Research — S-C (Mission-Type Creatability via Rich Step Model)

Consolidated grounding for #2724. Two squads fed this mission: a 3-lens content/architecture grounding squad (researcher-robbie, architect-alphonso, paula-patterns) and a 3-lens adversarial verification squad (paula-patterns, architect-alphonso, reviewer-renata) that stress-tested the direction change. All lenses were profile-loaded and read-only.

1. Direction change: retire template_set, don't keep projecting to it

S-B shipped template_set as a projected read-field on MissionType ("reuse the existing mechanism to limit blast radius"). The operator challenged this on reflection; four theories, all confirmed in code:

1. Partial missiontypes-as-doctrine. Epic #2721 and S-B's own code both label template_set "transitional." Authoring is graph-backed (step.yaml); consumption still reads a flattened dict, not the authority. Half-cutover. 2. Model collision. Dict MissionType.template_set (6 sites) vs scalar doctrine.template_set (40+ sites, charter/*), co-located in specify_cli/runtime/ (resolver.py:395 dict vs doctor.py:149/prompt_builder.py:441 scalar), separated only by a prose C-008 fence (docstrings). 3. Semantic confusion. MissionType exposes three views of one structure: steps (authority) + action_sequence (list projection) + template_set (dict projection). 4. Split-brain from contributions. The live raw-YAML fallback mission_type_repository.py:200-201 (raw.get("template_set")) + the still-authorable field models.py:259 would silently honor a community/pack-authored template_set whenever steps lack refs. The exact two-surface split-brain S-B set out to kill — dormant, not dead.

Key structural finding: the dict template_set has one behavioral consumer — resolve_configured_template (resolver.py:403) — and the flattening happens inside a thunk (_resolve_template_set_slot, mission_type_profiles.py:744) that already re-fetches the mission from the repository. The resolver never needed the flattened field.

2. Grounding squad — content + step topology (researcher-robbie)

Per-type step topology (each on its OWN names; do NOT assume specify/plan shape):

  • documentation: discover → audit → design → generate → validate → publish → accept (7 sequence) + retrospect (in_action_sequence: false, excluded).
  • research: scoping → methodology → gathering → synthesis → output (5 sequence) + retrospect (excluded). (research/mission-runtime.yaml also lists accept, but there is no mission-steps/research/accept/ dir — not part of the 16.)
  • plan: specify → research → plan → review (4 sequence). Name-collides with software-dev but is decomposition/decision (no code) — highest authoring cost.

16-file acceptance surface (retrospect ×2 excluded): all 18 mission-steps/<type>/<step>/prompt.md are 0 bytes; the 16 in scope are documentation(7) + research(5) + plan(4).

Guard tests/doctrine/missions/test_prompt_emptiness.py is mutated, not just satisfied:

  • _SEEDED_BLANK_STEPS (16 tuples) + _is_empty_or_dummy (flags zero-bytes / whitespace / TODO/PLACEHOLDER/FIXME).
  • TestSeededBlankPromptsAreNamedGaps marks each xfail(strict=False) → filling flips to XPASS (signal, not hard pass).
  • test_seeded_prompt_is_zero_bytes asserts st_size == 0 with no xfail → hard-RED on the first byte written. S-C must edit this file: shrink _SEEDED_BLANK_STEPS, drop each xfail, decrement the golden 16 (# golden-count: cardinality-is-contract), update _SEQUENCE_STEPS_BY_TYPE.

Content disposition: prompt.md ≠ guidelines.md. documentation/research ship rich guidelines.md → promote into the software-dev prompt shape. plan has no guidelines.md and empty templates/ → author-fresh (+ author its scaffold templates). DD-06 rationale: kitty-specs/mission-step-authority-01KXNZMT/traces/design-decisions.md:57-67.

3. Grounding squad — DRG + resolver architecture (architect-alphonso)

  • Node/URN convention (mirror exactly): template:<kebab> for #2712's 16 exemplars (template.graph.yaml, currently edges: []); mission-qualified template:<mission>/<name> via template_catalog.template_urn (module exists, not wired into generate_graph). Steps are action:<type>/<step> nodes (no new NodeKind).
  • Graph-back seam: new pass paralleling extract_mission_type_edges (extractor.py:864), minting from the step projection (deterministic) not discover_templates (filesystem scan → freshness flap). Relation.INSTANTIATES already exists (drg/models.py:82). Instantiates edges are action:-sourced → land in action.graph.yaml; template.graph.yaml stays nodes-only. Partition already modeled by the unit test test_extractor.py:822-878.
  • DRG delta (intentional): +N nodes / +N edges (N = steps with a template ref; software-dev already has 2 ungraphed: specify→spec, plan→plan). Orphans stay 10 (new template nodes have an instantiates in-edge). Bump _EXPECTED_NODE/EDGE_COUNT (test_extractor_projection.py:40-42) to 280+N/757+N; ceiling 14 (test_doctrine_regenerate_graph.py:78) untouched. Add a positive shipped-graph instantiates assertion (none exists today).
  • Resolve-by-URN = second lane: today resolution is name-based end-to-end (resolve_configured_template:346 → template_set[artifact_kind] → filename → resolve_template:318 → 5-tier filesystem). template_catalog.resolve_template_by_id already exists and terminates in the same Stage-2 resolve_template. The filesystem↔URN duality is a compatibility contract (ADR D3 line 104) — do not collapse.
  • #2660 invariant: resolver.py:378-384 typeless rejection + no software-dev inference; the new URN mode must inherit the identical fail-closed posture (no defaulting an unqualified URN to software-dev).

4. Grounding squad — boundary/regression risks (paula-patterns)

Ranked: (R1) template_set dict/scalar duality co-located in specify_cli/runtime/; (R2) collapsing name/URN kills the 5-tier override chain; (R3) fix by authoring, not by relaxing the :395 guard (#2689/#2660); (R4) emptiness test gameable by a filler sentence + a coupled-edit whack-a-field; (R5) cross-type contamination (per-type template_file; but note the spec-review correction: artifact_key is the SHARED "spec"/"plan" runtime contract, not per-type); (R6) instantiates has no emission path today → new emission trips orphan-residual + _ARCH_SHARD_N_FILES + golden-count markers. Correction (verified by the spec-review squad on branch feat/mission-step-creatability): templates are not skipped by the extractor (_SKIP_REF_TYPES is empty; _KIND_MAP["template"]=NodeKind.TEMPLATE) — the earlier "extractor.py:358-360 skips template targets" reading is stale. The real reason a new pass is required is that a step's template ref is a structured MissionStepTemplateRef field, not a references: list entry, so no existing edge pass traverses it.

5. Adversarial verification squad — the "small blast radius" claim (LAND, with corrections)

Verdict: LAND the template_set cutover in S-C; HOLD action_sequence for a separate slice. All claims survived attack; empirically proven byte-for-byte parity for all four types; *no mission_types/.yaml authors template_set** → the raw-YAML fallback is already dead code.

paula-patterns (refute: 1 consumer / no boundary widening) — CLAIM 1 SURVIVES (one behavioral production consumer, resolver.py:403), corrections: there is no ProfileBundle class (it's ResolvedMissionType); the :1001 display is MissionTypeProfile.template_set, a scalar, unrelated. CLAIM 2 conclusion holds but the mechanism was FALSE — MissionType is frozen, extra="forbid" with no .steps field; the thunk must resolve steps via MissionStepRepository.resolve_all_for_mission_type(type, pack_context=None). Surviving risks: (HIGH) stored field is schema-pinned by ~6 test files; (MED) mission_type.py:1491 --json serialization contract; (LOW) a future pack authoring template_set without steps must fail loudly, not be silently dropped.

architect-alphonso (refute: action_sequence symmetry is cheap) — CLAIM 3 SURVIVES (template_set cutover ~4-5 sites, resolver signature + fail-closed guards untouched; template_set is not imported by the DRG extractor — why it's cheap). CLAIM 4 PARTIALLY-REFUTED — action_sequence is NOT "template_set with more call sites": it drags in (1) the DRG shipped-graph extractor + NFR-002 freshness gate, (2) a hard non-empty validate_action_sequence invariant (all-types precondition), (3) an eager FSM hot-path field (ResolvedMissionType.action_sequence:332) with extends-chain resolution under <100ms (runtime_bridge_composition.py:186/321, decision.py:606). Net: ~4 sites (template_set) vs ~9-12 (both), and it crosses a categorically different gate → do NOT fold in; distinct slice.

reviewer-renata (refute: no correctness/fail-closed regression) — CLAIMS A/B/C all SURVIVE. Proven at runtime: stored == project_template_set(steps) byte-for-byte (software-dev {'plan':'plan-template.md','spec':'spec-template.md'}; the 3 types None). Fail-closed guards sit one layer above the change (untouched). _template_set_thunk is compare=False + template_set is a @cached_property → equality/determinism unaffected. Coupled edits the spec MUST guard: (1) mission_type.py:1491,1509 read the model field, UNTESTED → AttributeError on removal; migrate + dict()-wrap the MappingProxy before json.dumps + add test; (2) injection removal must be lockstep with field removal (extra="forbid"); (3) the new slot must pass pack_context=None and be cached (MissionStepRepository.default() is NOT memoised → naive rewrite = filesystem walk per resolution); (4) decline the ResolvedMissionType.template_set rename. KEEP TestSoftwareDevProjectionParity (test_softwaredev_roundtrip.py:68-112) — it's the enduring regression net; retire only the injection-half + the ~6 field-pin tests. Also: project_template_set doesn't sort (set-backed) → sort for determinism (S-C graph-backs).

6. Net decision → spec

  • Retire the persisted template_set field + dead fallback; consume via cached project_template_set(steps) (pack_context=None) — FR-001/FR-002.
  • Migrate the CLI reads + add the missing test — FR-003.
  • Author the 3 types on their own step names; own the emptiness-test coupled edit; genuine content — FR-004..008, NFR-004.
  • Graph-back via the step projection; resolve-by-URN as a second lane; intentional DRG delta; sort for determinism — FR-009..011, NFR-002.
  • Frozen: fail-closed (#2689/#2660), C-008 scalar fence, per-type shapes, name↔URN contract, keep the enduring parity guard, don't rename the resolved property — C-001..006.
  • Deferred: action_sequence symmetry (distinct slice), FR-009/FR-011 prior deferrals — C-007.
  • Open (→ /plan): Q1 artifact_kind trace at mission creation for the 3 types.

Key files: src/doctrine/missions/{models.py,step_projection.py,mission_type_repository.py}, src/charter/mission_type_profiles.py, src/specify_cli/runtime/resolver.py, src/specify_cli/cli/commands/mission_type.py, src/doctrine/drg/migration/extractor.py, src/doctrine/template_catalog.py, src/doctrine/{template,action,mission_type}.graph.yaml, tests/doctrine/missions/test_prompt_emptiness.py, tests/doctrine/missions/test_softwaredev_roundtrip.py, tests/doctrine/drg/migration/test_extractor_projection.py. ADR: docs/adr/3.x/2026-07-16-2-mission-type-step-authority-and-template-vocabulary.md (D3).