fix(sdd): batch small same-shape tasks into one dispatch

[obra]

Who is submitting this PR? (required)

Field	Value
Your model + version	Claude Fable 5 (claude-fable-5)
Harness + version	Claude Code 2.1.220
All plugins installed	superpowers 6.2.0, episodic-memory, linear, context7, superpowers-chrome, agent-sdk-dev, code-simplifier, github-triage, plugin-dev
Human partner who reviewed this diff	Jesse Vincent (@obra) — directed the treatment, the eval program, and this submission; reviewing the diff here

What problem are you trying to solve?

Subagent-driven-development pays a fixed overhead per task: a fresh implementer dispatch plus a task-review dispatch, each carrying the brief, the diff package, and a full agent context. That price is right for tasks that need their own judgment and test cycle. But real plans sometimes enumerate many tiny, same-shape edits — the same one-line fix across N files, a constant change, a field addition — as separate tasks, and the current skill text dispatches one implementer + one review per micro-task. In our mined-session baseline (174 SDD sessions), the median dispatch consumes roughly 79,500× more tokens in agent overhead than the useful output it produces; micro-task plans are where that ratio does the most damage.

What does this PR change?

skills/subagent-driven-development/SKILL.md (+8 lines, one file): a rule at the top of the Task Loop — when the plan lists several small, independent, same-shape edits, compose ONE dispatch brief listing every file and its change, send the batch to a single subagent, and review the result as one diff. The rule's own boundary reserves per-task dispatch for anything needing its own judgment, its own tests, or its own review surface.

Is this change appropriate for the core library?

Yes — it modifies core SDD controller behavior that applies to any project using subagent-driven-development. No dependencies, no domain-specific content.

What alternatives did you consider?

• Fix it at plan-authoring instead (writing-plans right-sizes tasks so micro-task plans never exist): registered as a competing hypothesis in our eval program and worth pursuing — but SDD also executes plans it did not author (imported, legacy, human-written), so the execution-side guard retains value regardless. Defense in depth, not either/or.
• Cheaper models per micro-task (already in the skill's Model Selection): reduces per-dispatch price but keeps the seat count — every micro-task still costs an implementer + reviewer round-trip with full context. Measured effect is much smaller than removing the seats.
• Do nothing: the mined baseline above says this is among the largest recurring cost shapes in real SDD sessions.

Does this PR contain multiple unrelated changes?

No — one file, one rule.

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related PRs: fix(sdd): dispatched subagents never dispatch subagents #2059–docs: codex-efficiency fix-cycle spec and plan (campaign record) #2064 and fix(sdd): rule and continue — non-catastrophic conflicts get ledgered rulings, not blocking questions #2077 (same eval program, complementary SDD fixes — this diff applies cleanly to dev independent of all of them; merge order is free). No open or closed PR adds dispatch batching to SDD.

Environment tested

Harness (e.g. Claude Code, Cursor)	Harness version	Model	Model version/ID
Claude Code (authoring)	2.1.220	Claude Fable 5	claude-fable-5
Codex CLI (containerized evals, quorum harness)	0.46 lineage	GPT-5.6-Codex	gpt-5.6-codex

New harness support (required if this PR adds a new harness)

N/A — no new harness.

Evaluation

Two pre-registered batteries, containerized quorum evals (fresh clone per rep, pinned scripted user replies, mechanical scoring):

• Where batching SHOULD bind — a 12-micro-task plan (same-shape small edits): treatment cut cost 73% and dispatches 87% vs control, with better completion (control reps dropped edits; the batched pass completed them all).
• Where batching must NOT bind (the ship gate) — a 5-task, 1100-line Go plan whose tasks are individually non-trivial (multi-file renderers with test cycles): treatment dispatch counts 14/14/13 vs control 16/13, i.e. zero batching; completion parity (5/5 reps pass, 7/7 post-checks each); cost parity ($7.96 vs $8.49 mean). The boundary language held 3/3.
• Full records (public): https://github.com/prime-radiant-inc/superpowers-autoresearch — logs/2026-08-01-queue-campaign.md (Task 7 fixture + MINE baseline, Task 12 battery verdict, 2026-08-02 ship-gate pre-registration and verdict) and reports/2026-08-queue-campaign.md §2.

https://claude.ai/code/session_0185AJr98gHx5EmwqNeft4Sy

[obra]

@arittr — as with #2077, we'd value an independent eval run on this one. Three behaviors worth testing (this is Claude, writing at Jesse's direction):

1. The batching case: an SDD plan of many tiny same-shape tasks (e.g. the same one-line fix across 10+ files as 10+ tasks). Success = ONE implementer dispatch covering the batch, all edits actually landing, and one review of the combined diff. Failure modes to watch: edits silently dropped from the batch, or the controller still dispatching per-task.
2. The boundary: a plan of individually non-trivial tasks (own tests, multi-file, judgment). Success = zero batching — dispatch counts identical to what you'd see without this text. In our gate battery (5-task Go plan) treatment matched control 14/14/13 vs 16/13.
3. Review integrity under batching: the batched diff must still get a genuine task review as one unit — batching the implementation must not become an excuse to skip or thin the review.

Full eval records are public in the autoresearch repo (linked in the PR body): the micro-task battery (73% cost / 87% dispatch reduction, better completion) and the boundary gate (3/3 no-batching, full parity).

[arittr]

LGTM

[arittr]

Independent validation from my environment, as requested — posting post-merge for the record. Setup: quorum A/B (3 reps × 7 scenarios, codex lane, dev vs the full 7-PR stack, both arms on the same binary — host codex 0.144.4, so V2-specific claims were validated separately in live probes on codex-cli 0.146.0 with subscription auth), plus a fresh 2-task SDD run on 0.146. Full numbers below for this PR.

This was the one PR in the set with no independent verification available — no should-bind scenario exists in any eval corpus we have. What we can say: across the full A/B, no false batching on non-trivial plans (dispatch structure at parity with dev), and the dropped-edit net you named as the top failure mode is now closed at the reviewer seat (78cc1892: batch reviews reconcile the diff against the brief's file list; a listed file with no hunk is a Missing finding).

Missing eval asset if you want the bind rate measured: a should-bind battery (10–15 same-shape micro-tasks → expect 1 implementer + 1 review, every edit present, one deliberately wrong edit spec that the review must catch) plus a must-not-bind boundary battery (individually non-trivial tasks → expect zero batching).

Validation performed by Claude Fable 5 (Claude Code) at my direction.