live: Codex TUI actor fanout and live observer dogfood

Goal

Make mimetic-cli dogfood itself with real local Codex TUI actors and a live observer, starting from the committed mimetic/ dogfood config.

Current state

mimetic run --dry-run works and produces public-safe synthetic bundles.
mimetic watch --no-open renders a static observer for a completed bundle.
mimetic run without --dry-run correctly fails closed with MIMETIC_LIVE_RUN_UNIMPLEMENTED.
The dogfood config explicitly marks real Codex TUI actors, 4x fanout, and live observer streaming as deferred.

Scope to spec first

Define the smallest safe local Codex TUI actor contract.
Start with 1x local actor before 4x fanout.
Add deterministic actor lifecycle events: spawn, prompt submitted, observation, artifact, verdict, exit, timeout, cancellation.
Stream events into the observer while the run is active.
Keep all generated runtime state under ignored .mimetic/.
Require explicit opt-in before provider spend, GitHub mutation, E2B usage, or external network calls.

Public-safety requirements

No PII, PHI, secrets, keys, tokens, raw private transcripts, private screenshots, or private upstream artifacts in committed files or public issue drafts.
Redact or block actor logs before they become observer artifacts.
Keep credentials as env var names and setup instructions only.
Fail closed if a run cannot prove redaction status.

Acceptance proof

pnpm mimetic:doctor
pnpm mimetic:run runs a real local 1x Codex TUI dogfood actor without --dry-run.
A separate command proves 4x persona fanout or the ticket is split after the 1x proof.
pnpm mimetic:watch opens or serves a live observer that updates during the run.
pnpm mimetic:verify --run latest verifies the completed bundle.
pnpm check passes.
The observer artifact and feedback draft are public-safe on review.

Added a local spec on branch/worktree codex/oss-meta-completion:

docs/architecture/local-codex-tui-actor.md

The spec scopes the smallest safe local Codex TUI actor before 4x fanout:

explicit opt-in for live local actor work;
deterministic lifecycle events (actor.spawned, actor.prompt.submitted, actor.observation, actor.artifact, actor.verdict, actor.exited, actor.timeout, actor.cancelled);
ignored runtime artifact layout under .mimetic/runs/<run-id>/;
redaction and stop conditions;
first-slice proof commands for a 1x actor.

Proof for the spec slice:

pnpm public-surface:scan
git diff --check

This intentionally does not implement the live local actor yet. The next implementation slice should keep non-dry-run fail-closed unless --actor codex-tui or an equivalent explicit opt-in is present, then prove 1x before splitting/attempting 4x fanout.

Updated #28 on PR #37 with the first implementation slice:

mimetic run --actor codex-tui --sims 1 --timeout-ms <ms> now explicitly opts into one local actor.
Non-dry-run still fails closed unless --actor codex-tui or MIMETIC_ENABLE_LOCAL_CODEX_TUI=1 is present.
The actor writes deterministic lifecycle events: actor.spawned, actor.prompt.submitted, actor.observation, actor.artifact, actor.verdict, and actor.exited/actor.timeout.
Actor output is redacted before persistence; the bundle links a sanitized transcript and actor trace under ignored .mimetic/ runtime state.
verify and watch --run ... --detach --no-open now accept/render live-mode actor bundles.
Linux real-TUI launch uses a script PTY wrapper because Codex TUI requires a terminal.

Proof:

pnpm check
pnpm public-surface:scan
git diff --check
MIMETIC_CODEX_ACTOR_COMMAND=... pnpm mimetic -- run --actor codex-tui --sims 1 --timeout-ms 5000 --run-id codex-local-fixture-proof --json
pnpm mimetic -- verify --run codex-local-fixture-proof --json
pnpm mimetic -- watch --run codex-local-fixture-proof --detach --no-open --json

Real local TUI attempt:

The first no-PTY attempt failed with stdin is not a terminal.
After adding the PTY wrapper, the TUI launched but timed out at Codex workspace trust before autonomous completion.
The timeout bundle still verified and rendered through Observer with public-safe sanitized evidence.

Next #28 slice: add an explicit Codex workspace-trust preflight/recovery command, or split a separate non-TUI codex exec actor mode while keeping the TUI contract honest about PTY/trust requirements. This comment intentionally excludes raw logs, secrets, and local private transcript content.

Follow-up on the #28 implementation slice in PR #37: added fail-fast Codex workspace-trust preflight.

Current behavior for the default real TUI actor:

If the Codex trust root is not explicitly trusted, mimetic run --actor codex-tui --sims 1 blocks before spawning the TUI.
The run still writes a public-safe blocked bundle with actor.preflight.blocked, actor.verdict, and actor.blocked lifecycle events.
That blocked bundle passes mimetic verify and renders with mimetic watch --run ... --detach --no-open.
Fixture actor proof still passes through MIMETIC_CODEX_ACTOR_COMMAND=..., so CI does not need real provider/auth state.

Final proof for this slice now includes:

pnpm check
pnpm public-surface:scan
git diff --check
MIMETIC_CODEX_ACTOR_COMMAND=... pnpm mimetic -- run --actor codex-tui --sims 1 --timeout-ms 5000 --run-id codex-local-fixture-proof --json
pnpm mimetic -- verify --run codex-local-fixture-proof --json
pnpm mimetic -- watch --run codex-local-fixture-proof --detach --no-open --json
pnpm mimetic -- run --actor codex-tui --sims 1 --timeout-ms 90000 --run-id codex-local-trust-preflight-proof-2 --json
pnpm mimetic -- verify --run codex-local-trust-preflight-proof-2 --json
pnpm mimetic -- watch --run codex-local-trust-preflight-proof-2 --detach --no-open --json

Remaining #28 work: add an explicit trust bootstrap or a separately-scoped codex exec actor mode, then prove autonomous completion rather than a preflight block. No raw logs, secrets, stream URLs, or private transcript content included here.

Added the next #28 slice in PR #38: noninteractive local codex-exec actor support.

What changed:

mimetic run --actor codex-exec --sims 1 is now a supported explicit live actor mode.
The actor runs codex exec with read-only sandbox, JSON event output, ephemeral session mode, and ignored project rules, then persists a sanitized JSONL transcript plus actor.json trace under .mimetic/runs/<run-id>/.
The resulting live bundle verifies and renders in Observer like the TUI actor bundle.
codex-tui remains the PTY/trust-gated visual actor contract; codex-exec is the autonomous noninteractive local completion path, not a replacement for TUI proof.
The committed mimetic/ dogfood config now reflects the current actor boundary instead of saying all live actors are unsupported.

Proof:

pnpm check
pnpm public-surface:scan
git diff --check
pnpm mimetic -- doctor --json
pnpm mimetic -- watch --json --no-open --run-id watch-post-37-baseline
pnpm mimetic -- lab oss --dry-run --json --no-open --run-id oss-dry-post-37-baseline
pnpm mimetic -- lab oss-smoke --limit 4 --json --run-id oss-smoke-post-37-baseline
pnpm mimetic -- lab oss --detach --no-open --json --run-id oss-live-post-37-baseline --repos developit/mitt,lukeed/clsx,sindresorhus/is-plain-obj,ai/nanoid --count 4
pnpm mimetic -- verify --run oss-live-post-37-baseline --json
pnpm mimetic -- run --actor codex-exec --sims 1 --timeout-ms 90000 --run-id codex-exec-real-proof-3 --json
pnpm mimetic -- verify --run codex-exec-real-proof-3 --json
pnpm mimetic -- watch --run codex-exec-real-proof-3 --detach --no-open --json

The live E2B OSS lab classified 4/4 target repos as passed and captured 4/4 screenshot fallbacks. The real local codex-exec actor passed in a bounded run and produced a verified Observer-renderable live bundle.

Remaining #28 work: TUI trust bootstrap, live Observer follow while actors are still running, and 4x fanout. No raw transcripts, stream URLs, sandbox IDs, or secret-bearing material included here.

Progress update for #28: PR #39 adds explicit noninteractive codex-exec fanout.

Public-safe proof from the branch:

codex-exec now supports --sims 1..4; codex-tui remains single-lane with trust preflight.
Four bounded read-only lane focuses are emitted: install/readability, public-safety/trust, Observer/evidence, and verification/release.
Fanout bundles now include 4 terminal streams, 4 sanitized transcript artifacts, 4 actor traces, per-lane lifecycle events, and aggregate review status.
The existing 1x codex-exec artifact contract is preserved: transcripts/codex-exec-sanitized.jsonl and actor.json.

Verification completed:

pnpm check -> 10 test files / 48 tests passed, build passed.
pnpm public-surface:scan -> 72 candidate text files checked.
git diff --check -> passed.
pnpm mimetic -- doctor --json -> passed.
Env-sourced real fanout run codex-exec-fanout-real-proof-env -> passed, 4/4 lanes.
pnpm mimetic -- verify --run codex-exec-fanout-real-proof-env --json -> 5/5 checks passed.
pnpm mimetic -- watch --run codex-exec-fanout-real-proof-env --detach --no-open --json -> Observer rendered.
Env-sourced 1x regression codex-exec-1x-regression-env -> passed and preserved original 1x artifact paths.

No raw transcripts, secrets, sandbox URLs, or target OSS mutations are included here.

PR: #39

PR #39 is merged to main as a996b589138ebaf88883e9002f76aee2573d6690.

Merged slice:

Explicit codex-exec --sims 1..4 fanout.
Per-lane sanitized terminal evidence, actor traces, lifecycle events, and aggregate review status.
Preserved 1x exec artifact paths.
Updated dogfood docs/coverage/scenario language.

Remote CI passed on Node 22.14.0 and Node 24. Local proof also passed with env-sourced real 4-lane fanout, verify, Observer render, and 1x regression.

Still open for #28:

Codex TUI trust bootstrap beyond fail-fast blocked bundles.
Live Observer follow while local actors are still running.
Deeper repeated OSS lab loops using the new fanout path.

Progress update for #28: PR #40 adds active-run Observer snapshots for local codex-exec.

Public-safe behavior added:

Local codex-exec runs now write a provisional running run.json and observer/observer-data.json before actor completion.
Each lane emits actor.running in events.ndjson.
The same Observer data path is refreshed after sanitized transcripts, actor traces, and final verdict events are available.
This is scoped to noninteractive codex-exec; Codex TUI trust bootstrap and TUI live-follow remain separate work.

Verification completed:

pnpm check -> 10 test files / 49 tests passed, build passed.
pnpm public-surface:scan -> 72 candidate text files checked.
git diff --check -> passed.
pnpm mimetic -- doctor --json -> passed.
Env-sourced real fanout run codex-exec-live-follow-real-proof-env -> passed, 4/4 lanes.
pnpm mimetic -- verify --run codex-exec-live-follow-real-proof-env --json -> 5/5 checks passed.
pnpm mimetic -- watch --run codex-exec-live-follow-real-proof-env --detach --no-open --json -> Observer rendered.
Metadata check: 4 actor.running events, 4 passed streams, final Observer active count 0.
Fixture regression proves the intermediate active state with a slow fake actor: running Observer data appears while actors are still active, then refreshes to passed after completion.

No raw transcripts, secrets, sandbox URLs, or target OSS mutations are included here.

PR: #40

PR #40 is merged to main as b37179386b9a68d1716938150f460b3ad697307b.

Merged slice:

Local codex-exec runs publish running run.json and observer/observer-data.json before actor completion.
Each exec lane emits actor.running events.
Final Observer data refreshes after sanitized transcript/trace artifacts and verdicts are available.
Dogfood coverage now distinguishes exec active-run snapshots from the still-open Codex TUI live-follow path.

Remote CI passed on Node 22.14.0 and Node 24. Local proof passed with the active slow-actor fixture, env-sourced real 4-lane fanout, verify, and Observer render.

Still open for #28:

Codex TUI trust bootstrap beyond fail-fast blocked bundles.
Codex TUI live-follow once trusted TUI launch is available.
Deeper repeated OSS lab loops using codex-exec fanout plus active Observer snapshots.

Progress update for #28: the post-#40 OSS lab loop found and fixed a result-classification gap in PR #41.

Post-merge lab run on current main:

pnpm check -> 10 test files / 49 tests passed, build passed.
pnpm mimetic -- doctor --json -> passed.
pnpm mimetic -- watch --json --no-open --run-id watch-post-live-follow-main -> Observer rendered.
pnpm mimetic -- lab oss --dry-run --json --no-open --run-id oss-dry-post-live-follow-main -> passed.
pnpm mimetic -- lab oss-smoke --limit 4 --json --run-id oss-smoke-post-live-follow-main -> 4/4 disposable public clones passed.
Env-sourced live lab oss-live-post-live-follow-main -> 3 lanes passed, 1 lane timed out waiting for remote bootstrap completion marker, 4/4 screenshots captured, bundle verified.

Finding:

The mixed live result still returned ok: true and review contract_proof_only, which made a timed-out live lane look too green.

PR #41 fix:

Live OSS meta-lab terminal failed, blocked, and timed_out lanes now propagate into top-level JSON ok and bundle review verdict.
Dry-run remains contract_proof_only.
Missing-key live attempts remain command-successful blocked evidence.
All-passed live terminal completions become review pass.

Proof for PR #41:

pnpm vitest run tests/oss-lab.test.ts -> 9 tests passed.
pnpm check -> 10 test files / 50 tests passed, build passed.
pnpm public-surface:scan -> 72 candidate text files checked.
git diff --check -> passed.

No raw remote logs, sandbox IDs, stream URLs, secrets, or target OSS mutations are included here.

PR: #41

PR #41 is merged to main as 025edbbf2bf5c300f09de90e8df2dbc5f0d70931.

Merged slice:

Live OSS meta-lab terminal failed, blocked, and timed_out lanes now propagate into top-level JSON ok and bundle review verdict.
Dry-run remains contract_proof_only.
Missing-key live attempts remain command-successful blocked evidence.
All-passed live terminal completions become review pass.

Proof:

pnpm vitest run tests/oss-lab.test.ts -> 9 tests passed.
pnpm check -> 10 test files / 50 tests passed, build passed.
pnpm public-surface:scan -> 72 candidate text files checked.
git diff --check -> passed.

GitHub registered no CI checks for PR #41; local proof was the available merge gate. No raw remote logs, sandbox IDs, stream URLs, secrets, or target OSS mutations were included.

Next loop should rerun the live OSS lab on 025edbb and confirm that any future timed-out lane is reported as non-green instead of neutral contract proof.

Post-PR #41 confirmation on fresh main 025edbb:

pnpm vitest run tests/oss-lab.test.ts -> 9 tests passed.
pnpm mimetic -- doctor --json -> passed.
Env-sourced live lab oss-live-post-classifier-main across developit/mitt, lukeed/clsx, sindresorhus/is-plain-obj, and ai/nanoid -> 4/4 lanes passed.
pnpm mimetic -- verify --run oss-live-post-classifier-main --json -> 5/5 checks passed.
pnpm mimetic -- watch --run oss-live-post-classifier-main --detach --no-open --json -> Observer rendered.
Metadata check: review verdict pass, 4 passed simulation statuses, 4 passed stream completions, 4 bootstrap passed events, and 4 screenshot fallbacks captured.

This confirms the positive side of the classifier: all-passed live OSS terminal completions now produce pass rather than neutral contract_proof_only.

Remaining #28 work is now concentrated on Codex TUI trust bootstrap and true TUI live-follow after trust, plus further repeated OSS lab hardening as new failures emerge.

Progress update for #28: PR #42 moves the Codex TUI boundary from false preflight block to trusted launch with timeout evidence.

What changed:

Codex workspace trust preflight now accepts an exact trusted project entry or an explicit trusted ancestor project entry in $CODEX_HOME/config.toml.
Missing trust still fails closed before spawn and writes a blocked bundle.
A new regression test exercises the default TUI command path with a fake codex binary on PATH, so preflight is not bypassed by MIMETIC_CODEX_ACTOR_COMMAND.

Proof:

pnpm vitest run tests/run.test.ts -> 12 tests passed.
pnpm vitest run tests/run.test.ts tests/dogfood.test.ts -> 15 tests passed.
pnpm check -> 10 test files / 51 tests passed, build passed.
pnpm public-surface:scan -> 72 candidate text files checked.
git diff --check -> passed.
Env-sourced real TUI proof codex-tui-trusted-ancestor-real-proof -> no preflight block; actor spawned and then timed out at the actor timeout.
pnpm mimetic -- verify --run codex-tui-trusted-ancestor-real-proof --json -> 5/5 checks passed.
pnpm mimetic -- watch --run codex-tui-trusted-ancestor-real-proof --detach --no-open --json -> Observer rendered.
Metadata check: actor.spawned, actor.prompt.submitted, actor.timeout; no actor.preflight.blocked.

This does not claim autonomous Codex TUI completion yet. Remaining #28 work: make the trusted TUI path produce a final verdict instead of timing out, and then add TUI live-follow after that.

PR: #42

PR #42 is merged to main as 55076a8f0fa3aa25d0f6394e526155af06efcc7e.

Merged slice:

Codex TUI trust preflight honors exact trusted project entries and explicit trusted ancestor project entries.
Missing trust still blocks before spawn and writes a verifiable blocked bundle.
Trusted launch now gets past the previous false preflight block.

Proof:

pnpm check -> 10 test files / 51 tests passed, build passed.
pnpm public-surface:scan -> 72 candidate text files checked.
git diff --check -> passed.
Remote CI passed on Node 22.14.0 and Node 24.
Env-sourced real TUI proof codex-tui-trusted-ancestor-real-proof -> no preflight block; actor spawned and timed out at the actor timeout.
pnpm mimetic -- verify --run codex-tui-trusted-ancestor-real-proof --json -> 5/5 checks passed.
pnpm mimetic -- watch --run codex-tui-trusted-ancestor-real-proof --detach --no-open --json -> Observer rendered.

Remaining #28 work is now narrower: make the trusted TUI path produce a final verdict instead of timing out, then add TUI live-follow after that.

Status update after merged PR #43 (f61348a): the trusted Codex TUI autonomous completion slice is now green, but I am reopening this issue because live TUI Observer follow remains a separate open slice.

What landed:

exact Codex project-root trust is required for default TUI launch; ancestor-only trust is now blocked before spawn because the real Codex TUI still prompts internally;
the local actor runner answers minimal terminal cursor/color queries, normalizes captured TUI output, and records sanitized evidence;
TUI actor completion now uses a nonce-bearing final marker: MIMETIC_ACTOR_VERDICT=<status> MIMETIC_ACTOR_NONCE=<run-nonce>;
the actor is terminated after the final marker instead of drifting to timeout;
TUI latest.json is published only after final run/review artifacts exist;
the TUI prompt is now read-only-sandbox aware, so it inspects committed config/existing evidence instead of trying write-producing commands.

Proof:

PR Complete trusted Codex TUI actor runs #43 CI passed on Node 22.14.0 and Node 24.
Local pnpm check passed with 54 tests, typecheck, and build.
pnpm public-surface:scan passed: 72 candidate text files checked.
pnpm mimetic -- doctor --json passed.
Real env-backed Codex TUI proof passed with run id codex-tui-readonly-real-proof:
- run --actor codex-tui --sims 1 --timeout-ms 240000 returned ok true;
- verify --run codex-tui-readonly-real-proof passed 5/5 checks;
- watch --run codex-tui-readonly-real-proof --detach --json --no-open returned ok true;
- sanitized metadata: review verdict pass, stream status passed, reason actor reported passed verdict marker, duration 35066ms, transcript bytes 217067, nonce marker passed=true / blocked=false / failed=false.

Remaining for this issue:

live Observer follow while the Codex TUI actor is still running;
if desired, a TUI-specific fanout design beyond the current 1x TUI + 1-4x codex-exec fanout split.

No target OSS repos, GitHub issues, or production/private artifacts were mutated by this proof.

Final status after merged PR #44 (4bdcc86): #28 acceptance is complete.

What is now covered:

real local 1x Codex TUI actor dogfood run without --dry-run;
exact project-root trust preflight for the TUI actor;
terminal startup responses, sanitized transcript capture, nonce-bearing final verdict markers, and marker-based actor shutdown;
active-run TUI Observer snapshots while the actor is still running;
final TUI Observer data refresh with transcript/verdict artifacts;
completed bundle verification;
1-4 lane noninteractive codex-exec fanout from prior PR Add Codex exec fanout #39, which satisfies the “prove fanout or split after 1x” acceptance path.

PR #44 proof:

CI passed on Node 22.14.0 and Node 24.
Local pnpm check passed with 54 tests, typecheck, and build.
pnpm public-surface:scan passed: 72 candidate text files checked.
pnpm mimetic -- doctor --json passed.

Real env-backed TUI live-follow proof, run id codex-tui-live-follow-final-proof:

Active snapshot while actor was running:
- review verdict contract_proof_only;
- stream/completion status running;
- lifecycle and events included actor.running;
- observer/observer-data.json stream status running;
- latest.json pointed to the valid active run;
- transcript artifact absent while active.
Final result:
- run --actor codex-tui --sims 1 --timeout-ms 240000 returned ok true;
- verify --run codex-tui-live-follow-final-proof passed 5/5 checks;
- watch --run codex-tui-live-follow-final-proof --detach --json --no-open returned ok true;
- sanitized metadata: review verdict pass, stream status passed, Observer stream status passed, reason actor reported passed verdict marker, duration 44256ms, transcript bytes 193232, nonce marker passed=true / blocked=false.

Remaining boundaries are now documented as future scope, not #28 blockers: raw interactive terminal streaming and any TUI-specific fanout beyond the current 1x TUI plus 1-4x codex-exec split.

No target OSS repos, GitHub issues, or production/private artifacts were mutated by this proof.

Navigation Menu

Repository navigation

live: Codex TUI actor fanout and live observer dogfood #28

Goal

Current state

Scope to spec first

Public-safety requirements

Acceptance proof

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

live: Codex TUI actor fanout and live observer dogfood #28

Description

Goal

Current state

Scope to spec first

Public-safety requirements

Acceptance proof

Activity

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

danielgwilson commented on Jun 2, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions