Scoping brief — evict runtime-mutable WP state into the event log (#2684 cluster)

Status: scoping input for /spec-kitty.specify. This brief is the research spine for a single, exhaustive 3.2.6 mission. It is grounded by a four-lane research squad (2026-07-19) against the current main (HEAD 874673ea3, which already carries the merged #2684 red-first test) plus the prior-art design corpus in the sibling wp-op-schema-design working tree. It is not a plan — it collects the decided facts and the still-open decisions so the spec author starts with zero undecided code facts. Every line number below was grep-verified on main; where the tracker's numbers had drifted, the correction is called out.

0. Decisions resolved (HiC, 2026-07-19)

The seven open decisions from §14 are now closed. These are binding inputs for /spec-kitty.specify; §14 is retained as the rationale record.

  1. Off-axis events → Option A, realized as ONE generic InnerStateChanged event. Non-transition mutations enter the log via a single annotation event carrying a typed partial delta (WPInnerStateDelta with optional typed fields — not a free dict[str, Any], to avoid re-introducing the split-brain). It bypasses validate_transition (no from_lane/to_lane) and is folded by the reducer with a per-field merge rule and no force_count increment. This is the reusable realization of the Option A annotation class — one event kind, not per-kind variants.
  2. Subtask granularity → subsumed by the delta. No separate per-subtask-vs-snapshot event: the InnerStateChanged delta carries subtasks: Mapping[str, Status], which expresses a single mark and a batch mark naturally, and (per the HiC framing) the same generic event also carries the PID refresh — a reusable state-delta pattern.
  3. Activity notes → fold into InnerStateChanged. A note delta field with append semantics; the reducer keeps a notes list on the snapshot; AC-4 renders the Activity Log from it. One off-axis event class for pid + subtasks + notes.
  4. tracker_refs → EVICT (runtime, event-sourced). map-requirements and move-task emit an InnerStateChanged delta (tracker_refs with union/append merge); FR-011 runtime append is preserved. Consistent with the eviction thesis.
  5. Review-cycle → evict ALL in this mission. Delete the dead verdict-field read fallbacks (workflow_cores.py:340-341, done_bookkeeping.py:104-105) and evict the actively-written review_artifact_override_* into InnerStateChanged. Governing principle: #2684 is the authoritative owner of the WP-metadata surface — deferred PR #2641 yields to the mission (its file-collision risk is moot; it was deferred precisely for treading on this active-design turf). AC-1 is fully complete within this mission. Update 2026-07-19: the same principle resolves the #2160 dependency — #2160's writer work is pr:deferred and yields to this mission, so this mission owns the implement.py:1730 shell_pid-writer restructuring outright (the former "hard co-sequence" in §10/§13 is retired — no external blocks/blocked_by gate; see spec FR-014).
  6. Static-model election → DEFER to a follow-up. Land + verify the eviction first (blocker B4: WPMetadata can't become a clean static-only projection until its runtime half is stripped); the election coordinates with #1619 and opens against a clean surface next.
  7. progress field → retire explicitly. Remove from MUTABLE_FIELDS/schema; document the removal in the migration notes (backfill no-op — no writer ever existed).

Net effect on the decomposition (§10): WP-A collapses to one InnerStateChanged event class + one typed WPInnerStateDelta + one reducer fold covering shell_pid(+baseline), subtasks (map/replace), notes (append), and tracker_refs (union) — simpler than per-kind annotations. WP-F (tracker_refs) is now decided (evict) and folds into WP-B/WP-D rather than standing alone.

1. The mission in one paragraph

tasks/WP##.md (and the tasks.md subtask surface) is two authorities glued into one YAML block: static design-intent authored once, and runtime-mutable state written on every lifecycle event. The runtime writes churn the dossier content hash, split subtask-completion truth between markdown and the event log, and force operators to hand-tick checkboxes before a WP can move to review. #2684 (P0) evicts all runtime-mutable state into the canonical append-only event log, leaving WP files holding only static intent. It is the execution vehicle for the #2093 authority ruling and must land in 3.2.6 (milestone #4 "3.2.x") — the reason a red-first repro (#2806, merged) already reddens main.

2. Cluster ticket map

Epic spine (native sub-issue edges, GraphQL-verified): #1799 (Charter & Doctrine) → #2400 (sub-epic: Metadata & profile authority) → #2093 (WP-metadata authority split — the ruling, P1) → #2684 (the mission, P0) → #2686 (projection-hash child, blocked-by #2684). Parallel: #1619 (unify mission execution context) → #2160 (coord task/status artifact authority — co-sequence anchor); #2017 (workflow guards epic); #1914 (governed/gate ops must be no-op-stable).

In the mission:

  • #2684 — the eviction (AC-1..6): shell_pid(+baseline), history[] (dead — delete), subtask-checkbox state, review-cycle fields, ## Activity Log, agent/assignee; the 4 shell_pid writers + move-task; delete legacy fallbacks; repoint model readers.
  • #2686 — hash the WP static projection, not raw bytes (tail of this mission or immediate follow-up; delivered in substance by AC-5).

Dependencies (land with/before, or share touched files):

  • #2160 — hard co-sequence. Restructures implement.py/workflow.py, including the implement.py:1730 shell_pid claim writer WP-D evicts. Set an explicit blocks/blocked_by edge on the writer-cutover WP.
  • #2647 (P1) — move-task cwd read-path; reconcile in the same move-task rewrite (write-side fix 5107c700f is on main — see §7).

Defer (adjacent tasks_move_task.py lane owners, but gate/placement concerns, not metadata eviction): #2549, #2626, #2300, #2573 (P0 pre-review-gate-hang), #1734, #2583, #2493, #2555, and the pre-review-gate sub-cluster (#2794/#2801/#2534/#2741/#2598/#2705/#2802/#2602). Coordinate lane ownership on tasks_move_task.py; do not fold in.

Closed precedents — do NOT re-do (cite the fix): #1862 & #1764 (analysis-freshness normalization analysis_report.py::_normalize_tasks_md), #2513 & #2576 (rollback uncheck), #2510 (coord husk fail-open), #2567/#2324/#2346/#2574 (subtask parser consolidation), #2580 (the 4th shell_pid writer routed through write_shell_pid_claim — confirms the "4 writers" inventory), #2575/#1231/#2369/#2512 (shell_pid-liveness precedent chain this generalizes), #2154/#2155 (via PR #2181, #2160 core), #2648/#2649/#2650 (#2533 write-side + degod/SSOT-consolidation of the exact files #2684 edits), #2740, #2504.

Out of scope but related: #2736 (P0 catfooding mission that re-confirmed #2684; its fix is sync bisection #2735, not eviction), #2644 (static task-graph authority — different field-class), #1341 (same append-only+derived-YAML pattern in the glossary domain — reference), #2533/#2602 (coord placement).

3. In-flight PR watchlist (conflict zones)

No open PR pre-empts the eviction itself — nothing in flight touches tasks_mark_status.py, tasks_transition_core.py, core/subtask_rows.py, migration/strip_frontmatter.py, frontmatter.py, or the status reducer/snapshot. Collision risk is concentrated in three PRs:

PR State Cluster files Conflict
#2641 CHANGES_REQUESTED tasks_move_task.py, tasks_materialization.py, review/{cycle,artifacts,arbiter}.py HIGH — the exact move-task/materialization/review surfaces WP-A/D rewrite
#2766 open, 3 CI reds workflow.py, workflow_executor.py, core/worktree_topology.py Med-High — workflow_executor.py hosts two shell_pid writers (:695, :1370)
#2800 MERGEABLE (nearest landing) merge/done_bookkeeping.py + move-task/mark-status/rollback tests Med — WP-E must delete done_bookkeeping.py:104-105; expect test churn collision

Also watch #2612 (review auto-commit path), #2492/#2611 (grazing move-task test files). Rebase the mission on main after #2641/#2766/#2800 land, or coordinate lane ownership up front.

4. The authority ruling (#2093)

Split by field-class, do not unify into one store:

  • STATIC design-intent — frontmatter STAYS canonical, never mirrored into events/meta.json: work_package_id, title, dependencies, requirement_refs, plan_concern_refs, owned_files, create_intent, authoritative_surface, scope, task_type, cross_cutting, priority, phase, the planning trio, agent_profile (authored assignment only), prompt body.
  • DYNAMIC runtime state — the event log / invocation record is the SOLE authority; frontmatter copies are retired, not left as fallback readers: lane (already retired — the precedent), shell_pid(+baseline), agent/assignee, history[], subtask checkbox state, review-cycle fields, activity_log, base_branch/base_commit/created_at.
  • AMBIGUOUS — agent_profile: authored assignment stays STATIC; the resolved binding (which profile actually ran, id + hash) records DYNAMIC on the status event / Op record (precondition for #2399's enforcement seam).

Target invariant (enforced by a refactor-stable architectural test — the generalization of the shipped phase-2 lane-authority guard): no consumer ever reads a dynamic frontmatter field as authority; static fields are never mirrored into events; agent_profile is split intent-vs-binding. #2684 accepts this verbatim and deliberately stops at "file holds only static intent," not "file is derived" (that is the 3.3.x YAML-flip — §9).

5. Field-by-field authority inventory (grep-verified on main)

T = write fires on a lane transition (fits a StatusEvent); OFF = off-axis (no lane change).

Field Current authority Writers (verified file:line) Key readers T/OFF Eviction note
shell_pid (+shell_pid_created_at) frontmatter tasks/WP##.md; co-write helper frontmatter.py:357 write_shell_pid_claim implement.py:1730 ✓; workflow_executor.py:695 (was :669); workflow_executor.py:1370 (was :1337); tasks_move_task.py:1770 (was :1638). Delete site: tasks_move_task.py:1747 claim-liveness stale_detection.py:402-403 → :221/:244; WorkPackage.shell_pid support.py:295; WPMetadata wp_metadata.py:251,361-363; summary_core.py:69 MIXED (claim=T; resume refresh=OFF, the pivotal case) in MUTABLE_FIELDS; carry (shell_pid,baseline) in policy_metadata on the claim event
history[] frontmatter list WPMetadata.history wp_metadata.py:259 frontmatter.py:176/:347 — ZERO live callers; wp_metadata.py:570 docstring-only build merge wp_metadata.py:585 (never fires); audit/shape_registry.py:113 N/A dead TRULY DEAD + MIS-FILED in STATIC_FIELDS (strip:56). Delete outright; no migration
subtask checkbox (- [x] T###) tasks.md WP-section rows (core/subtask_rows.py) tasks_mark_status.py:266 _resolve_checkbox; uncheck subtask_rows.py:192/:218 ← tasks_move_task.py:1806 done-inference emit.py:299→302 done==total; accept gate gates_core.py:112; dashboard scanner.py:961 OFF (the P0) hardest: per-subtask stream, no lane event. Needs subtask-completion projection in the reduced snapshot; re-point emit.py:302 + _guard_subtasks. This is the merged red test.
review_status / reviewed_by / reviewer* / review_feedback canonical = event log; frontmatter = legacy fallback only NO live runtime writers DELETE fallbacks: workflow_cores.py:340-341 (provenance "frontmatter"); done_bookkeeping.py:104-105 (DoneEvidence frontmatter-migration:{wp}) T (verdict) already in MUTABLE_FIELDS; delete the two read fallbacks, don't just stop writing
review_artifact_override_* (_at/_actor/_wp_id/_reason) frontmatter — ACTIVELY written tasks_materialization.py:58-61 and :125-128 review/artifacts.py:134-136,181-182 OFF NOT in MUTABLE_FIELDS — live writer the issue's set omits; needs an event/policy_metadata home
## Activity Log (body) WP##.md body; append task_utils/support.py:221 6 writers: workflow_executor.py:705,:1377; tasks.py:914; tasks_move_task.py:1777; external orchestrator_api/commands.py:1563; task_metadata_validation.py:135 (migration-only) tasks.py:1149-1150 validation; human narrative MIXED evict to event log; the cross-package writer commands.py:1563 must migrate too
agent frontmatter workflow_executor.py:694,:1369; tasks_move_task.py:1763. Delete site :1746 WorkPackage.agent support.py:291; wp_metadata.py:419-458 claim=T, reassign=OFF in MUTABLE_FIELDS; co-written with shell_pid at every claim → same vehicle
assignee frontmatter tasks_move_task.py:1761 (only writer) support.py:287; wp_metadata.py:246 OFF in MUTABLE_FIELDS
tracker_refs frontmatter (WP_FIELD_ORDER, "per DIR-012") but runtime-written tasks_map_requirements.py:428; tasks_move_task.py:1721→_mt_persist_tracker_refs:1787 wp_metadata.py:213,320-322 OFF FR-011 tension: cannot be both static AND derived. NOT in MUTABLE_FIELDS. Must re-decide (§8)

MUTABLE_FIELDS today (migration/strip_frontmatter.py:23-34) — the canonical set to extend, not fork: {lane, review_status, reviewed_by, review_feedback, progress, shell_pid, assignee, agent}. In-scope additions: shell_pid_created_at, history (move out of STATIC_FIELDS), tracker_refs (if evicted), review_artifact_override_*, reviewer_shell_pid. Note the Activity-Log body and subtask checkboxes live in tasks.md/body, not frontmatter keys — strip_mutable_fields (:130-155) only touches frontmatter, so they need a parallel eviction path. progress is in MUTABLE_FIELDS with no live writer/reader → explicitly retire it (document the removal).

6. Mechanism map (task-status / gate / event-log / hash)

  • Two markdown subtask authorities, both to re-point: (a) _guard_subtasks (tasks_transition_core.py:384) via _check_unchecked_subtasks (tasks_shared.py:412); (b) done-inference _infer_subtasks_complete (emit.py:279→302) feeding the FSM guard wp_state.py:370. Both count tasks.md rows (core/subtask_rows.py:39,111,164).
  • The reducer carries NO subtask state. reducer.py:48-65,118 folds only {lane, actor, last_transition_at, last_event_id, force_count}; StatusSnapshot.work_packages (models.py:322) has no subtask field. mark-status is coreless — it rewrites the checkbox (tasks_mark_status.py:302) and emits a free-text HistoryAdded note (:326, not a StatusEvent) that nothing folds back. The eviction must add a reduced, gate-queryable subtask-completion projection and re-point both (a) and (b) at it.
  • AC-5 content-hash churn root cause: the dossier hashes raw file bytes of both tasks.md and tasks/WP##.md (dossier/hasher.py:14; dossier/indexer.py:242; parity dossier/snapshot.py:25-53; manifest expected-artifacts.yaml:36). Runtime writes into those files churn content_hash_sha256 → parity churn → false drift. Eviction into status.events.jsonl (not a dossier artifact) stabilizes the hash.
  • FSM rejects self-edges (the constraint is real in code): StatusEvent mandates from_lane/to_lane (models.py:248-249); InProgressState.allowed_targets() excludes IN_PROGRESS (wp_state.py:351); check_transition returns Illegal transition: in_progress -> in_progress (:162,179), and force additionally demands actor+reason and increments force_count (reducer.py:64). So off-axis mutations cannot be clean transition events.

7. The pivotal ADR decision (non-transition mutations)

Three evicted mutations are off-axis (no lane change): (a) shell_pid refresh on resume of an already-in_progress WP, (b) mid-in_progress subtask marks, (c) activity-log notes. policy_metadata: dict|None (status/models.py:258, threaded through emit.py) can already carry (shell_pid, baseline) on the claim transition with no wire-schema change.

  • Option A — non-transition annotation event class (prior-art recommendation). A new record in the same log distinguished by absent from_lane/to_lane + annotation_kind (shell_refresh|subtask_marked|activity_note), typed payload, truthful at. Deltas: wp_state.py sanctioned self-edge / annotate() primitive bypassing check_transition; models.py event discriminator round-tripped in to_dict/from_dict; reducer.py precedence change (fold annotation after transition, last-writer-wins per field, no force_count increment). Highest blast radius (touches FSM authority + wire model + reducer core) but the only option that makes the log a true SSOT for (a)+(b)+(c) — and it is what delivers the subtask projection §6 needs.
  • Option B — fold onto transitions + documented behavior change. Carry (shell_pid,baseline) in policy_metadata of the next real transition only; near-zero FSM/reducer change. But resume emits no event → stale PID → claim-liveness degrades to the git-timestamp heuristic (stale_detection.py:254,331, path intact) → a resumed long-planning agent can be falsely flagged stale until its first commit. Complete only for (a); (b) and (c) still need a home.

Gate re-sourcing (re-point _guard_subtasks and _infer_subtasks_complete at a reduced-snapshot field) is required under BOTH options — it is the part the merged red test's control protects (the fix must re-source the gate, not delete it). This decision gates the whole mission and must be closed at spec time (candidate standalone ADR — §10).

8. Migration contract

Strict order — writer-first is prohibited: backfill → verify → reader cutover → writer cutover → delete fallbacks + land hash guard.

  • Why (the B3 clobber window): readers keep the frontmatter fallback until backfill is verified; cutting writers first strips e.g. shell_pid while stale_detection.py:402 still reads frontmatter → live claims read reclaimable, evidence clobbered on the next implement.
  • Seed-event determinism (AC-6): the reducer dedups by event_id, which must match ULID_PATTERN (reducer.py:139-149, models.py:70) — a content hash is NOT a valid ULID. Use a namespaced deterministic ULID from mission_id + wp_id + field so backfill is idempotent; re-runs must not double-seed.
  • Subtask reconstruction contract (timestamp honesty): checkboxes carry no at → clamp the reconstructed mark to the WP's claimed timestamp. "No data loss" is stated against count+value parity of the reduced snapshot, NOT literal temporal fidelity — say so in the spec. Verify (step 2) confirms reduced snapshot == pre-migration state by count+value.

9. Acceptance criteria (AC-1..6) — which prove the mission

  • AC-1 No implement/mark-status/move-task/review action writes tasks/WP##.md (bytes + mtime unchanged across a full lifecycle). move-task is an explicit target (the god-write, §11.1).
  • AC-2 Claim-liveness resolves from the reduced snapshot; a claimed WP with empty frontmatter is detected live.
  • AC-3 Done-inference resolves from subtask_marked annotations; lane gating identical. ← the already-merged red test test_issue_2684_subtask_completion_event_sourced.py pins this (RED today because emit.py/_guard_subtasks read tasks.md; GREEN when they read the snapshot).
  • AC-4 Activity Log / History / review sections render from events with no content loss (the annotation payloads add the prose home M7 needs).
  • AC-5 A full lifecycle produces a stable content hash (the churn fix; wired once, no mixed parity pool). Headline proof of the mission.
  • AC-6 Migration backfills the corpus idempotently (deterministic ULID seeds); honest reconstruction contract; no unrecoverable loss.

The "prove-it" quartet: AC-1 + AC-2 + AC-3 + AC-5.

10. Proposed decomposition (a PROPOSAL for the spec author)

Critical path A → {B, C} → D → E, with F parallelizable. WP-A's payload-shape choices (§ open decisions) must be closed at spec time, not deferred into implementation.

  • WP-A — Annotation event class + reducer fold (foundation). New off-axis record shape; validate_transition bypass; reducer precedence + force_count neutrality; typed snapshot slots (shell_pid, shell_pid_created_at, baseline, subtask_state, notes). FSM untouched. Closes subtask-granularity + activity-log-home decisions. Independent of #2160.
  • WP-B — Reader cutover. Repoint WorkPackage.{shell_pid,agent,assignee} (support.py:287-296), WPMetadata coercion (wp_metadata.py:361,580), stale_detection.py:402-403, emit.py:302 done-inference, and _guard_subtasks/tasks_shared.py:412 to the reduced snapshot; frontmatter fallback behind a flag until backfill verified. Turns the merged AC-3 red test green. Needs A.
  • WP-C — Migration (backfill + verify). Extend MUTABLE_FIELDS; emit seed transition + annotation events with deterministic namespaced ULIDs; subtask clamp; read-back count+value verify. Needs A; gates D.
  • WP-D — Writer cutover. Cut all writers off the file: the 4 shell_pid writers (implement.py:1730, workflow_executor.py:695&:1370, tasks_move_task.py:1770), agent/assignee writers, the 6 activity_log writers (incl. external commands.py:1563), subtask materialization + the tasks.md uncheck (tasks_move_task.py:1806), tracker_refs (:1787) per §8 decision. Must co-sequence with #2160. Needs C's verify gate. Delivers AC-1.
  • WP-E — Delete legacy fallbacks + land the AC-5 hash guard. Remove workflow_cores.py:340-341 and done_bookkeeping.py:104-105; wire the stable-hash guard once (no mixed pool); add the refactor-stable "no consumer reads a dynamic frontmatter field as authority" architectural test (#2093 invariant). Needs B + D.
  • WP-F — Resolve tracker_refs (author-immutable OR evict). Ship exactly one (§ open decision). Parallelizable once decided.

11. Highest-risk couplings (where naive eviction breaks a gate)

  1. move-task god-write — _mt_persist_wp_file rewrites assignee(:1761)+agent(:1763)+ shell_pid(:1770)+Activity-Log(:1777) atomically, and on planned rollback also clears agent/shell_pid(:1746-1747), unchecks subtasks(:1806), persists tracker_refs(:1787). Piecemeal eviction breaks the co-write atomicity the others rely on.
  2. done-inference vs checkboxes — evict checkbox state but leave emit.py:302 counting tasks.md → the done gate silently mis-fires (the #2684 P0).
  3. claim-liveness — evict shell_pid without redirecting stale_detection.py:402-403 → every claim mis-reads alive/dead → allocator/resume corruption.
  4. merge DoneEvidence fallback — strip reviewed_by/review_status while leaving done_bookkeeping.py:104-105 → merge loses done-evidence. Delete the fallback and guarantee canonical evidence together.
  5. external orchestrator writer — orchestrator_api/commands.py:1563 writes the Activity Log from outside the host CLI; a host-only rewire leaves it writing the retired surface.
  6. review-feedback provenance fallback — workflow_cores.py:340-341 returns provenance "frontmatter"; rejection-cycle gating for legacy WPs depends on it until deletion + backfill land together.

12. #2647 interaction (same cluster)

The write-side fix 5107c700f is on main: a modern coordination-less mission (single_branch/lanes/flattened) writes to worktree_root = repo_root on a caller-supplied, CWD-invariant destination_ref (coordination/transaction.py:783-799), no longer re-derived from Path.cwd() (the #2647 taint at :300-302). Eviction increases append-only write frequency on the target branch, riding this exact path. Every new emit site (Option A annotation, or Option B policy_metadata fold) MUST resolve destination_ref from stored topology/target branch, never Path.cwd(), or it silently reopens #2647. There is no runtime provenance check — the guard is the test_transaction_legacy_topology_routing regression test.

13. Sequencing with other tickets

  • #2160 — hard co-sequence (blocks/blocked_by on WP-D): it restructures the same implement.py:1730 shell_pid writer; the eviction lands with/behind it, never racing.
  • #1619 — neutral for the eviction (moves state to the already-shipped log, no new aggregate); only the final static-model election (out of scope, §B4) coordinates with it.
  • Op-debrief slice — independent (shares no code); do not couple.
  • #2686 semantic-only hash — follow-up that shrinks once nothing writes runtime state (co-move sync/body_upload.py TOCTOU; no mixed spec/raw parity pool across the un-migrated corpus).

14. Open decisions the spec MUST close — RESOLVED 2026-07-19 (see §0)

All seven are now closed; see §0 for the binding calls. Retained below as the rationale record.

  1. tracker_refs — author-immutable vs evict. Runtime-written today (tasks_move_task.py:1721, tasks_map_requirements.py:428, FR-011) yet listed static in the schema proposal. Cannot be both. Hardest classification call. → WP-F.
  2. Subtask-event granularity — per-subtask subtask_marked vs single subtasks_snapshot. → WP-A.
  3. activity_log home — first-class annotation event vs Tier-2 evidence sidecar. → WP-A.
  4. Review-cycle eviction timing — evict here vs fold into #2160's review-state work. Interacts with the #2160 co-sequence.
  5. Static-model election is deferred — enrich WPMetadata vs elect WorkPackageEntry. Real tension but out of scope; flag so the spec doesn't pull it in (this mission unblocks it via B4).
  6. progress field — retire explicitly (no writer/reader), don't silently drop.
  7. The pivotal ADR (Option A vs B) — §7. Everything rests on it.

15. Candidate NEW tickets — REQUIRES HiC APPROVAL, not filed

Per the governance boundary (no new spec-kitty tickets without HiC approval), drafted for review only:

  1. Standalone ADR/decision ticket for the self-edge / non-transition annotation-event design (§7) — currently buried in #2684's body.
  2. tracker_refs authority re-decision (§14.1) — small decision ticket; must resolve before writer cutover.
  3. Migration/backfill contract as its own sub-issue under #2684 (§8) — deterministic ULID seeding
    • honest reconstruction contract is substantial.
  4. 3.3.x YAML-authoritative / markdown-derived WP-prompt flip epic — explicitly out of #2684 scope; requires re-ratification by the #2400 owner; gated on this mission.

16. Sources

  • Research squad (2026-07-19): tracker archaeology + PRs; field-authority inventory; status/gate/ event-log mechanism map; prior-art synthesis. Code grounded on spec-kitty main 874673ea3.
  • Prior art (sibling wp-op-schema-design tree): docs/adr/3.x/2026-07-16-1-wp-runtime-state-authority-event-log-eviction.md; docs/architecture/wp-runtime-state-eviction.md; docs/plans/investigations/wp-runtime-state-eviction-scope.md; docs/plans/investigations/wp-op-schema-proposal.md (Part 4: B3/B4, M2/M4/M7).
  • Tracker: #2684, #2093, #2400, #2160, #2686 and the cluster in §2. Milestone #4 "3.2.x".
  • Merged anchor: tests/regression/test_issue_2684_subtask_completion_event_sourced.py (PR #2806).
Edit this page