Skip to content

Audrey 1.1.2 — self-extinguishing guard signals, delta packets, MCP Registry

Choose a tag to compare

View all tags
@Evilander Evilander released this 13 Jul 14:48
v1.1.2
7111c63
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Guard signal quality

  • Consolidation sweeps every real owning agent, so audrey dream converts episodes into semantic/procedural knowledge on stores that previously reported "evaluated 0 episodes".
  • Extinction learning: failure streaks retire once the tool succeeds afterward in scope; same-instant ties order by monotonic event id and fail toward warning. include_resolved keeps the diagnostic view.
  • Project-scoped failure signals with asymmetric directionality: cwd-less failures stay visible (cannot be proven foreign); only a provably-local success extinguishes a local streak. Shared semantic knowledge still travels across projects.
  • Autopilot-learned failure episodes are tool-matched at medium severity; user-authored risks keep full high-severity treatment. Capsule failure summaries cap at 240 chars.
  • Redaction preserves machine identifiers (mcp__* tool names, UPPER_SNAKE env vars) while still redacting title-cased passphrases and digit-bearing tokens.

Injection efficiency

  • Compact packet format (default): [memory_id confidence] "content" lines, one-line safety preamble, unchanged injection-safety invariants. AUDREY_PACKET_FORMAT=verbose reverts.
  • Session-delta injection (default): each memory injects once per session; SessionStart seeds the tracker, compaction clears it, state changes reinject. Measured live: a repeated 3.5k-char packet dropped to 0 on every prompt after the first. AUDREY_PACKET_DELTA=0 reverts.

MCP Registry

  • Audrey publishes to the official MCP Registry as io.github.evilander/audrey via GitHub OIDC, after the npm publish gates pass.

Full test suite: 853 passed. Hardened against an adversarial gpt-5.6-sol review (6 findings fixed pre-release).

Assets 2
Loading