feat(brain): section posteriors with Thompson Sampling (ADR-048 Phase 1)#505
feat(brain): section posteriors with Thompson Sampling (ADR-048 Phase 1)#505ohdearquant wants to merge 7 commits into
Conversation
…ivation (ADR-048 Phase 1)
- SectionType enum (10 knowledge-section types) in state.rs
- SectionPosterior struct with Beta(a,b) updates, ESS cap via (m,s) canonical form
- section.rs: combinatorial TS weight derivation using softmax-over-Gamma-sampled Beta
posteriors with temperature schedule (Marsaglia-Tsang gamma sampler)
- V20 migration: section_posteriors table with (profile_id, section) unique constraint
- BrainState: section_posteriors HashMap integrated into snapshots + restore
- create_profile handler: seed_priors parameter for role-based prior initialization
("implementer", "researcher", "tester" templates)
- 11 tests covering weight normalization, ESS cap mean preservation, role seed priors
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…orState Resolves all 3 CRIT and 7 MAJ findings from codex review of PR #505: - F-001/F-002/F-003: delete conflicting section.rs types; section.rs now delegates to SectionPosteriorState.weights() / deterministic_weights() - F-004/F-005: implement ADR-048 Correction 1 temperature schedule in sample_weights(): tau = tau_0 * (exploration_epoch / epoch_max), softmax over Thompson samples with numerically stable max-subtraction - F-005/F-007: exploit path uses tau_exploit=0.1 softmax over posterior means; apply_floor_and_renorm() iterates to guarantee all weights ≥ 0.05 - F-006: DEFAULT_ESS_CAP 50 → 100 (ADR-048 Correction 2) - F-009/F-010: lib.rs section_posteriors → section_states; feedback routes through SectionPosteriorFold (only updates when section_signals present, not uniform update to all sections) - Fix apply_ess_cap formula: scale = (cap - prior_ess)/(ess - prior_ess) so resulting ESS equals cap exactly - Add Default impls for SectionPosteriorState and SectionPosteriorFold 143 unit tests + 2 integration tests pass; clippy -D warnings clean. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Merge sequence
Merge #499 first — PRs #500, #502, #503 depend on it (GitHub auto-retargets on merge). |
…sertions DEFAULT_ESS_CAP was set to 100.0 by the prior agent commit, diverging from ADR-048 spec which defines 50.0. Corrected to match spec. V20 migration test assertions still referenced the old table name (section_posteriors) instead of the actual V20 tables (brain_profile_snapshots, brain_event_log). Updated all 4 test assertions to verify the correct table and index names. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implements the persistence layer that saves brain state to the V20 tables (brain_profile_snapshots, brain_event_log): - persist.rs: new module with append_brain_event(), upsert_snapshot(), load_latest_snapshot(), load_events_since(), ensure_loaded(), and persist_after_feedback() async functions - PersistenceTracker: tracks loaded namespaces and dirty event counts with configurable batch threshold (default 5) - Save path: after every feedback fold, append to brain_event_log; upsert brain_profile_snapshots when dirty count reaches threshold - Load path: ensure_loaded() lazily loads latest snapshot and replays newer events through BalancedRecallFold + SectionPosteriorFold - Wired into dispatch() for lazy initialization on first verb call - Persistence failures are non-fatal (logged to stderr); the shared event store remains the source of truth for event history Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
RuntimeConfig::default() now includes additional_embedding_models with ParaphraseMultilingualMiniLmL12V2, which gets registered in the DB and causes note-creation tests to attempt model loading. In CI, the ONNX model files don't exist, causing 5 integration tests to fail with "model initialization failed: IO error: No such file or directory". Fix: explicitly set additional_embedding_models: vec![] in all test RuntimeConfig blocks that set embedding_model: None. Also includes cargo fmt and deno fmt fixes for runtime.rs and ADR docs. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 025a6a279a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if let Some((snapshot, updated_at)) = snapshot_result { | ||
| let replay_events = load_events_since(sql.as_ref(), &namespace, updated_at).await?; |
There was a problem hiding this comment.
When a namespace has no row in brain_profile_snapshots, this branch is skipped and the shared BrainState is left as whatever namespace was previously loaded. In a long-running server, the first brain.* request for a fresh namespace after namespace A was mutated will see A's profiles/bindings/posteriors, and any feedback already appended before the first batched snapshot is also never replayed because load_events_since is only called inside this branch. Treat the no-snapshot case as a fresh BrainState for this namespace and replay the event log from the beginning.
Useful? React with 👍 / 👎.
| if let Some(ref ss) = p.section_signals { | ||
| data["section_signals"] = ss.clone(); | ||
| } |
There was a problem hiding this comment.
If section_signals contains an unknown section key or signal value, the handler still returns emitted: true and stores the payload, but interpret() parses it with .ok() and converts the whole map to None, so SectionPosteriorFold performs no update. For knowledge-compose feedback with a typo such as "formalisms" or "usefull", callers get a successful response while the section posterior is silently unchanged; validate this object here and return InvalidInput before appending the event.
Useful? React with 👍 / 👎.
ADR-048 Correction 2 specifies ESS cap = 100 (half-life ~140 events). The research-derived optimal C_opt ≈ 1/(2ε²) for ε=0.07 yields ~100. 50 was too aggressive for slowly-stabilizing sections. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
Superseded by clean rebase — new PR incoming (old branch had 20+ merge conflicts from cherry-picked CI fixes). |
1 participant
Summary
with temperature schedule (ADR-048 Correction 1 — NOT normalized-Beta which has linear regret)
section_posteriorstable with (profile_id, namespace, section) unique constraintbrain.create_profilehandler acceptsseed_priorsparam for role-based initializationTest plan
cargo clippy -p khive-pack-brain --all-targets -- -D warningscleancargo check --workspaceclean🤖 Generated with Claude Code