| description | Work a task end-to-end with lean context gathering, implementation, and verification |
| argument-hint | [task description | issue id/link] |
| disable-model-invocation | true |
| name | task |
| metadata | {"skiller":{"source":".agents/rules/task.mdc"}} |
Work Task
Handle $ARGUMENTS. Start from the source of truth, load extra skills only when
they earn their keep, and verify before calling the task done.
#$ARGUMENTS
Core Rules
- Read the task source first.
- Read local repo instructions and nearby implementation patterns before
editing.
- Search for existing patterns before inventing new ones.
- Prefer the best durable ownership fix over the smallest local patch.
- Prefer targeted tests and checks during iteration.
- Keep the user updated at milestones.
- Verify the actual result before claiming done.
- Do not default to research swarms, review swarms, browser proof, GitHub
comments, or compounding.
- Every GitHub PR requires a dedicated
task invocation and task plan for that
exact PR. A batch-level task, plan, auto, or autoclosure run does not
satisfy this requirement.
- Every task-run PR body must include
🧭 Task plan: docs/plans/<plan>.md. Before autoclosure, the plan must exist
at the PR head and identify the exact PR number or URL.
- For verified code-changing work, commit, push, and create or update a PR by
default. The
task skill is the explicit git permission. Only skip that path
when the user explicitly says not to, the work has no local patch, the task is
analytical/blocked/inconclusive, or a real blocker is recorded.
- Before calling a task blocked on a repo-wide gate, rule out local install
corruption once when the failure smells wrong for the diff.
- Before the first file mutation, capture a walkthrough diff baseline when UI
or rendered output may change. Use
.agents/skills/walkthrough/scripts/diff-baseline.mjs and record its path.
Intake
- Classify the input:
- Plain task text: the user prompt is the source of truth.
- File path or spec path: read it first.
- GitHub issue URL: fetch it with
gh issue view first.
- GitHub PR URL: fetch it with
gh pr view first.
- Bare GitHub issue like
#555: resolve it against the current gh repo
first, then fetch it with gh issue view.
- Read the full source-of-truth context before doing anything else.
- For GitHub issues and PRs, also read comments and attachments when available.
- If source evidence includes video or screen recording, load
video-transcripts, use or create the shared transcript cache through that
skill, and require normalized <video-transcripts> XML before
implementation. If the helper cannot produce it after a real attempt, stop
and report the blocker.
- Classify task shape:
- Testing or coverage work.
- Program or batch work.
- Ordinary one-shot work: bug, feature, refactor, docs, review, or
investigation.
- Classify heavyweight work:
- Heavyweight: architecture or public API redesign, breaking changes, major
cross-package refactors, benchmarking, profiling strategy, scalability
work, framework comparison, migration analysis, RFCs, proposals, or
spec-first major changes.
- Non-heavyweight: ordinary bugs, one-package features, docs-only edits,
routine test work, small refactors, or normal issue execution.
- If heavyweight, load
major-task immediately and let it own workflow.
- If non-heavyweight, classify complexity:
- Non-trivial: multi-step, research-heavy, phased, or likely more than a few
tool calls.
- Trivial: quick question, small edit, or work that does not need persistent
working memory.
- If non-trivial and measurable/auditable:
- load
autogoal before implementation
- create or update one
docs/plans goal plan from the dominant-risk primary
template plus touched-surface packs:
- docs-dominant work:
--template docs
- other normal work:
--template task
- supporting docs touched: add
--with docs
.agents/**, .claude/**, .codex/**, skills, hooks, commands,
prompts, or user-action tooling touched: add
GitHub Source Rules
Apply only when the source is a GitHub issue or PR.
- Treat the GitHub issue or PR as the source of truth.
- Use
gh for fetch, comment, and read-back.
- If useful, rename the thread to
<issue-number> <issue-title>.
- Prefer PR before issue comment for verified code-changing work unless
blocked or the user said not to.
- For GitHub comments after a fix, write for QA: one fixed-in-PR line plus
short verification steps. Do not mention internal files, tests, branch names,
or staging mechanics.
- Do not require PR creation, screenshots, or comments for analytical, blocked,
or inconclusive work.
Public Issue Challenge Gate
For public GitHub bug reports, behavior claims, technical diagnoses, or
suggested fixes, challenge the issue before implementation.
- Restate the reporter claim in falsifiable terms.
- Separate observed behavior from reporter interpretation and suggested fix.
- Reproduce through the lowest honest layer first:
- focused test/source-level repro when applicable;
- existing repo-owned automated browser or integration proof when useful;
- repo-approved Browser proof when tests or automation cannot model normal
app-surface behavior honestly;
- Chrome proof directly for native browser/profile/OS behavior such as
downloads, print/print-preview, file picker/uploads, clipboard, browser
dialogs/permissions, extension/profile state, or exact Chrome rendering;
- Computer Use proof when native Chrome/OS UI must be visually confirmed;
- screenshot or explicit visual-proof waiver only after the applicable
Browser/Chrome/Computer path cannot inspect the visual/native state.
- If no existing test can reproduce it, create the smallest honest repro or
harness before fixing.
- If the issue is not reproduced, invalid, or won't-fix, hard stop and report
the evidence. Do not code around a claim you cannot prove.
- If the issue is partially valid, discard the weak suggested fix and pivot to
the best long-term ownership fix for the valid behavior.
- Record the verdict in the task plan when one exists:
valid, not reproduced, invalid, wont-fix, partially valid, or
platform limitation.
Reproduction And Contradiction Harness
Before fixing reported behavior, list every source-stated case. Do not reduce a
multi-case report to the first reproduction.
| Case | Source claim | Smallest honest harness | Before | Expected after | Proof |
|---|
If source, tests, docs, and runtime disagree:
- State the contradiction precisely.
- Identify which owner can settle it: public API, runtime, generated output,
docs, or user direction.
- Build the smallest harness that makes both interpretations observable.
- Choose the durable owner, repair source and dependents, then rerun every
source-listed case.
Never “fix” a contradiction by adding fallback parsing, duplicate exports, or a
compatibility shim unless the task explicitly requires compatibility.
Implementation Readiness Triage
Classify before mutation:
ready: acceptance, owner, and harness are known;
repair-source: the source is usable after a bounded contradiction or
placeholder repair;
major: public API/architecture/migration/benchmark decisions belong to
major-task;
blocked: missing authority or external state prevents an honest next move;
invalid: the claim is disproven or outside the requested boundary.
Only ready and repaired sources proceed to implementation. Record the verdict
in the plan with evidence.
Browser Proof
- Use the fastest honest browser surface:
[@Browser](plugin://browser@openai-bundled) first for ordinary
browser-rendered app QA: route navigation, DOM checks, forms,
screenshots, responsive checks, and UI visual proof.
[@Chrome](plugin://chrome@openai-bundled) directly when native
browser/profile/OS behavior is in scope: downloads, print or print
preview, file picker/uploads, clipboard, browser permissions/dialogs,
extension/profile state, or exact Chrome rendering.
[@Computer](plugin://computer-use@openai-bundled) only when native
Chrome/OS UI must be visually inspected or interacted with and Chrome
automation cannot read it, such as print preview, save/open dialogs, or
permission sheets.
- Any task with visual output requires direct proof on the real affected
surface. Browser-rendered output requires Browser proof; native
download/print/file/browser-profile output requires Chrome proof, with
Computer Use when the proof is in native UI. Visual output includes UI
screens, generated PDFs, emails, screenshots, rendered docs, tables,
formatting changes, previews, downloads, print previews, and user-visible
browser states.
- If Browser hits a known limitation and native proof matters, switch to
Chrome/Computer instead of lowering confidence or asking the user to confirm
something the agent can inspect.
- Direct visual proof is part of the confidence score. Do not claim
95-100%
confidence for visual-output work unless the required visual cases were
directly inspected in the real affected browser, inbox, or external surface,
or the user explicitly waived that proof.
- Do not claim the browser tool is blocked until the Browser->Chrome->Computer
escalation has been tried when it applies.
Load Skills Only When Justified
Skill Diet
Default to task for normal work and major-task for heavyweight work. Load a
niche skill only when it owns a hard domain gate, command, or proof surface that
the active task would otherwise miss.
Do not keep repo-local skills for generic lifestyle, app-template, local git
ops, stale command stubs, or broad CE ceremony when task, major-task,
autogoal, autoreview, or a Convex-specific skill owns the workflow better.
If a generated skill is gone but skills-lock.json still references it, remove
it through npx skills remove <skill> -y first. If the CLI removes the agent
files but leaves stale lock entries, record that evidence before cleaning the
lock.
autogoal: measurable or auditable non-trivial work. Use the dominant-risk
primary template and touched-surface packs: docs-heavy work gets
--template docs, normal work gets --template task, and supporting docs,
browser, agent-native, or package/API surfaces add matching --with <pack>
rows. Review expectations stay in the primary template. Do not use root
planning files, hooks, .planning/**, or docs/goals/**.
major-task: heavyweight architecture, framework, migration, benchmark, or
proposal work.
testing: tasks primarily about tests, coverage, regression gaps, or suite
phases.
tdd: bugs and feature work where behavior-level automated coverage is sane.
video-transcripts: source evidence contains a video or screen recording.
- If requirements remain ambiguous after source and local context, ask the
smallest clarifying question or switch to a planning goal when the user wants
planning.
find-skills: a genuinely missing specialist capability after local skills
and source owners have been checked.
- Browser/Chrome/Computer: real UI or native browser behavior needs proof;
follow Browser Proof for escalation.
agent-browser-issue: browser automation is blocked by a reusable tool-side
issue.
changeset: published package work under packages/ needs release notes.
Prefer updating the active unreleased .changeset/*.md draft instead of
creating a parallel changeset when one already exists.
- Docs/content work: use
--template docs when docs dominate; use --with docs
when docs are a supporting touched surface. For www/**, keep matching
packages/kitcn/skills/kitcn/** content in sync.
- Git/PR shipping: when verified code should ship and repo policy permits it,
use normal
git/gh commands directly. Stage the entire current checkout
per repo policy when creating the PR, create the commit, push, create or
update the PR before issue comments. The task skill owns the PR body:
write the PR description from the task-style final handoff contract below.
Do not skip this merely because the user did not type a separate "open a PR"
sentence.
- Review skills: load only for risky, large, user-facing, or
architecture-sensitive changes.
- Agent-native surface: when changes touch , ,
AI/tooling surfaces, commands, or user actions an agent should perform, use
the autogoal agent-native pack and end with .
Review And Risk Gates
Keep this lighter than a full architecture review. A normal task should not
grow a scorecard, issue ledger, or pass calendar, but risky work still needs
real closeout pressure.
- Autoreview is a hard closeout gate for non-trivial implementation changes.
Load
.agents/skills/autoreview/SKILL.md, pick the target from the actual
diff state, and keep going until there are no accepted/actionable findings.
- The autogoal agent-native pack plus
autoreview are required when the task
changes .agents/**, .claude/**, .codex/**, skills, hooks, commands,
prompts, or user-action tooling.
- Source authority is workspace-local. A check run in the planning repo cannot
prove behavior owned by a sibling repo, package, app, browser route, or
GitHub surface. Record the cwd/tool that owns each proof.
- For public API, runtime, package-boundary, browser behavior, agent-action, or
command-contract changes, add a compact high-risk note before closeout:
realistic failure mode, proof plan, and why the chosen boundary is still the
right one.
- Trivial docs, wording, and no-local-patch tasks may mark these gates N/A with
a reason.
Execution Path
Bug
- Reproduce first when possible.
- Add a behavior-level regression test when sane.
- Fix the real ownership boundary, not every caller around it.
- If the best fix requires an API change, make it unless task constraints rule
it out.
- Re-run targeted checks and browser flow only when the bug lives there.
Feature
- Build the source-listed case matrix and readiness verdict.
- Use bounded TDD at the behavior boundary when automated proof is sane.
- Implement the smallest coherent slice that satisfies all acceptance cases.
- Prefer the cleanest long-term ownership design that fits the slice.
- Verify the user-facing outcome and every case row.
Testing Or Coverage Work
- Use the testing policy before choosing files or commands.
- Pick the smallest honest hotspot or ordered slice.
- Add or deepen focused tests instead of broad smoke coverage.
- Verify with targeted commands first.
Program Or Batch Work
- Respect explicit order.
- For every GitHub PR, invoke
task with that exact PR and use a dedicated
per-PR task plan. A coordinating batch plan may link those plans but cannot
replace them.
- Define done for the current slice before implementation.
- Complete one
task -> optional autoclosure slice cleanly before moving to
the next PR unless the user explicitly requested parallel work.
Refactor Or Chore
- Preserve behavior.
- Do not do fake TDD theater.
- Improve bad APIs or abstractions when that is the real fix.
- Run the narrowest regression checks plus relevant build, typecheck, or lint.
Docs Or Content
- Skip engineering ceremony.
- Verify links, examples, formatting, and rendered output as appropriate.
- When updating
www/**, update corresponding packages/kitcn/skills/kitcn/**
content in the same diff.
Review Or Investigation
- Read relevant diff, files, and surrounding context first.
- For reviews, report findings first, ordered by severity.
- For investigations, identify failure mode, probable cause, and next action
before changing code.
- Only implement changes if the user asked for them.
Verification
Keep verification mandatory and proportional.
- Run targeted tests for changed behavior.
- Run package/app build and typecheck when relevant.
- Run lint when code changed and repo policy expects it.
- Run Browser verification only for normal browser-rendered app/UI tasks. Use
Chrome directly for native downloads, print/print-preview, file
picker/uploads, clipboard, browser dialogs/permissions, extension/profile
state, or exact Chrome rendering. Use Computer Use when native Chrome/OS UI
must be visually confirmed.
- Run broader repo-wide gates only when repo instructions or change scope
justify them.
- Run verification in the workspace, package, app, route, or external system
that owns the changed behavior; record the cwd when that is not obvious.
- If
bun test, bun check, or bun typecheck fails with local-corruption
signals unrelated to the diff, run bun install once and rerun the exact
failing command before declaring the task blocked.
- If verified work changes published package code under
packages/, update or
create a changeset before PR/final handoff.
- If verified work changes package code under
packages/kitcn, run
bun --cwd packages/kitcn build.
- If work changes
kitcn init -t templates or scaffold sources, run
bun run fixtures:sync and bun run fixtures:check.
- If verified work changed code, commit it and create or update the PR before
GitHub sync-back and final handoff unless the user explicitly said not to.
Do not mark commit/PR gates N/A merely because the user did not ask for a PR;
the task skill requires shipping verified code. If commit or PR creation is
impossible after real attempts, record the blocker and stop instead of
silently handing off a local-only patch.
- A final response that says "No commit/PR created because you did not ask" is
wrong for verified code-changing
task work. Either create/update the PR or
name the explicit decline/blocker.
- If the task came from a GitHub item and reached a meaningful outcome, sync
back unless the user said not to.
Evidence-Bound Confidence
Confidence describes the exact delivered claim, not how good the diff looks.
95-100%: every source-listed case has fresh direct proof at the owning
layer, required visual proof is inspected, and no accepted finding remains.
80-94%: the core claim has direct proof but a bounded non-critical case is
inferred or externally unavailable; name it.
- below
80%: do not present the task as complete.
Link confidence to the case matrix, changed owner, and final repository gate.
Durable Learning Capture
When the task exposes a reusable repository-specific trap, add the smallest
durable note to the real owner: nearby test helper, rule, docs map, ADR, or
active plan. Capture the failure signature, correct owner, and next-different
move. Do not create a generic diary or duplicate existing doctrine.
Final Handoff
- Be extremely concise.
- Report PR/issue, confidence, tests, browser proof, outcome, caveats,
design choice, and verification only when applicable.
- For non-trivial task goals, close every relevant task-template gate before the
final response.
- If a PR exists, keep the PR description synced to the final handoff.
- For GitHub comments, write for QA or the issue owner, not internal
implementation history.
- For browser work, include the exact route and human verification steps.
- When UI or rendered output changed, run
.agents/skills/walkthrough/SKILL.md
after final proof and show its annotated images inline.
Task-Style PR Body
When a task run creates or updates a PR, the PR description must mirror the
task final handoff. Do not use a generic Summary / Verification PR body,
generic git-helper prose, or a generated badge footer unless the caller or repo
template explicitly asks for it.
Use the accepted task PR format from PR #270. The shape is not optional:
- Preserve any existing
<!-- auto-release:start --> block. If a changeset is
part of the diff and repo policy expects auto release, include that block.
- Use an emoji-prefixed issue/fix line, for example
🐛 Fixes #123 or 🐛 Fixes ➖ N/A. Never include a line that links to the
current PR itself; the current PR URL belongs in the final response, not in
its own description.
- Add
🧭 Task plan: docs/plans/<plan>.md; the path must resolve at the PR
head and the plan must identify the exact PR before autoclosure.
- Use an emoji confidence line, for example
🟢 95-100% confidence.
- Use this exact table header:
| Phase | 🧪 Tests | 🌐 Browser |
- Use
Reproduced and Verified rows. Mark passing proof with 🟢, repro or
failing proof with 🔴, and non-applicable browser/test cells with ➖ N/A.
- Use bold emoji section headings exactly in this family:
**✅ Outcome**, **⚠️ Caveat**, **🏗️ Design**, and
**🧪 Verified**.
The body should tell QA/reviewers what was fixed, how it was reproduced, how it
was verified, and why the chosen ownership boundary is right. It must not use
plain Fix:, plain Confidence:, ## Outcome, ## Verified, or a generic
Summary / Verification shape for task-run PRs. After editing, verify it
with gh pr view --json body before final handoff.
Success Criteria
- Source-of-truth context was read first.
- Relevant repo instructions and patterns were read before editing.
- GitHub items were fetched and summarized correctly when provided.
- Video evidence used
video-transcripts before implementation when required.
- Bare GitHub issues like
#555 were resolved against the current repo.
- The chosen fix addressed the highest-leverage ownership boundary available.
- Non-trivial measurable work loaded
autogoal and used the right primary
template plus touched-surface packs.
- Testing work loaded the testing policy before implementation.
- Only necessary skills were loaded.
- Batch work did not sprawl without explicit instruction.
- Every PR in a batch has its own
task invocation and task plan; no aggregate
autoclosure was used as a substitute.
- The PR body names the dedicated task plan, the plan exists at the PR head,
and it identifies the exact PR before autoclosure.
- Verification matched change scope.
- Verified code-changing work was committed and PR'd, or the user explicitly
declined that path, the work had no local patch, or a real blocker was
recorded.
- PR descriptions created by task runs used the PR #270 emoji task-style body
and were verified with
gh pr view --json body.
- Final handoff matched the task type and any task-template gate evidence.