Subagent Driven Development

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

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

Setup

Ensure the work happens in an isolated workspace: use superpowers: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.

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 plan owns a workspace: at skill start, run this skill’s bash scripts/sdd-workspace PLAN_FILE — it prints the plan’s git-ignored directory (under <repo-root>/.superpowers/sdd/), home to every artifact for THIS plan: ledger, briefs, reports, review packages. Another plan’s directory is never yours to read or write.
• Check for this plan’s ledger at <workspace>/progress.md. If its first line names your plan file, tasks with a Task <N>: 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 plan file — or a stray ledger at the old flat path .superpowers/sdd/progress.md — is another plan’s progress: leave it in place and start your own, fresh.
• Create the ledger with its identity as the first line: # SDD ledger — plan: <plan file path>.
• 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); if that happens, recover from git log.

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 Task 1, 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 Task 1. 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 (git rev-parse HEAD) before dispatching — the review package and fix-round diffs need it.

• Task brief: before dispatching an implementer, run this skill’s bash scripts/task-brief PLAN_FILE N — it extracts the task’s full text to 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 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 (opens in a new tab)

2. Handle the report

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

DONE: Generate the review package (bash scripts/review-package PLAN_FILE 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 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 this skill’s bash scripts/review-package PLAN_FILE 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.
• 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 (opens in a new tab)

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 in the progress ledger as you go (Task <N>: minor (deferred): <one-liner>), and point the final whole-branch review at that list 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 plan’s text requires — is yours to rule on: weigh the finding against the plan text, decide with the spec as the binding authority, and ledger the ruling 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 your harness cannot send another message to a 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.

Rounds 4-5 — dispatch a fresh implementer on a more capable model (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 bash scripts/review-package PLAN_FILE FIX_BASE HEAD where FIX_BASE is the head the previous review saw, and dispatch re-review-prompt.md (opens in a new tab) 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 ledger as deferred minors — they never extend the loop.

After each round, append to the ledger: Task <N>: 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. Adjudicate each open finding yourself — you hold the plan and the cross-task context the reviewer lacks:

• The reviewer is wrong, or the point is contestable: park it — Task <N>: 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, ledger it as Task <N>: 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.

Adjudicate only at the cap. Adjudicating earlier to end a loop is pre-judging with a different name. Every adjudication is a ledger entry — 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 — append the completion line to the ledger in the same message as your other bookkeeping:

• Task <N>: complete (commits <base7>..<head7>, review clean)
• Task <N>: complete (commits <base7>..<head7>, <K> parked) after a tripped breaker

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 bash scripts/review-package PLAN_FILE 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 superpowers:requesting-code-review’s code-reviewer.md (opens in a new tab). 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 (bash scripts/review-package PLAN_FILE FIX_BASE HEAD over the fix range, re-review-prompt.md (opens in a new tab)). 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 finishing-a-development-branch 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 plan’s workspace (rm -rf <workspace>) — the git history is the record now. Sibling directories belong to other plans; leave them alone.

Use superpowers:finishing-a-development-branch.

Common Rationalizations

Example Workflow