v0.24.0 — model routing and context archiving

A minor release focused on model-routing depth, context lifecycle
management, and operator diagnostics. It adds richer model/provider
metadata, local semantic routing support, a read-safe context archiving
skill, context-consumption reporting, and team/model assignment
infrastructure. No breaking changes are intended for existing consumers.

Added

• Added model-class assignment and richer model/provider characteristics:
  provider, model, version, effort, architecture, license, deployment,
  latency/speed, context length, token accounting, access, rate limits,
  and supported runtime parameters.
• Added get_model_for_class() and class-based routing metadata to the
  model recommender so callers can resolve fast, balanced, deep, and
  specialized model classes consistently.
• Added the context-archiving skill and MCP server for planning and
  creating archive tarballs, extracting archived payloads, and keeping
  archived entities queryable through the index.
• Added archive-aware index metadata via storage_location, including
  archived LogEntry event-query support.
• Added task-router v0.2 scoring with local n-gram semantic matching,
  scoring basis details, and low-confidence LLM-escalation metadata.
• Added context-consumption checkpoint reports and team identity /
  model-assignment bindings for role-based routing.
• Added optional sqlite-vec semantic search and FTS5-backed index search.
• Added a library-expert skill template plus skill-builder and
  skill-reviewer guidance for version-scoped library expertise.

Changed

• Improved processkit routing and MCP ergonomics across task-router,
  skill-finder, model-recommender, and related docs.
• Refined harness diagnostics and pk-doctor drift reporting, including
  server-header drift coverage.
• Updated model schemas in both context/ and src/context/ to carry
  the broader model-characteristics surface.

Fixed

• Fixed artifact metadata updates so callers can avoid restamping
  freshness-sensitive metadata.updated values.
• Restored the docs-site build after the v0.23.1 release line.

Notes

• v0.24.0 is a minor release because it introduces new
  backwards-compatible capabilities. The context-archiving MCP server is
  implemented, but harness registration remains gated on installer sync
  support for new per-skill MCP config entries.
• Pre-release validation on the release branch retains the known optional
  release-audit warnings for unused entity directories (actors/,
  scopes/, gates/) and the known pk-doctor preauth warning gated by
  aibox#55.

v0.23.1 — release-audit cleanup (0 ERRORs)

A focused content+code patch resolving the 106 ERRORs surfaced by the
first-run release-audit on v0.23.0. Hybrid validator + content fixes
per DEC-20260426_1859-MightyRobin (BACK-FierceOwl): three documented
validator relaxations close 96 false-positives, and ten real content
findings get backfilled. Net: 106 ERROR → 0 ERROR; 3 WARN unchanged
(actors / scopes / gates entity-dirs are deliberately unused — flagged
for v0.24.0).

Changed

• change(release-audit): excludelist aibox CLI migration prose docs.
  entity_files walker now skips *-to-*.md and INDEX.md under
  context/migrations/. Closes 16 A1 false-positives — these are aibox
  CLI prose, not Migration entities. Real Migration entities under
  pending/ and applied/ continue to be validated.
• change(release-audit): model the team-member directory layout.
  entity_files walker now understands the team-manager skill's
  alt schema: registers Persona as a known kind, special-cases the
  team-member.md filename stem to map onto TEAMMEMBER-{slug}, and
  walks relations/ subdirs as Persona-related sub-files. Closes 11
  A2 false-positives.
• change(release-audit): make metadata.processkit.layer optional
  for non-processkit-category skills.
  Layer is a processkit-DAG concept; library-style skills
  (category != "processkit") participate in no compositional layer
  graph. Strict layer enforcement preserved for the processkit DAG
  where it matters. Closes 69 C false-positives — across engineering
  (39), devops (11), data-ai (11), design (5), documents (3).

Added

• add(release-audit): 24 new pytest cases. test_release_audit.py
  gains TestMigrationProseSkipped (4 cases), TestTeamMemberLayout
  (3 cases), and TestLayerConditional (2 cases) covering the three
  validator changes above. Plus +203 lines total of fixture and
  helper plumbing.

Fixed

• fix(skills): backfill metadata.processkit.layer on 7
  processkit-category SKILL.md files. Each value computed from the
  skill's declared uses[] deps:
  • model-recommender: layer 2
  • skill-builder: layer 4
  • skill-finder: layer 4
  • skill-gate: layer 4
  • skill-reviewer: layer 4
  • task-router: layer 4
  • team-creator: layer 4
• fix(team-creator): author missing ## Full reference section.
  +213 lines reference content covering all team-creator commands,
  flags, and recipes.
• fix(team-manager): author missing ## Overview and
  ## Full reference sections. +316 lines covering the team-member
  directory layout, persona files, relations, and consistency checks.

Notes

• v0.23.1 is a patch release — additive validator allowances and
  documentation backfills only, no breaking changes. Consumers on
  v0.23.0 can adopt with no migration steps.
• pk-release-audit on this release: 0 ERROR / 3 WARN / 999 INFO
  (down from 106 ERROR / 3 WARN / 903 INFO on v0.23.0). The 3
  remaining WARNs are entity.missing-dir for actors/, scopes/,
  and gates/ — the project doesn't use those entity kinds yet.
• Incidentals rolled into the same release per FierceOwl's transition
  note: pk-doctor drift.py parser fix follow-up to processkit#13,
  drift-script provenance allowlist updates, and a refreshed aibox CLI
  migration prose for v0.20.0 → v0.21.1.
• DEC-20260426_1627-CuriousButter was corrupted at write-time
  (XML-style fields embedded inside the decision string plus
  trailing tool-call markup) and supersedes-replaced by
  DEC-20260426_1859-MightyRobin.

---

v0.23.0 — RoyalFern filters + 4 new validators/tools + 3 scaffolding fixes

A focused batch release: governance filter implementation that gives
the v0.22.1 RoyalFern schema fields runtime teeth, plus four scaffolding
bug-fixes (CalmArch / RapidDaisy / VastLark) and four new validators &
tools (TidyGrove / NobleBrook / AmberCliff). Eight WorkItems in total,
all detect-or-fix improvements with no breaking changes.

Implementation followed two waves: wave one (CalmArch + RapidDaisy +
VastLark + SolidWolf) shipped sequentially; wave two (TidyGrove +
NobleBrook + AmberCliff + BraveMeadow) ran as four parallel
sonnet-tier subagents, one per WI, against disjoint target dirs.

Added

• feat(model-recommender): query_models() RoyalFern filters (SolidWolf).
  BACK-20260426_1214 — query_models() gains 9 governance / data-privacy
  filter kwargs sourced from the RoyalFern fields landed in v0.22.1:
  phi_hipaa_eligible, pii_eligible, gdpr_eligible,
  training_on_customer_data (enum match), data_retention_days_max
  (numeric ceiling; accepts the literal "zero", rejects "unknown"),
  jurisdiction_country_in (HQ membership), legal_regime_in (any-match
  on applicable_legal_regimes), data_residency_in (any-match on
  data_residency_regions), and max_latency_p50_ms. Missing-field
  semantics: a model lacking the relevant field is REJECTED — conservative
  governance default. _profile_block surfaces the new fields on every
  result so callers can verify what filter they got hits on. The four
  query_models() examples in the model-recommender SKILL.md (HIPAA,
  jurisdiction, training, retention) now return non-empty result sets
  against the live 34-model roster (16/20/10/15 hits respectively).
• feat(release-audit): new skill (TidyGrove).
  BACK-20260409_1830 — context/skills/processkit/release-audit/
  ships a detect-only validator that runs four passes before a release
  tag: entity-file frontmatter compliance (apiVersion / kind / id-vs-
  filename), SKILL.md structural compliance (required frontmatter
  fields + four required body sections), MCP-server tool annotation
  presence, and skill cross-reference resolution. Single Markdown
  report, exit 0 / 1, command pk-release-audit. First live run
  surfaces 105 pre-existing ERRORs (legacy aibox migration prose,
  team-member sub-files using an alt schema, non-processkit skills
  missing the layer field) — flagged for separate triage and not
  blocking this release.
• feat(pk-doctor): skill_dag check (NobleBrook).
  BACK-20260409_1738 — new pk-doctor check walks every SKILL.md,
  builds the dependency graph from each metadata.processkit.uses[],
  and validates: missing-references (every used skill exists), cycles
  (iterative DFS with 3-colour marking), and layer constraints (a
  skill at layer N may only reference skills at layer ≤ N). On the
  live tree the first run found 0 cycles, 0 missing refs, and 5 layer
  violations — all five resolved in this release as a content-only
  layer bump (see Fixed below).
• feat(skill-finder): user-facing catalog MCP tool (AmberCliff).
  BACK-20260410_1840 — new catalog() tool on
  processkit-skill-finder. Signature:
  catalog(category, tag, keyword, columns, sort_by, output) with
  output ∈ {"markdown", "json", "yaml"} and column projection on
  any frontmatter field. Defaults to a Markdown table of every skill
  sorted by name. Reuses _all_skills() — no new SQLite plumbing.
  SKILL.md gains a "Catalog queries" section with trigger phrases
  ("list skills", "show me all skills in X", "skills as JSON") and
  five example invocations.

Changed

• change(skill-gate): prefix-based PreToolUse matcher (VastLark).
  BACK-20260425_1235 — check_route_task_called.py switches from an
  explicit allowlist to a prefix-based matcher. Any tool whose name
  starts with create_, transition_, link_, record_, open_,
  update_, apply_, reject_, add_, end_, import_,
  reactivate_, deactivate_, release_, reserve_, supersede_,
  start_, evaluate_, or skip_ is now gate-locked, plus an
  explicit allowlist for log_event (verb+noun shape). Closes the
  auto-renew gap that surfaced in the 2026-04-25 session when only
  create_binding traffic ran across a midnight boundary and the
  marker expired despite continuous compliant writes. 40 new test
  cases in test_hooks.py.
• change(workitem-management + all entity-writing MCPs): YAML
  literal-block-scalar serialization (CalmArch).
  BACK-20260425_1755 — context/skills/_lib/processkit/frontmatter.py
  adopts a custom SafeDumper subclass with a str representer that
  emits | block-scalar style for any string containing a newline,
  normalizing trailing whitespace to a single \n (canonical clip-
  chomp shape). parse() re-attaches the trailing \n the regex
  consumed at the \n--- boundary. Markdown bodies with pipe-tables,
  fenced code blocks, and backslashes now round-trip safely instead
  of failing PyYAML re-parse. The fix lives in the shared lib so
  every entity-writing MCP (workitem-management, decision-record,
  event-log, …) benefits. 8 new tests cover the QuickBison repro
  shape.
• change(model-recommender): _entity_to_legacy() propagates RoyalFern fields.
  Pass-through for vendor_model_id, knowledge_cutoff,
  latency_p50_ms, jurisdiction, and data_privacy from
  Model.versions[] into the legacy-shape entry that query_models()
  iterates over. Required to make the SolidWolf filters above work
  end-to-end against the real model roster.
• change(skill-finder): SKILL.md "Catalog queries" section.
  Documents the new catalog() tool, trigger phrases, and example
  invocations.

Fixed

• fix(event-log): log_event MCP now validates required actor field
  (RapidDaisy).
  BACK-20260425_1755 — log_event() calls
  schema.validate_spec("LogEntry", spec) before write. Calls
  missing or with empty actor return {"error": "..."} and write
  no file. Previously the schema-required field was silently omitted,
  producing LogEntry files that pk-doctor's schema_filename check
  flagged as ERRORs after the fact (and required a hand-fix). 3 new
  tests cover missing-actor, valid-actor, empty-actor cases. Docstring
  • module preamble updated to mark actor as required.
• fix(skills): layer-value bumps to satisfy uses[] declarations.
  Surfaced by the new NobleBrook skill_dag check on its first live
  run. Five layer violations resolved by promoting the affected skills
  to a layer ≥ their highest declared dependency:
  • engineering/changelog: layer 2 → 3 (uses release-semver at 3).
  • processkit/status-briefing: layer 2 → 4 (uses agent-management at 4).
  • processkit/status-update-writer: layer 2 → 4 (uses agent-management at 4).
  • product/onboarding-guide: layer 2 → 4 (uses agent-management at 4).
    Behavior unchanged; only the frontmatter metadata.processkit.layer
    values change.

Verified (no changes needed)

• chore(owner-profiling): reference files complete (BraveMeadow).
  BACK-20260410_1049 — verified
  owner-profiling/references/observable-signals.md (105 lines, 37
  signals × 6 categories) and interview-protocol.md (128 lines)
  are present, complete, spec-compliant against
  NOTE-20260410_1046-StoutSwan, and mirror-clean. No file changes.

Notes

• v0.23.0 is a minor release — additive features and pure bug
  fixes, no breaking changes. Consumers on v0.22.1 can adopt with no
  migration steps.
• The release-audit skill's first live run reports 105 ERRORs of
  pre-existing tree issues; these are NOT blockers for v0.23.0 and
  will be triaged in a separate v0.23.x content-cleanup pass.
• Two cross-project bugs were filed upstream during the prep window
  (no processkit-side fix): aibox#56 (CleverRiver — same-version
  migrations) and aibox#57 (FierceWren — content-diff overwrites
  locally-added entity content). Workaround in place: run git status
  after every aibox sync.
• pk-doctor green at release time: 0 ERROR / 2 WARN / 11 INFO. The
  two WARNs are the unchanged context/.processkit-provenance.toml
  drift (known v0.22.0 stamping bug, acknowledged in aibox v0.20.0
  release notes and tolerated). One additional INFO comes from the
  new skill_dag check.

---

v0.22.1 — RoyalFern Model schema (jurisdiction + data_privacy + 3 more)

A focused point release shipping the RoyalFern Model schema enhancement
(DEC-20260425_2256): five additive optional fields on
Model.spec_schema.versions[] — jurisdiction, data_privacy,
knowledge_cutoff, vendor_model_id, and latency_p50_ms — backfilled
across all 34 model entities and documented under the Governance
dimension in the model-recommender skill. No breaking changes; existing
callers are unaffected because every new field is optional.

The query_models() MCP tool does not yet implement filtering on these
fields — that work is tracked as BACK-20260426_1214-SolidWolf and
targeted for v0.23.0. Until that lands, the new fields are read via
get_profile() / direct file inspection.

Added

• feat(schema): RoyalFern Model schema fields.
  context/schemas/model.yaml gains five optional fields on
  spec_schema.versions[]:
  • jurisdiction — vendor HQ country (ISO-3166-1 alpha-2),
    applicable_legal_regimes (slugs: EU-GDPR, US-CLOUD-Act,
    US-HIPAA, CN-DSL, CN-PIPL, UK-DPA, CA-PIPEDA), and
    data_residency_regions (e.g. ["us-east","eu-west"]). Lossless
    replacement for ad-hoc governance_warning free-text callouts.
  • data_privacy — vendor data-handling posture: dpa_available
    (boolean or DPA URL), data_retention_days (integer or
    zero/unknown), training_on_customer_data (never/opt-in/
    opt-out/always/unknown), pii_eligible, phi_hipaa_eligible,
    gdpr_eligible, and sub_processors_url.
  • knowledge_cutoff — vendor-published training-data cutoff date.
  • vendor_model_id — exact provider API model identifier
    (e.g. claude-opus-4-7-20251031), preferred dated/pinned form.
  • latency_p50_ms — approximate p50 first-token latency in
    milliseconds; quantitative companion to dimensions.speed.
    All fields are optional and additive; pre-RoyalFern Model entities
    remain valid without modification.

Changed

• chore(models): backfill RoyalFern fields across all 34 model entities.
  Every MODEL-*.md under context/models/ (Anthropic, OpenAI, Google,
  xAI, Meta, Mistral, DeepSeek, Cohere, Microsoft, Alibaba, MiniMax)
  now carries the new fields where vendor-published values exist;
  unknown / omitted where they don't. Backfill done by two parallel
  shard subagents to bound context cost; both passes verified against
  the schema.
• docs(model-recommender): Governance dimension expanded.
  context/skills/processkit/model-recommender/SKILL.md adds a
  structured-fields block under the Governance dimension and four
  example query_models() filter usages
  (phi_hipaa_eligible=true,
  jurisdiction.vendor_hq_country in ["US","EU"],
  data_privacy.training_on_customer_data="never",
  data_privacy.data_retention_days=0). Note: the SKILL.md docs are
  the spec; the corresponding query_models() filter implementation
  ships in v0.23.0 (tracked as BACK-20260426_1214-SolidWolf).

Notes

• v0.22.1 is a patch release — five additive optional schema
  fields with no behavioural changes to any existing tool. Consumers
  on v0.22.0 can adopt with no migration steps.
• Two upstream aibox bugs surfaced during this session and were
  filed for the aibox project: aibox#56 (CleverRiver — same-version
  migrations) and aibox#57 (FierceWren — content-diff overwrites
  locally-added entity content). Neither has any processkit-side
  fix in this release; both are tracked as cross-project WorkItems
  (BACK-20260425_1711-CleverRiver, BACK-20260426_1205-FierceWren).
• pk-doctor green at release time: 0 ERROR / 2 WARN (both expected
  drift WARNs for context/.processkit-provenance.toml — known
  v0.22.0 stamping bug acknowledged in aibox v0.20.0 release notes
  and tolerated by the host).

---

v0.22.0

[v0.22.0] — 2026-04-25

Two new pk-doctor checks (migration_integrity, server_header_drift)
and the processkit-side half of preauth-spec shipping (WildGrove
Phase A) — plus a host CLI bump to aibox 0.19.2 with mcp_config_hash
recorded in the lock file. Closes the v0.21.0-deferred RapidSwan
pk-doctor check, lands the schema-malformed-migration detection that
motivated CleverRiver, and exports an 18-pattern preauth bundle for
derived projects (consumed by aibox#55).

A small one-shot cleanup pass also clears 3 pre-existing pk-doctor
schema ERRORs (1 WorkItem with mangled YAML quoting, 2 LogEntries
missing actor); the producer-side MCP fixes are tracked as
RapidDaisy and CalmArch for a future release.

Added

• feat(pk-doctor): migration_integrity check (CleverRiver). New
  check flags two malformed-migration patterns: same-version-with-content
  (a Migration whose from_version == to_version but whose
  affected_groups / affected_files are non-empty) and
  affected-files-empty (table rows present in the body but the
  affected_files array empty). Both fire as WARN, cite CleverRiver,
  and recommend reject_migration on the offending file. Surfaced the
  malformed MIG-20260425T164303 that motivated the WI. CleverRiver
  itself stays open as a cross-project tracker for the upstream
  aibox-sync diff-generator fix.
• feat(pk-doctor): server_header_drift check (RapidSwan). New
  WARN-level check that hashes the PEP 723 inline metadata block
  (# /// script … # ///) of every MCP server.py and compares
  against a manifest baseline. On drift, lists the affected skill
  slugs and recommends restarting the harness so uv re-resolves the
  pinned venv. context/.processkit-mcp-manifest.json is extended
  with a per-skill per_server_header field; aggregate_sha256
  semantics are unchanged (still computed over mcp-config.json
  files only — aibox#54 contract preserved). Per
  DEC-20260424_0127-QuickPine (SharpBrook split: BraveBird shipped
  the schema-reload half in v0.21.0; RapidSwan ships the dep-drift
  half here). Closes
  BACK-20260424_0128-RapidSwan-pk-doctor-server-header.
• feat(skill-gate): preauth spec asset (WildGrove Phase A).
  context/skills/processkit/skill-gate/assets/preauth.json ships an
  18-pattern bundle of server-wildcard permission entries and 18
  matching enabledMcpjsonServers entries that derived projects merge
  into their harness settings to pre-authorise every processkit MCP
  server. Phase A is processkit-side only — aibox#55 picks up Phase
  B (the merge logic on aibox sync).
• feat(pk-doctor): preauth_applied check (WildGrove Phase A).
  WARNs in derived-project context when the preauth spec exists in
  the processkit tree but the corresponding entries are missing from
  the harness settings file, with a "run aibox sync once aibox#55
  ships" hint. No-op in the processkit dogfood repo.

Changed

• chore(aibox): host CLI 0.18.7 → 0.19.2 + record mcp_config_hash
  in lock file. Bumped the harness CLI; aibox.lock now carries an
  mcp_config_hash field recording the sha256 of the merged
  .mcp.json at sync time, giving pk-doctor and aibox-sync a stable
  reference for drift detection independent of per-skill
  manifest hashes.

Fixed

• chore(repair): clear 3 pre-existing pk-doctor schema ERRORs.
  One-shot direct edit on a WorkItem whose YAML quoting was mangled
  by a fragile MCP serializer, plus two LogEntries missing the
  required actor field (written by a broken log_event MCP path
  that didn't populate the field). No tooling changes here — the
  producer-side MCP fixes are tracked as RapidDaisy
  (log_event missing actor validation) and CalmArch
  (workitem MCP description quoting) for a future release.
• chore(migrations): reject MIG-20260425T164303. Malformed
  migration (rejected via reject_migration per the
  migration_integrity finding above). Filed CleverRiver as the
  upstream-side aibox-sync diff defect that produced it.

External

• projectious-work/aibox#55 — aibox-side preauth merge (WildGrove
  Phase B). The processkit-side spec stays in-tree; preauth_applied
  WARNs in derived projects until the matching aibox sync logic
  ships.

---

v0.21.0 — retro batch + pk-doctor portability fixes

Five fixes spanning the v0.19.2 retro action items and two
correctness bugs in pk-doctor that surfaced when the owner ran
/pk-doctor inside an aibox-derived project and got a falsely-clean
report. Both pk-doctor bugs were silent-zero failures: the checks
silently no-op'd in derived projects and hid every entity-hygiene
problem they were supposed to catch.

The two heavier retro items (SharpBrook MCP schema hot-reload,
SnappyBird append-only LogEntry repair) were worked through design
discussions (DISC-DaringBird, DISC-WiseLily) and split or narrowed
per DEC-QuickPine and DEC-BrightHawk.

Added

• feat(_lib + 4 servers): reload_schemas MCP tool. Each of
  the schema-active servers (workitem-management, decision-record,
  event-log, artifact-management) gains a thin tool that calls a
  shared helper in _lib/processkit/schema.py::reload_caches(),
  clearing the in-process load_schema and
  state_machine.load lru_cachees so a disk-level schema or
  state-machine edit becomes visible without a server restart.
  Returns {ok, cleared: {schemas, state_machines}}. PEP 723
  dep-header edits are NOT addressed (the uv-resolved venv is
  pinned at process start — tracked separately by
  BACK-RapidSwan for v0.22.0+). Closes
  BACK-20260424_0128-BraveBird-reload-schemas-mcp-tool per
  DEC-20260424_0127-QuickPine.
• feat(model-recommender): 20 specialist + expert seed bindings.
  The v0.19.0 roster declared a 5-level seniority ladder
  (junior → specialist → expert → senior → principal), but the
  default-bindings pack seeded only 3 levels. As a result,
  resolve_model(<role>, "specialist"|"expert") returned
  no viable model across all 10 seeded roles — 20 missing
  combinations. KeenFern adds seeds inheriting from the nearest
  covered neighbour: specialist → junior's (model, effort band),
  expert → senior's (model, effort band). MANIFEST.yaml grew from
  30 → 50 seeds. New regression test
  test_default_bindings_coverage.py asserts every seeded role
  covers all 5 ladder levels and fires loudly on any future gap.
  Closes BACK-20260424_0134-KeenFern-fill-model-assignment-seniority.
• docs(AGENTS.md): "Sub-agent delegation" section. Codifies
  the read-only / mutating split for harness Agent-tool sub-agents
  (delegate read-only; keep mutating writes on the main session).
  Closes BACK-20260424_0038-ToughAnt-ephemeral-sub-agent-defaults.
• docs(AGENTS.md): "LogEntry repair" Skill-guards entry.
  Documents the narrow hand-edit escape hatch for schema-invalid
  LogEntries and points to pk-doctor --fix=schema_filename as
  the preferred path. Part of MightyFjord.

Changed

• feat(skill-gate): compliance-ack TTL is now an idle timeout.
  _any_valid_marker() rewrites the matching marker's
  acknowledged_at to now on every successful gate check.
  Sessions actively making compliant writes never expire mid-flow
  (the v0.19.2 midnight-span pain pattern); idle sessions
  (> _ACK_LIFETIME_HOURS, default 12 h) still must re-acknowledge.
  Touch failures are non-fatal. Closes
  BACK-20260424_0038-SwiftLynx-compliance-contract-acknowledgement-ttl.

Fixed

• fix(pk-doctor): schema_filename schemas-dir fallback (HappyReef).
  schema_filename.py resolved the schemas dir as
  <repo_root>/src/context/schemas/ only — a path that exists
  exclusively in the processkit dogfood repo. In every derived
  project (aibox-installed) schemas live at <repo_root>/context/ schemas/, so the check silently walked 0 entity files and
  reported 0 ERROR / 0 WARN regardless of how many malformed
  entities were on disk. The new _resolve_schemas_dir() tries
  the dogfood path first, then falls back to context/schemas/,
  and surfaces a single WARN when neither exists. Verified
  against an aibox checkout: the check now walks 186 entity files
  and reports the real findings (filename-date mismatches on
  _0000 placeholder filenames, missing-actor ERRORs, etc.).
  Closes BACK-20260425_1041-HappyReef-pk-doctor-schema-filename.
• fix(pk-doctor): migrations layout fallback (DeepMoss).
  Companion to HappyReef. The migrations check looked only in
  context/migrations/pending/ (processkit dogfood convention).
  Derived projects (aibox) keep pending migrations at the top
  level of context/migrations/ and only move applied ones into
  applied/. The new _candidate_pending_paths() probes
  pending/ first; if absent, walks context/migrations/*.md
  minus applied/, INDEX.md, and aibox-CLI upgrade-doc
  filenames (<YYYYMMDD>_<HHMM>_<from>-to-<to>.md). _list_pending
  also filters by kind: Migration so non-Migration markdown
  alongside migrations is ignored. Closes
  BACK-20260425_1041-DeepMoss-pk-doctor-migrations-detect.
• fix(pk-doctor): --fix=schema_filename for the CalmAnt-class
  pattern (MightyFjord). New run_fix on
  schema_filename.py patches LogEntries missing the required
  actor field by inserting actor: system, validates
  post-patch, and rolls back on failure. Behind --fix=schema_filename --yes, never default. Closes
  BACK-20260424_0128-MightyFjord-pk-doctor-fix-schema per
  DEC-20260424_0128-BrightHawk.
• fix(skill-gate): remediation text uses on-disk contract version.
  _remediation_msg() parses the leading <!-- pk-compliance vN -->
  marker rather than hard-coding v1. Part of SwiftLynx.
• fix(retrospective): /pk-retro --auto-workitems no longer fails
  with ModuleNotFoundError: No module named 'mcp'. Added
  mcp[cli]>=1.0 and jsonschema>=4.0 to pk_retro.py's PEP 723
  header so uv run --script pk_retro.py resolves the in-process
  MCP loader's transitive imports automatically. Closes
  BACK-20260424_0038-WildLake-pk-retro-auto-workitems.
• fix(skills): resolve cross-category retrospective skill basename
  collision. product/retrospective is renamed to
  product/sprint-retrospective; the processkit-category
  retrospective (release-scope, /pk-retro) keeps the bare name.
  Closes #11.

Deferred

• BACK-20260424_0128-RapidSwan-pk-doctor-server-header — pk-doctor
  server_header_drift check (PEP 723 dep drift detection,
  WARN-only). Targeted v0.22.0+.
• BACK-20260424_0037-SharpBrook-mcp-servers-cache-schemas
  (cancelled; superseded by BraveBird + RapidSwan per
  DEC-20260424_0127-QuickPine).
• BACK-20260424_0038-SnappyBird-data-repair-path-for
  (cancelled; replaced by the narrow MightyFjord run_fix per
  DEC-20260424_0128-BrightHawk; general data-fix migration kind
  not built — revisit only if >2 more recurrences in a calendar
  quarter).

---

v0.20.0 — retro follow-up: WildLake + ToughAnt + SwiftLynx + #11

Retro follow-up batch: fixes the three lightest action items from the
v0.19.2 retrospective (WildLake, ToughAnt, SwiftLynx). The two heavier
retro items — SharpBrook (MCP schema hot-reload) and SnappyBird
(data-repair path for append-only LogEntries) — are deferred to
v0.21.0 pending the design discussions opened as DISC-DaringBird and
DISC-WiseLily. See DEC-20260424_0101-SolidBadger-split-v0-20-0.

Added

• docs(AGENTS.md): "Sub-agent delegation" section. Pairs with the
  existing "Skill guards" list and codifies the read-only / mutating
  split for harness Agent-tool sub-agents: delegate Read, search
  Bash, and MCP query_* / get_* / search_* / list_*; keep
  Write, Edit, new mkdir, MCP create_* / transition_* /
  record_* / link_* / open_* / log_event, and git mutations
  on the main session. Closes
  BACK-20260424_0038-ToughAnt-ephemeral-sub-agent-defaults.

Changed

• feat(skill-gate): compliance-ack TTL is now an idle timeout.
  _any_valid_marker() in check_route_task_called.py now rewrites
  the matching marker's acknowledged_at to now on every
  successful gate check. A session that is actively making
  compliant writes no longer expires mid-flow (the v0.19.2
  midnight-span pain pattern); a session that goes idle for longer
  than _ACK_LIFETIME_HOURS (12h) still must re-acknowledge. Touch
  failures are non-fatal. Closes
  BACK-20260424_0038-SwiftLynx-compliance-contract-acknowledgement-ttl.

Fixed

• fix(skill-gate): remediation text uses the on-disk contract
  version. _remediation_msg() parses the leading
  <!-- pk-compliance vN --> marker in
  assets/compliance-contract.md rather than hard-coding v1. The
  v0.19.2 hard-code told callers to acknowledge v1 even though the
  on-disk contract was already v2. Part of SwiftLynx.
• fix(retrospective): /pk-retro --auto-workitems no longer fails
  with ModuleNotFoundError: No module named 'mcp'. pk_retro.py
  declared only pyyaml in its PEP 723 header, but its in-process
  MCP loader imports server.py modules from artifact-management,
  event-log, and workitem-management — all of which require
  mcp[cli] and jsonschema. Added both to the PEP 723 header so
  uv run --script pk_retro.py resolves them automatically. Closes
  BACK-20260424_0038-WildLake-pk-retro-auto-workitems.
• fix(skills): resolve cross-category retrospective skill
  basename collision. product/retrospective is renamed to
  product/sprint-retrospective (directory, name:, and id: SKILL-sprint-retrospective); the processkit-category
  retrospective (release scope, /pk-retro) keeps the bare name.
  Closes #11. The missing pk-doctor/commands/ directory
  (#10) and the duplicate pk-resume.md command
  (#12) were already resolved in v0.19.2 (90c980f and 87e185e
  respectively) and are closed here for bookkeeping.

Deferred to v0.21.0

• BACK-20260424_0037-SharpBrook-mcp-servers-cache-schemas — MCP
  schema hot-reload / reload_schemas tool. Design discussion:
  DISC-20260424_0101-DaringBird-how-should-mcp-servers.
• BACK-20260424_0038-SnappyBird-data-repair-path-for — data-repair
  path for malformed append-only LogEntries. Design discussion:
  DISC-20260424_0101-WiseLily-how-do-we-allow.

---

v0.19.2 — release-hygiene + derived-project-install bulletproofing

Release-hygiene and derived-project-install bulletproofing. Closes
four v0.19.2 WorkItems (SteadyCedar, BraveDove, ToughMeadow,
HappyFinch) and ships the processkit-side half of TrueQuail; the
aibox-side reconcile is tracked at
projectious-work/aibox#54.

Added

• feat(pk-doctor): commands_consistency check. Walks every
  context/skills/processkit/*/SKILL.md and ERRORs when a declared
  commands[].name has no matching commands/<name>.md file (WARNs
  on the inverse — stray files not declared in metadata). Prevents
  recurrence of the v0.19.1 slip where pk-doctor shipped without its
  own commands/pk-doctor.md. Closes
  BACK-20260423_1103-ToughMeadow-pk-doctor-skill-missing.
• feat(pk-doctor): mcp_config_drift check. Reads
  context/.processkit-mcp-manifest.json, recomputes per-skill
  mcp-config.json sha256es, and reports manifest staleness (WARN)
  or — in a derived-project context (aibox.lock + .mcp.json at
  repo root) — missing processkit servers in .mcp.json's
  mcpServers map (ERROR with a "run aibox sync" hint). Surfaces
  the exact failure mode that motivated TrueQuail.
• feat(release-semver): scripts/generate-mcp-manifest.py. Writes
  context/.processkit-mcp-manifest.json + src/ mirror with a sha256
  per-skill-config and an aggregate hash. Wired into
  scripts/build-release-tarball.sh so every release tarball ships a
  fresh manifest. Stable contract for downstream installers (aibox)
  to detect per-skill-config drift independent of the processkit
  version delta — see
  DEC-20260423_2049-VastLake-truequail-split-processkit-ships and
  projectious-work/aibox#54. Closes the processkit-side of
  BACK-20260423_0829-TrueQuail-aibox-installer-reconcile-mcp.
• feat(pk-doctor): ship commands/pk-doctor.md. v0.19.1 declared
  the slash command in SKILL.md metadata but forgot the file, so
  /pk-doctor was never registered in derived projects. This release
  ships the file in both trees and adds commands_consistency above
  as the prevention mechanism.
• docs(AGENTS.md): "MCP config manifest" contract section.
  Documents the manifest shape + path so aibox (and any other
  processkit installer) can implement against a stable interface.

Changed

• chore(skills): reconcile SKILL.md commands: metadata to /pk-
  namespace. 14 skills had stale <skill>-<verb> names declared in
  metadata while actually shipping /pk-<verb>.md files. This
  release renames the metadata to match shipped filenames (and for
  team-creator, promotes the commands: block out of provides: to
  metadata.processkit.commands: so the new
  commands_consistency check can see it). Drops three entries for
  commands that were declared but never shipped
  (model-recommender-profile, owner-profiling-refine,
  skill-reviewer-bulk-gotchas). No behavioural changes to any
  command. Closes
  BACK-20260423_2055-HappyFinch-skill-md-commands-metadata.

Fixed

• fix(model-recommender): add missing pyyaml>=6.0 to MCP server's
  PEP 723 header. The resolve_model call failed with
  ModuleNotFoundError: No module named 'yaml' when the harness
  invoked uv run server.py (no --script). Smoke-tested live
  returning an 8-layer routing trace for
  ROLE-product-manager@senior. Closes
  BACK-20260423_0829-SteadyCedar-model-recommender-mcp-server.
• fix(schemas): widen identity-class ID pattern to
  ^(ACTOR|TEAMMEMBER)-[a-zA-Z0-9_-]+$ across five schemas. The
  v0.19.0 TeamMember rollout left workitem.assignee,
  decisionrecord.deciders[], discussion.participants[],
  artifact.owner, and metric.owner still requiring ACTOR-*
  only, so MCP writes carrying a TEAMMEMBER-* subject failed
  jsonschema validation. Chose alternation over a clean flip to
  avoid retroactively invalidating ~90 residual ACTOR-* references
  in 44 entity files. Smoke-tested live post-harness-restart with
  both a create_workitem(assignee="TEAMMEMBER-cora") and a
  record_decision(deciders=["TEAMMEMBER-cora"]). Closes
  BACK-20260422_1643-BraveDove-schema-drift-workitem-assignee.
• fix(log): add missing required actor field to
  LOG-20260422_1643-CalmAnt-workitem-created. Pre-TeamMember MCP
  server wrote a LogEntry without populating actor, so
  pk-doctor --category=schema_filename reported a schema ERROR.
  actor: system accurately reflects the unattributed machine-origin
  of the event. One-off direct edit; no MCP tool exists to patch
  append-only LogEntries.

External

• projectious-work/aibox#54 — aibox-side reconcile-on-manifest-drift
  tracking issue. BACK-20260423_0829-TrueQuail stays open in the
  processkit backlog until that PR lands.

---

v0.19.1 — local-only release bulletproofing (skill flow + pk-doctor release_integrity)

Local-only release bulletproofing — no CI workflows. Addresses the
v0.19.0 post-mortem finding that a tag push does not create a GitHub
Release. Supersedes DEC-20260422_0926-MerryArch (CI workflow
approach, rejected on vendor-lock + cost grounds) with
DEC-20260422_1348-SnowyWolf (skill flow + doctor detection).

Added

• feat(pk-doctor): 6th check category release_integrity — walks every local v* git tag, probes GitHub via gh release view for a matching Release, and WARNs on any tag without one. INFO when gh is unavailable or the tag set is empty. Opt-out via PK_DOCTOR_SKIP_RELEASE_INTEGRITY=1; tag-scan cap via PK_DOCTOR_RELEASE_INTEGRITY_MAX (default 50). Each WARN carries a ready-to-paste gh release create <TAG> command with CHANGELOG extraction inlined. See DEC-20260422_1348-SnowyWolf-local-only-release-bulletproofing.

Changed

• chore(release-semver): collapse /pk-release into a single bulletproof flow. SKILL.md (both trees) rewritten with a 9-step recipe that prepares, publishes, and verifies in one turn — the release is not considered complete until gh release view vX.Y.Z succeeds (step 8). /pk-publish retained as a recovery alias for historical tags that are missing a Release. New Gotcha: "git push --tags is not a GitHub Release." Closes BACK-20260422_0925-MightyOtter-v0-19-1-release.

Fixed

• fix(release flow): v0.19.0 tag was pushed without a corresponding GitHub Release — the prepare phase of release-semver completed but the publish phase was skipped. v0.19.0 Release was created manually; v0.19.1 ships the prevention (collapsed flow) and detection (release_integrity check) mechanisms.

---

v0.19.0 — TeamMember model + 51-role catalog + Model artifacts + binding routing

[v0.19.0] — 2026-04-22

Added (v0.19.0 architecture refactor)

• feat(team-manager): new skill replacing actor-profile. Persistent
  participants are now TeamMembers (humans, named AI personas,
  services), each living as a directory tree:
  context/team-members/<slug>/ with team-member.md (entity), persona.md
  (loaded as persona prompt), card.json (A2A v0.3 Agent Card), and six
  memory tiers (knowledge/, journal/, skills/, relations/,
  lessons/, private/). Ad-hoc worker invocations are not
  persisted — they are ephemeral (role, seniority) dispatches. Skill
  ships 17 MCP tools (lifecycle, name pool, memory tree, export/import,
  consistency), a curated 59-name international name pool, and 10
  consistency check categories. Replaces actor-profile (clean break,
  no backward compat). See DEC-20260422_0233-SpryTulip.

• feat(team-member memory): file-based tiered memory. Six tiers
  (working, episodic, semantic, procedural, relational, lessons) all
  stored as Markdown with YAML frontmatter (tier, source,
  sensitivity, confidence, importance, created, last_reinforced,
  scope). Default consolidation cadence: per-task + daily journal +
  weekly importance-triggered promotion (Park-style). Each team-member
  has a private/ subdirectory for developer-local notes
  (gitignored via .gitignore.example shipped at repo root).

• feat(team-member export/import): export_team_member produces a
  versioned tarball excluding journal/, relations/, and
  private/ by default; sensitivity-tagged files (pii, confidential)
  are redacted. import_team_member validates the A2A card signature
  field and creates the entity tree on import.

• feat(role catalog): expanded from 8 to 51 curated roles under
  context/roles/. Function-grouped (engineering-software, platform-infra,
  data-ml, security, architecture, design-ux, marketing, sales-customer,
  finance, legal-compliance, executive, etc.). Seniority no longer
  baked into slugs — pure ordinal attribute with ladder
  junior → specialist → expert → senior → principal.
  Bindings (not role files) decide what each rung maps to in
  (model, effort). See DEC-20260422_0234-BraveFalcon.

• feat(model artifacts): models become first-class entities under
  context/models/, one file per (provider, family) with
  versions[] nested. 34 model artifacts replace the monolithic
  model_scores.json registry (which becomes a compiled cache).
  Each model carries equivalent_tier in the provider-neutral
  T-shirt capacity ladder (xs / s / m / l / xl / xxl, extensible
  in both directions). Effort enum normalised to
  [none, low, medium, high, extra-high, max] (aliases extra-high → xhigh
  at the Anthropic adapter boundary). See DEC-20260422_0234-LoyalComet.

• feat(bindings): model-assignment binding type + 8-layer resolver.
  model-recommender.resolve_model(role, seniority?, team_member?, scope?, task_hints?)
  returns ranked (model, version, effort) candidates via the precedence
  ladder: task-pin → team-member preference → project veto → capability
  filter → role+seniority → role default → project bias → shim fallback.
  Tie-breakers: project-preferred provider → cost → recency → reliability.
  Effort clamping, version pinning, stale-binding skip, in-module result
  caching, explain-mode trace. New /pk-explain-routing slash command
  for debugging. Default binding pack at
  context/skills/processkit/model-recommender/default-bindings/MANIFEST.yaml
  ships 30 starter bindings (10 roles × 3 seniorities) materialised into
  context/bindings/.

• feat(pk-doctor): 5th check category team_consistency. Wraps
  team-manager.check_all() and surfaces the 10 team-consistency check
  codes in pk-doctor's standard report (schema drift, tier-missing,
  dangling refs, name collision, name-pool compliance, orphan files,
  sensitivity placement, private-dir gitignore, memory file headers,
  card staleness).

• chore(.gitignore.example): bundle of all proposed processkit
  ignores (cache/state, context/**/private/, model cache, harness
  configs, OS/IDE/Python noise) shipped at repo root for adoption by
  new processkit projects.

Removed

• actor-profile skill superseded by team-manager. 8 role-class
  actors removed (ACTOR-developer, ACTOR-assistant, ACTOR-pm-claude,
  ACTOR-{sr,jr}-{architect,researcher,developer}); identity-class
  ACTOR-20260421_0144-AmberDawn-legacy-historical-backfill removed
  (not meaningful); ACTOR-20260421_0144-ThriftyOtter-owner migrated
  to TEAMMEMBER-thrifty-otter. 7 legacy role files
  (ROLE-{developer,project-manager,senior-architect,senior-researcher, junior-architect,junior-developer,junior-researcher}) removed,
  superseded by entries in the curated 51-role catalog. 8 legacy
  role-assignment bindings removed.

Fixed

• fix(mcp): systemic auto-log actor fix — every entity-mutating MCP tool (create_, transition_, link_, update_, apply_, reject_, start_, end_, open_, record_, supersede_, deactivate_) now passes actor= to log helpers, producing schema-valid LogEntries. Covers the bug pattern previously fixed for create_actor in WarmGrove. Includes backfill of 6 pre-fix-emission LogEntries from this session and drift-allowlist entry for template-only scripts/ subdirs.
• fix(actor-profile): create_actor MCP tool now emits actor.created LogEntry with spec.actor = new actor id (self-attribution); previously the emitted LogEntry was schema-invalid (missing required field). Closes BACK-20260421_0156.

Changed

• chore(mcp): skill-consultation prompt added to every entity-mutating MCP tool description across all processkit servers (provider-neutral replacement for pre-tool-use hooks). Closes BACK-20260411_0802-SolidCrow.
• chore(AGENTS): slim root AGENTS.md to <60 lines of core content; domain-specific instructions moved into their owning skills; explicit if/then skill guards added. Closes BACK-20260411_0802-EagerSpruce.
• chore(grooming): rename 4 process files and 1 artifact to match canonical metadata.id prefixes (PROC-/ART-). pk-doctor WARN residual eliminated (5 → 0). Closes pk-doctor Phase 2 filename-rename intent.

Added

• Renamed skill: morning-briefing → status-briefing (provider-neutral; time-of-day agnostic). Content migration emitted for downstream projects.

• New skill: retrospective — /pk-retro generates post-release blameless retrospectives (4 signals: release_summary, timeline, workitems, drift; dual-emit Artifact+LogEntry; --auto-workitems for proposed follow-ups; --verbose for full narrative including Appendix A raw signal dumps). See BACK-20260420_1340-LoyalFrog-add-pk-retro-skill.

• feat(pk-doctor): new health-check aggregator skill + /pk-doctor slash command (Phase 1). Detect-only by default; 4 checks: schema+filename validation against src/context/schemas/, sharding (logs YYYY/MM, migrations state-bucket), stale pending migrations, src/context drift. --fix/--fix-all opt-in; fixes route through existing MCP write tools (no hand-edits). Each run emits a doctor.report LogEntry via event-log MCP. See DEC-20260420_1631-WiseGarnet and BACK-20260420_1631-ProudGlade.

• feat(migration-management): new MCP server exposing 5 tools
  (fixes #9).
  list_migrations, get_migration, start_migration,
  apply_migration (with implicit-start when called from pending),
  and reject_migration. Each write-side tool stamps the appropriate
  timestamp (started_at / applied_at / rejected_at), moves the
  Migration file between pending/, in-progress/, and applied/
  subdirectories, refreshes context/migrations/INDEX.md (preserving
  the Applied-table Notes column and the ## CLI Migrations tail
  section), and writes a migration.* event via the side-effect log
  helper. Rejected migrations park under applied/ and are listed in
  a dedicated ## Rejected INDEX.md section. Adds
  spec.started_at to the Migration schema (optional,
  backward-compatible). See
  DEC-20260420_1342-WarmClover (shape) and
  DEC-20260420_1353-ProudReef (5 refinements).

---