rather than 25 hand edits; every hunk was reviewed individually. zip() sites take strict=False to preserve today's truncating behaviour rather than change it silently. No SIM105 rewrites — the broad except Exception handlers keep the agent loop alive and are ignored with a written rationale.
• Local build caveat (not a code issue). On Fedora, uv sync fails to build python-olm: CMake 4 rejects its cmake_minimum_required, and system Python 3.14 lacks dev headers. Use CMAKE_POLICY_VERSION_MINIMUM=3.5 uv sync --python 3.12. GitHub runners ship CMake 3.x, so CI is unaffected today.

Verification

ruff check .              All checks passed!
ruff format --check .     86 files already formatted
pytest -q                 124 passed
import wintermute.main    ok

Closes #236
Closes #180

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.