v2.9.6 (2026-09-03)
Harden benchmark prompt-input and rationale recognition
- Accept both frozen skill-path spellings in the prompt-input marker: VIRTUAL_SNAPSHOT/skills (Codex <=0.146) and VIRTUAL_HOME/.agents/skills/aegis (Codex >=0.147). methodPackMarkerCount now sums both.
- Widen minimum-change rationale from minimum/smallest change to bounded family minimum|minimal|smallest + change|fix|edit|patch with up to two short modifiers. Harden negation, quotation/reference, commitment ordering, negative-predicate, performance and modifier-bound handling.
- Add regression coverage for marker spelling, synonym handling, negation/reference/attribution, and bounded-modifier limits.
Advisory subagent delegation lifecycle and event signals
- subagent-driven-development: add Slice Lifecycle and Result Trust -- choose route only after minimum baseline is understood, keep route fixed per slice, bind SubagentContextPacket to baseline/checkpoint, mark stale on workspace change, treat child output as evidence/proposal, keep partial/cancelled/blocked/stale visibly non-conclusive.
- dispatching-parallel-agents: add Event-Based Opportunities -- after baseline lock, parallel read-only investigation may be considered for conflicting evidence, high-risk compatibility/impact review, retry with materially different context/model, or separate host/install surfaces. Safety gates unchanged.
- writing-plans: add advisory reconsideration after minimum baseline and dispatching-parallel-agents trigger prompt reinforcement. No new artifact, role, or runtime authority.
- .gitignore: add .autoloom/ local-only directory.
Verification boundary
- Version-bump audit passed: 8 declared manifests in sync at 2.9.6; no undeclared references.
- Targeted unit checks for the two benchmark fixes passed.
- Full e2e remains environment-bound on this Windows host and is not claimed as passed here.