repo-build-pr

name

repo-build-pr

description

End-to-end implementation loop for the os-clovy repo: study the prompt, ask the clarifying questions that change what gets built, plan and architect on the session model while delegating bulk implementation and review to cheaper strong models, isolate work in git worktrees, validate with deterministic checks plus agent-driven live app walkthroughs, record and upload QA video evidence through os-platform, open a draft PR, run the automated Greptile and Codex review loop with judgment, and mark it ready. Use when the user invokes /repo-build-pr (or $repo-build-pr in Codex), or asks to build, implement, ship, or fix something in os-clovy from a feature prompt, bug report, screenshot, PR comment, or tracker task id, optionally with a cross-harness implementer directive like "/repo-build-pr JUN-200 with codex".

Repo build PR

Use this skill for the end-to-end implementation loop in open-software-network/os-clovy. The goal is not only to make code changes. The goal is to understand the prompt, isolate the work in worktrees, ship a coherent PR, and run the automated review loop with judgment.

To hand this entire workflow to a different harness as orchestrator (e.g. dispatch a whole build to Codex from Claude Code, or vice versa), use the repo-orchestrate skill instead — it wraps this workflow in a per-harness runner with a stop-before-publish default.

Skill map

This is the repo's entry-point skill; it orchestrates and defers. Each fact lives in exactly one owner — load the owner at the step that needs it, never restate its commands here:

• repo-review — pre-publish review battery: axes, per-harness runners, disposition triage, convergence loop.
• repo-delegate — single-brief implementation dispatch to another harness (the building block of the cross-harness implementer mode).
• repo-orchestrate — hand this whole workflow to another harness.
• agent-e2e-qa — live QA process: surface decision tree, video recording/compression, os-platform upload, PASS/FAIL/BLOCKED evidence format.
• browser-test-tauri-fe — browser-surface technique: fake Tauri IPC bridge, Playwright/CDP driving, screenshot suites, PR-embeddable GIFs.
• os-platform / os-task-prep — tracker reads, task diagnosis and enrichment (intake step 3).
• repo-retrospect — post-cycle: mine the session, PR threads, and user feedback for lessons and fold them into the owning skill files.

Intake

Treat everything after /repo-build-pr (or $repo-build-pr in Codex) as the build prompt. If the user did not use the literal command but asks to build, implement, ship, or fix something in the repo, use this skill anyway.

1. Read the prompt carefully and restate the concrete objective, constraints, and likely affected surface area. A trailing implementer directive (with codex in Claude Code, with claude in Codex) selects a cross-harness implementer — see Model orchestration.

2. Read repo instructions before editing:

   • AGENTS.md
   • any referenced project plan or spec relevant to the task

   A handoff doc, prior analysis, or the Issue's own "Agent diagnosis" is evidence, not a conclusion. Use its file:line citations to skip re-discovery; re-derive the root cause yourself. Ranked hypotheses in these documents are frequently wrong, and they are wrong in the direction of the mechanism nearest the symptom. Cheapest test: name the one observation the reported symptom forces to be true, then find the code path that produces it. PR #676 — the handoff ranked a clipboard-restore race first; the transcript's presence in the dictation tab proved the paste command had already been sent, so the bug lay past that point entirely (it was a focus-drift TOCTOU). PR #660 — the same shape, the symptom's obvious layer was not the cause.

3. If the prompt names a tracker task (a JUN-xxx id or an os-platform issue reference — issues live on os-platform, org june, per docs/agents/issue-tracker.md): fetch it via the os-platform skill and validate it is actually implementable against the codebase (root cause, affected files, acceptance criteria — apply the os-task-prep diagnosis discipline if the issue is thin, and update the issue with what you learn). Then take it before writing code: assign it to the current user and set status to in-progress through the documented platform API (append-only, probe-then-verify). Include Closes <TASK-ID> in the eventual PR body.

4. Inspect the current checkout with git status -sb. A dirty checkout is fine, but never implement in it: all work happens in a worktree branched from freshly fetched origin/main (see Worktree strategy).

5. Fetch the target base branch. Use origin/main unless the user explicitly names another base.

6. Search the codebase with rg and read the narrowest relevant files before deciding on the implementation.

Grill the plan against the docs

Before asking the user anything, interrogate the draft plan against the repo's own documentation. Two goals: consume questions the docs already settle, and surface where the docs are stale or silent so this build can fix them.

Sources, in order: CONTEXT.md (every term the plan uses must match the glossary; check the _Avoid_ lists), docs/adr/ for the touched area (does the plan re-litigate an accepted decision? will it need a new ADR per the AGENTS.md three-part test?), the subsystem doc via docs/index.md, spec/ rules in scope, and any specs/NNN-*/ feature spec.

For each load-bearing decision in the plan, classify:

• Docs decide it — adopt the documented answer and cite it in the plan; not a question.
• Docs contradict the plan — one of them is wrong. Decide which with evidence; if the doc is stale, queue the doc fix. If the plan is wrong, fix the plan.
• Docs are silent — that is a real clarifying question. Carry the "checked X, Y — silent" note into the question so the user sees why it is genuinely open.

Queue every doc delta found here into the build itself: CONTEXT.md term updates ride in the same change (AGENTS.md convention), ADR-worthy decisions get an ADR, stale subsystem-doc claims get corrected. Small doc fixes belong in this PR; large rewrites become a linked follow-up issue.

Clarifying questions

Before writing any code, ask the questions whose answers change what gets built — the survivors of the docs grill. A wrong guess at this stage costs an entire build-review cycle; a question costs the user seconds. Ask them as ONE batch up front (AskUserQuestion in Claude Code, a single numbered list in Codex), with a recommended option per question so the user can mostly confirm.

Worth asking:

• product behavior or UX choices with more than one defensible shape (what should the user see, where does the control live, what happens on failure)
• scope boundaries: what is explicitly in and out, one PR or several, feature-complete or minimal first cut
• acceptance criteria when the prompt implies but does not state them (what makes this done, what must keep working)
• anything irreversible or outward-facing: schema migrations, API contract changes, billing, released-channel behavior, data deletion
• conflicts between the prompt and what the code or tracker Issue actually says

Not worth asking:

• anything the repo, issue, or git history already answers - the docs grill above should have consumed these
• choices with an obvious conventional default - pick it and note it in the PR
• details that do not change the diff

Do not trickle questions throughout the build; front-load them. If answers do not come or the task is explicitly AFK, take the conservative path, state each assumption prominently in the PR body, and flag the ones a reviewer should double-check.

Model orchestration

Assume the session is running on the most capable model available in the current harness. That model is expensive: it is the architect and orchestrator only, and delegates everything else. (Model names version faster than this skill is edited — reason in tiers, not versions.)

The session model keeps the work that determines whether the PR is right:

• intake, scoping, and the implementation plan
• architecture and the contracts between parallel tracks (command names, request/response shapes, file ownership)
• judgment calls: review-feedback triage, tradeoffs, anything ambiguous or irreversible
• verification: reading delegated diffs, adversarially re-checking claimed results, deciding what is actually done

Everything else is delegated. Delegated work runs on a cheaper but still strong model — one tier below the orchestrator, never the session model itself — selected via the Agent tool's model option (or --model for headless claude -p; the corresponding lower tier via codex exec from Codex). That covers both implementation (writing code against a specified contract, test authoring, mechanical refactors, merge-conflict resolution with clear instructions, QA recording, PR housekeeping) and the review battery (Standards / Spec / adversarial axis runs).

Delegation rules:

• Write each brief like a contract: exact scope and file ownership, the interface to build against, validation commands that must pass, repo conventions to follow, and an instruction to report deviations instead of improvising around them.
• Run implementers in parallel only when their file ownership does not overlap; define shared contracts up front so independently built halves meet.
• Never trust a delegated report on its own. Verify against the diff and test output, and route confirmed defects back to the agent that owns that code with the evidence.
• Right-size the overhead: if the brief would be longer than the diff, skip delegation and do the work directly on the top model.
• Expect subagents to die on transient failures (API overload, timeouts). Resume the same agent so it keeps its context instead of respawning from scratch; the same applies when sending follow-up scope or defect reports to an agent that already knows the code.
• Do not delegate the plan, the contracts, or the final go/no-go. If the orchestrating model finds itself writing bulk code, delegate; if a subagent starts making architectural decisions, pull them back up.

Cross-harness implementer (with codex / with claude)

When the build prompt carries an implementer directive (e.g. /repo-build-pr JUN-200 with codex), the named harness implements every delegated brief AND runs every review axis; the session model keeps everything that determines whether the PR is right — intake, the tracker lifecycle, architecture, chunking, verification of delegated diffs, triage of review findings, and publish.

1. Plan first, as above (including clarifying questions — the session is interactive even when the implementer is not). Then split the plan into small, independently verifiable chunks: each chunk gets its own brief file written to the contract standard above, plus the narrowest gate command that proves it.
2. Dispatch each chunk to the implementer in the active worktree via repo-delegate:

   .agents/skills/repo-delegate/scripts/run-codex.sh -t <chunk-brief.md> -C <worktree> -g "<chunk gate>"
   

   (run-claude.sh when orchestrating from Codex.) The delegate edits and runs the gate but never commits, and its runners require a clean tracked tree — so verify the diff, then commit each chunk before dispatching the next. Atomic commits fall out of the loop naturally.
3. Chunks run sequentially in one worktree. Genuinely independent tracks get separate worktrees (strategy below), one delegate stream per worktree.
4. Verify every chunk yourself against the diff and real gate output; a delegate report is a claim, not evidence. Route defects back as a follow-up brief that references the original chunk. For multi-chunk builds, run the adversarial review axis on each chunk right after committing it rather than only in the end-of-build battery — a defect a chunk introduces (a moved line whose consumers changed, a guard a dispatch branch bypasses) is cheapest to catch before later chunks and review rounds build on it (PR #633: a chunk-A clock-anchor regression surfaced only in external review hours later).
5. A cross-harness implementer mode is single-harness for all delegated work, whichever direction it runs: the pre-publish battery ALSO runs on the implementer harness — every axis via repo-review/scripts/run-codex.sh -a <axis> for with codex, run-claude.sh -a <axis> for with claude — with no review sub-agents on the orchestrating side. This deliberately accepts self-review by the implementer harness (see the carve-out in docs/agents/collaboration.md); the counterweights are the regression-gated fixture tests in every fix brief and the session model's own verification and finding triage, which never delegate. Exception: repo-review's convergence re-runs alternate the reviewing harness even in this mode (see its convergence loop) — the single-harness convention covers who implements and who runs the first battery, not who re-reviews the fixes.

Worktree strategy

Always isolate implementation work from the user's active checkout.

• Create a dedicated worktree under .worktrees/<branch> at the repo root — gitignored, and inside the workspace so a Codex delegate's sandbox can write it (a sibling ../ path sits outside the sandbox, so with codex mode cannot write it). Then copy the gitignored local environment files into it. Resolve the main checkout path first (via the common git dir, so this is correct even when the current directory is itself a worktree), because fresh worktrees do not inherit .env or clovy-api/.env:

  MAIN="$(cd "$(git rev-parse --git-common-dir)/.." && pwd)"
  git fetch origin main
  BRANCH="<user>/<short-description>"   # prefix with the user's handle, e.g. jakub/… or jun/…
  WT="$MAIN/.worktrees/<short-description>"
  git worktree add -b "$BRANCH" "$WT" origin/main
  cd "$WT"
  cp "$MAIN/.env" .env 2>/dev/null || true
  cp "$MAIN/clovy-api/.env" clovy-api/.env 2>/dev/null || true
  

  These files are gitignored and exist only in the main checkout. The app, the local dev token, and the QA video upload all depend on them. In particular, clovy-api/.env holds the os-platform API key (CLOVY__ISSUE_REPORTS__OS_PLATFORM_API_KEY, with the June-era alias, or OS_PLATFORM_API_KEY) that the video upload step reads, so without this copy prepare_qa_video.py --upload fails inside the worktree.
• Use one worktree for simple or medium tasks.
• Use multiple worktrees or subagents only when the prompt naturally splits into independent tracks, such as frontend plus backend exploration, competing implementation strategies, or a broad bug hunt.
• Keep one final integration branch and one final PR unless the user explicitly asks for multiple PRs.

Before editing, tell the user which worktree or worktrees you are using and why.

After the PR merges or the build is abandoned, remove the worktree so the convention dir does not accumulate stale trees (git worktree list shows how many are live): git worktree remove "$WT" — reconcile or copy out any leftover local changes first, and only add --force once you have — then git worktree prune. Do not remove the user's active checkout or an unrelated session's worktree.

Implementation

Follow the repo's existing patterns first.

• Keep edits scoped to the prompt and nearby supporting tests.
• Use apply_patch for manual file edits.
• Do not revert unrelated user changes.
• For UI work, follow AGENTS.md: sentence-case labels, design tokens from src/styles/tokens.css, and icons from central-icons or central-icons-filled only.
• Do not add new dependencies, abstractions, or global behavior unless they are clearly needed for the prompt.
• Commit only after reading the final diff and confirming every changed file belongs to the PR.

Validation

Run the smallest checks that prove the change, then broaden based on blast radius. The Makefile owns the commands (make help lists every target; the pinned Rust toolchain and exact flags live there, not here) — call the narrowest target rather than copying its recipe, so this skill cannot drift from CI.

• Full gate: make verify — the CI-parity gate (Biome, typecheck, web tests, and fmt/clippy/test for both Rust crates). A green make verify should mean green CI; run it before publish for any non-trivial diff.
• Frontend-only: make check typecheck test-web, or a single vitest file while iterating.
• Tauri Rust: make tauri-test (add make tauri-lint if shared behavior changed).
• Clovy API: make clovy-api-test — uses the pinned toolchain from the Makefile.
• Docs or skill-only: validate the skill's structure and that it is symlinked per AGENTS.md; skip expensive app builds unless touched files require them.

Hosted PR CI intentionally skips slower local-signoff gates. After the final commit is pushed, run make local-ci from a clean branch. That command detects changed paths, runs frontend typecheck/Vitest and macOS Tauri Rust fmt/clippy/tests when relevant, and posts the required signoff/frontend and signoff/rust-macos statuses. Do this before final review or ready-for-review; if local tooling or hardware blocks it, record the blocker and add the relevant escape-hatch label (run-frontend-ci or run-macos-ci) so hosted CI covers the gap.

Judge vitest by failure count, not exit code (hud-meeting teardown noise can exit non-zero at zero real failures — see AGENTS.md). If a check cannot run because of local tooling, missing services, or credentials, say exactly what blocked it and what evidence still supports the PR.

Never read a gate's exit code through a pipe: make verify 2>&1 | tail -40 reports tail's status, so a failed gate prints exit=0. Redirect to a log and echo $? (make verify > verify.log 2>&1; echo $?), or read the log's last target — make runs prerequisites in order and halts on the first failure, so the final target passing is itself the proof.

Live app walkthroughs

Use $agent-e2e-qa as the default human-like validation layer whenever the change affects a user-visible workflow or would be hard to trust from code and terminal output alone. Load that skill before running the walkthrough — it owns the surface decision tree (web preview / background browser video / native Tauri / Chrome handoff), the recording, compression, and os-platform upload pipeline, and the evidence format. For lightweight browser-only visual verification — screenshots or a PR-embeddable GIF without a full QA charter — load $browser-test-tauri-fe instead; it owns the technique of driving the Vite frontend in Chromium with a faked Tauri IPC bridge.

Run an agent-driven walkthrough for changes that touch:

• app UI, onboarding, settings, HUDs, trays, native windows, permissions, or visual layout
• agent conversations, prompt flows, streaming states, error states, or background runs
• auth, account, checkout, external browser handoff, file upload/download, or other integration paths
• bug fixes with a reproducible user sequence
• behavior that reviewers can understand faster by seeing it operate

Skip live walkthroughs for narrow docs-only, test-only, build config, pure refactor, or low-level utility changes when no user-visible behavior is affected. Say why it was skipped in the PR validation notes.

Pick the least invasive surface per $agent-e2e-qa's decision tree. Do not perform live billing, enter credentials, record microphone audio, or expose private data without explicit user confirmation.

Treat walkthrough failures as validation failures. Fix the issue, rerun the relevant deterministic checks, and rerun the live walkthrough before asking for final review. If the live surface is blocked by permissions, credentials, hardware, or unavailable services, include BLOCKED evidence and the remaining risk.

What this workflow requires of the walkthrough (the commands and pipeline live in $agent-e2e-qa):

• Record video when human reviewers would benefit from seeing the result (visual/UI changes, native interactions, agent behavior, fixed bug repros, "the test is the demo" flows); prefer the background browser helper over foreground screen capture so the run does not fight the user's screen.
• When PR sharing was authorized, compress and upload the video to os-platform and put the remote URL or PR comment in the validation evidence — a local path is not sufficient. The upload key rides in clovy-api/.env, which fresh worktrees lack (see Worktree strategy); a missing key is a BLOCKED upload with the local path kept as evidence.

Pre-publish review pass

Green checks and a passing walkthrough are necessary, not sufficient: they prove the code does what its tests say, not that the diff is free of defects the tests never imagined. For any non-trivial diff, run the repo-review battery locally before opening the draft PR (load .agents/skills/repo-review/SKILL.md; $repo-review in Codex):

1. Size and dispatch the battery per repo-review's Sizing the battery and dispatch rules — do not restate them here (that section owns the trivial/standard/high-stakes thresholds and which axis runs on which harness). The build-specific wiring: in-session (native) builds run the sub-agent axes on the orchestrating harness and send at least the adversarial axis to the harness that did not write the diff; a with codex / with claude implementer runs every axis on the implementer harness (the single-harness convention in Model orchestration). Baseline is origin/main for a PR branch.
2. Triage every finding to a disposition per the battery's aggregate step — fix-now, deliberate (amend the spec file), pre-existing parity (follow-up, checked against the fixed point), or refuted (with evidence). Verify before fixing; plausible-sounding findings that cannot name a failure scenario are noise.
3. Route confirmed defects back to the implementer agent that owns the code, with the evidence, re-run the relevant validation, then re-run the adversarial axis until it approves (the battery's convergence loop).

Record the sizing you chose in the PR validation notes, and say so when a trivial diff took only the minimal pass.

Publish

Use a draft PR for the first publish.

1. Review git diff and git status -sb.
2. Stage only intended files.
3. Commit with a terse message.
4. Push the branch:

   git push -u origin "$(git branch --show-current)"
   

5. Run the local signoff gate for the pushed commit:

   make local-ci
   

6. Open a draft PR against the chosen base. The PR body should include:
   • task ID from the prompt or live issue data, including Closes <TASK-ID> when a tracker Issue exists
   • what changed
   • why it changed
   • validation run
   • live agent walkthrough evidence, os-platform video URLs or PR comments, or the reason no live walkthrough was useful
   • assumptions taken on clarifying questions that went unanswered, flagged for reviewer attention
   • known gaps or skipped checks
7. Watch initial CI with:

   gh pr checks --watch
   

Do not mark the PR ready yet.

Review loop

After the draft PR exists, wait for automated review from Greptile and Codex within the current session when practical. This can be slow. Poll for up to 30 minutes before concluding no automated review is available, unless the user asks to stop sooner or the session is otherwise blocked.

Use gh to inspect review state:

gh pr view <number> --comments --json comments,reviews,reviewRequests
gh pr checks <number> --watch

Poll about every 30 seconds so feedback is picked up quickly. Triage every bot at all three surfaces, not just inline threads: inline review comments (gh api .../pulls/<n>/comments), review bodies (gh pr view <n> --json reviews — Octopus hides findings in collapsed <details> tables there), and the bot's summary comment (Greptile places outside-the-diff findings only in its summary). A round is not "addressed" until every finding at every surface has a disposition. Keep the user updated while waiting, but do not start duplicate polls or spam the PR with repeated bot pings.

For inline review threads, use GraphQL through gh api graphql when gh pr view is not enough. Inspect recent repo PRs if the current bot handles or re-trigger comments are unclear. At the time this skill was written, recent reviews used:

• Greptile summary/comment author: greptile-apps
• Codex review author: chatgpt-codex-connector
• Codex review trigger comment: @codex review

If a long review loop lets origin/main move under the branch (the mergeability check fails, or a reviewer flags drift), rebase onto freshly fetched origin/main, re-run the gate for the touched surfaces, and git push --force-with-lease before requesting the final review — a stale base can turn a green PR red after merge.

Classify every bot comment before acting:

• Relevant and correct: implement it.
• Correct but out of scope: reply with a concise rationale and leave it for a separate PR.
• Incorrect: reply with the evidence and do not change code.
• Duplicate: note the existing fix or prior response.

Do not apply bot feedback mechanically. The user explicitly wants judgment: address feedback only when it is relevant and good.

After fixing accepted feedback:

1. Re-run the relevant validation.
2. If the follow-up changed user-visible behavior, rerun the relevant $agent-e2e-qa walkthrough and refresh PR video evidence when reviewers benefit from seeing the new result.
3. Commit and push follow-up changes.
4. Re-check PR comments, review threads, and CI.
5. Request final review from Greptile and Codex using the repo's current trigger convention. For Codex, post the exact PR comment @codex review. If Greptile's convention is unclear, leave a clear PR comment tagging the observed Greptile identity and asking for another pass.
6. Mark the PR ready for review only after the final review request is posted and there are no known local blockers:

   gh pr ready <number>
   

Do not merge the PR unless the user explicitly asks.

Stop conditions

Stop and ask the user for help only when progress is blocked by authentication, missing secrets, inaccessible external services, or a product decision that cannot be inferred safely.

If automated reviewers do not respond after the full 30-minute polling window, leave the PR as draft or tell the user exactly what is still pending. Do not pretend a final review happened.