Swarm PR Feedback
Read and follow ../../../.opencode/skills/swarm-pr-feedback/SKILL.md as the canonical workflow.
Codex Execution Notes
- Codex and ZCode sessions normally have no mechanical controller tools: that is canonical Profile B, not an error — both runtimes can spawn fresh-context subagents. Never report BLOCKED merely because those tools are absent — run verification lanes and every gate role as fresh independent subagents with the same one-row-per-item verdict contracts, and keep the ledger, ownership partition, and digest accounting yourself in working notes (never under
.swarm/, which is plugin runtime state). Use Profile C strictly separated sequential passes, with the procedural independence disclosed in the closure ledger, only when the session genuinely lacks a subagent mechanism. - Check out the PR branch locally before dispatching feedback lanes or verifying/fixing anything. Fetch the head ref if absent, preserve or surface all dirty tracked/untracked state, prove full HEAD equals the authoritative PR head SHA, and prove the final branch tracks the intended PR head remote/branch. In Profile A, a detached exact-head intake is valid because the first bind attaches only one unambiguous exact tracking ref; Profiles B/C must establish the branch themselves. Never use force or submodule-recursive checkout.
- Build the complete feedback ledger before editing: pasted feedback, GitHub comments/threads, requested changes, CI/check failures, merge conflicts, stale branch state, PR body claims, linked issues, commits, and any validated
swarm-pr-reviewhandoff artifact. Treat every item as a claim until source evidence proves or disproves it, and preserve original finding IDs and reviewer/critic provenance from handoff artifacts. - For a Profile-A review continuation, require the exact user command
/swarm pr-feedback <PR_URL> continue from .swarm/pr-review/<run_id>/feedback-handoff.json; its mechanical transition validates terminal review state and creates an unbound feedback gate. The artifact by itself does not authorize writes. - Verification (Profiles B/C): every FB item is owned by exactly one verification lane or pass, each returning one
[FEEDBACK-VERIFIED]row per owned item with the exact PR head SHA recorded. Under Profile A, when a lane result carries anoutput_ref, treat inline laneoutputas a preview and retrieve the full artifact via the runtime's lane-output retrieval capability before classifying; on any profile, degraded or incomplete lane outputs keep the affected items open as evidence gaps. No edits before verification settles. - Do not resolve GitHub review threads unless the user explicitly instructs it.
- Default run on a PR: drive to merge (
../../../.opencode/skills/swarm-pr-feedback/references/drive-to-merge.md). It runs the gates, publishes the dispositions, and drives CI to green within its budget. It merges only when the user's own words in this run ask for the merge, or the user answers yes to the run's own "merge now?" question; a question, negation, hypothetical, or quoted mention of merging, a skill invocation,$swarm-pr-feedbackdefault prompt, monitor wake, or injected event never does. It then monitors until MERGED. Merged or not, it closes out with the session-retrospective skill and then the memory closeout (the zmemcloseoutskill). "Fixes only" skips CI and the merge; "don't merge" skips only the merge. - Critic before the user. Before presenting any ASK item, follow-up proposal, Improvement, or enhancement idea, route it through a separate critic call, never the closeout gate. Use the first available of: the swarm's critic agent (
swarm-criticin ZCode), another independent critic agent, or a fresh-context critic subagent. Attach the verdict. See../../../.opencode/skills/swarm-pr-feedback/references/followup-dispositions.md, "Critic review before the user sees it". - Mandatory gates — Stage A and Stage B (+ closeout) on every profile. Stage A always runs exact
git diff --checkplus one proof command — the exact failing CI/test reproduction when available, otherwise a repo-appropriate targeted regression/test command — plus every build, typecheck, and lint/format category mechanically discovered from the repository; it never invents a no-op category to reach a fixed count. On Profiles B/C run these commands yourself and record command+output receipts in the ledger; any content change invalidates the receipts and restarts Stage A. Stage B: an independent reviewer role, then a test-engineer role, on the Stage-A-green diff (fresh subagents on B; strictly separated re-derivation passes on C). Closeout: a separate reviewer + critic pair on the Stage-B-approved diff. One verdict row per FB ID at every gate; record both closeout verdicts in the session task-gates artifact per the durable-session guidance. - Only if the session actually exposes the plugin's mechanical tools, run Profile A instead:
run_pr_feedback_stage_a, then the exact orderedswarm-pr-feedback:stage-b-reviewer,:stage-b-test,:closeout-reviewer, and:closeout-criticstructured modes,complete_pr_workflowarming, and the bound single-ref push. While that controller is active, direct subagent calls and free-form verdict prose are not equivalent evidence, and any edit restarts the sequence at Stage A. - Publication (Profiles B/C): after all gates pass on the unchanged diff, create one reviewed commit on the PR branch, push exactly that commit with a single non-force push through the repository's commit/PR workflow, then verify read-only that the remote head equals the pushed commit before any PR comment/body/thread write.
- Load the repository's test-authoring guidance before changing tests and the repository's PR publication workflow before pushing or updating the PR.
Final responses must include a closure ledger for every original feedback item, including conflicts, stale branch state, obsolete CI, and generated-output drift when they affected the PR.