AGENT SKILL

grillspec-subagent-execution

Use when executing implementation plans with independent tasks in the current session

License 许可
MIT
所属仓库
forbxpg/grillspec
最近更新
2026-10-21
Stars
★ 0
Forks
0

01 前言 Front Matter

name
grillspec-subagent-execution
description
Use when executing implementation plans with independent tasks in the current session

02 资源文件 6

03 SKILL.md 正文

Subagent-Driven Development

Execute plan by dispatching a fresh implementer subagent per task, a task review (spec compliance + code quality) after each, and a broad whole-branch review at the end.

Why subagents: You delegate tasks to specialized agents with isolated context. By precisely crafting their instructions and context, you ensure they stay focused and succeed at their task. They should never inherit your session's context or history — you construct exactly what they need. This also preserves your own context for coordination work.

Core principle: Fresh subagent per task + task review (spec + quality) + broad final review = high quality, fast iteration

Narration: between tool calls, narrate at most one short line — the ledger and the tool results carry the record.

Continuous execution: Do not pause to check in with your human partner between tasks. Execute all tasks from the plan without stopping. The only reasons to stop are the four named below, or all tasks complete. "Should I continue?" prompts and progress summaries waste their time — they asked you to execute the plan, so execute it.

Rulings, not stalls. A running plan does not wait on a human. Conflicts, ambiguities, plan defects, a cap you would have asked to exceed — decide them. The spec is the binding authority, the plan is its argument, and your judgment settles what neither answers. Record every decision in the ledger as Ruling: <what you decided> — <why> — <what it costs if wrong>, and keep going. A wrong ruling costs rework your human partner can see and undo; a session parked on a question costs their whole day and buys nothing.

Four things stop you, and only these: an irreversible or destructive operation; a security-sensitive action; a side effect outside this worktree that norms say you ask about first (a merge, a push to a shared branch, a publish); and a plan so broken that every path forward is a guess. For those, stop and ask.

When to Use

digraph when_to_use {
    "Have implementation plan?" [shape=diamond];
    "Tasks mostly independent?" [shape=diamond];
    "Partner chose inline, or no subagent tool?" [shape=diamond];
    "grillspec-subagent-execution" [shape=box];
    "grillspec-inline-execution" [shape=box];
    "Manual execution or brainstorm first" [shape=box];

    "Have implementation plan?" -> "Tasks mostly independent?" [label="yes"];
    "Have implementation plan?" -> "Manual execution or brainstorm first" [label="no"];
    "Tasks mostly independent?" -> "Partner chose inline, or no subagent tool?" [label="yes"];
    "Tasks mostly independent?" -> "Manual execution or brainstorm first" [label="no - tightly coupled"];
    "Partner chose inline, or no subagent tool?" -> "grillspec-inline-execution" [label="yes"];
    "Partner chose inline, or no subagent tool?" -> "grillspec-subagent-execution" [label="no"];
}

vs. Executing Plans (inline):

  • Fresh subagent per task (no context pollution) instead of one context doing every task
  • Review after each task (spec compliance + code quality) instead of only at the end
  • Costs a fresh context per task and per review; inline costs one context plus one final reviewer
  • Both run in this session, share the same plan workspace and ledger, and never pause between tasks

The Process

The whole flow as a diagram is in process.md; read it once at setup. The sections below are the authority.

Setup

Ensure the work happens in an isolated workspace: use grillspec-worktrees to create one or verify the existing one. Never start implementation on a main/master branch without your human partner's explicit consent. In a new worktree, run grillspec use <change-id> first: the active change is recorded per working tree.

Conversation memory does not survive compaction. In real sessions, controllers that lost their place have re-dispatched entire completed task sequences — the single most expensive failure observed. Track progress in a ledger file, not only in todos.

  • Each change owns a workspace: at skill start, run grillspec exec workspace — it prints the change's git-ignored directory (<repo-root>/.grillspec/exec/<change-id>/), home to every artifact for THIS change: ledger, briefs, reports, review packages. Another change's directory is never yours to read or write.
  • Check for this change's ledger at <workspace>/progress.md. If its first line names your change, tasks with a Task <N.M>: complete line are DONE — do not re-dispatch them; resume at the first task without one. A task whose last line is a fix round is mid-loop: resume the loop at the next round. A ledger whose first line names a different change is another change's progress: leave it in place and start your own, fresh.
  • grillspec exec workspace creates the ledger with its identity as the first line: # grillspec ledger — change: <change-id>. The plan's own checkboxes are the other record: a task's box is ticked by grillspec exec done only when its tests pass.
  • The ledger is your recovery map: the commits it names exist in git even when your context no longer remembers creating them. After compaction, trust the ledger and git log over your own recollection.
  • git clean -fdx will destroy the workspace (it's git-ignored scratch). Git cannot bring back the ledger, briefs or rulings: recover progress from the plan's ticked boxes and git log, and re-ask your human partner about any ruling you cannot reconstruct.

Read the plan once, note its context and Global Constraints, and create a todo per task. If the plan names a Spec, read that too: the spec is the authority the plan argues from, and conflicts inside the plan resolve against it. A plan with no reachable spec gets a ledger note saying so — rulings made without one are provisional.

Before dispatching the first task, scan the plan once for conflicts, writing down what you checked as you check it:

  • tasks that contradict each other or the plan's Global Constraints
  • anything the plan explicitly mandates that the review rubric treats as a defect (a test that asserts nothing, verbatim duplication of a logic block)

The scan's output is a table, not a verdict. One row for every pair of tasks that share a file or an interface: the two tasks, what one produces against what the other consumes, and what you found. One row for every task: whether its own text agrees with itself — the tests it specifies against the code it specifies, the files it creates against the files it later touches. "The scan is clean" without those rows is not a scan you ran.

Write the table to the ledger. Rule on everything you find before execution begins — each finding against the plan text that mandates it — and record each ruling in the ledger. If the scan is clean, proceed without comment. Rule on each conflict it surfaces — the spec is the binding authority, the plan is its argument — record the ruling beside its row, and dispatch the first task. The review loop remains the net for conflicts that only emerge from implementation.

Model Selection

Use the least powerful model that can handle each role to conserve cost and increase speed.

Mechanical implementation tasks (isolated functions, clear specs, 1-2 files): use a fast, cheap model. Most implementation tasks are mechanical when the plan is well-specified.

Integration and judgment tasks (multi-file coordination, pattern matching, debugging): use a standard model.

Architecture and design tasks: use the most capable available model. The final whole-branch review is one of these — dispatch it on the most capable available model, not the session default.

Review tasks: choose the model with the same judgment, scaled to the diff's size, complexity, and risk. A small mechanical diff does not need the most capable model; a subtle concurrency change does. Scoped re-reviews of small fix diffs take a cheap-to-mid tier.

Fix-loop escalation (rounds 4-5): use a model at least one tier above the implementer that got stuck.

Always specify the model explicitly when dispatching a subagent. An omitted model inherits your session's model — often the most capable and most expensive — which silently defeats this section.

Turn count beats token price. Wall-clock and context cost scale with how many turns a subagent takes, and the cheapest models routinely take 2-3× the turns on multi-step work — costing more overall. Use a mid-tier model as the floor for reviewers and for implementers working from prose descriptions. When the task's plan text contains the complete code to write, the implementation is transcription plus testing: use the cheapest tier for that implementer. Single-file mechanical fixes also take the cheapest tier.

Task complexity signals (implementation tasks):

  • Touches 1-2 files with a complete spec → cheap model
  • Touches multiple files with integration concerns → standard model
  • Requires design judgment or broad codebase understanding → most capable model

The Task Loop

Batch small same-shape work. When the plan lists several tasks that are each a small, independent edit of the same kind — the same one-line fix, constant change, or field addition repeated across files — do not dispatch one subagent per task. Compose ONE dispatch brief listing every file and its change, send the whole batch to a single subagent, and review its diff as one unit. Reserve one-dispatch-per-task for work that needs its own judgment, its own tests, or its own review surface.

Everything you paste into a dispatch prompt — and everything a subagent prints back — stays resident in your context for the rest of the session and is re-read on every later turn. Hand artifacts over as files.

Waiting on dispatched subagents: never poll a wait interface with short timeouts, and never sit in one silent, open-ended wait either. While you have local work — ledger updates, packaging the next review, reading reports — keep working; child results arrive on their own. When you are genuinely idle, wait in bounded stretches (five to ten minutes, where your platform allows), and between stretches post one line of status and reconcile your live children: list them, and chase any that finished without reporting. A bounded stretch keeps nearly all of a long wait's efficiency while guaranteeing a stuck or lost child is noticed within minutes, not at the end of the session.

1. Dispatch the implementer

Record BASE before dispatching — the review package and fix-round diffs need it. grillspec exec start <N.M> prints it.

  • Task brief: before dispatching an implementer, run grillspec exec start <N.M> (the task number from the plan, such as 2.1) — it extracts the task's full text, with the plan's header, to a uniquely named file and prints the path and BASE. Compose the dispatch so the brief stays the single source of requirements. Your dispatch should contain: (1) one line on where this task fits in the project; (2) the brief path, introduced as "read this first — it is your requirements, with the exact values to use verbatim"; (3) interfaces and decisions from earlier tasks that the brief cannot know; (4) your resolution of any ambiguity you noticed in the brief; (5) the report-file path and report contract. Exact values (numbers, magic strings, signatures, test cases) appear only in the brief. Never make a subagent read the whole plan file.
  • Report file: name the implementer's report file after the brief (brief …/task-N-brief.md → report …/task-N-report.md) and put it in the dispatch prompt. The implementer writes the full report there and returns only status, commits, a one-line test summary, and concerns.
  • A dispatch prompt describes one task, not the session's history. Do not paste accumulated prior-task summaries ("state after Tasks 1-3") into later dispatches — a real session's dispatch hit 42k chars of which 99% was pasted history. A fresh subagent needs its task, the interfaces it touches, and the global constraints. Nothing else.
  • The dispatch carries the no-subagents contract (it is in the implementer template): the implementer never dispatches subagents — not helpers, and never a reviewer. Review arrives from you, after the report. In real sessions, every reviewer a worker spawned duplicated the task review the controller dispatched anyway — a full extra review seat per task.
  • If an earlier task parked a finding in the area this task touches, carry a pointer to that ledger entry in the dispatch.
  • Record the implementer's agent identity from the dispatch result — fix-loop rounds 1-3 resume this agent.
  • Never dispatch multiple implementation subagents in parallel (conflicts).

Template: implementer-prompt.md

2. Handle the report

Implementer subagents report one of four statuses. Handle each appropriately:

DONE: Generate the review package (grillspec exec review-package BASE HEAD — it prints the unique file path it wrote; BASE is the commit you recorded before dispatching the implementer — never HEAD~1, which silently drops all but the last commit of a multi-commit task), then dispatch the task reviewer with the printed path.

DONE_WITH_CONCERNS: The implementer completed the work but flagged doubts. Read the concerns before proceeding. If the concerns are about correctness or scope, address them before review. If they're observations (e.g., "this file is getting large"), note them and proceed to review.

NEEDS_CONTEXT: The implementer needs information that wasn't provided. Provide the missing context and re-dispatch.

BLOCKED: The implementer cannot complete the task. Assess the blocker:

  1. If it's a context problem, provide more context and re-dispatch with the same model
  2. If the task requires more reasoning, re-dispatch with a more capable model
  3. If the task is too large, break it into smaller pieces
  4. If the plan itself is wrong, rule on the correction, ledger it, and re-dispatch with the ruling carried in the dispatch

Never ignore an escalation or force the same model to retry without changes. If the implementer said it's stuck, something needs to change.

If the implementer asks questions — before starting or mid-task — answer clearly and completely, provide additional context if needed, and don't rush it into implementation.

3. Review the task

Per-task reviews are task-scoped gates. The broad review happens once, at the final whole-branch review. Never skip the task review, and never accept a report missing either verdict — spec compliance AND task quality are both required. Implementer self-review never replaces the task review; both are needed.

  • Hand the reviewer its diff as a file: run grillspec exec review-package BASE HEAD and pass the reviewer the file path it prints (or, if grillspec cannot run: git log --oneline, git diff --stat, and git diff -U10 for the range, redirected to one uniquely named file). The output never enters your own context, and the reviewer sees the commit list, stat summary, and full diff with context in one Read call. Use the BASE you recorded before dispatching the implementer — never HEAD~1, which silently truncates multi-commit tasks. Never dispatch a task reviewer without a diff file.
  • Reviewer inputs: the task reviewer gets three paths — the same brief file, the report file, and the review package — plus the global constraints that bind the task.
  • The global-constraints block you hand the reviewer is its attention lens. Copy the binding requirements verbatim from the plan's Global Constraints section or the spec: exact values, exact formats, and the stated relationships between components ("same layout as X", "matches Y"). The reviewer's template already carries the process rules (YAGNI, test hygiene, review method) — the constraints block is for what THIS project's spec demands.
  • Do not add open-ended directives like "check all uses" or "run race tests if useful" without a concrete, task-specific reason
  • Do not ask a reviewer to re-run tests the implementer already ran on the same code — the implementer's report carries the test evidence
  • Do not pre-judge findings for the reviewer — never instruct a reviewer to ignore or not flag a specific issue. If you believe a finding would be a false positive, let the reviewer raise it and adjudicate it in the review loop. If the prompt you are writing contains "do not flag," "don't treat X as a defect," "at most Minor," or "the plan chose" — stop: you are pre-judging, usually to spare yourself a review loop. The task reviewer may report "⚠️ Cannot verify from diff" items — requirements that live in unchanged code or span tasks. These do not block the rest of the review, but you must resolve each one yourself before marking the task complete: you hold the plan and cross-task context the reviewer lacks. If you confirm an item is a real gap, treat it as a failed spec review — it enters the fix loop with the other findings.

Template: task-reviewer-prompt.md

4. The fix loop

When the task reviewer reports anything to fix, read fix-loop.md now and follow it exactly before you complete the task.

5. Complete the task

When the review comes back clean — or every open finding is parked with a ruling at the cap — run grillspec exec done <N.M> <BASE> -- <the task's verify command>. It ticks the task's box in the plan and appends Task <N.M>: complete (…) to the ledger only when the tests pass; then add the review outcome to the ledger in the same message as your other bookkeeping:

  • Task <N.M>: review clean
  • Task <N.M>: <K> parked after a tripped breaker

grillspec exec done runs one command, without a shell: the task's main test command (usually the test suite). Run any other check the task's verify: line names yourself, before it, and note the result in the ledger. If the suite runs longer than your shell tool allows, run exec done in the background and wait for it.

Then commit the tick, so it travels with the branch: git add <plan file> and git commit -m "chore(plan): N.M done".

Then mark the todo complete and move on. Never move to the next task while the review has open Critical/Important issues that are neither fixed nor parked-with-ruling at the cap.

Final Review

The final whole-branch review gets a package too: run grillspec exec review-package MERGE_BASE HEAD (MERGE_BASE = the commit the branch started from, e.g. git merge-base main HEAD) and include the printed path in the final review dispatch, so the final reviewer reads one file instead of re-deriving the branch diff with git commands. Dispatch on the most capable available model (see Model Selection), using grillspec-requesting-review's code-reviewer.md. Point it at the ledger's deferred-minor and parked lines so it can triage which must be fixed before merge.

If the final whole-branch review returns findings, dispatch ONE fix subagent with the complete findings list — not one fixer per finding. Per-finding fixers each rebuild context and re-run suites; a real session's final-review fix wave cost more than all its tasks combined. Then run exactly one scoped re-review of the fix wave (grillspec exec review-package FIX_BASE HEAD over the fix range, re-review-prompt.md). Adjudicate any residual findings as in the task loop's breaker: park with rulings, or rule on the load-bearing ones and ledger what you decided. Only the four classes above stop you here. There is no second fix wave — residual load-bearing findings surface to your human partner when grillspec-finishing presents the options.

Finish

Before you delete anything, collect every ledger line containing Ruling: — preflight rulings, parked findings, breaker adjudications, all of them — into your final message under "Rulings I made", in the order you made them, each with what it costs if wrong. The list is exhaustive: if the ledger holds a ruling, the list holds it. That list is the only place the decisions you took on your human partner's behalf reach them — they read it and rework whatever you got wrong. A ruling that dies with the workspace was a decision made in secret.

When the final whole-branch review is clean and its fixes are merged, delete this change's workspace (rm -rf <workspace>) — the git history is the record now. This is the one deletion you do without asking: the workspace is this change's own scratch, and its rulings are already in your final message. Sibling directories belong to other plans; leave them alone.

Use grillspec-finishing.

Common Rationalizations

Excuse Reality
"Close enough on spec compliance" Reviewer found spec gaps = not done. Fix or hit the cap and adjudicate — those are the only exits.
"I'll fix it myself, dispatching is overhead" Controller fixes pollute your context and skip review. Resume the implementer.
"One more round will converge" Past the cap, rounds don't converge — the failure is structural. Adjudicate and route.
"The reviewer will just find something new anyway" Scoped re-reviews verify fixes; they cannot wander. New findings on untouched code go to the ledger, not the loop.
"This finding is obviously wrong, I'll drop it" You adjudicate only at the cap, and every ruling is a ledger entry. Silent discards are forbidden.
"The fix was small, skip the re-review" Unreviewed fixes are how regressions land. Every round ends with a scoped re-review.
"Reviews slow the loop down" The loop without reviews is just unverified churn. Reviews are the loop's brakes and steering.
"Ledger bookkeeping is overhead" The ledger is what survives compaction. Controllers without one have re-dispatched entire completed task sequences.
"The implementer spawned its own reviewer — free extra assurance" It's a duplicate seat reviewing the same diff; the task review is the gate. A worker-spawned reviewer is a defect to flag, not rigor.

Example Workflow

A full worked run is in example.md.

04 同仓库更多 Skills