feat: replace coding-agent hero demo

[ThomasK33]

Summary

• Replace the recursive coding-agent README demo with an external Outer Camera Hero Demo for real Codex and Claude Code TUIs.
• Add the TypeScript Hero Demo generator, deterministic unit coverage, pinned demo tools, and a named mise regeneration task.
• Promote the regenerated curated artifact set with 16:9 outer WebMs/thumbnails/transcripts, inner agent-tty cast/WebM proofs, prompts, final file proofs, and manifest hashes.
• Use Codex gpt-5.5 and Claude claude-opus-4-7 defaults for the promoted demo, with CLI/env overrides for maintainers.
• Update README, dogfood catalog, domain glossary, ADR, and PRD language to document the narrowed Hero Claim Boundary.

User-facing / automation-facing behavior

• README now links the new dogfood/agent-uses-agent-tty/ Hero Demo bundle.
• mise run demo:agent-uses-agent-tty -- --agent both --runs 3 --record-seconds 180 --promote regenerates and promotes the real-agent demo when local Codex/Claude auth is available.
• Ordinary CI remains credential-free and does not install live-demo-only recorder tools; the demo task invokes task-pinned VHS/ttyd/ffmpeg only when run manually.

Validation

• mise run format-check
• mise run lint
• mise run typecheck
• npx vitest run test/unit/tools/hero-demo.test.ts
• npm run validate-bundle:canonical
• custom text leak scan over promoted text artifacts
• file dogfood/agent-uses-agent-tty/artifacts/*-thumbnail.png dogfood/agent-uses-agent-tty/artifacts/*-outer.webm dogfood/agent-uses-agent-tty/artifacts/*-inner-nvim.webm
• mise run build
• mise run install-smoke
• GIT_CONFIG_GLOBAL=/dev/null mise run test

Note: plain mise run test in this workspace hits the user's global Git SSH signing config, which points at a missing 1Password signing binary. The full test suite passed with GIT_CONFIG_GLOBAL=/dev/null.

Dogfood / visual artifacts

• Bundle entrypoint: dogfood/agent-uses-agent-tty/README.md
• Manifest: dogfood/agent-uses-agent-tty/manifest.json
• Promotion summary: dogfood/agent-uses-agent-tty/promoted-run-summary.md
• Outer videos: dogfood/agent-uses-agent-tty/artifacts/codex-outer.webm, dogfood/agent-uses-agent-tty/artifacts/claude-outer.webm
• 16:9 thumbnails: dogfood/agent-uses-agent-tty/artifacts/codex-thumbnail.png, dogfood/agent-uses-agent-tty/artifacts/claude-thumbnail.png
• Inner product proofs: dogfood/agent-uses-agent-tty/artifacts/*-inner-nvim.cast, dogfood/agent-uses-agent-tty/artifacts/*-inner-nvim.webm, dogfood/agent-uses-agent-tty/artifacts/*-final-file-proof.txt

Design-doc deviations

None. The implementation follows the PRD/ADR direction: VHS is the external Outer Camera, agent-tty remains responsible for the inner proof artifacts, and the old recursive README bundle is removed.

---

📋 Implementation Plan

PRD: External Outer Camera for the Coding-Agent Hero Demo

Problem Statement

The current coding-agent demo is valuable self-dogfood, but it is too unstable and visually noisy for the README-facing story. It asks agent-tty to record an outer Codex or Claude Code TUI while that real coding agent creates an inner agent-tty session and exports proof artifacts. This recursive setup couples too many unstable layers: live coding-agent UI behavior, trust prompts, update notices, nested terminal sessions, renderer timing, exit behavior, thumbnail timing, and recording post-processing.

As a maintainer or reviewer, I want the README demo to be a polished, stable Hero Demo that shows real coding-agent TUIs using agent-tty to produce inner proof artifacts. I do not need the README visual to prove that agent-tty recorded the outer coding-agent TUI. That stronger self-dogfood claim makes the demo harder to regenerate and harder to review.

Solution

Replace the recursive README-facing demo with a Promoted Hero Demo that uses an external Outer Camera for the visible Codex and Claude Code TUIs, while keeping agent-tty responsible for the inner terminal proof artifacts.

The new flow will use a Node/TypeScript Hero Demo Generator. The generator prepares isolated inputs for Manual Demo Regeneration, creates runner scripts and raw VHS tapes, invokes VHS as the Outer Camera, validates the resulting Curated Hero Artifact Set, runs automated text leak checks, records tool/model settings, and writes a Promoted Hero Run Summary.

The scenario is an Exploratory Hero Demo: the real coding-agent TUI loads the packaged agent-tty skill, inspects the CLI as needed, chooses its own command flow, drives an inner Neovim workflow through agent-tty, and exports inner proof artifacts. The prompt supplies success criteria, required output paths, final text, and a configurable fixed review window, but no prewritten helper script or exact command sequence. The generated VHS tape owns startup waits, the configurable fixed review window, and exit keypresses during recording. Raw VHS tapes, recorder logs, and disposable workspaces are Debug-Only Raw Demo Files and are ignored by default.

The existing recursive bundle is removed once the new Hero Demo is promoted. The README claim is narrowed through the Hero Claim Boundary: the outer TUI is presentation, and the inner agent-tty artifacts are the product proof.

User Stories

1. As a README reader, I want to see a polished recording of Codex using agent-tty, so that I quickly understand the coding-agent workflow.
2. As a README reader, I want to see a polished recording of Claude Code using agent-tty, so that I understand the workflow is not tied to one coding-agent CLI.
3. As a README reader, I want the demo recording to be visually stable, so that startup waits and nested recorder noise do not distract from the product story.
4. As a README reader, I want a clickable PNG thumbnail for each outer recording, so that I can preview the coding-agent TUI before opening the WebM.
5. As a README reader, I want the outer WebM to show the real coding-agent TUI, so that the demo feels authentic rather than simulated.
6. As a README reader, I want the demo copy to clearly state that agent-tty produced the inner proof artifacts, so that I understand what the product is proving.
7. As a README reader, I want the demo copy not to imply that agent-tty recorded the outer coding-agent TUI, so that I do not misunderstand the proof boundary.
8. As a maintainer, I want a named Hero Demo Promotion Command, so that I can regenerate the demo without reverse-engineering several scripts.
9. As a maintainer, I want the regeneration command to support Codex-only runs, so that I can debug Codex behavior without paying the Claude Code cost.
10. As a maintainer, I want the regeneration command to support Claude-only runs, so that I can debug Claude Code behavior without paying the Codex cost.
11. As a maintainer, I want the regeneration command to support both agents in one promotion run, so that the promoted summary is coherent.
12. As a maintainer, I want the regeneration command to require three successful Codex runs and three successful Claude Code runs before promotion, so that a one-off smoke does not replace the README demo.
13. As a maintainer, I want a Hero Demo Partial Pass to fail promotion, so that the README does not end up with one fresh agent recording and one stale or missing recording.
14. As a maintainer, I want Codex and Claude Code model and effort defaults, so that routine regeneration has a simple happy path.
15. As a maintainer, I want model and effort overrides, so that I can adapt when coding-agent defaults or model availability changes.
16. As a maintainer, I want the resolved tool versions, model names, and effort settings recorded in the run summary, so that future reviewers can understand what produced the artifacts.
17. As a maintainer, I want VHS, ttyd, and ffmpeg pinned as repo tools for named demo tasks, so that regeneration is reproducible without forcing ordinary CI to run live-agent demos.
18. As a maintainer, I want ordinary CI not to require Codex or Claude Code credentials, so that the normal project validation remains portable.
19. As a maintainer, I want CI to keep validating checked-in bundle metadata and normal package behavior, so that the promoted artifacts do not drift silently.
20. As a maintainer, I want raw VHS tapes and recorder logs to remain ignored by default, so that the promoted bundle stays curated and low-noise.
21. As a maintainer, I want raw VHS tapes and recorder logs available locally after a failed run, so that I can debug prompt or UI drift.
22. As a maintainer, I want disposable workspaces to stay out of the curated artifact set, so that local paths and sensitive environment details do not leak.
23. As a maintainer, I want a strict Hero Demo Leak Check, so that account details, auth warnings, tokens, and absolute home paths do not get checked in.
24. As a maintainer, I want automated text scanning over transcripts, logs, summaries, and generated text artifacts, so that obvious leakage is caught before visual review.
25. As a maintainer, I want human visual review of PNG and WebM outputs, so that visual-only leakage and poor framing are caught before promotion.
26. As a maintainer, I want benign update or product notices to be allowed if they do not dominate the recording, so that live-agent artifacts do not require brittle post-processing.
27. As a maintainer, I want auth conflicts and billing/account lines to block promotion, so that checked-in visuals do not expose account-sensitive details.
28. As a maintainer, I want generated transcripts for the outer coding-agent TUIs, so that reviewers can audit what happened without watching the video frame-by-frame.
29. As a maintainer, I want inner agent-tty cast artifacts, so that reviewers can inspect terminal replay data from the product-generated proof.
30. As a maintainer, I want inner agent-tty WebM artifacts, so that reviewers can visually review the inner terminal workflow.
31. As a maintainer, I want final file proof from the inner workflow, so that the demo proves the agent-controlled TUI did the intended work.
32. As a maintainer, I want the prompt checked in with the promoted artifact set, so that reviewers can see exactly what the real coding agent was asked to do.
33. As a maintainer, I want a summary checked in with the promoted artifact set, so that the promotion evidence is easy to find.
34. As a maintainer, I want only one selected promoted run per agent checked in, so that the bundle does not grow with all trial outputs.
35. As a maintainer, I want the summary to prove that three runs per agent passed, so that the reliability bar remains visible even when only selected artifacts are promoted.
36. As a maintainer, I want the canonical manifest to include sha256 and byte counts for promoted artifacts, so that artifact drift is detected.
37. As a maintainer, I want the canonical manifest to exclude debug-only raw files, so that validation focuses on reviewer-facing outputs.
38. As a maintainer, I want the catalog and README to point to the new curated artifacts, so that reviewers do not follow obsolete recursive-flow links.
39. As a maintainer, I want the old recursive bundle deleted after replacement, so that there is one maintained coding-agent demo path.
40. As a maintainer, I want the deletion to be reflected in docs and catalog language, so that future contributors know the recursive proof was intentionally removed.
41. As a contributor, I want the domain glossary to distinguish Hero Demo, Outer Camera, Hero Demo Generator, and Recursive Dogfood Proof, so that implementation discussions use precise language.
42. As a contributor, I want the ADR to explain why the outer recording moved outside agent-tty, so that I do not reintroduce the recursive flow accidentally.
43. As a contributor, I want tests for the generator's planning and validation logic, so that refactors do not break artifact selection or leak checks.
44. As a contributor, I want tests that avoid live Codex or Claude Code calls, so that the generator's deterministic behavior can be verified in CI.
45. As a contributor, I want smoke/debug modes for the generator, so that I can iterate without running the full promotion bar.
46. As a contributor, I want promotion mode to fail loudly with actionable messages, so that missing artifacts, failed waits, or leak matches are easy to diagnose.
47. As a reviewer, I want the promoted artifacts to be small enough to review in GitHub, so that the demo remains practical to inspect.
48. As a reviewer, I want WebM to remain the primary video format, so that the new demo matches the current README pattern.
49. As a reviewer, I want PNG thumbnails to be captured from meaningful TUI states, so that the README preview shows the real coding-agent UI after useful output appears.
50. As a reviewer, I want the inner and outer artifacts to be clearly named by agent, so that I can compare Codex and Claude Code without guessing.
51. As a reviewer, I want the summary to say which artifacts were selected for promotion, so that I know which of the successful runs became the README-facing artifacts.
52. As a reviewer, I want the summary to say which successful trial outputs stayed debug-only, so that I understand the promotion evidence without checking in every file.
53. As an automation maintainer, I want named demo tasks to stay outside ordinary CI, so that the repository's default validation remains fast and credential-free.
54. As an automation maintainer, I want canonical bundle validation to continue covering the promoted demo, so that a checked-in artifact cannot change without manifest updates.
55. As an automation maintainer, I want the generator to use isolated homes and temporary workspaces, so that regeneration does not mutate the maintainer's real agent-tty home or editor state.
56. As an automation maintainer, I want cleanup to avoid deleting live agent-tty sessions it does not own, so that demo regeneration remains safe.
57. As an automation maintainer, I want failures to preserve debug-only raw files locally, so that a maintainer can inspect the exact VHS tape, transcript, and logs.
58. As an automation maintainer, I want successful promotion to copy only curated artifacts into the canonical bundle, so that sensitive debug files do not slip into review.
59. As a product maintainer, I want the demo to preserve authenticity by using real Codex and Claude Code, so that the README demonstrates actual coding-agent TUIs rather than fixtures.
60. As a product maintainer, I want the option to evolve to an active PTY proxy controller later if raw VHS tapes flake, so that the current design does not foreclose a more robust control layer.

Implementation Decisions

• Build a Node/TypeScript Hero Demo Generator as the deep module that owns setup, raw tape generation, runner generation, invocation, artifact validation, leak checks, promotion selection, and summary generation.
• Keep the generated VHS tape as the recorder-time control layer. It owns screen waits and keypresses while VHS records the outer coding-agent TUI.
• Keep VHS as the Outer Camera. It records the visible Codex and Claude Code TUIs and emits WebM, PNG screenshot/thumbnail, and transcript-like output.
• Keep the scenario as an Exploratory Hero Demo. The real agent discovers the agent-tty skill and CLI, chooses its own command flow, drives an inner Neovim workflow through agent-tty, and exports the inner proof artifacts to required paths.
• Reuse the existing canonical coding-agent demo bundle identity and replace its contents wholesale with the Promoted Hero Demo.
• Delete the old recursive agent-tty-records-the-agent flow after the new Hero Demo is promoted.
• Narrow the README claim through the Hero Claim Boundary: the outer TUI is presentation, and the inner agent-tty artifact set is the product proof.
• Expose regeneration through a named mise task that delegates to the generator.
• Pin VHS, ttyd, and ffmpeg in the named demo task, but do not install live-demo-only recorder tools or run live-agent regeneration in ordinary CI.
• Provide smoke/debug modes for one-off agent runs and a promotion mode that requires the full Hero Demo Promotion Bar.
• Require three successful local regenerations for Codex and three successful local regenerations for Claude Code before promotion.
• Treat a Hero Demo Partial Pass as a failed promotion. Both agents must pass before README, catalog, manifest, or promoted artifacts are replaced.
• Check in one selected promoted run per agent plus a Promoted Hero Run Summary. Keep extra successful trial outputs ignored as Debug-Only Raw Demo Files.
• Keep WebM as the primary outer video format and PNG as the README thumbnail format.
• Maintain a Curated Hero Artifact Set containing outer WebM, outer PNG thumbnail/screenshot, outer transcript, inner agent-tty cast, inner agent-tty WebM, final file proof, prompt, summary, and canonical manifest entries.
• Keep raw generated VHS tapes, recorder logs, and disposable workspaces out of the curated set by default.
• Maintain canonical manifest validation with sha256 and byte counts for all promoted artifacts.
• Record resolved Codex and Claude Code versions, models, effort levels, tool versions, and relevant generator settings in the promoted run summary.
• Provide default model and effort settings for Codex and Claude Code, with CLI or environment overrides for maintainers.
• Apply strict account-scrub policy during promotion: names, emails, billing/account lines, auth warnings, tokens, and absolute home paths block promotion.
• Allow generic update or product notices only when they do not dominate the recording.
• Use automated text scanning for transcripts, generated text artifacts, summaries, and logs that are candidates for promotion.
• Require human visual review of PNG and WebM outputs before promotion.
• Preserve debug-only raw files after failures so maintainers can inspect UI drift, wait failures, or recorder issues.
• Do not publish a new public CLI JSON contract for this feature. The work is a maintainer-facing demo-generation workflow and checked-in artifact replacement.
• Do not require ordinary users to install VHS, ttyd, or ffmpeg to use agent-tty. Those tools are for Manual Demo Regeneration.

Testing Decisions

• Tests should focus on external behavior of the Hero Demo Generator rather than implementation details.
• The generator should expose testable planning and validation units so most coverage can run without live Codex or Claude Code credentials.
• The tape-generation behavior should be tested by asserting generated tape intent from stable inputs: selected agent, dimensions, output names, startup waits, configurable fixed review window, and cleanup actions.
• The runner-generation behavior should be tested by asserting that agent-specific commands include expected model/effort settings, prompt wiring, workspace isolation, and environment variables for required artifact paths.
• The artifact validation behavior should be tested with fixture files that simulate complete, missing, empty, and mismatched curated artifact sets.
• The promotion selection behavior should be tested with fixture run records for full pass, partial pass, insufficient run count, and selected-run output.
• The Hero Demo Leak Check should be tested with fixture transcripts, summaries, and logs containing allowed generic notices and disallowed account-sensitive patterns.
• The summary-generation behavior should be tested from structured run results so it records tool versions, model settings, selected artifacts, run counts, and promotion outcome.
• The canonical manifest generation or update logic should be tested with fixture artifacts to verify sha256 and byte counts match on-disk content.
• The named demo task should have a lightweight smoke check that validates the command shape without requiring live-agent credentials.
• Live Codex and Claude Code regeneration should remain manual dogfooding, not CI-required tests.
• Manual dogfooding should run the promotion command for both agents with three runs each, then inspect WebM, PNG, transcript, inner cast/WebM, file proof, prompt, summary, and manifest.
• Manual visual review should verify that no names, emails, account/billing lines, auth warnings, tokens, or absolute home paths appear in promoted visuals.
• Manual visual review should verify that benign update/product notices, if present, do not dominate the thumbnail or video.
• Existing canonical bundle validation should remain the prior art for manifest completeness, sha256, and byte-count checks.
• Existing CLI and artifact tests remain prior art for validating JSON envelopes, artifact existence, and isolated homes, but the new generator should avoid depending on live CLIs in automated tests.
• A good test should treat VHS invocation as an external boundary. Unit tests should not assert private helper ordering when an externally visible generated tape, summary, or validation result is the behavior under test.
• A good test should verify failure messages are actionable for missing tools, missing auth, missing artifacts, failed leak checks, partial pass, and manifest mismatch.

Out of Scope

• Keeping the recursive outer agent-tty coding-agent recording as a maintained parallel proof path.
• Claiming that the README Hero Demo proves agent-tty can record the outer coding-agent TUI.
• Running live Codex or Claude Code regeneration in ordinary CI.
• Building an active PTY proxy controller in the first implementation.
• Switching the Hero Demo to fixture or mock coding-agent TUIs.
• Using asciinema/agg as the primary outer recording pipeline in this refactor.
• Publishing raw VHS tapes, recorder logs, or disposable workspaces as reviewer-facing artifacts by default.
• Changing public agent-tty CLI JSON contracts, protocol schemas, or artifact formats for end users.
• Adding new product features to agent-tty recording/export behavior beyond what the demo generator needs.
• Post-processing visuals with blur/crop/redaction as the default approach to secrecy or polish.
• Supporting arbitrary coding-agent CLIs beyond Codex and Claude Code in the initial promoted demo.
• Guaranteeing deterministic model output from live coding-agent services.

Further Notes

The VHS prototype passed a one-run-per-agent smoke: VHS launched real Codex and Claude Code TUIs, waited on screen regexes, captured PNG screenshots, and rendered compact WebM/ASCII evidence. That prototype proved feasibility, not promotion-level reliability. The promotion bar remains three successful local regenerations per agent plus automated leak checks and human visual review.

The implementation should keep the current ADR and glossary terminology aligned with the resulting code and docs. In particular, use Hero Demo Generator rather than “controller”, and use Exploratory Hero Demo rather than “Nested Helper Proof” unless the implementation intentionally returns to a deterministic helper design.

If raw VHS tapes fail the promotion bar due to optional prompts or UI drift, the design can evolve to an active PTY proxy controller while preserving the same Hero Claim Boundary and curated artifact policy.

---

Generated with mux • Model: openai:gpt-5.5 • Thinking: xhigh

[ThomasK33]

/coder-agents-review

[coder-agents-review]

This is a first-pass (Netero) mechanical review only. The full review panel has not yet reviewed this PR. The panel will review after these findings are addressed.

The change replaces the recursive coding-agent demo with a VHS-based Hero Demo Generator, adds deterministic unit tests, pins demo tools, and promotes a curated artifact set. The artifact structure is well-organized, the leak-check pipeline is thoughtful, and the test coverage on the deterministic boundary is solid.

Severity breakdown: 1 P2, 2 P3, 1 Nit, 1 Note.

The P2 is the main item: RunRecord.passed promises graceful partial-failure handling, but the implementation hard-aborts on any run failure, making the field and its downstream filters dead logic. Either make partial failure real (try/catch in the run loop) or remove the dead field.

"Every RunRecord that reaches the records array has passed: true. The filter can never exclude anything. The summary renderer formats 'pass'/'fail' per record, but 'fail' is unreachable." (Netero)

🤖 This review was automatically generated with Coder Agents.

[ThomasK33]

/coder-agents-review

[ThomasK33]

/coder-agents-review

[coder-agents-review]

AGTTY-1 (P2, RunRecord.passed dead logic) was addressed in 3d3958a with try/catch in the run loop and a failedRunRecord() factory. Good fix.

However, further review is blocked until the author responds to or pushes fixes for the remaining findings:

• AGTTY-2 (P3) Local invariant at line 97 duplicates src/util/assert.ts export. Still present, no response.
• AGTTY-3 (P3) mkDebugRoot unused _bundleDir parameter at line 1061. Still present, no response.
• AGTTY-4 (Nit) CuratedArtifact and RunRecord exported with no external consumers. Still present, no response.
• AGTTY-5 (Note) Post-promotion leak check tautological for pre-sanitized text artifacts. No response.

For each: fix in this PR, or respond on the thread explaining why the current code is acceptable. The full review panel will run once all findings have a response.

🤖 This review was automatically generated with Coder Agents.

[coder-agents-review]

First full panel review (11 reviewers). All five prior findings (AGTTY-1 through AGTTY-5) are resolved. The architecture is sound: VHS as the outer camera, the claim boundary is well-documented, and the generator's pure/IO split is clean.

Severity breakdown: 1 P2, 8 P3, 3 Nit (Nits not posted as inline comments).

The P2 is a sanitizer/leak-checker regex divergence that will cause spurious promotion failures with real agent greetings. The P3s cover test gaps, magic numbers, usage string errors, type duplication, and missing schema validation.

"The sanitizer demands a trailing ! that the leak detector does not. Real agent greetings without ! survive sanitization, then the leak check flags them, and promotion aborts for a 'leak' the sanitizer was supposed to have removed." (Hisoka)

Nits not posted inline: AGTTY-6 (safeDemoToolVersion duplicates safeToolVersion), AGTTY-16 (no doc comments on exported functions), AGTTY-17 (duplicate mise.toml comment about mise install --locked), AGTTY-18 (sentence vs expectedText naming inconsistency).

🤖 This review was automatically generated with Coder Agents.

[ThomasK33]

/coder-agents-review

[coder-agents-review]

Second panel round (8 reviewers). All 9 prior findings (AGTTY-7 through AGTTY-15) independently verified fixed by 4+ reviewers. The sanitizer regex fix, named constants, schema validation, type deduplication, and usage string correction are all solid.

No P0-P2 findings remain. 4 new P3 observations noted inline for future improvement; none block merge.

The architecture is clean: VHS as outer camera, agent-tty as inner proof, hero claim boundary well-documented. The generator's pure/IO split is thoughtful, and the deterministic test boundary covers all planning and validation logic. The fix-chain quality across rounds has been strong: each commit targeted root causes, no regressions introduced.

"Non-promote runs never clean debug roots. If you run 50 debug iterations, /tmp grows by 50 disposable workspaces. The flag exists, but it only fires in the code path where you'd want to keep them." (Hisoka)

🤖 This review was automatically generated with Coder Agents.

[ThomasK33]

/coder-agents-review

[coder-agents-review]

Third panel round (5 reviewers). All 4 prior findings (AGTTY-19 through AGTTY-22) verified fixed by 3+ reviewers. No P0-P2 findings remain. No new P3+ from the panel.

One Netero observation (AGTTY-25): quote() has a latent shell-quoting bug for inputs containing single quotes. No current caller triggers it (all callers pass paths and model names without single quotes), so it's noted inline but does not block.

The PR has gone through 5 rounds across 25 findings. 17 fixed by author, 1 acknowledged, 6 dropped by orchestrator (Nits/Notes below posting threshold), 1 latent observation noted. Fix quality has been consistently strong throughout.

🤖 This review was automatically generated with Coder Agents.

[ThomasK33]

/coder-agents-review

[coder-agents-review]

Fourth panel round (5 reviewers). AGTTY-25 (quote() shell-quoting bug) verified fixed in f208273. Netero and 4 of 5 panel reviewers found no new issues. Bisky flagged 2 P3 test coverage observations inline; neither blocks merge.

All 25 prior findings resolved across 6 rounds (18 fixed, 1 acknowledged, 6 dropped). The fix chain has been clean throughout, with each commit targeting root causes and no regressions introduced.

The two new P3s are about test assertions that don't cover key behavioral lines: the tape's runner invocation and the runner's artifact-path env vars. Both gaps mean you could delete the critical lines and tests would still pass. Worth adding, but the live-demo workflow catches both immediately.

🤖 This review was automatically generated with Coder Agents.

[ThomasK33]

/coder-agents-review

[coder-agents-review]

Fifth panel round (5 reviewers). AGTTY-26/27 (test assertion gaps) verified fixed in 2cc286b. Netero and all 5 panel reviewers found zero new findings.

All 27 findings across 7 rounds are resolved: 20 fixed by author, 1 acknowledged, 6 dropped by orchestrator. No open items remain. Fourth consecutive APPROVE.

🤖 This review was automatically generated with Coder Agents.