Token tooling repairs for the GPT-6 findings on #289/#297/#300/#303: probe exit status and lifecycle, context-mode Codex binding, retained coverage evidence

[seathatflowsinourveins]

Summary

This PR repairs the GPT-6 cross-family findings on merged #289, #297, #300 and #303, the token-efficiency wiring and coverage check, and retains the evidence behind them. Numbers below come from the committed artifacts.

• 289-1: Codex context-mode project binding.
  • Problem: the Codex plugin's MCP server declares cwd ".", which Codex resolves against the plugin root. The server then chose the newest Codex session log's worktree, not the session's own.
  • Fix: adoption/templates/codex.config.template.toml registers [mcp_servers.context-mode] at user scope with no cwd. It runs the pinned upstream start.mjs with CONTEXT_MODE_PLATFORM=codex and disables the plugin's own server while keeping its hooks and skills.
  • Evidence: in evidence/artifacts/context-mode-codex-binding-20260926/results.json, a local integration check across real servers answering ctx_execute pwd, the plugin form bound 2 of 4 sessions and the bare CLI 2 of 4, while the repaired user-scope start.mjs form bound 4 of 4.
  • Upstream tests: context-mode's own resolver tests at the pin are retained (upstream-resolver-tests.txt).
  • Host: templates only; this host is not reconfigured.
• 289-2: probe exit status. scripts/adoption_status.py --pinned-versions counts a version match only when the probe exits 0.
• 289-3: probe lifecycle.
  • Probes run in their own session and process group, and the whole group is killed on timeout or interruption.
  • SIGTERM and SIGHUP handlers are installed only when the disposition is SIG_DFL, so nohup and background jobs keep an inherited SIG_IGN.
  • Interrupts are held from before Popen until the group is killed and reaped. A test injects an interrupt at the start of every line the probe runs.
  • tests/test_adoption_status.py now has 72 tests. The failing-first controls are retained under evidence/artifacts/token-coverage-20260926/.
• 289-4: retained coverage evidence. evidence/artifacts/token-coverage-20260926/ holds the coverage runs with provenance, plus a failing wiring control. With an empty CODEX_HOME, the run reports complete: false. Statements that were never retained are marked as such.
• 303-1 / 303-2: headroom MCP offline claim. The template sets HF_HUB_OFFLINE=1 and TRANSFORMERS_OFFLINE=1 next to HEADROOM_OFFLINE and DO_NOT_TRACK. The decision addendum states that tiktoken's o200k_base still downloads on a cold cache (bounded by HEADROOM_TIKTOKEN_LOAD_TIMEOUT_SECONDS).
• 300-1: already resolved on main by memory-lifecycle-v2 repairs: ai-memory's own upstream suites at v2.3.2/v2.4.0/v2.4.1, repaired fixture (GPT-6 reviewed) #330's rewrite of blueprints/memory-lifecycle-v2/README.md, which says stack.json pins 2.4.1 and the landscape winner stays 2.3.2.
• 297-1: the dashboard wording ("7 versions unchecked") lands in the separate lane:shared dashboard PR.

New-machine inputs change here (scripts/adoption_status.py and the Codex templates). Each doc unit that describes the changed behaviour carries a "changed after v2026.09.26" marker.

Review

1. GPT-6: reviewed the four merged PRs.
2. Opus: repair, then an independent review, a finisher and a verification.
3. GPT-6 verification pass: marked all findings resolved and found 1 medium (the interrupt window around Popen), fixed at both ends, plus 1 low (a per-request deadline in binding_check.py), also fixed. The as-run original is kept as binding_check.original.py.
4. Independent headless review: its provenance notes are all addressed.

Residuals

• No host coverage run used the final scripts/adoption_status.py (sha256 2b372896…). The last runs used earlier revisions, and the later changes affect only an interrupted probe, which the unit tests and the retained every-line control cover.
• binding_check.py after the deadline fix has not been re-run against real servers. The retained results came from binding_check.original.py.
• Some first-run provenance was never retained and stays marked as such: invocation, exit status and stderr of the 00:04:37Z and 02:39:52Z runs, and the review's model attribution.
• Applying the template to this host (Codex user config and the headroom HF offline variables) is still pending.

Verification

• scripts/validate.py, host_receipts.py validate, component_matrix.py --check, new_host_grand_list.py --check, evidence_manifest.py --check and validate_catalogs.py: all pass.
• Full python -m unittest on this change before its final replay onto main: Ran 6058 tests, OK (skipped=635). After the replay, the related modules and every validator were re-run and pass.

SOTA sources

• Codex MCP server configuration (mcp_servers, cwd, per-plugin server overrides): https://developers.openai.com/codex/mcp and https://github.com/openai/codex/blob/rust-v0.155.1/codex-rs/core/src/plugin_config.rs
• context-mode project-dir resolution at the pin: https://github.com/mksglu/context-mode
• Python subprocess.Popen(start_new_session=True) and signal: https://docs.python.org/3/library/subprocess.html#popen-constructor and https://docs.python.org/3/library/signal.html
• Hugging Face offline mode (HF_HUB_OFFLINE, TRANSFORMERS_OFFLINE): https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.