effort-sizing

Complexity Scoring

What this does

effort-sizing is a gate, not a scorer. It never computes a complexity score itself. Its entire job is: decide whether the caller needs a score at all, dispatch the sizing-analyst agent when it does, coerce whatever comes back into a fixed schema, and hand the caller a routing verdict. Treat every line of judgment about how complex a task is as belonging to the agent — this skill only formats and routes.

When to invoke

Call this skill only when the caller's own heuristic-first routing logic cannot confidently classify a task on its own. The decision of whether to invoke this skill at all belongs to the caller, not to this skill.

• sdlc-engine calls this skill as a standing step in its flow.
• sdlc-brief and sdlc-direct call this skill only as a fallback, when their own lightweight heuristics fail to produce a confident route.

Do not invoke this skill to re-check a routing decision the caller has already made with confidence, and once invoked, do not re-judge the verdict it returns — see Ownership boundary below.

Inputs

Field	Source	Notes
task_description	Verbatim from Phase 1 (story-intake) output	Pass through unmodified — do not summarize or rewrite it
feature_area	Short keyword summary	e.g. "provider integration SSO"
repo_path	Absolute path to the current checkout	Gives the agent filesystem context for its estimate

Operating steps

1. Confirm the gate applies. This skill assumes the caller already decided its own heuristics couldn't resolve the route. Do not second-guess that decision — proceed straight to dispatch.
2. Dispatch the agent. Invoke sizing-analyst via the Agent tool with:

   subagent_type='sizing-analyst'
   task_description='<input>'
   feature_area='<input>'
   

3. Validate the response against the output schema below. Confirm score is an integer 6–36, breakdown has all six sub-scores each 1–6 and summing to score, and routing is one of the three allowed literals.
4. On a malformed response, retry once with a stricter prompt that restates the exact schema and forbids prose outside the JSON object.
5. On a second malformed response, do not error. Fall back to a fixed default so the caller can keep moving:
   • score: 18
   • routing: "brainstorming"
   • rationale: state plainly that the fallback fired and why (agent output did not parse after retry)
6. Check for the split signal. If the agent's response carries the SPLIT REQUIRED flag, override routing to split-required regardless of the numeric score, and halt — see Routing outcomes below.
7. Map score to routing using the thresholds table, unless step 6 already forced split-required.
8. Write the result to <run_dir>/complexity.json. This is an R1 (run-artifact write) operation — no repository files are touched, no external side effects occur.
9. Return the normalized {score, breakdown, routing, rationale} object to the caller. Do not add commentary, re-score, or re-route what the agent produced or what the fallback set.