Which Shape

/which-shape

Recommend an operating loop, not a single tool. /which-skill answers "which skill?" after the work shape is clear. /which-shape answers "what kind of work are we doing?"

This is advisory-only in v1. It never invokes the recommended skills.

Forms

Before invoking, check the conversation for the discriminating dimensions — recurrence? approved proposal in hand? unknown or inherited repo? scope width? durable choice? — and carry those words verbatim into the task string. The matcher is keyword-based, so the paraphrase IS the routing decision.

/which-shape "this inherited repo feels slow and chaotic"
/which-shape "unknown project; where should an agent start?"
/which-shape "this bug keeps coming back"

Running the script is mandatory — the recommendation must come from a real run, never composed from this file. Paste the script's Project context: and Confidence: ... (score=N) lines verbatim in your reply.

PROJECT_ROOT="$PWD"
(
  cd "$PROJECT_ROOT/.agents/skills/which-shape"
  python3 scripts/route.py \
    "this inherited repo feels slow and chaotic" \
    --project-root "$PROJECT_ROOT"
)

Use --json for machine-readable output and --skip-log for tests.

Treat the script output as a lexical prior, not the decision. If conversation evidence contradicts the cue match (negation, paraphrase, sarcasm), override it, say so, and name the cue collision in your reply so telemetry can see the override.

When a status projection exists (.engineering/local/status.json, ADR 0037 — generated by scripts/status.py), the rationale additionally cites up to three projection-derived signals (pending approvals, stale artifacts, staged queue items, in-flight plans). The live project-context read stays authoritative: a stale or malformed projection is dropped silently, and with the file absent the output is byte-identical to an ungrounded run. --status <path> overrides the default location.

Output

The recommendation includes:

• shape id and title;
• confidence and rationale;
• first next command;
• short loop sequence;
• stop/reassess condition;
• alternatives.
• an on-demand handoff for the first skill plus any companions declared by the shape, with exact local guide/tool paths, shared source inventory, validated language/fact/outcome capability rows for that closure, and a fresh non-context-sub-agent default;
• a pinned ambient-install command only as an explicit optional alternative when every selected closure member has passed install evidence; otherwise the output reports why that optional path is unavailable.

If the script reports confidence: low — or any rationale line reads "fallback shape candidate" — do not present a single shape. Present the top 2-3 alternatives by score and ask one discriminating question (see the pre-run dimensions under Forms) before routing.

Registry

The explicit shape registry lives in shapes.json — the sole shape inventory; this file deliberately does not mirror the list. Keep it small and loop-level. Do not mirror the whole skill catalog.

When adding or materially repurposing skills, run /check-ecosystem-consistency and review whether shapes.json needs an update. Add a skill to a shape only when it changes the operating loop; purely tactical skills can stay out after that review is captured in the ecosystem state.

Boost weights are registry data too: every shape declares a boost: block in shapes.json (simple cues/weight/rationale, or a small schema-validated rules form for conditional boosts; boost: {} is an explicit opt-out), plus the narrow_signal and context_exempt flags. The scorer holds no per-shape table — adding a shape never requires editing scripts/route.py. After editing shapes.json, run route.py --validate — it checks the schema, including every boost block.

Project Context

The router reads .engineering/project/adapter.yml, .engineering/project/profile.yml, and .engineering/project/open-questions.md when present.

Missing project context is a routing signal, not a universal blocker. Broad unknown-project prompts should route to project-intake; narrow typos and concrete bugs should still route directly.

Telemetry

The router logs event_kind: recommendation events to .claude/skill-use/log.jsonl with outcome: unscored by default — a recommendation is not evidence of usefulness at the moment it is made. Projection and compaction keep these separate from actual skill-run useful rates, and the compaction digest's overridden% counts only scored events.

The correction path runs when a human notices a misroute: rerun the same route with an explicit outcome (the rerun appends a second recommendation event; it does not amend the first):

cd .agents/skills/which-shape
python3 scripts/route.py \
  "this inherited repo feels slow and chaotic" \
  --outcome overridden \
  --human-override "wrong-shape: should have started with project-intake"

Relationship To /which-skill

Use /which-shape first when the operating mode is unclear. Use /which-skill once the shape is known and the question is tactical.

Bad pattern: asking /which-skill to decide how to onboard or stabilize an unknown repo. That is exactly what /which-shape exists to handle.