Skip to content

Fix C06 C08 C09 live qualification harness - #18

Merged
KeyffMS merged 4 commits into
mainfrom
fix/c06-c08-c09-deterministic-harness
Sep 2, 2026
Merged

Fix C06 C08 C09 live qualification harness#18
KeyffMS merged 4 commits into
mainfrom
fix/c06-c08-c09-deterministic-harness

Conversation

@KeyffMS

@KeyffMS KeyffMS commented Sep 2, 2026

Copy link
Copy Markdown
Owner

Summary

Replaces the remaining harness-blocked C06, C08, and C09 paths with deterministic live probes. C13 and baseline 2.2 are intentionally unchanged.

C06 — real PreToolUse + mandatory postcondition

  • installs the real PlanAnvil release and real plan-anvil-guard.py PreToolUse hook;
  • wraps the real hook only to record sanitized event/tool-name structure;
  • requires a direct apply_patch tool call to produce a real PreToolUse observation;
  • separately performs a controlled mutation outside the Codex hook lifecycle and requires Git postcondition detection without assuming hook interception.

C08 — deterministic PreCompact stop and repair

  • creates a real active PlanAnvil generation run with no valid checkpoint;
  • uses model_auto_compact_token_limit = 200 and model_auto_compact_token_limit_scope = "body_after_prefix", redundantly applied by CLI;
  • forces genuine automatic compaction through a large tool-result payload;
  • requires the real plan-anvil-compaction.py PreCompact hook to return continue=false with checkpoint/recovery semantics and no PostCompact;
  • creates a real generation checkpoint and repeats the probe, requiring PreCompact + PostCompact with no stop to prove the blocker is repairable rather than permanent.

C09 — repeated genuine compaction and recovery

  • starts with a schema-valid real generation checkpoint;
  • uses a 1000-token body_after_prefix automatic-compaction budget plus large deterministic payload segments;
  • records real PreCompact/PostCompact/PreToolUse events while delegating to the real product hooks;
  • requires at least two completed compaction cycles;
  • validates the real checkpoint against canonical state/Git before and after compaction;
  • requires a real tool call after the second PostCompact to prove the session did not enter a permanent stop loop.

Safety / scope

  • no sandbox or approval bypass;
  • no danger-full-access, privileged container, or extra capabilities;
  • model-tool network remains disabled;
  • self-hosted qualification remains workflow_dispatch-only from main;
  • this PR itself uses only the ordinary hosted CI matrix;
  • C13 remains tracked separately in Analyze C13 SubagentStart assumptions under codex exec --ephemeral #17.

@KeyffMS
KeyffMS merged commit 5c2c270 into main Sep 2, 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