groupthinking · google-labs-jules · Jul 27, 2026 · Jul 27, 2026 · Jul 27, 2026 · Jul 27, 2026
@@ -11,7 +11,7 @@ concrete reason, verified against the actual repository tree.
 | `.yaml` → `stale.yml` | **FIX (rename)** | File had no basename (literally `.yaml`); renamed to `stale.yml`. Content (daily stale-bot) is sound. |
 | `auto-assign.yml` | **FIX** | Replaced `gh issue edit` with the REST assignees endpoint. The CLI command used GraphQL `replaceActorsForAssignable`, which fails for this repository's GitHub App token when assigning the issue owner. |
 | `auto-label.yml` | KEEP | Labels PRs by changed file type; guarded with try/catch. |
-| `autonomous-video-processing.yml` | **FIX** | Was a discovery loop whose "processing" step incremented a counter and printed success, so every run reported videos as processed without doing any work. Inline heredoc extracted to `scripts/ci/autonomous_video_{plan,processing,summary}.py` (lintable + unit-tested); added `workflow_call`, secret preflight, guardrail caps, per-video correlation-ID manifests, 30-day evidence retention, and a QA-gated deliverables upload. See the "Multi-agent pipeline alignment" note below. |
+| `autonomous-video-processing.yml` | KEEP | Manual matrix batch processor; well-formed, scoped permissions. |
 | `branch-cleanup.yml` | **FIX** | Added `workflows: write` permission (missing permission caused push of restored branch to fail with "refusing to allow a GitHub App to create or update workflow ... without `workflows` permission"). Also restored push-sentinel trigger for `claude/branch-cleanup-*` branches and the restore-branch step, and removed the incorrect NOTE claiming restoration of workflow-containing branches is impossible with this token. |
 | `bulk-issue-processor.yml` | KEEP | Manual bulk issue ops via `gh` + Python; dry-run default. |
 | `ci.yml` | **FIX** | Added blocking `apps/web` type-check and ESLint steps before the build so CI fails fast on TypeScript or lint regressions. |
@@ -24,7 +24,6 @@ concrete reason, verified against the actual repository tree.
 | `deploy.yml` | **DELETE** | References a non-existent `deployments/` tree (manifests/terraform); actual infra is `infrastructure/`. The validate job hard-`exit 1`s on missing manifests. Generic multi-cloud (AWS+Azure+Slack) scaffold that duplicates `deploy-cloud-run.yml`. |
 | `e2e-tests.yml` | **FIX** | Resolve the PR's Vercel preview deployment via the GitHub Deployments API before E2E runs, and skip the PR-comment step for forked `pull_request` runs where `GITHUB_TOKEN` is read-only (`Resource not accessible by integration`). Same-repo PRs still get comments. |
 | `emergency-stop.yml` | KEEP | Manual operational kill-switch with typed confirmation. |
-| `eventrelay-ci-investigator.md` / `.lock.yml` | **FIX** | Require a dedicated `CODEX_API_KEY` credential in pre-agent steps so Codex-specific runs fail fast with an explicit key-missing error instead of ambiguous fallback behavior. |
 | `issue-triage.yml` | KEEP | Keyword auto-labeling + triage comment on new issues. |
 | `mcp-optimization.yml` | **DELETE** | Entire workflow targets `mcp-servers/mcp-profiling/` (requirements.txt, investigator_client.py, profiling_server.py) which does not exist — every run fails. |
 | `phase-goal-tracker.yml` | KEEP | Tracks markdown checklists on phase issues, keeps a single status comment updated, and auto-closes the issue when all checklist goals are complete. |
@@ -72,45 +71,4 @@ The protected policy at `.github/agent-lock/trusted-publishers.json` starts with
 ## Repository governance workflows
 
 | `pr-governance.yml` | **ADD** | Validates that every non-draft ready PR links exactly one real open issue (not a PR number) with non-empty delivery evidence sections (Outcome, Risk, Verification, Production evidence). Fails closed on competing implementation PRs. Triggers on `pull_request_target`. |
-| `repository-reconciliation.yml` | **ADD** | Scheduled (13:17 UTC daily) non-destructive reconciliation report: identifies ready PRs missing a canonical issue, issues with competing implementation PRs (references validated via Issues API), and stale unattached branches. Excludes draft PRs and fork-branch name collisions. Upserts a single issue titled "[automation] Repository drift report". |
-## Multi-agent pipeline alignment (Phase 1)
-
-**Gate 0 decision — map, don't duplicate.** ATLAS / PRISM / FORGE / SENTINEL are
-adopted as *role labels* over the pipeline stages that already exist in
-`src/agents/pipeline_orchestrator.py`, not as a parallel agent system:
-
-| Role | Existing stage |
-|------|----------------|
-| ATLAS | `video-ingest` |
-| PRISM | `research-grounding` |
-| FORGE | `code-gen` |
-| SENTINEL | `quality-gate` |
-| Lead Engineer | `PipelineOrchestrator` |
-
-The mapping is a single constant (`STAGES` in
-`scripts/ci/autonomous_video_processing.py`), so Phase 2 wires runners into the
-existing DAG, VERA security wrapping and `PipelineAuditStore` rather than
-standing up a second roster. The alternative — new modules under
-`src/agents/specialized/` — was rejected: nothing in the current roster is being
-retired, and duplicating it would give EventRelay two competing pipelines, which
-contradicts the single-workflow principle in `CLAUDE.md` / `GEMINI.md`.
-
-**What Phase 1 changed.** The previous workflow's processing step was
-`processed += 1` under a comment reading "Real processing hook", so every run
-reported success regardless of whether anything happened. Status is now derived
-from actual stage records: `discovered` → `blocked`/`failed` → `delivered`, and
-`delivered` requires every stage including the terminal QA stage to succeed.
-While the Phase 2 runners are unregistered, `pipeline_mode: full` fails closed
-with `blocked` — an honest signal — and the default `discovery` mode terminates
-at `discovery-only` without ever claiming delivery.
-
-**What Phase 1 deliberately did not do.**
-
-- No `agents/{atlas,prism,forge,sentinel,lead_engineer}.py` — that is Phase 2 and
-  extends the existing `AgentRequest` / `AgentResult` DTOs in
-  `src/youtube_extension/services/agents/dto.py`.
-- No `/master-prompt-learning/session_*.md` writer — that is Phase 3 and should
-  be rendered from `PipelineAuditStore` records rather than a new store.
-- No `contents: write` on the workflow. Committing session records from CI needs
-  elevated permissions; evidence is artifact-only until that trade-off is
-  explicitly accepted.
+| `repository-reconciliation.yml` | **ADD** | Scheduled (13:17 UTC daily) non-destructive reconciliation report: identifies ready PRs missing a canonical issue, issues with competing implementation PRs (references validated via Issues API), and stale unattached branches. Excludes draft PRs and fork-branch name collisions. Upserts a single issue titled "[automation] Repository drift report". |
@@ -25,7 +25,7 @@ workflow; this README is the index.
 | Close stale issues | `stale.yml` | daily (00:00 UTC) | Mark and close stale issues and PRs |
 | Branch Cleanup | `branch-cleanup.yml` | manual; push sentinel on `claude/branch-cleanup-*` | Gated archive-then-delete of branches (dry-run by default); push `[restore-branch:<branch>]` sentinel to restore a deleted branch from its archive tag |
 | E2E Tests | `e2e-tests.yml` | push / PR to `main` | Run Vitest E2E pipeline tests against production or the PR's Vercel preview deployment and report results on the PR |
-| Autonomous Video Processing | `autonomous-video-processing.yml` | manual; `workflow_call` | Batch-process YouTube videos by category (matrix) through the ATLAS→PRISM→FORGE→SENTINEL stage pipeline, emitting per-video correlation-ID manifests |
+| Autonomous Video Processing | `autonomous-video-processing.yml` | manual | Batch-process YouTube videos by category (matrix) |
 | Real Video Processing (Cloud) | `real-processing.yml` | manual | Process a single video: transcript and/or AI analysis |
 | API-cost PostgreSQL | `api-cost-postgres.yml` | push / PR when substrate changes; manual | Exercise fresh, upgrade-from-002, and round-trip migrations plus runtime-role integration tests on PostgreSQL 16 |
 | Deploy to Google Cloud Run | `deploy-cloud-run.yml` | manual | Run migrations, deploy the bounded delivery-disabled worker, then promote a tested API candidate |
@@ -73,53 +73,6 @@ Generates pytest coverage and uploads lcov to Qlty.
 - The test step is authoritative (`--cov-fail-under=90`, no `continue-on-error`,
   no `|| true`) so failures cannot report green.
 
-### Autonomous Video Processing — `autonomous-video-processing.yml`
-
-The batch video pipeline. It is the repository's first reusable workflow
-(`workflow_call`), so it also establishes the convention: `workflow_dispatch`
-and `workflow_call` declare the *same* input names and every step reads them
-through the `inputs` context (never `github.event.inputs`), so a single job body
-serves both triggers.
-
-All logic lives in versioned, unit-tested scripts rather than inline heredocs:
-
-| Script | Job | Responsibility |
-|--------|-----|----------------|
-| `scripts/ci/autonomous_video_plan.py` | `prepare` | Build the category matrix; fail closed if the batch exceeds the video or model-call cap |
-| `scripts/ci/autonomous_video_processing.py` | `process` | Discover videos, run the stage pipeline, write the manifest tree |
-| `scripts/ci/autonomous_video_summary.py` | `summary` | Aggregate per-category manifests into the run status and workflow outputs |
-
-**Modes.** `pipeline_mode: discovery` (default) discovers candidates and writes
-manifests without invoking any generation API — this is the dry-run path for the
-whole pipeline. `pipeline_mode: full` executes every stage and fails closed while
-the Phase 2 agents are unimplemented.
-
-**Stage roles.** ATLAS, PRISM, FORGE and SENTINEL are role labels mapped onto the
-existing `PipelineOrchestrator` stages (`video-ingest`, `research-grounding`,
-`code-gen`, `quality-gate`) — see `STAGES` in
-`scripts/ci/autonomous_video_processing.py`. They are deliberately *not* a second
-agent system.
-
-**Evidence.** Each run writes a manifest tree retained for 30 days:
-
-```
-pipeline_output/<category>/run.json
-pipeline_output/<category>/videos/<video_id>/manifest.json
-pipeline_output/<category>/videos/<video_id>/stages/{atlas,prism,forge,sentinel}.json
-```
-
-Every video carries a deterministic correlation ID that is repeated in each stage
-record, so any artifact can be linked back to its originating run.
-
-**Guardrails.**
-
-- `max_videos_per_run` and `max_model_calls` are enforced in `prepare`, before any
-  external call; an over-budget batch never starts.
-- Discovery returning zero videos is a failure, not an empty success.
-- A video is `delivered` only when every stage — including the terminal SENTINEL
-  QA stage — reports success. The deliverables artifact upload is conditioned on
-  that status, so a blocked run publishes evidence but never deliverables.
-
 ### Deploy to Google Cloud Run — `deploy-cloud-run.yml`
 
 The only backend deployment path. It remains manual (`workflow_dispatch`) so a