subagent-driven-dev

Subagent-Driven Development

Execute plan wave-by-wave. Within each wave, dispatch all implementer subagents in parallel inside the same feature-branch worktree, then run two-stage review (spec compliance, then code quality) per task. Move to the next wave only after every task in the current wave is approved and committed.

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.

Why parallel waves: Sequential execution of independent tasks wastes wall-clock time. The plan's ## Execution Waves summary already proves which tasks are safe to run concurrently — honor it.

Why one shared worktree (Pattern B): Per-agent worktrees add filesystem ceremony and a merge step. Instead, the wave's implementers all work in the feature-branch worktree on strictly disjoint file sets (enforced by the plan's Worktree-safe: declarations). To prevent git-index races, implementers never run git — they edit files and report changed paths; the controller stages and commits per task in a deterministic order after the wave finishes.

Core principle: Parallel-by-default within a wave + disjoint file sets + controller-only git = maximum throughput, zero races.

Isolation: Pattern B is the opt-out

This skill's wave-based controller-only-git approach is intentionally not worktree-isolated — it relies on provable file-set partitioning per wave. For any other parallel-agent dispatch (review fanouts, ad-hoc multi-agent edits), default to isolation: "worktree" on the Agent tool calls — see the uberdev:dispatching-parallel-agents skill.

When to Use

digraph when_to_use {
    "Have implementation plan?" [shape=diamond];
    "Tasks mostly independent?" [shape=diamond];
    "Stay in this session?" [shape=diamond];
    "uberdev:subagent-driven-dev" [shape=box];
    "uberdev:execute-plan" [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?" -> "uberdev:subagent-driven-dev" [label="yes"];
    "Stay in this session?" -> "uberdev:execute-plan" [label="no - parallel session"];
}

vs. uberdev:execute-plan (parallel session):

• Same session (no context switch)
• Fresh subagent per task (no context pollution)
• Two-stage review after each task: spec compliance first, then code quality
• Faster iteration (no human-in-loop between tasks)

The Process

High-Level Flow

1. Read plan once. Extract every task's full text and the ## Execution Waves summary.
2. Create TodoWrite with one todo per task, labeled with its wave (e.g., [wave-2] Task 4: ...).
3. Verify clean baseline: git status is clean; you're on the feature branch in the feature-branch worktree.
4. For each wave (sequential): a. Dispatch every implementer in the wave in a single message with multiple Agent tool calls. All run in the current worktree. Each implementer gets an explicit allowlist of files it owns and an explicit denylist of files owned by sibling tasks. b. Implementers never run git. They edit files, run their tests, and report Status + changed file paths + test results. c. Wait for all wave implementers to report. d. For each completed implementer (in task ID order, sequential): controller stages only that task's reported paths with git add <paths> and commits with the task-specific message. e. Run the project's full test command in the worktree once after all wave commits land. If it fails, identify which task regressed and re-dispatch that task's implementer with the failure context. Re-test until green. f. For each committed task: dispatch spec reviewer (parallel across the wave). g. Loop spec fix-up per task until all spec reviewers approve. Fix dispatches still don't run git — controller amends the task's commit (or creates a fix-up commit) using the implementer's reported new paths. h. Dispatch code quality reviewers (parallel). Same fix-loop pattern. i. Mark every task in the wave complete in TodoWrite.
5. After the final wave, dispatch the whole-implementation final reviewer.
6. Hand off to uberdev:finish-branch.

Parallel Dispatch Pattern

[wave-1] →  Agent(T1, edits files only)  ┐
            Agent(T2, edits files only)  ├─ all in ONE message, shared CWD
            Agent(T3, edits files only)  ┘
                ↓ wait for all three
            controller: git add <T1 paths> && git commit  (sequential, deterministic)
            controller: git add <T2 paths> && git commit
            controller: git add <T3 paths> && git commit
            controller: run full test suite
                ↓
            spec reviewers (parallel) → fix loop → re-reviews
            quality reviewers (parallel) → fix loop → re-reviews
                ↓ no merge step — already on feature branch
[wave-2] →  Agent(T4, edits files only)  ┐
            Agent(T5, edits files only)  ┘  (parallel, depend on wave-1 commits)
            ...

File-Ownership Enforcement

Before dispatching a wave, build the wave's ownership map:

T2 owns: src/recovery.ts, tests/recovery.test.ts
T3 owns: src/progress.ts, tests/progress.test.ts
T4 owns: src/telemetry.ts, tests/telemetry.test.ts

Every implementer prompt receives its own allowlist + the union of sibling-owned paths as a denylist. If two tasks claim the same file, the wave decomposition is wrong — bump one to the next wave before dispatching.

Per-Task Inner Loop (unchanged)

digraph per_task {
    rankdir=TB;
    "Implementer (in worktree)" [shape=box];
    "Implementer questions?" [shape=diamond];
    "Answer & re-dispatch" [shape=box];
    "Spec reviewer" [shape=box];
    "Spec OK?" [shape=diamond];
    "Implementer fixes spec" [shape=box];
    "Code quality reviewer" [shape=box];
    "Quality OK?" [shape=diamond];
    "Implementer fixes quality" [shape=box];
    "Task complete" [shape=box style=filled fillcolor=lightgreen];

    "Implementer (in worktree)" -> "Implementer questions?";
    "Implementer questions?" -> "Answer & re-dispatch" [label="yes"];
    "Answer & re-dispatch" -> "Implementer (in worktree)";
    "Implementer questions?" -> "Spec reviewer" [label="no"];
    "Spec reviewer" -> "Spec OK?";
    "Spec OK?" -> "Implementer fixes spec" [label="no"];
    "Implementer fixes spec" -> "Spec reviewer";
    "Spec OK?" -> "Code quality reviewer" [label="yes"];
    "Code quality reviewer" -> "Quality OK?";
    "Quality OK?" -> "Implementer fixes quality" [label="no"];
    "Implementer fixes quality" -> "Code quality reviewer";
    "Quality OK?" -> "Task complete" [label="yes"];
}

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, design, and review tasks: use the most capable available model.

Task complexity signals:

• 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

Handling Implementer Status

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

DONE: Proceed to spec compliance review.

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, escalate to the human

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

Prompt Templates

• ./implementer-prompt.md - Dispatch implementer subagent
• ./spec-reviewer-prompt.md - Dispatch spec compliance reviewer subagent
• ./code-quality-reviewer-prompt.md - Dispatch code quality reviewer subagent

Example Workflow

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

[Read plan file once: docs/uberdev/plans/feature-plan.md]
[Extract all 5 tasks + Execution Waves summary:
   wave-1: T1 (schema)
   wave-2: T2, T3, T4 (parallel — different files)
   wave-3: T5 (depends on T2,T3,T4)
]
[Create TodoWrite labeled by wave]

=== WAVE 1 ===

Task 1: Hook installation script (alone in wave-1)
T1 owns: scripts/install-hook.sh, tests/install-hook.test.sh

[Dispatch implementer in shared worktree, full task text + allowlist + "no git commands"]

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

You: "User level (~/.config/uberdev/hooks/)"

Implementer: "Got it. Implementing now..."
[Later] Implementer:
  - Edited scripts/install-hook.sh, tests/install-hook.test.sh
  - Tests 5/5 passing
  - Self-review: Found I missed --force flag, added it
  - Status: DONE — paths: [scripts/install-hook.sh, tests/install-hook.test.sh]

[Controller: git add scripts/install-hook.sh tests/install-hook.test.sh && git commit -m "feat: install-hook script"]
[Run full test suite — green]

[Dispatch spec compliance reviewer]
Spec reviewer: ✅ Spec compliant - all requirements met, nothing extra

[Dispatch code quality reviewer]
Code reviewer: Strengths: Good test coverage, clean. Issues: None. Approved.

[Mark Task 1 complete]

=== WAVE 2 ===

Ownership map:
  T2 owns: src/recovery.ts, tests/recovery.test.ts
  T3 owns: src/progress.ts, tests/progress.test.ts
  T4 owns: src/telemetry.ts, tests/telemetry.test.ts

[Single message, three Agent calls in parallel — same shared worktree, no git permitted]
  Agent(T2: Recovery modes,    allow=[T2 paths], deny=[T3+T4 paths])
  Agent(T3: Progress reporting, allow=[T3 paths], deny=[T2+T4 paths])
  Agent(T4: Telemetry hooks,    allow=[T4 paths], deny=[T2+T3 paths])

[Wait for all three implementers to report back with their changed paths]

[Controller, sequential — one commit per task in task ID order]
  git add <T2 paths> && git commit -m "feat: recovery modes"
  git add <T3 paths> && git commit -m "feat: progress reporting"
  git add <T4 paths> && git commit -m "feat: telemetry hooks"

[Run full test suite — green]

[Single message: spec reviewers for T2, T3, T4 in parallel]
[Loop: any failed spec review → re-dispatch that task's implementer (no git); controller amends or fix-up commits using reported paths; re-review until ✅]

[Single message: code quality reviewers for T2, T3, T4 in parallel]
[Loop: any failed quality review → same fix pattern → re-review until ✅]

[Mark Tasks 2, 3, 4 complete]

=== WAVE 3 ===

[T5 alone — depends on wave-2 commits being on the branch]
[Dispatch implementer in shared worktree]
[Controller commits → run full suite → spec review → fix loop → quality review → fix loop → mark complete]

=== FINAL PASS ===

[Dispatch whole-implementation final reviewer]
Final reviewer: All requirements met, ready to merge.

[Hand off to uberdev:finish-branch]

Advantages

vs. Manual execution:

• Subagents follow TDD naturally
• Fresh context per task (no confusion)
• Parallel-safe (subagents don't interfere)
• Subagent can ask questions (before AND during work)

vs. uberdev:execute-plan:

• Same session (no handoff)
• Continuous progress (no waiting)
• Review checkpoints automatic

Efficiency gains:

• No file reading overhead (controller provides full text)
• Controller curates exactly what context is needed
• Subagent gets complete information upfront
• Questions surfaced before work begins (not after)

Quality gates:

• Self-review catches issues before handoff
• Two-stage review: spec compliance, then code quality
• Review loops ensure fixes actually work
• Spec compliance prevents over/under-building
• Code quality ensures implementation is well-built

Cost:

• More subagent invocations (implementer + 2 reviewers per task)
• Controller does more prep work (extracting all tasks upfront)
• Review loops add iterations
• But catches issues early (cheaper than debugging later)

Red Flags

Never:

• Start implementation on main/master branch without explicit user consent
• Skip reviews (spec compliance OR code quality)
• Proceed with unfixed issues
• Dispatch multiple implementers without explicit file allowlists/denylists — they will trample each other's edits
• Let implementer subagents run any git command (add, commit, stash, restore) — that's the controller's job
• Use git add -A or git add . to stage a task's commit — always pass the implementer's reported paths explicitly
• Dispatch implementers from different waves in parallel — wave-N depends on wave-(N-1) being committed first
• Skip the post-wave full-test-suite run — without it, a regression introduced by parallel edits hides until much later
• Run tasks sequentially when the plan declares them in the same wave (defeats the whole point)
• Make subagent read plan file (provide full text instead)
• Skip scene-setting context (subagent needs to understand where task fits)
• Ignore subagent questions (answer before letting them proceed)
• Accept "close enough" on spec compliance (spec reviewer found issues = not done)
• Skip review loops (reviewer found issues = implementer fixes = review again)
• Let implementer self-review replace actual review (both are needed)
• Start code quality review before spec compliance is ✅ (wrong order)
• Move to next task while either review has open issues

If subagent asks questions:

• Answer clearly and completely
• Provide additional context if needed
• Don't rush them into implementation

If reviewer finds issues:

• Implementer (same subagent) fixes them
• Reviewer reviews again
• Repeat until approved
• Don't skip the re-review

If subagent fails task:

• Dispatch fix subagent with specific instructions
• Don't try to fix manually (context pollution)

Integration

Required workflow setup (run before this skill):

• Isolated worktree — git worktree add .worktrees/<feature-name> -b <branch-name> (verify .worktrees/ is in .gitignore; add and commit if not). Run the project's setup command (npm install / cargo build / pip install -r requirements.txt / go mod download) and the project's test command to verify a clean baseline before starting.

Related skills:

• uberdev:write-plan — creates the plan this skill executes
• uberdev:execute-plan — alternative for parallel-session/inline execution

Subagents follow TDD discipline within each task: write a minimal failing test for the new behavior FIRST, run it to see it fail for the expected reason, write the simplest code that makes it pass, run again to see green, then refactor while green. The implementer-prompt.md template enforces this; the spec and code-quality reviewers verify it was actually applied.

Code review dispatch: the code-quality reviewer in this skill dispatches the bundled uberdev:code-reviewer agent (see plugins/uberdev/agents/code-reviewer.md) — no separate "requesting-code-review" skill is needed, the agent's own prompt encapsulates the review template.

Finishing the development branch: after all tasks pass review and the final-pass code review approves, invoke uberdev:finish-branch to verify tests, present the 4-option close-out (merge / PR / keep / discard), execute the chosen one, and clean up the worktree.