GanyuanRan/Aegis v2.13.0

v2.13.0: task outcomes and focused baselines

on GitHub

Task outcomes guide decisions

• Define three shared principles inside the process baseline: preserve intent
  and boundaries, distinguish task and project goals, and deliver a complete,
  usable result supported by evidence. Their abstraction does not grant higher
  instruction priority or override concrete acceptance criteria.
• Carry the relevant checks into entry routing, design, planning, debugging,
  delegation, continuation, review, and completion. Preserve the requested
  object, authorized scope, actions, and deliverables across each stage.
• Separate successful internal steps from completed user delivery. Keep missing
  data and blocked actions visible, distinguish task completion from product
  capability acceptance, and reject hardcoded single-case repairs.

Focused baselines with clear ownership

• Reduce the two main baselines from 2,734 to 467 lines, about 74% fewer UTF-8
  bytes. The current-doc index routes readers to the relevant owner and section
  instead of making the whole baseline tree a default reading bundle.
• Define shared principles once in the process baseline. Keep engineering rules
  in execution guidance, observable acceptance criteria in the quality baseline,
  and operational actions and output contracts in their skills.
• Preserve necessary entry summaries for skill-copy installations. Raise the
  entry warning target to 4,500 bytes and hard ceiling to 6,000 bytes; the current
  entry uses 3,805 bytes. Other skill and route-bundle limits are unchanged.
• Check canonical owners and references instead of duplicated paragraphs. Add
  representative outcome cases to the workflow matrix, now containing 110 samples.

Upgrading and evidence limits

• Synchronize all eight package and plugin version fields at 2.13.0. Update
  through the existing Git or host plugin installation route, then restart or
  reload the host as required. Both manual global routing-prefix mirrors remain
  compatible; no manual re-copy is needed. When upgrading a copy-skills
  installation from before v2.12.3, follow the updater precaution in v2.12.3 below.
• Five paired native Codex scenarios passed on the previous and skill-changing
  candidate versions, covering selected-result reporting with an unseen variant,
  read-only diagnosis, partial-data delivery, a tiny rename, and focused baseline
  retrieval. Those runs predate the final documentation ownership refinement;
  they do not establish general behavioral or performance improvements.
• Byte reductions describe maintained documentation, not measured model
  attention, latency, token cost, or cross-host reliability. Existing host
  compatibility limits remain explicit. Aegis remains a Method Pack with no
  new runtime authority, universal goal-framing step, or official npm channel.

Release verification

• The full release check, bash tests/e2e/run-all.sh --full --host-profile fast, passed on the release source tree in an isolated Linux filesystem.

• Fresh representative Codex 0.162.0 smokes passed for natural-language skill routing and explicit skill requests, using gpt-6-astra at medium reasoning. These establish sampled skill loading, not general task success.

• OpenCode base compatibility, Codex plugin synchronization, version consistency, and undeclared-version audits passed. All 82 Python helper tests ran successfully with one filesystem-dependent skip.

• The missing-owner negative probe still rejects lost communication guidance. Broader live-host behavior remains subject to the published compatibility limits.

• GitHub CI passed on both the release PR and merged main commit. The merged source tree matches the tested release candidate exactly.

Full changelog: v2.12.3...v2.13.0