Imported from jwp23/joe-bag-of-tricks (plugins/joe-bag-of-tricks/skills/subagent-driven-development/SKILL.md). Install upstream with npx skills add jwp23/joe-bag-of-tricks --skill subagent-driven-development. Copyright stays with the author.

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 — bd state 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 as bd note <task-id> "Ruling: <what you decided> — <why> — <what it costs if wrong>" (epic-level rulings go on the epic), 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];
    "Stay in this session?" [shape=diamond];
    "subagent-driven-development" [shape=box];
    "executing-plans" [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?" -> "Stay in this session?" [label="yes"];
    "Tasks mostly independent?" -> "Manual execution or brainstorm first" [label="no - tightly coupled"];
    "Stay in this session?" -> "subagent-driven-development" [label="yes"];
    "Stay in this session?" -> "executing-plans" [label="no - parallel session"];
}

vs. Executing Plans (parallel session):

• Same session (no context switch)
• Fresh subagent per task (no context pollution)
• Review after each task (spec compliance + code quality), broad review at the end
• Faster iteration (no human-in-loop between tasks)

The Process

digraph process {
    rankdir=TB;

    subgraph cluster_per_task {
        label="Per Task";
        "Claim task:\nbd update <task-id> --claim" [shape=box];
        "Run task-brief <task-id>" [shape=box];
        "Dispatch implementer agent\n(mechanical|integration|complex)\n(./implementer-prompt.md)" [shape=box];
        "Implementer asks questions?" [shape=diamond];
        "Answer questions, provide context" [shape=box];
        "Implementer implements, tests,\ncommits, self-reviews, writes report file" [shape=box];
        "Create beads for discovered work\n(if implementer reported any)" [shape=box];
        "Run review-package BASE HEAD,\ndispatch task reviewer\n(./task-reviewer-prompt.md)" [shape=box];
        "Spec OK and quality approved?" [shape=diamond];
        "Finding conflicts with the task's design?" [shape=diamond];
        "Rule on the conflict,\nbd note the ruling" [shape=box];
        "Fix round R of 5: R<=3 resume implementer;\nR>=4 fresh implementer, next agent tier up" [shape=box];
        "Dispatch scoped re-review\n(./re-review-prompt.md)" [shape=box];
        "All findings addressed?" [shape=diamond];
        "R = 5?" [shape=diamond];
        "Dispatch adjudicator: all open\nfindings as one question packet" [shape=box];
        "Any load-bearing finding?" [shape=diamond];
        "Rule and continue; stop only if\nevery path forward is a guess" [shape=box];
        "Park findings on the task bead\nwith rulings (bd note)" [shape=box];
        "Close task bead:\nbd close <task-id> --reason \"commits <base7>..<head7>, review clean\"" [shape=box];
    }

    "Setup: worktree, render whole tree\n(bd children <epic-id> --pretty),\npre-flight review" [shape=box];
    "Load feature tasks:\nbd children <feature-id> --json" [shape=box];
    "More tasks in feature?" [shape=diamond];
    "bd children <feature-id>: all closed?\nClose feature" [shape=box];
    "More features remain?" [shape=diamond];
    "bd children <epic-id>: all closed?\nClose epic" [shape=box];
    "Run review-package MERGE_BASE HEAD,\ndispatch final code reviewer\n(../requesting-code-review/code-reviewer.md)" [shape=box];
    "Final findings? ONE fix dispatch,\none scoped re-review, adjudicate residuals" [shape=box];
    "Final review clean: delete the SDD workspace" [shape=box];
    "Use finishing-a-development-branch" [shape=box style=filled fillcolor=lightgreen];

    "Setup: worktree, render whole tree\n(bd children <epic-id> --pretty),\npre-flight review" -> "Load feature tasks:\nbd children <feature-id> --json";
    "Load feature tasks:\nbd children <feature-id> --json" -> "Claim task:\nbd update <task-id> --claim";
    "Claim task:\nbd update <task-id> --claim" -> "Run task-brief <task-id>";
    "Run task-brief <task-id>" -> "Dispatch implementer agent\n(mechanical|integration|complex)\n(./implementer-prompt.md)";
    "Dispatch implementer agent\n(mechanical|integration|complex)\n(./implementer-prompt.md)" -> "Implementer asks questions?";
    "Implementer asks questions?" -> "Answer questions, provide context" [label="yes"];
    "Answer questions, provide context" -> "Implementer implements, tests,\ncommits, self-reviews, writes report file";
    "Implementer asks questions?" -> "Implementer implements, tests,\ncommits, self-reviews, writes report file" [label="no"];
    "Implementer implements, tests,\ncommits, self-reviews, writes report file" -> "Create beads for discovered work\n(if implementer reported any)";
    "Create beads for discovered work\n(if implementer reported any)" -> "Run review-package BASE HEAD,\ndispatch task reviewer\n(./task-reviewer-prompt.md)";
    "Run review-package BASE HEAD,\ndispatch task reviewer\n(./task-reviewer-prompt.md)" -> "Spec OK and quality approved?";
    "Spec OK and quality approved?" -> "Close task bead:\nbd close <task-id> --reason \"commits <base7>..<head7>, review clean\"" [label="yes"];
    "Spec OK and quality approved?" -> "Finding conflicts with the task's design?" [label="no"];
    "Finding conflicts with the task's design?" -> "Rule on the conflict,\nbd note the ruling" [label="yes"];
    "Rule on the conflict,\nbd note the ruling" -> "Fix round R of 5: R<=3 resume implementer;\nR>=4 fresh implementer, next agent tier up";
    "Finding conflicts with the task's design?" -> "Fix round R of 5: R<=3 resume implementer;\nR>=4 fresh implementer, next agent tier up" [label="no"];
    "Fix round R of 5: R<=3 resume implementer;\nR>=4 fresh implementer, next agent tier up" -> "Dispatch scoped re-review\n(./re-review-prompt.md)";
    "Dispatch scoped re-review\n(./re-review-prompt.md)" -> "All findings addressed?";
    "All findings addressed?" -> "Close task bead:\nbd close <task-id> --reason \"commits <base7>..<head7>, review clean\"" [label="yes"];
    "All findings addressed?" -> "R = 5?" [label="no"];
    "R = 5?" -> "Fix round R of 5: R<=3 resume implementer;\nR>=4 fresh implementer, next agent tier up" [label="no - next round"];
    "R = 5?" -> "Dispatch adjudicator: all open\nfindings as one question packet" [label="yes - breaker trips"];
    "Dispatch adjudicator: all open\nfindings as one question packet" -> "Any load-bearing finding?";
    "Any load-bearing finding?" -> "Rule and continue; stop only if\nevery path forward is a guess" [label="yes"];
    "Any load-bearing finding?" -> "Park findings on the task bead\nwith rulings (bd note)" [label="no"];
    "Park findings on the task bead\nwith rulings (bd note)" -> "Close task bead:\nbd close <task-id> --reason \"commits <base7>..<head7>, review clean\"";
    "Close task bead:\nbd close <task-id> --reason \"commits <base7>..<head7>, review clean\"" -> "More tasks in feature?";
    "More tasks in feature?" -> "Claim task:\nbd update <task-id> --claim" [label="yes"];
    "More tasks in feature?" -> "bd children <feature-id>: all closed?\nClose feature" [label="no"];
    "bd children <feature-id>: all closed?\nClose feature" -> "More features remain?";
    "More features remain?" -> "Load feature tasks:\nbd children <feature-id> --json" [label="yes"];
    "More features remain?" -> "bd children <epic-id>: all closed?\nClose epic" [label="no"];
    "bd children <epic-id>: all closed?\nClose epic" -> "Run review-package MERGE_BASE HEAD,\ndispatch final code reviewer\n(../requesting-code-review/code-reviewer.md)";
    "Run review-package MERGE_BASE HEAD,\ndispatch final code reviewer\n(../requesting-code-review/code-reviewer.md)" -> "Final findings? ONE fix dispatch,\none scoped re-review, adjudicate residuals";
    "Final findings? ONE fix dispatch,\none scoped re-review, adjudicate residuals" -> "Final review clean: delete the SDD workspace";
    "Final review clean: delete the SDD workspace" -> "Use finishing-a-development-branch";
}

Setup

Ensure the work happens in an isolated workspace: use using-git-worktrees to create one or verify the existing one. Never start implementation on a main/master branch without your human partner's explicit consent.

Durable progress rides on bd, not a ledger file. 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. A separate ledger file would duplicate state a beads-only-tracking project already tracks for free, and bd's globally unique task IDs mean no two plans can ever collide over the same record.

• At skill start (and after any compaction), reload the hierarchy with bd children <epic-id> --pretty. That renders the WHOLE tree at every depth, closed issues included, and ends with a Total: N issues footer. bd children --json returns ONE level — a controller that walked it once took five features for the task layer and never saw their seven task children. Use --json afterwards for per-issue detail, never to learn the tree's shape.
• When you reduce bd JSON into a summary you will read later, keep issue_type, status, and parent on every row. The controller above had "issue_type": "feature" in its raw dump and dropped that field while reformatting; nothing afterwards ever contradicted it.
• Tasks bd already shows closed are DONE — do not re-dispatch them; resume at the first task not closed. A task that is in_progress with fix-round notes is mid-loop: resume the loop at the next round.
• Every ruling, deferral, and fix round is a bd note <task-id> "..." — it appends to the task's notes without changing status. Closing carries the commit range as the reason. Those breadcrumbs name commits that exist in git even when your context no longer remembers creating them.
• After compaction, trust bd show / bd children and git log over your own recollection.
• The .joe-bag-of-tricks/sdd/ workspace (briefs, reports, review packages) is disposable scratch, not the record — losing it (git clean -fdx included) costs nothing; bd and git are the only durable state.

Pre-Flight Plan Review

Before claiming Task 1, scan the bd hierarchy once for conflicts. The bd hierarchy IS the plan — there is no markdown plan file to read instead.

The scan's first command is bd children <epic-id> --pretty. Read the depth off the tree it prints and the issue count off its Total: footer before reading anything else. It is one command that cannot be half-run; a per-feature walk can, and its partial result looks exactly like its complete one. Then pull the detail you need per issue with bd show/bd children --json — the depth question is already answered.

The epic is the plan's argument for its spec: if bd show <epic-id> names a spec (--spec-id, pointing at the living design doc), read that too — conflicts inside the plan resolve against the epic's spec fields (description/design) and that doc. An epic with no reachable spec gets a bd note saying so — rulings made without one are provisional.

Scan for, writing down what you checked as you check it:

• tasks that contradict each other or the epic's Global Constraints (bd show <epic-id> — the design field set during writing-plans)
• anything a task's design explicitly mandates that the review rubric treats as a defect (a test that asserts nothing, verbatim duplication of a logic block)
• steps telling the implementer to invoke a skill or run a / command — implementer agents carry no Skill tool, so the step cannot be followed as written. Rule it into "follow the procedure in <path>" where that file is readable from the worktree; where it is not, put what the step needs into that dispatch's task-specific instructions yourself. Either way the rewrite reaches the implementer only through item (4) of the dispatch below — task-brief copies the task verbatim, so a bd note alone never gets there
• runs of consecutive mechanical tasks touching the same package that could be ONE dispatch and ONE review (see "Batch small same-shape work" below). Batch them unless a task's design needs its own review surface. Four separately dispatched refactors of the same shape cost four cold starts, four full test gates, and four reviewer seats for one reviewable unit of work — and nothing later in the loop will notice the waste, which is why it is scanned for here

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.

Record the table as a bd note on the epic, opening with the count: Scan: N issues across D levels — N from the Total: footer, D from the tree's depth, and every leaf-level issue accounted for in the rows below. A conflict table is a judgment artifact: a scan of 5 of 13 issues produces one that looks exactly like a complete scan. The count is the part of the note anyone can falsify in one command, so it is the part that has to be there.

Rule on everything you find before execution begins — each finding against the task text that mandates it, the spec as the binding authority — and record each ruling beside its row. If the scan is clean, proceed without comment. The review loop remains the net for conflicts that only emerge from implementation.

Model Selection

Use the least powerful agent that can handle each role. Start cheap, escalate on failure.

Implementers dispatch as plugin agent types — each definition pins its own model and reasoning effort, so the dispatch names an agent and passes no model. Reviewers dispatch as general-purpose subagents with an explicit model param, drawn from "haiku" | "sonnet" | "opus". That is the reviewer tier list, not the Agent tool's full set — the session's roster may offer higher tiers. The adjudicator dispatches as an agent type like the implementers do, so it pins its own model and takes no model param. Use this table:

Most implementation tasks are mechanical when the plan is well-specified. Plans from writing-plans include code snippets, file paths, and acceptance criteria — enough context for implementer-mechanical to succeed.

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 sonnet as the floor for reviewers, and implementer as the floor for implementers working from prose descriptions. When the task's plan text contains the complete code to write, the implementation is transcription plus testing: dispatch implementer-mechanical for that task. Single-file mechanical fixes also take the cheapest tier.

Complexity signals for implementers:

• Touches 1-2 files with a complete spec → implementer-mechanical
• Touches multiple files with integration concerns → implementer
• Requires design judgment or broad codebase understanding → implementer-complex

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 opus; a subtle concurrency change does — escalate the task reviewer to opus for those.

Fix-loop escalation (rounds 4-5): re-dispatch one step up the implementer ladder (implementer-mechanical → implementer → implementer-complex) from the agent that got stuck. The ladder tops out at implementer-complex.

Escalation is the safety net: If implementer-mechanical reports BLOCKED, re-dispatch implementer. If implementer reports BLOCKED, re-dispatch implementer-complex. Never retry the same agent type without changing something (see step 2 below).

Always name the agent type or the model explicitly when dispatching. An implementer dispatch names its agent type, which carries the model and effort with it. A reviewer dispatch names its model, and an omitted model inherits your session's model — often the most capable and most expensive — which silently defeats this section.

Escalation. You cannot reliably see what you are missing. That is a property of models, not of tiers — an orchestrator on the top tier is as blind to its own gaps as one on a mid-tier model, so never escalate because a call feels hard. Escalate when one of these fires, each detectable by counting or comparing:

Name the governing decisions up front. Trigger 2 is checked against a list you name in every dispatch you compose — the brief is generated from bd and cannot carry it: the task's own design, plus whichever recorded decisions bear on the work. Decisions recorded during the run join the list as they are written. Checking against every doc in the project instead is a scan you will silently skip, which defeats the trigger.

When several triggers fire at once, that is still ONE dispatch — one question packet naming every fired trigger. Dispatch joe-bag-of-tricks:adjudicator, passing the artifact file paths, the governing-decision paths, and the narrow question. It pins its own model; pass no model param. Never dispatch it as a fork — a fork inherits your whole session.

Record the ruling as a bd note (Ruling: <what> — <why> — <cost if wrong>). The ruling stands, the run continues, and it surfaces in the "Rulings I made" roll-up at Finish — your human partner audits after, not during. Only the four stop classes interrupt the run.

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. The Pre-Flight scan is where you look for these; batching after the first of them is already dispatched saves nothing.

Overlap a clean-expected review with the next implementation. The loop is otherwise strictly serial — implement, review, close, claim next — and on a run of mechanical tasks about a quarter of the wall clock is you waiting on reviews that come back clean. Once task N's implementer has reported DONE and you have dispatched N's reviewer, you may claim and dispatch task N+1 without waiting for that review, when BOTH hold:

• N+1 is a mechanical dispatch — its brief carries the code it needs — and
• N's review is expected clean: N was itself mechanical AND its implementer returned plain DONE.

Never overlap when N went to implementer-complex, when N returned DONE_WITH_CONCERNS / BLOCKED / NEEDS_CONTEXT, or when N is already in a fix round. Those are precisely the reviews that open findings, and N+1 would be building on a base you are about to change.

Still exactly one implementer writing at a time. The overlap is review-against-implement, never implement-against-implement — two writers in one worktree is the conflict this skill has always forbidden, and nothing here relaxes it.

Record the overlap on N+1 before you dispatch it: bd note <N+1-id> "Overlap: dispatched while <N-id>'s review was open — base included unreviewed work". An audit trail that hides a moving base is worse than no overlap.

If N's review then opens findings: do not rebase, and do not touch the worktree while N+1's implementer holds it. Wait for N+1 to report, then run N's fix round as ordinary commits on top of N+1's. Scope its re-review with FIX_BASE = the HEAD you record immediately before dispatching that fix — NOT the head the first review saw, which now has N+1's commits sitting after it and would put N+1's whole diff in front of a re-reviewer scoped to N's findings. Close N only when N's own loop is clean; N+1 still waits on N+1's review.

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. The same goes for command output you'll consult more than once: capture the test suite's output to a scratch file and re-grep the file — re-running the suite to re-read its output pays the wall-clock cost again for nothing.

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 — bd notes, 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

Claim the task (bd update <task-id> --claim) and record BASE (git rev-parse HEAD) before dispatching — the review package and fix-round diffs need it.

• Task brief: run this skill's scripts/task-brief <task-id> — it reads the task straight out of bd (bd show <task-id> --json: title, description, design, notes) into a uniquely named file and prints the path. 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 bd issues directly.
• Report file: name the implementer's report file after the brief (brief …/task-<task-id>-brief.md → report …/task-<task-id>-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 task's bead 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).
• Subagents never create or close bd issues, push to a remote, or create PRs. You manage all of those.

Template: implementer-prompt.md

2. Handle the report

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

Every bd note or close reason you write for this task is the implementer's own paste-ready bd note text from its report, verbatim or trivially extended — never your own re-narration of what the report already says.

DONE: Generate the review package (scripts/review-package BASE HEAD, from this skill's directory — 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 the same agent type
2. If the task requires more reasoning, re-dispatch one step up the implementer ladder (implementer-mechanical → implementer → implementer-complex)
3. If the task is too large, break it into smaller pieces
4. If the plan itself is wrong, rule on the correction, record it (bd note <task-id> "Ruling: ..."), and re-dispatch with the ruling carried in the dispatch

Never ignore an escalation or force the same agent type 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.

If the implementer reported discovered work, create a bead for it now: bd create --title="..." --type=task --priority=2 --deps discovered-from:<task-id>. Ignoring discovered work is a silent discard. bd create prints the new bead's ID — take it from that output and hold it in your own tracking, not bd list | grep. A bead you just created is never lost; going back to find it is wasted work, and grepping a list you already know the answer to is a sign that ID slipped out of your context.

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 this skill's scripts/review-package BASE HEAD and pass the reviewer the file path it prints (or, without bash: 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.
• Reviewer return contract: the reviewer writes its FULL report to a file (brief …-brief.md → review …-review.md) and returns only the two verdicts plus Critical/Important findings as one-liners (≤15 lines). Reviewer prose returned inline is permanently resident in the controller's context and re-read every turn after — across a multi-task epic that is thousands of tokens of pure bloat. The findings list must still be complete enough to run the fix loop without opening the file.
• The global-constraints block you hand the reviewer is its attention lens. Copy the binding requirements verbatim from the epic's Global Constraints (bd show <epic-id> — the design field set during writing-plans) 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. Name the governing-decision paths in that block and ask for any conflict with them: the reviewer is what performs trigger 2's comparison.
• 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 closing the task: 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

The loop triggers when the review reports spec ❌, any Critical or Important finding, or a ⚠️ item you confirmed as a real gap.

Before the loop starts, two routes leave it immediately:

• Record Minor findings on the task bead as you go (bd note <task-id> "Minor (deferred): <one-liner>"), and point the final whole-branch review at the closed tasks under the epic (bd show <task-id>) so it can triage which must be fixed before merge. A roll-up nobody reads is a silent discard. Minor findings never enter the loop.
• A finding labeled plan-mandated — or any finding that conflicts with what the task's design requires — is yours to rule on: weigh the finding against the task text, decide with the spec as the binding authority (via the adjudicator dispatch when a structural trigger fires — see Model Selection), and record the ruling as a bd note before you act on it. Do not dismiss the finding because the plan mandates it, and do not dispatch a fix that contradicts the plan without a recorded ruling.

Everything else enters the loop. A fix round is one fix dispatch plus one scoped re-review. Five rounds maximum per task:

Rounds 1-3 — resume the original implementer. Send it the open findings verbatim. Its context is intact: it knows the task, the code, and its own choices. If you cannot send another message to the live subagent, dispatch a fresh implementer carrying the brief path, the report-file path, and the findings — the report file is the persistent memory either way.

Never dispatch a fork for a fix round. A fork inherits the controller's entire session context and runs on the controller's model — the most expensive possible dispatch for what is usually the smallest diff of the task. A real session paid ~250k tokens forking a 2-line fix that a resumed implementer or a fresh cheap-tier dispatch would have done for ~5k. Resume the implementer; when that's impossible, a fresh implementer with the brief and report paths is the fallback — never a fork.

Rounds 4-5 — dispatch a fresh implementer one step up the agent ladder (per Model Selection), with the brief path, the report-file path, the open findings, and this framing: "A prior implementer attempted this task [N] times; you own it now. Read the report file for what was tried." A loop that survives three resumes usually means the implementer cannot see its own problem — fresh eyes and a capability bump in one move.

Every round, either way: the implementer fixes, re-runs the tests covering the amended code, appends its fix report to the same report file, and returns the short contract. Before re-dispatching the reviewer, confirm the fix report contains the covering tests, the command run, and the output; dispatch the re-review once all three are present. Name the covering test files in the fix message — a one-line fix does not need the whole suite.

The re-review is scoped. Run scripts/review-package FIX_BASE HEAD where FIX_BASE is the head the previous review saw, and dispatch re-review-prompt.md with the findings list, the brief, the report file, and the printed diff path. The re-reviewer verdicts each finding ADDRESSED or NOT ADDRESSED and flags new breakage in the fix diff only. New Critical/Important breakage in the fix diff joins the open findings list. Out-of-scope observations go to the task bead as deferred minors — they never extend the loop.

After each round, append to the task bead: bd note <task-id> "Fix round <R>/5: <X> addressed, <Y> open — <finding one-liners>; commits <a7>..<b7>"

Never fix findings yourself in the controller session — your context stays clean for coordination, and controller fixes skip review.

The breaker. When round 5's re-review still leaves findings open, stop dispatching fixes. Trigger 3 has fired: dispatch the adjudicator once with every open finding as one question packet (see Escalation). You route each ruling it returns — you hold the plan and the cross-task context it lacks:

• The reviewer is wrong, or the point is contestable: park it — bd note <task-id> "Parked: <finding> — Ruling: <why the code stands>". The final review sees both sides.
• Real, but nothing downstream builds on it: park it the same way, with a ruling that says it's real and deferred.
• Real and load-bearing — a later task builds on it, or it reveals a plan defect: rule on the smallest change that unblocks the dependent work, record it (bd note <task-id> "Ruling: <finding> — <what you decided and why>"), and carry it into the next task's dispatch. Parking a structural failure silently lets every dependent task build on it. Stop only when the defect leaves every path forward a guess.

Escalate only at the cap. Escalating earlier to end a loop is pre-judging with a different name. Every ruling is a bd note — a silent discard is forbidden.

5. Complete the task

When the review comes back clean — or every open finding is parked with a ruling at the cap — close the task bead with the commit range as the reason:

• bd close <task-id> --reason "commits <base7>..<head7>, review clean"
• bd close <task-id> --reason "commits <base7>..<head7>, <K> parked" after a tripped breaker

Never close a task before the reviewer's report shows spec ✅ (or every ⚠️ resolved) and task quality Approved. 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 — an overlapped dispatch (see The Task Loop) is allowed only while N's review has returned nothing yet and none is expected; the moment a finding lands, N+1 is the last task you dispatch until N is clean.

6. Roll up the hierarchy

When every task bead under a feature is closed, close the feature (bd close <feature-id> --reason "All tasks complete"); when every feature is closed, close the epic.

Run bd children <id> before every parent close and refuse while any row is not ✓. It prints the whole subtree, so it catches a grandchild too. Nothing else does: bd close accepts a parent over open children silently and exits 0 (verified, bd 1.1.2 — --force's "unsatisfied gates" are formula-step gates, not parent/child), and bd epic status / bd epic close-eligible count DIRECT children only — an epic whose one feature is closed reports "1/1 children closed, Eligible for closure" while that feature's own task is still open. Closing a parent over an open child hides unfinished work; four such closes are what let a whole task layer ship undelivered.

Final Review

The final whole-branch review gets a package too: run scripts/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 model: "fable" (top tier; if Fable is not in this session's roster, use the top tier that is) using requesting-code-review's code-reviewer.md. Point it at the deferred-minor and parked notes on the epic's closed tasks 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 (scripts/review-package FIX_BASE HEAD, re-review-prompt.md). Adjudicate any residual findings as in the task loop's breaker: park with rulings on the relevant beads, or rule on the load-bearing ones and record what you decided. Only the four stop classes above stop you here. There is no second fix wave — residual load-bearing findings surface to your human partner when finishing-a-development-branch presents the options.

Finish

Before you delete anything, collect every ruling from bd — preflight rulings, parked findings, breaker and adjudicator rulings, all of them: walk the epic's hierarchy and grep the notes for Ruling: (bd show each closed task, or bd children <epic-id> --json and read the notes fields) — 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 a bd note holds a ruling, the list holds it. Design/plan-conflict rulings lead the list, individually flagged — each with the finding, the plan text it collided with, and which you ruled governs. Those are the rulings your human partner most needs to see: they overrode plan text on their behalf, and the summary is where they understand what shifted and direct changes or rework. A ruling left only in a bead was a decision made in secret.

When the final whole-branch review is clean and its fixes are merged, delete the SDD workspace (rm -rf "$(git rev-parse --show-toplevel)/.joe-bag-of-tricks/sdd") — bd and git history are the record now.

Dispatch the joe-bag-of-tricks:branch-shepherd agent with this one branch (its worktree path, and the PR number if one already exists) to run finishing-a-development-branch's tail — push, PR, CI, CodeRabbit, squash-merge, cleanup — unattended, rather than conducting it yourself step by step. Dispatch it in the background and keep working while it runs; it reports back one outcome line for the branch.

For interactive, step-by-step delivery instead — watching each stage and deciding as you go — use finishing-a-development-branch directly; its procedure is what branch-shepherd runs on your behalf.

Common Rationalizations

Example Workflow

You: I'm using Subagent-Driven Development to execute this plan.

[Setup: worktree verified]
[bd children <epic-id> --pretty — 2 features, 5 tasks, Total: 8 issues across
 3 levels; none closed, fresh start]
[Pre-flight scan noted on the epic: "Scan: 8 issues across 3 levels — ..."]

--- Feature 1 (bd-feat1): Hook system ---

[Load feature tasks: bd children bd-feat1 --json]

Task 1 (bd-abc): Hook installation script
  Complexity: 1-2 files, clear spec with code snippets → implementer-mechanical

[bd update bd-abc --claim; record BASE]
[Run scripts/task-brief bd-abc; prints .../task-bd-abc-brief.md]
[Dispatch joe-bag-of-tricks:implementer-mechanical with brief path + report path + context]

Implementer: "Before I begin - should the hook be installed at user or system level?"

You: "User level (~/.local/share/my-app/hooks/)"

Implementer: [Later]
  - Implemented install-hook command
  - Added tests, 5/5 passing
  - Self-review: Found I missed --force flag, added it
  - Committed
  - Wrote full report to .../task-bd-abc-report.md; returned a <15-line status summary

[Run scripts/review-package BASE HEAD; dispatch task reviewer (model: sonnet)
 with the brief, report, and diff-package paths]
Task reviewer: Spec ✅ - all requirements met, nothing extra.
  Strengths: Good test coverage, clean. Issues: None. Task quality: Approved.

[bd close bd-abc --reason "commits abc123f..def456a, review clean"]

Task 2 (bd-def): Recovery modes
  Complexity: multi-file, integration with hook system → implementer

[bd update bd-def --claim; task-brief; dispatch joe-bag-of-tricks:implementer]
Implementer: reports DONE, discovered work: "Found edge case in error path"
[bd create --title="Edge case in error path" ... --deps discovered-from:bd-def]
[Run review-package BASE HEAD; dispatch task reviewer (model: sonnet)]
Task reviewer: Spec ❌:
  - Missing: Progress reporting (spec says "report every 100 items")
  Issues (Important): Magic number (100)

[Fix round 1: resume the implementer with both findings]
Implementer: Added progress reporting, extracted PROGRESS_INTERVAL constant.
  Re-ran test/recovery.test.js — 10/10 passing. Fix report appended.

[Run review-package FIX_BASE HEAD; dispatch scoped re-review (model: haiku)]
Re-reviewer: Missing progress reporting — ADDRESSED (src/recovery.js:41).
  Magic number — ADDRESSED (src/recovery.js:7). New breakage: none.
  Verdict: all findings addressed.

[bd note bd-def "Fix round 1/5: 2 addressed, 0 open; commits d4e5f6a..b7c8d9e"]
[bd close bd-def --reason "commits ghi789b..jkl012c, review clean"]

[bd children bd-feat1 — every row ✓; bd close bd-feat1 --reason "All tasks complete"]

--- Feature 2 (bd-feat2): Verification ---
...
[bd children <epic-id> — all 8 rows ✓; bd close <epic-id> --reason "All features complete"]

[Run scripts/review-package MERGE_BASE HEAD; dispatch final code reviewer
 (model: fable; top-available-tier fallback if not in roster) with the printed path —
 requesting-code-review's code-reviewer.md]
Final reviewer: All requirements met. Deferred minors triaged: none block merge.

[Delete the SDD workspace — the record now lives in bd and git]

Done! Using finishing-a-development-branch.

Integration

Required workflow skills:

• using-git-worktrees - REQUIRED: Set up isolated workspace before starting
• writing-plans - Creates the bd task hierarchy this skill executes
• requesting-code-review - Code review template for the final whole-branch review
• finishing-a-development-branch - Reference procedure for delivery; run directly for interactive delivery, or unattended via branch-shepherd (see Dispatches)

Subagents should use:

• test-driven-development - Subagents follow TDD for each task

Dispatches:

• implementer-mechanical / implementer / implementer-complex agents (haiku / sonnet / opus) - One per task, chosen per Model Selection; also the fix-loop and BLOCKED escalation ladder
• branch-shepherd agent (sonnet) - Runs the finishing-a-development-branch tail unattended for the finished branch (Finish)
• adjudicator agent (fable) - One-shot ruling when an escalation trigger fires (Escalation)

Alternative workflow:

• executing-plans - Use for parallel session instead of same-session execution
• dispatching-parallel-agents - Compose with this skill when an epic's independent tasks run inside a parallel batch: each parallel item can itself be an SDD task-brief