Skip to content

v4.13.1 — Auto-orchestration fast-follows: concurrency + the adapt loop

Choose a tag to compare

View all tags
@alexherrero alexherrero released this 02 Jun 00:28
· 214 commits to main since this release
v4.13.1
fb1992c

PATCH. Three fast-follows on the v4.13.0 push-surface, found in a post-release sweep before moving on.

Fixed

  • Atomic state writes (concurrency-safety). auto_orchestration.save_state wrote the shared <vault>/_meta/auto-orchestration-state.json with a plain write_text — but that file is read+written by every repo/agent against the one shared vault, and the operator runs concurrent agents. A reader could catch a torn/partial write and degrade to an empty shape (a lost cooldown → a spurious re-fire). Now writes to a pid-unique temp + os.replace (atomic on POSIX and Windows), so a reader only ever sees a complete file. Verified under 4 concurrent writers — 0 torn reads.

Added

  • The idle chain's staged adapt candidates now surface. The idle chain stages Pass-1 candidates under _meta/skill-discovery-cache/adapt-state/ — but nothing told you they were there, so the discover→adapt→evaluate loop staged into a black hole (Pass-2 never prompted). The SessionStart briefing now reports "N skill candidate(s) staged for adapt-evaluation (/memory adapt-skills)", counting only candidates without a watchlist entry yet so the count clears as you evaluate. The adapt-evaluator contract now deletes the consumed Pass-1 scratch JSON after judging — so LOW-dropped candidates clear too, and a latent "re-judge every LOW candidate each run" inefficiency is closed.

Changed

  • memory-reflect-stop/hook.md brought in line with the .sh. The doc still described the long-superseded task-3 "mines but does NOT save" scaffold; in reality the hook has routed (HIGH → canonical paths, MEDIUM/LOW/ideas → _inbox/) since task 5, and gained the V4 #23 .reflected phase-dispatch dedup guard so a session is never reflected twice.

Internal

  • +6 tests (auto-orchestration 27 · briefing 36); full suite 556 green, 4-OS. Adversarial review: NO ISSUES FOUND (empirically verified the atomic write under 4 concurrent writers + the staged-count shifted-state snapshot transitions).

Cross-references

Assets 2
Loading