16 — Codebase Reassessment (Debbie/Pedro fan-out)

Phase: 2 → 3 bridge · Date: 2026-06-03 · Method: 6 parallel investigators (Debugger Debbie ×4 forensic, Python Pedro ×2 structural) validating the dialectic's sharpened hypotheses (15) against real code-paths. This is the evidence base for the consolidated model and the 06 mapping.

Headline: the code confirms all of the dialectic's refutations, and adds five findings that change the 06 concretization. The structural core (Mission≠MissionRun) is real and contained; several "domains" are actually shared kernels / seams; and two of the proposed homes in CLAUDE.md / earlier notes are stale or non-existent.

Verdict table

H Hypothesis Verdict One-line
H1 Context = Shared Kernel / OHS, not Execution subdomain CONFIRMED CQRS split is by I/O-role, not domain; write-side serves status+governance
H2 Status/kanban is a shared bounded context CONFIRMED 70 consumers (~24 exec / ~27 plan); the integration seam — but boundary leaks
H3 Actor is cross-domain CONFIRMED + fragmented Only LLM-step-executor is execution-bound; no single Actor type (3 vocabularies)
H4 Prompt = governed projection, not DTO CONFIRMED Published Language; three distinct projections exist
H5 MissionRun can't name its Mission CONFIRMED Snapshot stores type+uuid only; inputs["mission_slug"] is write-only dead code
H6 Model maps onto the package graph MOSTLY — 3 corrections runtime↔specify_cli is bidirectional/unenforced; canonical runtime ≠ where notes say

Per-hypothesis evidence

H1 — Context is a Shared Kernel fronted by two OHS facades

Shared Kernel = core/paths.py + workspace/root_resolver.py + mission_metadata.resolve_mission_identity; OHS facades = resolve_action_context (core/execution_context.py:220) + resolve_mission_read_path (missions/_read_path_resolver.py:94). Consumed by Governance-read (doctrine.py:98, context.py:263), Mission Mgmt (kanban/dashboard/tasks), Acceptance, Merge, Runtime, Sync, Orchestrator. The CQRS split is real (get_status_read_root READ vs get_main_repo_root/canonicalize_feature_dir WRITE) but partitioned by I/O role — the write-side canonicalizer serves status/emit.py + charter writes, not execution; resolve_action_context fuses implement/review with tasks*/accept in one body (execution_context.py:23-32). → Context = Shared Kernel + OHS; Execution is a consumer, never owner.

H2 — Status/kanban is a shared context (and the integration seam)

70 consumer files, balanced ~24 execution / ~27 planning, on the same five primitives (read_events, reduce, materialize, get_wp_lane, Lane/StatusEvent). The runtime.next loop imports nothing else from the planning side — the status event log is literally the contract by which planning and execution communicate WP state. New debt finding: the 060-cleanup public API (status/__init__.py __all__) is bypassed — ~245 deep-submodule imports vs 6 facade imports, many reaching non-exported internals (lifecycle_events, locking, adapters, work_package_lifecycle, emit.build_status_event). The context is conceptually clean but operationally porous; an import-boundary enforcement test (mirroring tests/architectural/test_shared_package_boundary.py) is the action item.

H3 — Actor is cross-domain and fragmented

Human/operator effects change in Governance (charter/interview.py, generator.py, activation_engine → config.yaml) and Mission Mgmt (merge.py:160 _resolve_merge_actor; RACI accountable-must-be-human P0 invariant runtime/next/_internal_runtime/schema.py:78-85; HiC gate retrospective/gate.py:305). External systems mutate governance via the tracker ACL (tracker/config.py:39 doctrine_mode="external_authoritative"). The runtime is itself an actor-kind. Only the LLM-agent-as-step-executor is execution-bound (and even that slot is vacated by the human-in-charge sentinel). New finding: there is no unified Actor type — three vocabularies across packages (human|llm|service in runtime/decisions; human|agent|runtime in retrospective; free-form str in status/emit). Any "Actor domain" claim must name which metamodel.

H4 — Prompt is a Published Language; there are three projections

_build_wp_prompt (prompt_builder.py:142-289) splices 4 domains (Mission Mgmt template+WP, Governance charter+profile directives, Context workspace/branch, Actor identity) into flat text → temp file → returns only the path. Not a DTO. The three projections of one governed invocation, previously conflated:

  1. Executor Prompt — rendered text (prompt_builder.py) → LLM agent. (Published Language)
  2. ActionContext.to_dict() JSON (cli/commands/agent/context.py:111) → agent-context CLI/shim. (a real DTO, different consumer)
  3. OperationalContext (frozen VO, charter/invocation_context.py:155) → logged/threaded at the decision boundary, not passed to the prompt builder. → Consolidating these three is a design target in its own right.

H5 — MissionRun degeneracy (contained fix)

MissionRunSnapshot (schema.py:523-536) and MissionRunRef (engine.py:92-97) store run_id(uuid) + mission_key(type) only — no mission_id/mission_slug. start_mission_run mints uuid4().hex (engine.py:196); the only run↔mission link is the external forward index feature-runs.json (slug→run). inputs["mission_slug"] is written but never read (zero readbacks in src/). Fix blast-radius: 2 schema classes + ~6 in-engine snapshot-copy sites (silent-drop risk) + 2 bridge call sites + feature-runs.json + additive legacy migration — contained to runtime/next/; the external MissionRunStartedPayload event is out of scope. Separable from #1619.

H6 — Package-graph corrections (important for 06)

  • Layer order (tests/architectural/test_layer_rules.py, conftest.py:33-54): kernel ← doctrine ← charter ← specify_cli is the spine; runtime/ and glossary/ are siblings at the charter level, not a single chain.
  • runtime ↔ specify_cli is bidirectional and unenforced: runtime imports specify_cli 56×; the only hard rule forbids runtime → specify_cli.cli / specify_cli.next (test_layer_rules.py:200-220). Not a clean downward dependency.
  • Canonical internal runtime = src/runtime/next/_internal_runtime/ (engine.py, schema.py, planner.py). src/specify_cli/next/ is a deprecation shim (__deprecated__ = True, removed in 3.3.0). The CLAUDE.md / earlier-notes reference to specify_cli/next/_internal_runtime is STALE — do not anchor the model there.
  • MissionStatus aggregate does not exist — status is the event-log (Lane + StatusEvent, status/models.py). ActionContext confirmed at core/execution_context.py:44.
  • mission_runtime/ does not exist (net-new). A net-new top-level package would fail the meta-guard test_no_unregistered_src_packages until registered in _DEFINED_LAYERS.
  • Aside: dashboard/ is not layer-registered (would flag the meta-test) — separate follow-up.

Emergent findings (beyond the original hypotheses)

  1. Status/kanban is a first-class shared context — surface it explicitly; add import-boundary enforcement (H2). Filed: #1664.
  2. Actor metamodel is fragmented across 3 vocabularies — a unification candidate (answers the 12 §7 "shared Actor type?" question) (H3).
  3. Three parallel context projections (Prompt PL / ActionContext DTO / OperationalContext VO) — consolidation target (H4).
  4. MissionRun→Mission reference gap — small, contained (H5). Filed: #1663.
  5. Stale/absent homes: canonical runtime is runtime.next (not specify_cli.next); mission_runtime/ and MissionStatus are net-new and constrained by the layer meta-guard (H6). CLAUDE.md stale runtime path: fixed in this branch.

Implications for the consolidated model + 06 mapping

  • Confirm: Mission≠MissionRun; MissionType ∈ Governance(doctrine); the execution spine.
  • Reclassify: Context → Shared Kernel + OHS; Status/kanban → its own shared context (the planning↔execution seam); Actor → cross-domain (+ fragmented); Prompt → Published Language (+ two sibling projections).
  • Re-home: anchor Execution on runtime/next/_internal_runtime/; treat mission_runtime/ + MissionStatus as net-new subject to layer registration; respect the bidirectional runtime↔specify_cli reality.
  • Net-new vs existing (for 06): Governance=charter/⊕doctrine/ (exists); Status context=status/ (exists, needs boundary enforcement); Execution=runtime/next/_internal_runtime/ (exists); Context=ActionContext (exists, to harden); MissionStatus aggregate + mission_runtime/ umbrella (net-new).
Edit this page