sdd-triage over-decomposes: per-task PR overhead exceeds the work delivered

[norrietaylor]

Problem

sdd-triage decomposes a feature tracking issue into task sub-issues at the granularity of a single function or component. Each task becomes one dispatched sdd-execute run, one PR, one CI cycle, one sdd-validate pass, one sdd-review pass, and one merge decision. When the resulting change is 40–90 lines across 2–4 files, the fixed per-task pipeline cost dwarfs the implementation.

Evidence

A representative feature was split into 6 tasks. Three of the resulting implementation PRs:

PR diff	Files	What it did
+98 / -0	2	add one helper function + its unit test
+97 / -0	4	add one component + one helper + i18n keys + test
+70 / -7	3	wire the above into one tab

These three are a single cohesive change (~265 lines) split across three task issues, three agent runs, three PRs, three CI cycles, three reviews, three merges. The triage pass that produced the split consumed ~25M tokens; each sdd-execute run adds ~1M+ on top.

Root cause

The decomposition rule optimizes for task isolation — one file / one requirement per task, clean depends on edges — with no lower bound on task size and no accounting for the fixed per-PR cost. A requirement that maps to a 40-line diff still earns its own task, PR, and full validate/review/merge cycle.

Cost shape

Per task, incurred regardless of diff size:

• 1 dispatched sdd-execute run (agent tokens + CI minutes)
• 1 PR open + CI pipeline
• 1 sdd-validate pass
• 1 sdd-review pass (+ up to SDD_MAX_REVIEW_ITERATIONS revise cycles)
• 1 merge (human or auto)

For a sub-100-line PR this is mostly overhead. The ratio of pipeline cost to delivered diff is the metric to drive down.

Proposed direction

Give triage a task-sizing target and let it bundle sub-threshold work:

• Add a minimum-task-size / "merge small siblings" rule: tasks under a diff or file-count threshold that share a file set or a dependency chain collapse into one task.
• Decompose by cohesive unit of review — a slice a human would review in one sitting — rather than one task per function or file.
• Expose a triage granularity knob (e.g. SDD_TRIAGE_MIN_TASK) so an operator can tune decomposition per repo.
• Route features whose entire scope is small down a single-PR fastpath instead of a spec → triage → N-task fan-out.

Acceptance

• A feature whose total scope is ~250 lines across a handful of files produces 1–2 tasks, not 6.
• The sizing rule is documented in sdd-triage and its threshold is configurable.
• No regression in dependency correctness: bundled tasks still respect blocked by ordering.