Harden memory prompt and relation hygiene #128

Merged

kleb merged 17 commits from feature/harden-memory-post-audit into main

2026-08-30 23:52:08 +02:00

kleb commented

2026-08-30 22:23:48 +02:00

Collaborator

Summary

redact high-confidence credential shapes on memory writes and add confirmed targeted prompt replacement
stop pending relations from suppressing V2 recall, ignore canonical project names during relation matching, and make candidate judgment explicit
prune only truly empty hook sessions, bound CLI session listings, and add cleanup-only orphan repair
cover API, CLI, MCP, recall, relation, and lifecycle behavior

Verification

bun run check:quality
964 Bun tests passed
238 Vitest tests passed
targeted memory suites passed after final lifecycle safeguards
blinded 72-run gpt-5.6-luna comparison: base 33/36, feature 35/36, zero base-only wins

The Luna misses were confined to an exact-marker capture grader. Every missed trace still called memory_save successfully and reported completion; all ordinary memory CRUD, retrieval, judgment, pause/resume, and session-end trials passed.

## Summary - redact high-confidence credential shapes on memory writes and add confirmed targeted prompt replacement - stop pending relations from suppressing V2 recall, ignore canonical project names during relation matching, and make candidate judgment explicit - prune only truly empty hook sessions, bound CLI session listings, and add cleanup-only orphan repair - cover API, CLI, MCP, recall, relation, and lifecycle behavior ## Verification - `bun run check:quality` - 964 Bun tests passed - 238 Vitest tests passed - targeted memory suites passed after final lifecycle safeguards - blinded 72-run `gpt-5.6-luna` comparison: base 33/36, feature 35/36, zero base-only wins The Luna misses were confined to an exact-marker capture grader. Every missed trace still called `memory_save` successfully and reported completion; all ordinary memory CRUD, retrieval, judgment, pause/resume, and session-end trials passed.

kleb added 1 commit

2026-08-30 22:23:48 +02:00

Harden memory prompt and relation hygiene 28b61d1fd8

kAgent commented

2026-08-30 22:23:48 +02:00

Collaborator

kReview review

Verdict: no findings

The reviewed changes have low risk; no concrete correctness, security, or lifecycle defect is established by the visible diff and repository evidence.

Reviewed by kReview at ed39fed81d. This comment is conservative and based only on the PR diff, metadata, and supplied repository context.

Est. cost ~$24.93 total (15.6M in / 79.2k out) · this run ~$0.30 (59.5k in / 707 out) / gpt-5.6-sol.

## kReview review **Verdict:** no findings The reviewed changes have low risk; no concrete correctness, security, or lifecycle defect is established by the visible diff and repository evidence. _Reviewed by kReview at `ed39fed81d`. This comment is conservative and based only on the PR diff, metadata, and supplied repository context._ _Est. cost ~$24.93 total (15.6M in / 79.2k out) · this run ~$0.30 (59.5k in / 707 out) / gpt-5.6-sol._

kleb added 1 commit

2026-08-30 22:27:11 +02:00

Limit automatic secret redaction f7b0cca557

kleb added 1 commit

2026-08-30 22:31:33 +02:00

Cover authorization and prompt route edges 1655d5b357

kleb added 1 commit

2026-08-30 22:41:09 +02:00

Make prompt redaction recovery-safe f4a1d1ca94

kleb added 1 commit

2026-08-30 22:46:07 +02:00

Scrub prompt-derived summary content 79226b6f5f