Record Codex native skill-event overturn condition

[seathatflowsinourveins]

Scope

Codex 0.161.0 defines and emits the native OTel event codex.skill_invocation, meeting the skills decision's capability overturn condition. Append a dated source correction and replace the README's stale event-denial statements.

The rollout scan remains a complementary input until the deployed collector preserves invocation type, scope and plugin attributes when present, and native coverage of effective RTK-wrapped reads is verified. No parser or collector behavior changes.

• Base: 9f7b38b23d15e47ceb6f68c6651b86088f621797.
• Lane: lane:foundation; draft for co-op review.
• Scope: two documentation files and their native-helper hash refresh in manifests/evidence.json.
• The decision's original 1,767-line prefix is byte-identical, preserving the original lines 280-281 and existing line citations.

SOTA sources

• openai/codex rust-v0.161.0, commit 979011409de0a60b52f179721948e65531d26144: native event emitter, documented contract, shell command normalization/dispatch and the indexed skill-document detector.
• Collector allowlist at the base: keeps skill.name/turn.id and omits the three native attributes.
• Directly read rollout records: metadata record 1 establishes cli_version = 0.161.0; tool-call record 66 establishes the requested skill-document read. Sanitized file/date/line locators are in the addendum; the exact filename mapping stays in a private observation receipt outside the checkout.
• Existing host_receipts.register_file API refreshes the two registered documents. No new registration code or test runner.

Evidence-class table

Claim	Evidence class	Evidence
Native event capability exists by 0.161.0	source_review	Pinned Codex emitter and contract
Client version and skill-read request appear in retained rollout records	source_review	Direct record 1/66 reads; no raw conversations retained
Current repository collection policy omits the three native attributes	source_review	Base-pinned collector source; active host revision is not asserted
Documentation preserves existing table/line contracts and parser fixtures	local_integration / synthetic	Named profile and skill-usage modules
Registered artifacts satisfy native integrity/scope checks	local_integration	Validator exit 0

Local commands run

Each module used nice -n 10 ionice -c2 -n7 timeout 600 python3 -m unittest <module>:

• tests.test_install_claude_profile: 109 tests, pass, one skip.
• tests.test_skill_usage: 203 tests, pass, 19 skips.
• Counted once: 312 tests, 292 passed, 20 skipped; zero failures/errors.
• nice -n 10 ionice -c2 -n7 timeout 600 python3 scripts/validate.py: exit 0; 70 components, 10,900 hashed files, four profiles, 239 receipts.
• git diff --check: exit 0.

Changed test expectations: none. An independent source critic verified the distinction between native capability, observed rollout requests, collector retention and successful execution. The addendum makes no fresh telemetry-count claim; the optional bounded local Loki query failed. Initial privacy/whitespace validation failures are retained in the private observation record; the final public locator is sanitized and native gates are unchanged.

Decision record

Append-only addendum in docs/decisions/2026-09-25-skills-trial-and-usage.md; README links to it.

Host evidence

No host receipt, account, service, active configuration or provider execution changed. PR #887 is unchanged.

Checklist

• Exact base, source pins and lane label.
• Original decision bytes and line citations preserved.
• Native-helper registration, named-module checks and native validator pass.
• No raw rollout content or local session identifier published.
• Co-op reads and exact-head hosted acceptance.
• CC cue and 5f landing.