feat: relevance-ranked decision delivery, SessionStart + edit-time injection with usage feedback#776
Merged
Merged
Conversation
SessionStart now scores active decisions against the session's likely working set (dirty/staged files, branch-vs-main diff, previous session's edits, branch-name tokens) expanded one hop through import edges and co-change partners, and injects the top few under a hard ~400-token cap with a relevance floor: relevance or silence, never top-confidence globally. Repo-wide session rules (no node links) carry a flat base relevance since SessionStart is their only delivery path. PostToolUse Edit/Write additionally surfaces the edited file's governing decision once per session per decision, under a strict per-session cap. Dedup is an atomic INSERT OR IGNORE into the sessions.db sidecar because the JSON session state is written read-modify-write by concurrently racing hook processes and loses updates. Shown-decision ids land in a new injections table; at update time the miner judges each one against that session's mined user corrections: contradiction bumps staleness (the evolution amended move), silence relaxes it (reaffirmed). Binary on purpose for v1; the followed vs ignored split is validation-gated backlog. All hook lookups are raw read-only stdlib sqlite3. Measured on this repo: edit notice ~15ms, SessionStart decision scoring ~135ms (git subprocess dominated), no LLM, no network. Dogfood-driven guards: link node ids are compared in POSIX (Windows extraction stores backslashes), top-level module links never fire, and legacy rows that echo the title into decision/rationale are not repeated.
Cold-start dogfood promoted a few one-off task instructions to active alongside the durable rules. Dismissals handle the records themselves, but the block now also caps always-eligible unlinked rules at two per session so a future mis-promotion costs one slot instead of the whole block, and working-set-linked decisions can never be crowded out.
|
✅ Health: 7.6 (unchanged) 📋 At a glance 🚨 Change risk: high (riskier than 83% of this repo's commits · raw 9.6/10)
🩹 Review priority (files here with the most recent bug-fix history — defects cluster, so review these first)
🔎 More signals (3)🔥 Hotspots touched (5)
2 more
🔗 Hidden coupling (1 file)
💀 Dead code (1 finding)
📊 Full report · ⭐ Star Repowise · 📥 Install bot · Last updated 2026-07-11 08:32 UTC |
RaghavChamadiya
changed the title
swati510
approved these changes
Jul 11, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
The decision layer now reaches the agent at the moment it matters, ranked by relevance to what the session is actually doing. Two delivery moments, both pure indexed-SQLite lookups (no LLM, no network):
SessionStart block. The augment hook scores the repo's active decisions against the session's likely working set: dirty/staged files, files changed on the branch vs main, the previous session's edited files, and branch-name tokens, expanded one hop through import edges and co-change partners. Scoring is hop-weighted overlap x confidence x freshness with a relevance floor and a hard ~400-token cap. Nothing clears the floor, nothing is injected: relevance or silence, never top-confidence-globally. Repo-wide rules mined from user corrections (no file links, so no other delivery path) compete at a flat base relevance.
Edit-time notice. When the agent edits a file with a governing decision (
decision_node_links, file links plus module-prefix links), it gets a one-line "governed by: <title> because , confirmed across N sessions" notice, once per session per decision under a strict per-session cap.Usage feedback v1. Every injected decision id is recorded in the
.repowise/sessions/sessions.dbsidecar (newinjectionstable). At update time the session miner judges each shown decision against that session's mined user corrections: a contradicting correction bumps staleness (the evolutionamendedmove), silence relaxes it (reaffirmed). Deliberately binary for v1.Noise guards (dogfood-driven)
Running the hooks live on this repo while building surfaced four failure modes, each now guarded and tested:
packages) "governed" every file in the tree; module links now require at least two path segments.INSERT OR IGNOREinto the sidecar; the state list is only a fast path.Latency
Measured on this repo (hook budget: no LLM, indexed lookups):
git status --porcelain --branchcall); SQLite lookups are ~40ms. Runs once per session.Tests
31 new tests: SessionStart relevance/floor/cap/silence, one-hop expansion via graph edges and co-change partners, branch tokens, global rules, edit-notice dedup (including the state-loss race), path normalization, module-depth and echo guards, injection recording, and the followed/contradicted feedback pass including the min-age gate and schema compatibility between the hook-side raw table and the staging store.