No MCP server ran on central, so remote agents reached dev-hub only through a
reverse tunnel back to the workstation — routing a request for a service on
their own machine out to another host and back.
Adds a gated mcp Deployment and ClusterIP Service running the same image with
`-m mcp_server.server`, API_BASE defaulted to the in-cluster API Service, and
tcpSocket probes. No Ingress: the MCP layer proxies an unauthenticated API and
must not be reachable from outside the cluster.
Two fixes were needed before the manifests could work:
- server.py hardcoded host="127.0.0.1". A Service routes to the pod IP, so a
loopback bind is unreachable. Now MCP_HOST, still defaulting to loopback so
local runs do not silently expose an unauthenticated proxy.
- The container runs `-m mcp_server.server`, not the file path, so /app lands
on sys.path rather than /app/mcp_server.
mcp.enabled stays false in the deploy values: the running image predates
MCP_HOST, so enabling it before the tag is bumped would ship a pod that never
becomes reachable.
Refs CUST-WP-0067-T08
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 2583210@bnt-lap001
Assistant-Session: f2bff2d5-e9b2-4338-92ca-10282a927006
History is archived fleet-side, so the read surface has no remaining job.
- api/routers/suggestions.py: whole prefix 410s with a pointer to intake and
to the archive; the router drops from 176 lines to a stub
- mcp_server: the 6 suggestion tools removed outright rather than stubbed --
a retired tool that still appears in the tool list costs every agent
session context on every call, which is the opposite of retiring it
- write_idempotency: 5 /suggestions rules dropped
- dashboard: suggestions.md deleted, nav entry removed, reference.md and
wsjf-triage.md updated; docs/suggestions.md rewritten as archive pointer
- tests: two tests pinned the old read-live behaviour and now pin the
retirement contract instead
Tables stay: they are retire/archive in SHR-INV-0001 and are captured by the
final dump at T06.
Untouched, and worth knowing during cutover: ui-feedback.md / todo.md
'suggestions' are Shift+click dashboard feedback backed by technical_debt, a
different entity that shares the word. E3 (dashboard-meta) is that page; its
owner is state-hub-until-cutover so it retires at the T06 window, not now.
Full suite 612 passed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Founder-reviewed decision (WorkOrchestrationArchitectureDraft.md v0.2
section 8 item 6): the fresh intake work-record entity replaces
suggestions, not a rename-bridge. All 5 mutation endpoints (create, vet,
decline, promote, bump-relevance) now 410 with a pointer to POST
/intakes/ and the migration doc; GET/list stay live for the historical
record (10 rows migrated to file-backed intake records in the-custodian,
see that repo's intake-legacy-suggestions-migration.md and CUST-IN-0001
through CUST-IN-0010).
Removed dead code the retirement makes unreachable: Task/TaskPriority/
TaskStatus/normalize_task_status imports (only used by the deleted
promote body), the suggestion_relevance.bump_relevance import, and the
_ALLOWED_*_FROM stage-guard sets + _reject_stage helper (only used by
the deleted vet/decline/promote bodies). WSJF ranking (compute_wsjf,
cost_of_delay, suggestion_sort_key) stays -- still exercised by the
surviving GET /suggestions/?rank=wsjf read path.
MCP tool docstrings (create_suggestion, vet_suggestion,
decline_suggestion, promote_suggestion_to_task,
bump_suggestion_relevance) updated to point at the replacement
(create_intake/route_intake/close_intake) rather than silently 410ing
with no guidance.
tests/test_suggestions.py rewritten: verifies all 5 mutations 410,
GET/list still work for historical rows (seeded directly via the DB
session since creation is retired -- there's no other way to get
historical data into the table anymore), 404 still behaves normally on
unknown ids. Live-verified against the running dev API: POST 410s,
GET with include_terminal=true still returns all 10 migrated-and-declined
historical rows. No regressions: full repo suite green (563 tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Fresh hub entity per the founder-reviewed decision (not a suggestions
rename-bridge): kind: intake per canon/standards/work-record-types_v0.1.md,
lifecycle open -> vetted -> routed -> closed(promoted|declined|absorbed).
- api/models/base.py::new_uuid7 -- dependency-free RFC 9562 UUIDv7
generator (48-bit ms timestamp, version/variant bits, random remainder);
existing tables keep new_uuid (UUIDv4) unchanged, this is opt-in for new
work-record entities per the identity-layering canon
- api/models/intake.py: Intake + IntakeNote ORM models, mirroring
Decision's shape (topic/workplan/repo scope, lane, status, outcome,
promoted_to back-link); CHECK constraints enforce scope-required,
closed-requires-outcome, promoted-requires-promoted_to at the DB level
- migrations/a7c3e9f1b4d2: intakes + intake_notes tables, 3 enum types
- api/routers/intake.py: list/create/get/patch + /route + /close + /notes
actions, mirroring decisions.py's pattern (409 on invalid transitions,
progress event on close)
- api/schemas/intake.py: Pydantic create/update/route/close/note schemas
- mcp_server/server.py: create_intake, list_intakes, route_intake,
close_intake tool wrappers
- tests/test_intake.py: 12 tests against the real Postgres test DB
(create/list/scope-validation, full lifecycle incl. 409s and the
promoted-requires-promoted_to constraint, notes, UUIDv7 verification)
Verified live against the running dev API + DB (not just pytest): applied
the migration, restarted the MCP server, and ran a full create -> route ->
close cycle over the real REST endpoints. No regressions: full existing
suite (test_routers_core, test_suggestions, test_mcp_smoke,
test_mcp_write_tools, test_mcp_registration, test_consistency_check,
test_consistency_sweep) all green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Removes the redundant open_workstreams mirror field from StateSummary (schema,
router, MCP get_domain_summary reader, dashboard consumers + empty-state stub,
and tests). Consumers already preferred open_workplans, so this is the
low-risk half of T04.
Deferred (still have live callers — not yet retirement-ready):
- workstream_id query/body field alias on preferred routes — external
session-close curls/scripts fleet-wide still send it.
- flows/workstream.yaml — /flows/workstream/{id} routes are still served and
exercised by tests; retire only once no callers remain.
Staged on branch state-wp-0070-legacy-retirement — do not merge until the 7th
documented zero-usage window is captured.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Meter-gated removal (7 consecutive zero-usage windows reached). Deletes:
- MCP tools: create_workstream, list_workstreams, update_workstream,
update_workstream_status
- MCP resource: state://workstreams/{topic_slug}
- Dead helpers: _LEGACY_MCP_* maps, _legacy_mcp_deprecation,
_meter_legacy_mcp, _attach_legacy_deprecation, _update_workplan_legacy_impl
TOOLS.md keeps a retired→preferred migration map (per backlog). Removed the
now-dead STATEHUB_MCP_LEGACY_METER guard in conftest. Retargeted the
create_workplan error-skip test to the preferred tool; dropped alias-only tests.
Staged on branch state-wp-0070-legacy-retirement — do not merge until the 7th
documented zero-usage window is captured.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Alias tests monkeypatch _get/_post but _meter_legacy_mcp uses _client()
directly, so every pytest run recorded real usage and reset zero-window
retirement streaks. STATEHUB_MCP_LEGACY_METER=off skips metering; set in
tests/conftest.py. Verified meter counts flat across the alias suite.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
T03: MCP alias retirement gate was unstartable — deprecation notices existed
but no mcp:* interfaces were ever registered or metered. Add
_meter_legacy_mcp() posting /legacy-meter/usage (kind mcp_tool, component
state-hub.mcp) from create_workstream, list_workstreams, update_workstream,
update_workstream_status, and state://workstreams/{topic_slug}; register all
five phase-3 backlog keys so zero-window streaks accrue from today.
T04: trace residual POST /progress/ workstream_id (3 calls/8h, unknown) to
.custodian-brief.md labelling workplans "workstream_id:"; brief generator now
prints "workplan_id:". Briefs refresh on each repo's next fix-consistency run.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stop dual-publishing org.statehub.workstream.completed; return 410 Gone for
legacy DELETE /workstreams/{id}. Migrate fix-consistency, MCP adhoc bootstrap,
and dashboard token summary to /workplans/. Add legacy-meter evidence capture
script and pytest snapshot; update docs and close out the workplan.
_emit_progress_event strips legacy workstream_id from POST /progress/
payloads. Task and decision side-effect events now set workplan_id
explicitly so automatic progress logging stays workplan-first.
Wire hub-core progress legacy-meter hook for GET /progress/?workstream_id.
MCP list_tasks, list_blocked_tasks, and list_human_interventions now call
REST with workplan_id only so internal clients do not inflate legacy usage.
Introduce flows/workplan.yaml and /flows/workplan/* routes alongside the
legacy workstream flow. Dashboard, README, and MCP flow docs now prefer
open_workplans with open_workstreams fallback until dual-key retirement.
Legacy MCP tools and the workstreams resource now return _deprecation payloads
pointing to workplan-named replacements. get_domain_summary uses workplan-first
prose and exposes a workplans key alongside the legacy workstreams field.
Add C-25..C-30 fix-consistency checks for blocked-workplan inbox sweeps,
stale unread triage, workplan ID prefix/collision lint, and SCOPE freshness.
Extend brief generation and get_domain_summary with inbox hygiene warnings.
Complete workplan_id aliases on remaining MCP tools and retry transient
_api_get failures to reduce false stale-reference errors under load.
Introduce canonical MCP_SERVER_NAME constants, shared registration helpers,
and a migrate_mcp_config.py script for ~/.claude.json upgrades. Registration,
patch, and custodian CLI checks accept both dev-hub and legacy state-hub during
transition. API root metadata and session-protocol template reflect the new name.
Replace the fixed 15s TTL on GET /state/summary with per-table revision
watermarks, stale-while-revalidate background refresh, and a progress-tail
section split. SQLAlchemy write hooks invalidate core or progress sections
on mutation. Adds tests, benchmark script, and operator docs.
Adds first-class tracking for API and interface mutations across the
agent ecosystem. Breaking changes are documented, affected repos are
notified via inbox, and agents discover pending changes at session
start via the dispatch endpoint.
- Migration q4l5m6n7o8p9: interface_changes table
- Model/schema: InterfaceChange with draft→published→resolved lifecycle
- Router: POST/GET/PATCH /interface-changes/, /publish, /resolve actions
(auto-notify affected repo agents on publish; progress event on origin)
- Dispatch: GET /repos/{slug}/dispatch now returns pending_interface_changes
- MCP tools: register_interface_change, list_interface_changes,
publish_interface_change, resolve_interface_change
- Dashboard: /interface-changes page with type badges, planned calendar,
published cards, and draft table
- EP-CUST-ICR-001 registered: webhook subscriptions (deliberately deferred)
First record: trailing-slash normalisation (2026-04-26), published,
affecting repo-registry — visible in repo-registry dispatch immediately.
223 tests passing.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Tier 1 (exact counts) now defaults to note="measured" instead of null,
signalling the counts were read from the Claude Code status bar.
Callers can pass note="userbased" when a human provided the numbers.
measured — agent read exact counts from the Claude Code status bar
userbased — counts provided by a human
workplan — prorated from workplan total across task count
heuristic — server fallback, 1000/500, no agent input
Added token_note field to TaskUpdate schema and exposed note param on
update_task_status and record_interactive_task MCP tools.
TOOLS.md documents the full taxonomy. 185 tests pass.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
New tool for capturing ad-hoc work done outside formal workplans.
Finds or creates a persistent 'interactive-<repo>' workstream for the
repo, creates the task, marks it done, and records a token event using
the three-tier logic — all in a single call.
Seeded two example events on interactive-the-custodian:
- Three-tier token recording on task done (8000/3500)
- Add record_interactive_task MCP tool (4500/1800)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Token events are now always created when update_task_status is called
with status="done", using the best available data:
Tier 1 (best): exact tokens_in + tokens_out passed by agent
Tier 2: workplan_tokens_in + workplan_tokens_out prorated
across workstream task count (note="workplan")
Tier 3 (fallback): heuristic 1000 in / 500 out (note="heuristic")
Non-done status changes never create a token event.
MCP tool updated with workplan_tokens_in/out params and tiered docs.
Ralph-workplan skill files updated with the three-tier guidance.
184 tests pass.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
POST /topics/ was already implemented in the REST API but had no MCP
wrapper, so agents couldn't create topics (e.g. inter_hub) via MCP.
Tool follows the same pattern as create_domain.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add `Optional` to typing imports in mcp_server/server.py — it was used
in 13 annotations but never imported, crashing FastMCP v3 at startup
- Remove legacy tunnel/tunnel-daemon/tunnel-loop/tunnel-status/tunnel-stop
targets from Makefile; ops-bridge (tunnels-up/status/check) supersedes them
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a structured dispute mechanism when capability request routing is wrong:
- New `routing_disputed` status with four DB columns (dispute_reason, disputed_by,
dispute_suggested_domain, disputed_at) via Alembic migration m0h1i2j3k4l5
- POST /capability-requests/{id}/dispute — any party can flag misrouting with a reason
and optional suggested domain; notifies custodian + current fulfilling domain
- POST /capability-requests/{id}/reroute — custodian re-routes to correct domain via
catalog_entry_id or direct slug; appends audit trail to routing_note; resets to requested
- Two new MCP tools: dispute_capability_routing and reroute_capability_request
- Dashboard: amber disputed-banner at top of Summary, routing_disputed Kanban column,
dispute details (reason, suggested domain, raised-by) shown on disputed cards
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add `routing_note` column (migration l9g0h1i2j3k4) to persist why a request was routed to a given domain
- Fix substring-match bug in `_route_capability`: use `\b` word-boundary regex so 'postgres' no longer matches inside 'postgresql'
- Include `title` in keyword scoring for better routing accuracy
- Return `routing_note` string from `_route_capability` and store it on the request
- Add `PATCH /capability-requests/{id}` endpoint + `CapabilityRequestPatch` schema to correct mutable metadata (catalog_entry_id, priority, blocking_task_id, fulfilling_workstream_id)
- Add `patch_capability_request` MCP tool wrapping the new endpoint
- Add 105 lines of routing tests (word-boundary, title-match, multi-entry scoring, broadcast fallback)
- Add `tunnels-up`, `tunnels-status`, `tunnels-check` Makefile targets for ops-bridge managed tunnels
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Implements the 14-criterion DoI checklist as a runnable gate with API,
MCP tools, CLI script, and dashboard integration.
Core components:
- api/doi_engine.py — async engine evaluating all 14 criteria (asyncio.to_thread
for non-blocking HTTP self-calls), shared by API and CLI
- api/schemas/doi.py — DoICriterion, DoIReport, DoISummaryEntry schemas
- api/routers/repos.py — GET /repos/{slug}/doi + GET /repos/doi/summary
- scripts/check_doi.py — CLI: make check-doi REPO=<slug> / check-doi-all
- mcp_server/server.py — check_repo_doi(), get_doi_summary() tools
Dashboard (repos.md):
- DoI tier badge per repo (None/Core/Standard/Full) colour-coded red→green
- Domain block shows lowest DoI tier across its repos
- DoI KPI card in summary row
- DoI filter in All Repos Table
- Link to Repository DoI policy page
Also fixes: TPSC snapshots 500 error (missing nested selectinload for
catalog_entry relationship in list_snapshots endpoint).
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Stale host_paths entries (wrong username, old machine) were silently overriding
the correct local_path, causing FileNotFoundError on tools like list_kaizen_agents.
Extracts _resolve_repo_path(repo) helper that tries host_paths[hostname] first
but validates the path exists on disk before trusting it, then falls back to
local_path. Both candidates support ~ expansion. Applied to all 4 call sites:
_kaizen_agents_dir, validate_repo_adr, check_repo_consistency, ingest_sbom_tool.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Introduces a capability catalog (CUST-WP-0022) so domains can advertise what
they provide and agents can request capabilities from other domains with
auto-routing, lifecycle tracking, and task-unblocking on completion.
- New models: CapabilityCatalog, CapabilityRequest with full lifecycle
(requested → accepted → in_progress → ready_for_review → completed/rejected/withdrawn)
- Migration i6d7e8f9a0b1: capability_catalog + capability_requests tables
- Router /capability-catalog and /capability-requests with accept/status endpoints
- 7 new MCP tools: register_capability, list_capabilities, request_capability,
accept_capability_request, update_capability_request_status,
list_capability_requests, get_capability_request
- StateSummary gains open_capability_requests count
- Dashboard: capability-requests.md page + docs/capabilities.md + docs/scope.md
- SCOPE.md: three seed capabilities documented (MCP registration, state tracking, SBOM)
- scope.template: Provided Capabilities section with example block
- scripts/ingest_capabilities.py + make ingest-capabilities[/-all] targets
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Agents had no way to look up task UUIDs by workstream; they were stuck
unable to call update_task_status without already knowing the UUID.
list_tasks() wraps GET /tasks with workstream_id filter, returning
[{id, title, status, priority}] for all matching tasks.
FR raised by kaizen-agentic worker on COULOMBCORE while syncing
KAIZEN-WP-0002 task IDs. Marked merged in contributions.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
