Verification record

Issue #1 compatibility recheck — 2026-09-14

On macOS, installed codex-cli 0.154.0 help still describes --ephemeral as disabling session-file persistence. A bounded read-only probe ran with codex exec --ephemeral --json --sandbox read-only -C <checkout> and inherited model/reasoning settings. It requested one standby harness_explorer, with no product reads, writes, or recursive delegation and no retry on spawn failure.

The CLI exited 0 and reported child ID /root/compatibility_probe and its READY response. The captured JSONL contains a completed wait event and that final report, but no spawn metadata independently confirming the dispatch. This is a reported success, not proof of a fixed CLI version, native parallelism, or a complete Harness workflow. The earlier 0.153.4 failure below remains a separate observation. Use the actual first-spawn check described in compatibility for each session.

The same installed CLI also completed the updated ordinary-session regression:

python3 scripts/live_smoke.py run \
  --target _workspace/live-tests/issue-1-20260914 \
  --baseline-file /tmp/codex-harness-issue-1-20260914-baseline.json --timeout 900

The runner exited 0 using its caller-held protected-file baseline. It verified rejection of the deliberately blocked first run, preservation of that run and alpha's output, reuse of alpha, rerun of beta and QA, and both final tests passing. Native tool metadata separately confirmed that alpha's child ID returned before beta was spawned, with exactly two harness_worker children and one harness_qa; the first child was reused without an extra probe. The local verifier still labels transport records as reported; this recheck does not independently audit all message contents or prove behavior under the original missing-thread failure. Local evidence is retained under the ignored fixture's _workspace/ and .harness/runs/issue-1-inputs/.

For this change, independent offline QA passed 94 tests, root and example structural validation, git diff --check, and the changed documentation's local links, anchors, and shell examples. These checks remain separate from native execution evidence.

2026-09-07 baseline

Checked 2026-09-07 (Asia/Seoul) against upstream revision cceac68ea1d0ad198ef4b7b906cd238375836387, using Python 3.13.7 and codex-cli 0.153.4 on macOS.

Automated checks

• python3 -m unittest discover -s tests -v: 94 tests passed.
• python3 scripts/validate.py --project .: 5 agents, 1 skill, 0 errors.
• python3 scripts/validate.py --project examples/service-migration: 2 agents, 1 skill, 0 errors; the example task graph also passed validation.
• The local skill-creator validator accepted the factory and example skills.
• The local plugin-creator validator accepted .codex-plugin/plugin.json and the standard skills/ payload.
• git diff --check: passed.

Tests exercise installation preservation, idempotency, dry runs, conflicts, symlink destinations, rollback, logical role-name collisions, Unicode TOML round trips, installed helper portability, invalid skill metadata, missing references, task dependency cycles and simultaneous write ownership. Runtime regressions additionally cover required check evidence, result/artifact/input freshness, context refresh, immutable prior runs, contract updates through a new plan template, retries, native lifecycle observations, capacity, and stop acknowledgments. Communication tests include 24 concurrent writer processes, Unicode and multiline payloads, correlated replies, corrupt-tail preservation, and separate exports. Live-test evidence checks reject changed checker code and unrelated delivery records; those offline tests do not simulate native agents.

GitHub Actions passed on Ubuntu with Python 3.11 and 3.13 for the first published migration commit 53660b1: CI run. Both jobs ran the automated suite, native artifact validation, and whitespace checks. The authenticated live Codex regression remains outside CI.

Live write, blocker, communication, and resume regression

Ran the opt-in fixture with codex-cli 0.153.4 in an ordinary workspace-write session:

python3 scripts/live_smoke.py run \
  --target _workspace/live-tests/native-20260907

Parent thread: 01a0799c-0bc4-7261-8dae-c1bb7e028575.

Observation	Result
Native workers	Actual harness_worker spawns returned /root/alpha_worker and /root/beta_worker
First output	Alpha wrote alpha.py and passed its acceptance test
Deliberate blocker	Beta observed UNRESOLVED in its spec versus expected BETA, captured the failing test, and returned blocked
Peer exchange	Alpha asked about the function contract and beta's blocker; the parent recorded the correlated answer and reported native send/receipt observations
Failure gate	First-run completion exited 1, identifying blocked beta and pending QA
Partial resume	After the controlled spec correction, alpha was reused with 0 new attempts; the same beta worker read a new packet and wrote/tested beta.py
Independent QA	Actual harness_qa spawn returned /root/verify_qa; both tests passed
Final completion	Resumed run passed with 0 completion errors
Preservation	Previous run records and the reused alpha file retained their checkpoint hashes
Lifecycle	Final native observations showed completed child turns; recorded as idle, with no closure claim
Communication records	13 first-run + 13 resumed-run events, including lifecycle entries; JSONL and Markdown saved separately

The primary session independently reran both product tests and the stronger evidence verifier against an external baseline covering 30 protected fixture files. The baseline was captured during the live run and checked afterward; future run invocations capture their full baseline before launching Codex. Sanitized native collaboration tool metadata was also checked against all three reported role IDs and final lifecycle observations. Message bodies and model analysis are not copied into that metadata extract.

Local evidence lives under _workspace/live-tests/native-20260907/_workspace/:

• verification.json: independent state, preservation, correlation, and test checks.
• native-tool-metadata.json: inspected native role dispatch and lifecycle metadata.
• native-report.json, native-events.jsonl, native-result.md: session observations and CLI output.
• communications/first.jsonl, communications/resumed.jsonl, and corresponding .md exports.
• checkpoint.json: the rejected first completion and previous-run hashes.

These artifacts are local and Git-ignored. The standalone verifier labels native IDs and message delivery as reported; it does not authenticate native transport or measure execution overlap. The separate metadata review verifies role dispatch and final lifecycle observations. This run is a small functional regression, not a throughput benchmark or a guarantee for every Codex client. The later resume --plan-file option and additional boundary guards were covered by the offline CLI tests; this live run exercised input-file changes with the original task graph.

Reproduce with a new disposable target. An authenticated local Codex CLI is required; the live command is deliberately outside unattended CI.

Earlier read-only native subagent smoke

Ran codex exec --strict-config --json --sandbox read-only from the final checkout with a bounded prompt: start harness_explorer and harness_reviewer before waiting, assign each one independent file to read, collect their actual results, and make no project edits or recursive delegations.

Final parent thread: 01a0781b-3657-7811-8ba0-a3ac06a8f0d7.

Observation	Result
Skill discovery	Runtime listed codex-harness:harness
Explorer	/root/smoke_explorer read .codex/config.toml and returned concurrency 3
Reviewer	/root/smoke_reviewer read the factory skill and returned all six architecture patterns
Scheduling	Both custom subagents started before the first wait; both returned results
Role substitution	None; the requested custom roles were used
Project mutations	None during this read-only smoke

This confirms custom-role discovery and a small parallel read workflow on the tested local CLI. It does not benchmark throughput or certify every client, write sandbox, domain workflow, or three-worker workload. The service migration example is a validated starter template, not an already migrated application.

An earlier --ephemeral probe discovered the roles but could not spawn them: the runtime returned collab spawn failed: no thread with id. The ordinary session above passed. Use an ordinary session for the documented live smoke; this observation does not establish the behavior of other CLI versions.

For local project invocation use $harness; when loaded as the plugin, the skill picker can display the namespaced codex-harness:harness name.