Skip to content

Fix C01 C05 C09 C14 live qualification harness - #14

Merged
KeyffMS merged 5 commits into
mainfrom
fix/codex-capability-harness-c01-c05-c09-c14
Sep 1, 2026
Merged

Fix C01 C05 C09 C14 live qualification harness#14
KeyffMS merged 5 commits into
mainfrom
fix/codex-capability-harness-c01-c05-c09-c14

Conversation

@KeyffMS

@KeyffMS KeyffMS commented Sep 1, 2026

Copy link
Copy Markdown
Owner

Summary

Fixes the four run #5 capability outcomes that were blocked by qualification-harness setup rather than by the Codex 0.152.0 runtime contract.

C01 — repository skill discovery

  • seeds a real repository skill under .agents/skills/fixture-capability/SKILL.md outside the model sandbox;
  • creates a tracked nested working directory;
  • runs explicit activation both from repository root and nested cwd in read-only mode.

C05 — immutable reviewer handoff

  • creates a real immutable bundle plus recorded bundle digest and named artifact hash;
  • exercises fresh, stale-target, changed-bundle, missing-bundle, and escaped-path cases separately;
  • makes the previously missing stale-bundle rejection auditable.

C09 — checkpoint / compaction / recovery

  • installs the actual PlanAnvil distribution into the disposable fixture;
  • creates a real PlanAnvil planning run and schema-valid generation checkpoint before Codex starts;
  • configures model_auto_compact_token_limit to trigger genuine Codex auto-compaction;
  • wraps the real PreCompact/PostCompact/SessionStart hooks only to record structural invocation evidence while preserving their output and return code;
  • requires two real compaction/recovery cycles and prohibits direct hook invocation or simulated events.

C14 — planning worktree isolation

  • installs PlanAnvil into a separate disposable source Git repository under the qualification worktrees auxiliary writable root;
  • runs the real plan_anvil.py start controller against that repo;
  • verifies source branch/HEAD/index/status/source sentinel remain unchanged while a distinct planning branch/worktree is created;
  • does not disable the Codex sandbox or use danger-full-access.

Scope

Only C01, C05, C09 and C14 are overridden. C02 and C13 remain untouched because run #5 may represent real Codex behavior drift and must be investigated independently.

The self-hosted Codex qualification workflow remains workflow_dispatch-only and will not run on this PR. Standard PlanAnvil CI is the validation gate before merge.

@KeyffMS
KeyffMS merged commit 1d092f0 into main Sep 1, 2026
7 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant