df87fcf059
* [sweep] lifecycle: 6 confirmed gaps fixed (cancel-ceo-gate, claim_pr_review gate, needs_team_match, valid_next_verbs narrowing, pr_reviewer unclaim, complete side_effect ordering)
* [chore] logical-gaps: route-layer force gate + privileged-field gate + pre-task audit attribution
tasks.py (5 gaps):
- _HATCH_OVERRIDE_STATES expanded to 7: a privileged PATCH INTO a gate
state (completed/cancelled/awaiting_{qa,documentation,pr_review,
pm_review,ceo_approval}) now requires explicit force — the panel hatch
is no longer a quiet click that drops a task into/out of a human gate.
- _RESURRECT_SOURCE_STATES: a privileged PATCH OUT of a terminal status
(completed/cancelled) resurrects finished work and likewise requires
force, audited as an override.
- _PRIVILEGED_UPDATE_FIELDS gate: a bare task owner (UPDATE_OWN, no
ASSIGN) cannot self-reassign / re-team / re-parent / re-depend /
re-block / rewrite-plan / re-project its task — those structural fields
are PM-gated; the REST surface must not bypass the verb-layer's
reassign/delegate/triage gate. A 403 names the touched fields + the
verb to use instead.
- pre-task create denial: a role that cannot create tasks is now logged
via log_task_creation_denial (distinct task_creation target_type +
attempted payload) instead of a 'N/A' task_id that coerced to NULL and
left the role-escalation attempt unattributable.
audit.py:
- split log_task_action_denial (5-param, under PLR0913) from
log_task_creation_denial (4-param) — the create path has no task_id;
the non-UUID sentinel (N/A) is preserved in details[target_id_raw]
rather than dropped to a NULL target_id indistinguishable from any
other NULL-target denial.
tests:
- test_tasks_routes.py: parametrized admin-override gate (force
required for gate + terminal states, force succeeds).
- test_tasks_route_privileged_fields.py: dev owner 403 on
assigned_to/team/parent_task_id, 200 on dev-facing description.
- test_audit.py: pre-task attribution via log_task_creation_denial +
non-UUID sentinel preservation.
* [chore] logical-gaps: kanban board column coverage + status-class fixes (6 gaps)
models/kanban.py:
- DEV_COLUMNS: cover all 15 lifecycle statuses (was 7; dropped BACKLOG,
PAUSED, VERIFYING, NEEDS_REVISION, AWAITING_PR_REVIEW, AWAITING_PM_REVIEW,
AWAITING_CEO_APPROVAL, CANCELLED). A dev whose task bounced to
needs_revision or sits in a gate used to see their own task vanish.
- PM_COLUMNS: add the gate/revision/paused/cancelled/backlog columns so the
cell PM sees the QA->docs->PR-review->PM-review->CEO chain on its board.
- QA_COLUMNS: drop the 'In Review'->VERIFYING mapping. VERIFYING is the dev's
self-verification (task still with the dev, not with QA); it misrepresented
dev mid-verification as active QA work.
services/kanban.py:
- _build_flat_board: add an 'Other' fallback column for any task whose status
matches no configured column, so total_cards == sum(card_count) and no card
is built-then-silently-dropped (the vanished-card leak).
- get_qa_board: drop VERIFYING from qa_statuses (consistent with the column
change).
- get_documenter_board: scope to task_type=documentation so a dev IN_PROGRESS
code task sharing the cell team no longer appears under 'Gathering'.
- get_main_pm_board_flat: widen the status filter to include PENDING/CLAIMED/
COMPLETED and route those to the incoming/distributed/done columns, which
were structurally always empty under the in-flight-only filter.
tests/integration/test_kanban_service.py: parametrized coverage of every
dropped dev status, PM gate/revision states, QA excludes VERIFYING,
documenter excludes dev code tasks, flat Main PM incoming/distributed/done
populated, and the 'Other' fallback invariant.
* [chore] logical-gaps: lifecycle-enforcement validators + status-class fixes (5 gaps)
enforcement/task_lifecycle.py:
- drop the spurious VERIFYING->awaiting_documentation legacy edge. The
canonical exit is submit_qa -> awaiting_qa -> (qa_pass) ->
awaiting_documentation; the direct edge bypassed the entire QA review hop
(ungated — no role gate existed for it).
- is_waiting_state: add awaiting_pr_review. The PR-review gate parks the PM on
the reviewer; it is a waiting state. The hard-coded set was never updated
when AWAITING_PR_REVIEW was added to the enum, so the gate status was
miscategorized as active.
foundation/_validate_lifecycle.py:
- _check_status_enum_coverage: replace the tautology (STATUS_GRAPH keys every
Status by construction) with a real bidirectional check — every non-terminal
Status is the source of a transition (catches orphan states), and every
source/target referenced is a real Status member (catches stray-string
targets).
- _check_terminal_exits: split the {COMPLETED, CANCELLED} reachability into a
COMPLETED-path requirement + a cancel-exit requirement. The cancel fan-out
made the old check structurally trivial — a status whose sole exit was cancel
passed with no real forward completion path.
- _check_status_enum_parity (new, registered): cross-check spec.Status against
models.base.TaskStatus at import so the ORM column type and the lifecycle
map cannot drift (TaskType had this guard; Status did not).
tests: verifying->awaiting_documentation rejected, self-fail preserved,
awaiting_pr_review is waiting, mutually-disjoint classification invariant,
status enum parity, stray-string-target / orphan-source / cancel-only-exit
validator rejections.
* [chore] logical-gaps: stream-bus poison-pill ACK + dead-letter, periodic reclaim, cancelled-handler marker cleanup (3 gaps)
stream_bus.py:
- _handle_message isolates Event.from_json in its own try/except; an
undecodable payload (unknown EventType, bad UUID/timestamp, malformed
JSON) is dead-lettered then ACKed instead of falling through to the
broad except that only logged — a poison pill stayed pending forever
and re-failed on every reclaim. (gap: stream-bus-malformed-event-poison-pill)
- _reclaim_loop spawned alongside _listen_loop in start_listening (cancelled
in disconnect). XREADGROUP '>' delivers only NEW messages, so a runtime
handler failure left its message pending and unretried until a restart;
the loop re-runs recover_pending every 60s so the idempotency-guarded
replay actually fires. (gap: stream-bus-no-runtime-reclaim-loop)
- _run_handler_guarded marker cleanup catches BaseException so a handler
cancelled mid-flight (asyncio.CancelledError is BaseException-derived
since 3.8) clears its SET-NX marker; otherwise the guard suppressed the
very redelivery that would complete the work. (gap: stream-bus-cancelled-
handler-keeps-idempotency-marker)
TDD: 4 red->green tests in tests/unit/events/test_bus.py.
* [chore] logical-gaps: verb_runner trailing-None side-effect guard + actor_agent_id threading (3 gaps)
_verb_runner.py:
- run_intent skips the side_effects loop when a TRAILING composed action
returned None (its source-status check failed under a concurrent
transition). Previously the loop ran unconditionally on the None task
and _do_push_branch(None)/_do_pr_merge(None) crashed with a
NoneType AttributeError, turning the clean INVALID_STATE the
entry/intermediate guards give into a 500/respawn loop. The trailing
None now flows to the caller's `if task is None` handler. Latent today
(no shipped intent has both a None-capable compose and trailing
side_effects) but the runner is generic. (gap: runner-side-effects-fire-
on-trailing-none-task)
- _do_push_branch / _do_create_pr / _do_create_root_pr forward
actor_agent_id=agent.id into git_service (push_branch / create_pr),
matching _do_pr_merge. Without it, a verb on a task whose assigned_to
was cleared before the side effect falls through to created_by and
pushes from / opens a PR against the wrong workspace.
(gap: side-effect-handlers-drop-actor-agent-id)
- _do_escalate_to_ceo forwards actor_agent_id=agent.id so the
awaiting_ceo_approval audit row attributes to the specific PM/Board
agent. (gap: do-escalate-to-ceo-drops-actor-agent-id)
task.py: escalate_to_ceo gains actor_agent_id param, passed as
audit_agent_id to _validate_and_set_status and recorded as
escalated_by_agent_id in the event payload + log. escalate_to_ceo_for_agent
forwards agent.agent_id.
_impl.py: the main_pm complete->escalate path forwards
actor_agent_id=main_pm_agent_id.
TDD: 5 red->green tests (synthetic trailing-None intent, actor forwarding
for push_branch/create_pr/create_root_pr/escalate_to_ceo) + real-DB audit
test asserting the awaiting_ceo_approval row carries the actor UUID.
Updated 3 board escalate_to_ceo tests to assert the forwarded actor.
* [B-REL] release executor: idempotent half-landed retry + commit-scoped CI + decoupled workflow
Three confirmed gaps in the release fail-closed pipeline (#87/#318/#402):
#87 publish_failed retry duplicates changelog: execute() only short-circuits
on an existing tag. A publish_failed outcome (commit pushed + CI green, no
tag) left no tag, so a retry re-ran apply_version_bumps + write_changelog_entry
(re-inserting the entry above the already-present heading -> duplicate) and
commit_and_push (a second chore(release) commit). Add ReleaseOps
.release_commit_sha(version) detecting a prior release commit on the branch
(clone already at the target version); when present, skip the bump/changelog/
gate/commit pipeline and rejoin the shared CI -> publish tail on the existing
commit. No second commit, no duplicate entry.
#318 wait_for_ci polls branch-latest, not the release commit: a later push to
master during the ~40min wait made the latest run's head_sha != the release
sha forever, exhausting _CI_MAX_POLLS -> false ci_failed on a release whose
own CI was green. Thread head_sha through get_latest_ci_conclusion /
_fetch_latest_ci_run (GitHub actions/runs?head_sha=) so the gate polls the
release commit's own run; a concurrent push can no longer mask it.
#402 release CI gate reuses self_heal_ci_workflow: that setting documents an
empty-string mode for single-workflow repos which, inherited here, degraded
the fail-closed gate to the all-workflows mode git.py itself flags as
unreliable. Add release_ci_workflow (default ci.yml) and _resolve_release_
ci_workflow(); the release gate always resolves a NAMED workflow, never None.
Refactor: bundle the CI-fetch per-project inputs into a _CiRunQuery dataclass
so _fetch_latest_ci_run stays under the arg-count gate; unify the half-landed
path into execute's shared tail (drops a separate _publish_existing, one
return path). TDD red->green; ruff/mypy clean.
* [chore] logical-gaps: a2a service hierarchy gate (typed, unconditional) + persist skill on message row (3 gaps)
create_a2a_notification gated A2A hierarchy only when both ends resolved
(`if from_agent and target_agent:`), so an unattributed (from_agent falsy)
or unresolvable-target request slipped past the hierarchy matrix and
dispatched with from_agent='unknown' / to_agent='' — and a denial came back
as a bare ValueError indistinguishable from the missing-task_id ValueError.
Require both ends present, then validate via the shared typed
validate_a2a_access path (A2AAccessDeniedError + route_hint) so the legacy
notification surface enforces the same who-may-talk-to-whom invariant as the
conversation path.
send() accepts skill= and the gateway callers (qa/doc/pr_gate) pass it
expecting the receiver to learn which capability the message is about, but
send_chat_message never read it from options — silently dropped. Persist a
nullable skill column (migration 054) on a2a_messages, wire it through
send_chat_message + _msg_to_model + the A2AChatMessage model, and fix the
send() docstring (it claimed 'recorded in message metadata').
TDD: 4 red→green (skill recorded on message + surfaces in inbox; permission
denied raises typed A2AAccessDeniedError with route_hint; self-A2A raises
typed; missing from_agent raises instead of silent dispatch). 103 a2a
integration tests green; ruff/mypy clean; migration 054 verified
upgrade/downgrade on throwaway PG.
* [chore] logical-gaps: release-proposal already_published closes proposal + heartbeat-lock-loss cancels execute (2 gaps)
approve() closed the proposal only on status=='published'. A retry that finds
the tag already shipped returns 'already_published' (is_already_published),
so if a prior publish's route commit failed / HTTP 504'd, the proposal stayed
non-terminal forever — every retry returned already_published and never
closed it; only a manual cancel unstuck it. Close on both published and
already_published: the release shipped either way.
_heartbeat_loop returned silently when the lock was no longer owned (a >TTL
Redis outage let the mutex expire mid-execute), leaving executor.execute
running UNGUARDED — a concurrent approve (once Redis returns) could then
acquire the lock and _prepare_release_clone rm -rf the in-flight shared
release clone while the first execute was still mid-run_gate, re-opening the
very rm -rf-clone race the mutex+heartbeat exist to prevent. Run execute as a
task; on lock-loss the heartbeat sets a flag and cancels it, and approve()
turns the CancelledError into a structured 'lock_lost' result (an external
cancellation of approve itself still propagates — distinguished by the flag).
TDD: 2 red→green (already_published → COMPLETED not wedged; heartbeat lock-loss
→ lock_lost + execute cancelled, proposal not completed). 8 concurrency tests
green; ruff/mypy clean.
* [chore] logical-gaps: release approve async dispatch (202) — kill the 40min synchronous HTTP 504
The approve route ran the whole fail-closed execute inline: clone(600s) +
gate(1800s) + CI poll(2400s) + publish(300s) ≈ up to 85min worst case. nginx
(the single :3000 entry point, ~60s read timeout) 504'd long before it
finished, so the CEO's approve always appeared to fail even when the release
succeeded server-side — the structured ReleaseResult was unreachable over the
wire. dispatch_approve spawns the execute in a background task with a fresh
session (built from the request session's engine) and the route returns 202
'accepted' immediately; _INFLIGHT_APPROVES tracks the dispatched task for
observability (self-cleans via done-callback; the Redis mutex still refuses a
double-execute on a second click). The panel already polls GET /proposal every
30s, so it observes the final status (COMPLETED on published/already_published,
else the proposal stays open for retry); the card's approve toast now treats
'accepted' as an info 'dispatched, running in the background' instead of the
old 'Release halted' warning.
TDD: 2 route tests red→green (approve returns 202 'accepted' + the proposal
transitions to COMPLETED / stays PENDING once the background faked execute
completes; the dispatched task is awaited while the executor patch is live).
83 release tests green; ruff/mypy clean; panel typecheck+lint+format+test
green.
* [chore] mcp-servers: normalize exception bodies to Envelope + lift task_id/correlation_id on circuit_open (#232 #359 #57)
flow_server/do_server: the non-404 JSON path returned exception-handler bodies
raw (dict `error` from roboco/generic/http exception handlers, or a 422
`detail` list) — neither is the Envelope wire format the agent is prompted to
trust (string error kind + message + remediate + missing), so on any
service/validation failure the agent got no remediate and flailed until the
breaker tripped. _normalize_exception_envelope lifts the body into a real
Envelope (code -> counted string kind via _classify_dict_error_code, NOT_FOUND
-> not_found, message lifted, remediate synthesized, missing=[]; 422 -> incomplete_input with the validation detail preserved). The synthesized
envelope still flows through the breaker so a 500/422 storm trips it.
_record_and_check_circuit: the circuit_open substitution dropped task_id /
correlation_id from the top level (the SDK's envelope omits them); lift them
from the original rejection so the agent's envelope contract and ops audit-join
of the trip event still work, not just nested in inner.
intake_server._post_event: capture the relay response body under `detail` on
non-success so the grok intake agent gets the real reason (e.g. 'session not in
MegaTask scope' on a 422) instead of an opaque http_422 token with no
remediation.
TDD red->green; ruff + mypy clean; 157 mcp/SDK-breaker tests pass.
* [chore] a2a-routes: authenticate send_message responder + gate cancel task (PM-only) (#116 #423)
send_message took the responder identity from a client-supplied
metadata.from_agent, so any caller could spoof anyone (e.g.
from_agent='ceo') in the task's notes and in the spawn/notification
routed back to the original requester. Stamp the authenticated caller's
slug as the responder instead (CurrentAgentContext).
cancel_task was ungated: no auth dependency and no role check, so any
agent (or any caller) could cancel a task the lifecycle rule reserves to
PM roles (Any -> cancelled: PM roles only) — and the cascade-cancel of
all non-terminal descendants ran with a hardcoded cell_pm role and no
recorded actor. Add require_any_authenticated_agent + a PM-or-above gate,
and thread the authenticated role (into the cascade role gate) and slug
(into the cancellation note) into A2AService.cancel_task.
Tests: send_message ignores a spoofed from_agent and records the
authenticated slug; cancel rejects a developer (403) and a missing auth
header; a PM cancel threads role + slug into the service; the pre-existing
cancel success/already-terminal/not-found tests now run under a PM context
(the success test's body was missing the A2A 'name' field and false-passed
on a 422 — now genuine).
* [chore] work-session-routes: ownership check on mutating routes + stamp merge_pr merged_by from auth (#158 #271)
Every mutating work-session route keyed off session_id alone after the
role gate, so any developer could commit into / abandon / complete a
peer's active session (breaking the single-active-WorkSession invariant
and stranding that task) and any PM could merge any cell's PR — the REST
surface bypassed the verb layer's active-claimant gate entirely. Add a
shared _assert_ownership guard: dev ops require session.agent_id to be
the caller; PM merge_pr requires a cell PM to own the session's task cell
(main PM / CEO / board coordinate every cell), 404 for a missing session.
merge_pr took merged_by from the request body, so any PM could record a
PR merge under another agent's id, corrupting the merge audit trail the
completion/CEO-approval chain and metrics rely on. Drop the body param
and stamp the authenticated caller's agent_id as merged_by (the
MergePRRequest schema is gone with it).
Tests: a second dev's token hitting a peer's /commits and /abandon -> 403
(session left active); a foreign-cell PM -> 403, same-cell PM -> 200; a
spoofed body merged_by is ignored and the persisted row records the PM.
* [chore] ci-watch/dep-update dedupe: normalize git_url + treat empty-string workflow as default (#148 #1267)
The per-repo open-task dedupe filtered ProjectTable.git_url == git_url
(exact), while the orchestrator collapses its poll set by repo_key
(lower / strip trailing '/' / drop '.git'). Two projects whose git_url
differs only by those accidentals (a monorepo's cell-projects, or a
re-registered canonical project) defeated the one-open-task-per-repo
invariant and opened duplicate fix / dep-update tasks. Extract
roboco.utils.converters.repo_key as the single source and match the
dedupe query on its SQL mirror (regexp_replace(rtrim(lower(...)))).
The ci_watch (git_url, workflow) dedupe used func.coalesce(ci_watch_workflow,
default), but SQL COALESCE only substitutes for NULL — a project saved with
ci_watch_workflow='' (reachable via panel/API) yielded coalesce('', default)
= '' != default, so the DB diverged from the engine/orchestrator (which
collapse '' to the default via Python truthiness) and opened a duplicate
fix task every red cycle. Wrap with func.nullif(..., '') so an empty string
collapses to the default too.
Tests: a ''-workflow + NULL-workflow project on one repo dedupe to one task;
git_url accidentals (.git suffix / trailing slash) dedupe across both
ci_watch and dep_update. The orchestrator _repo_key now delegates to repo_key.
* [chore] admin_set_status: attribute the blocked-restore to the admin actor + emit override row (#2176)
admin_set_status taking a BLOCKED task to pending/in_progress with a
pre-block snapshot returned early via _apply_pre_block_restore, which
emitted its audit row with agent_role=None and audit_agent_id=restored_owner
(the pre-block dev) — the admin actor_id/actor_role were dropped entirely.
Because this branch runs with force=false (pending/in_progress aren't hatch
destinations), the distinguishing task.admin_override row (written only on
the non-restore path, gated by force) was never written, so an operator
could silently re-own a blocked task with no trace of who triggered it.
Thread actor_id/actor_role into _apply_pre_block_restore (admin_set_status
passes them with admin_override=True) so the transition audit row attributes
the re-owning to the admin, and emit a task.admin_override row (forced=False,
restore=True) on this branch independent of the force flag. The in-band
unblock(restore=True) path passes no actor and keeps the legacy attribution
(restored owner) with no override row.
Test: admin PATCH status=pending on a BLOCKED task with a snapshot attributes
every audit row to the admin (not the restored dev) and emits the override
row.
* [chore] converters: typed InvalidIdentifierError from require_uuid + log the orchestrator drop (#25)
require_uuid raised a bare ValueError('UUID value cannot be None'), so a
malformed/None identifier propagated as an opaque error callers either let
500 or broad-catch-and-silently-swallow — the orchestrator reaper call site
wrapped it in a bare except-Exception return with NO log, dropping a bad
task_id_str invisibly. Introduce InvalidIdentifierError(ValueError) and
raise it from require_uuid for both None and unparseable input; it stays a
ValueError subclass so existing except-ValueError / except-Exception callers
are unaffected, but typed so a caller can handle a bad identifier distinctly.
The reaper now catches the typed error, logs at warning, and no-ops — the
drop is visible instead of swallowed.
Tests: None and an unparseable string both raise InvalidIdentifierError; it
subclasses ValueError (back-comat).
* [sweep] notification_delivery: list_system_notifications over-fetch-then-slice for pending_ack_only
The SQL limit was applied before the post-fetch 'not fully acked' Python
filter. A window of newer fully-acked ack-required rows filled the limit
and masked older unacked notifications the operator still needs to act on
(the pending-ACK queue silently under-reported; a CEO-approval notification
could be hidden by newer already-acked noise). pending_ack_only now drops
the SQL limit, filters in Python, then slices to limit; the non-pending
branch keeps the SQL limit unchanged.
* [sweep] proactive: drop vestigial code-patterns surface from context package
Code indexing was removed, so _find_code_patterns always returned [] yet
build_context_package still called it, ContextPackage.code_patterns stayed
a live field, _build_summary advertised 'Found N code patterns', and
_count_items counted it — a permanently-empty slot the system claimed to
populate. The dead method, its call, the summary line, and the count
reference are removed. The code_patterns field itself is retained
(always-empty, serialized in to_dict and the optimal route response) for
API/schema back-compat, marked deprecated in its docstring.
* [sweep] migration 052: integration-test the task_cell_projects unique constraint
The UNIQUE(task_id, team) 'one project per cell per task' invariant was
only exercised through SimpleNamespace stubs that never touch a DB
session, so the real Postgres constraint was unverified. If it were
mis-declared or dropped, two same-team rows could coexist and
_resolve_subtask_project would non-deterministically return one, cutting
a subtask's branch/PR against the wrong repo. Adds an integration test
that inserts two same-(task_id, team) rows and asserts IntegrityError on
uq_task_cell_projects_task_team, plus a positive different-teams case.
* [sweep] pr_gate: classify MegaTask root-subtask as root so its root->master PR gets COMMENT (#608)
_post_gate_review_to_pr identified a root->master PR by absence of a
parent_task_id. A MegaTask root-subtask opens its own root->master PR into
the project's master (submit_root, parent='master') but carries
parent_task_id=umbrella, so is_root was False and the gate posted APPROVE
(pr_pass) / REQUEST_CHANGES (pr_fail) instead of COMMENT. The APPROVE could
satisfy a single-approval master branch-protection rule and let a non-CEO
merge via the GitHub UI before the CEO, against the documented invariant
that only the CEO acts on master. is_root now also covers
is_batch_root_subtask (batch_id set + parented); a non-batch cell-PM
coordination root keeps batch_id=None so it stays a cell->root PR
(APPROVE/REQUEST_CHANGES). Extends the _task test helper with a batch_id
kwarg.
* [sweep] enforcement: complete the status-class partition + coverage invariant (#247)
is_waiting_state already covered awaiting_pr_review (the primary fix), but
the doc's coverage invariant was missing: backlog and pending fell through
ALL three predicates (terminal/active/waiting), so a future enum addition
could silently land in no category. is_waiting_state now also covers
pending (waiting for a claim) and backlog (waiting on PM activation), so
is_terminal_state / is_active_state / is_waiting_state partition the whole
Status enum. Adds test_status_classification_covers_every_enum_member
asserting every Status member is classified by exactly one predicate, so
an enum addition that drifts the partition fails the build.
* [chore] test-suite: unblock the quality gate (mypy + 2 behavior fixes)
12 mypy errors across 5 test files: drop banned type:ignore comments
(lifecycle_spec monkeypatch uses cast(Any, ...); the ignores were unused),
wrap SQLAlchemy-typed ids with cast(UUID, ...) for AgentContext / WorkSession
args (AgentTable.id is Mapped[sqla UUID], not uuid.UUID), annotate **kw: Any,
and cast(Any, svc) for a method-assignment mock.
test_cancel_descendants_cascades_for_authorized_pm: the child was parked in
awaiting_ceo_approval, which the spec gates to CEO-only cancel
(lifecycle.py:378-389) — a cell_pm cascade correctly refuses it (the #103
refuse path). Use a PM-cancelable in_progress child so the positive-cascade
assertion holds; the refuse case is already covered by its sibling test.
test_a2a_message_auth: /message/send now resolves the authenticated
responder slug via get_agent_context (a DB lookup, #116). This is a DB-free
unit test of the token gate + route body, so stub get_agent_context in the
fixture — the gate (require_any_authenticated_agent) still runs real and
401s on a missing/forged token before that dependency resolves.
* [chore] complexity: split 5 C-rank blocks to <=B for the xenon gate
No behavior change; each C-rank function factored into a helper so the
complexity gate (xenon --max-absolute B) holds.
- lifecycle.can_invoke_action: extract the team-match check into
_check_team_match.
- a2a.cancel_task: extract _status_value_of + _apply_cancel_note.
- task._apply_pre_block_restore: extract _restore_block_ownership (status/
owner restore + snapshot clear) and _emit_admin_override_audit (#2176).
- release_proposal.approve: extract _finalize_release_lock (heartbeat/
execute cancel + mutex release) out of the finally.
- kanban.get_main_pm_board_flat: dict-dispatch the column routing instead
of a 7-branch if/elif ladder (status wins over team; in-flight + no cell
team falls through to Coordination, #196).
* [chore] lifecycle artifacts: regenerate to match the spec (foundation-check)
The rendered artifacts (docs/rag/lifecycle, panel/lib/lifecycle.json, the
_generated role-prompt fragments) had drifted from the spec — the prior
sweep commits (cancel-CEO gate, claim_pr_review preconditions, pr_reviewer
unclaim, complete merge-first ordering) changed spec data without
regenerating, and the foundation-check render+diff stage never ran because
mypy failed earlier in the gate. make foundation-check now passes.
* [fix] chat: wire live message delivery end-to-end (MESSAGE_SENT)
send_message persisted messages but never broadcast them, there was no
MESSAGE_SENT event type or bridge forwarder, and the panel session view
had no websocket subscription — the live chat path was dead end-to-end.
- add EventType.MESSAGE_SENT and publish it best-effort on every persisted
send (a bus outage logs, never rolls back the durable row)
- bridge _handle_message_event forwards to /ws/sessions/{id} and
/ws/channels/{id}; subscribe it in register_websocket_bridge_handlers
- panel useSessionStream subscribes the session view; the page invalidates
the transcript + session-detail queries on each message.new so the held
(staleTime Infinity) views refresh live without the manual Refresh
* [fix] chat: return session task_links in one read; drop panel N+1
GET /sessions/{id} ran a bare select and session_to_response omitted
task_links, so it always returned them empty — the panel worked around it
with a triple-fetch (get session, get-tasks-for-session which re-fetched
the same endpoint, then a task GET per link), and the links never showed.
- add get_session_with_links(_or_raise) that eager-loads task_links -> task
- add session_to_response_with_links; GET /sessions/{id} uses both
- panel useSession now relies on the single populated response; remove the
dead getTasksForSession + per-task fetch and the unused tasksApi import
* [fix] chat: validate reply_to against the effective session; guard closed-session composer
Posting to a closed session transparently redirects the message to the
group's active session (intended for agents holding stale refs), but
reply_to was validated against the requested session, not the one the
message lands in — letting a cross-session reply slip through — and the
panel silently posted there too, so the message vanished from the view.
- validate reply_to against session.id (the effective, possibly-redirected
session), not req.session_id
- panel: render a "session is closed" notice instead of the composer for a
non-active session; if a send still lands elsewhere (stale status), toast
that it went to the active session rather than letting it appear to vanish
* [fix] chat: close session/group/message read IDOR; fix doubled 404s
get_session and the messages-list took an agent id but never used it, and
get_group took none at all — any authenticated agent could read any private
channel's group, session, and message transcripts. Three NotFoundError sites
also passed a full sentence as resource_type, yielding "... not found not found".
- add require_group_read_access / require_session_read_access (channel
member / silent observer / privileged, mirroring list_group_sessions_for_agent)
and get_session_with_links_for_agent; enforce on GET /sessions/{id},
GET /sessions/{id}/tasks, GET /messages, GET /groups/{id} (-> 403 on deny)
- fix the three doubled-404 sites to the NotFoundError(resource_type, resource_id) form
Also folds two gate fixes for the prior chat commits: cast session.id to UUID
for the reply_to validation, and ruff import/format touch-ups.
Note: POST /messages intentionally still skips the channel write-ACL on the
HTTP (human-CEO/panel) path — the CEO is not in writers for 8/11 channels, so
enforcing it there would block the panel; the gateway/agent path enforces it.
* [fix] secretary: harden live chat — stuck spinner, mid-reply clobber, reload
The Secretary live chat had three live-behaviour bugs: a dropped SSE
connection left a permanent "thinking…" spinner (openStream set no
transport-error handler, so the no-data error Event was swallowed by the
JSON-parse guard and streaming never reset); sending mid-reply wiped the
accumulation buffer and pushed a user message without guarding the in-flight
turn, abandoning/duplicating the reply; and the chat lived only in React
state, so a reload wiped it.
- route the dual-purpose `error` listener: server-sent JSON → handleEvent,
transport error (no data) → reset streaming, surface a notice, close stream
- guard send while streaming (streamingRef); disable the composer Send/Enter
while a reply is in flight
- persist sessionId + messages to localStorage (TTL'd) and, on mount, restore
+ re-attach the stream once the backend confirms the session is still alive
(mirrors the intake/prompter durability)
* [chore] groups: extract group-read helper to keep module rank A
The get_group IDOR access-check added try/except branches that tipped the
module to xenon rank B. Extract the service-error→HTTP mapping into a small
helper so get_group stays lean and the module is rank A again (behaviour
unchanged; covered by the groups route tests).
* [fix] chat: correct panel session-task mutation endpoints
linkTask/unlinkTask posted to /add-task and /remove-task (with a body), but
the backend exposes POST /sessions/{id}/tasks and DELETE
/sessions/{id}/tasks/{task_id} (path param) — so every call 404'd. updateTaskLink
targeted /update-task, a route that does not exist at all. Point linkTask and
unlinkTask at the real routes and drop the phantom updateTaskLink. All three were
unused, so no behaviour changes today — this removes a latent 404 trap.
* [docs] chat: document live message delivery (MESSAGE_SENT / message.new)
Document the live transcript-update path the chat-subsystem fixes wired:
- docs/api/websockets.md: add the message.new event-types row (carried on
/ws/sessions + /ws/channels from EventType.MESSAGE_SENT) and note the
forwarder sets type:"message.new"
- docs/panel/communications-and-journals.md: the session transcript updates
live; a closed session is read-only (composer disabled)
- CLAUDE.md: name message.new on the per-resource streams and make
MESSAGE_SENT the worked example of the add-a-live-event recipe
The internal roboco_map slices (gitignored) were updated in place to match.
* [docs] reconcile published docs with code since v0.13.0
Drift caught by the doc-reconciliation pass (all verified against HEAD):
- CLAUDE.md + rag: pr_reviewer gained the unclaim verb (16b71be8)
- rag permissions/task-states/task-tools: awaiting_ceo_approval -> cancelled is
CEO-only, not PM+CEO (16b71be8 cancel-ceo-gate; lifecycle.py:373-382)
- deploy/env-reference: ROBOCO_APP_VERSION default 0.9.0 -> 0.14.0 (config.py:31);
add ROBOCO_RELEASE_CI_WORKFLOW row (2759edf7, config.py:454)
- deploy/data-and-migrations: 44->54 revisions, head 054_a2a_message_skill
- optional/autonomous-maintenance: CI-watch dedupe is per (repo, workflow) (d34bc1a7)
---------
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
47 KiB
CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
Licensing
RoboCo is licensed under AGPL-3.0 (see LICENSE). Copyright (c) 2026 Renzo Franceschini. Do NOT reintroduce an MIT or other license reference anywhere (README, headers, package metadata) — the project is AGPL.
Contributions require a signed Contributor License Agreement (CLA.md), automated via the CLA Assistant workflow (.github/workflows/cla.yml). The CLA preserves the option to dual-license / offer a commercial edition later; keep copyright assignment language intact. See CONTRIBUTING.md.
Project Overview
RoboCo is an AI Agentic Company - a virtual organization of 25 AI agents + 1 human CEO, designed to operate as a complete software development workforce. The system implements a structured organizational hierarchy with formal communication protocols, task management, and quality controls.
Core Architecture
CEO (Renzo - Human)
|
+-- Intake (on-demand interviewer: chats only with the CEO to draft a task)
+-- Secretary (on-demand chief-of-staff: reads company state, runs gated CEO directives)
+-- PR Reviewer (read-only: the main reviewer — inbound external/fork + internal PRs, and the root→master in-path gate)
|
+-- Board (3 agents)
+-- Product Owner
+-- Head of Marketing
+-- Auditor (silent observer, reports to CEO)
|
+-- Main PM (coordinates all cells)
|
+-- Backend Cell (6 agents: 2 Devs, 1 QA, 1 PM, 1 Documenter, 1 PR Reviewer)
+-- Frontend Cell (6 agents: 2 Devs, 1 QA, 1 PM, 1 Documenter, 1 PR Reviewer)
+-- UX/UI Cell (6 agents: 2 Devs, 1 QA, 1 PM, 1 Documenter, 1 PR Reviewer)
Hardware Infrastructure
- Olares One (Powerhouse): Intel Ultra 9 + RTX 5090, runs Claude Code instances and AI inference - NOT YET ARRIVED
- UGREEN NAS (Warehouse): 36TB RAID6, 128GB RAM, hosts PostgreSQL, Redis
- Pi Cluster (Operations): Monitoring, notifications, smart home
Development Standards
Python (Backend)
# Package manager
uv
# Before any commit
uv run ruff format .
uv run ruff check .
uv run mypy roboco/
uv run pytest
# Coverage target: 80%
TypeScript (Frontend)
# Package manager
pnpm
# Before any commit
pnpm format
pnpm lint
pnpm typecheck
pnpm test
# Coverage target: 80%
Technology Stack
| Layer | Technology |
|---|---|
| API Framework | FastAPI |
| Database | PostgreSQL + asyncpg |
| Vector Store | PostgreSQL + pgvector (in-house engine) |
| RAG Engine | in-house (asyncpg + pgvector, hybrid retrieval) |
| Cache/Queue | Redis |
| Container Runtime | Docker + Docker Compose |
| Cloud LLM | Claude API (claude-opus-4-6) + xAI Grok (official grok CLI, SuperGrok subscription) |
| Local LLM | Ollama (glm-5.2:cloud for RAG/hybrid retrieval) |
| Embeddings | qwen3-embedding:0.6b (1024 dim) |
| Frontend | Next.js 16 + TypeScript + Tailwind + Radix UI (in panel/) |
| Edge / Proxy | nginx (single entry point on port 3000) |
Multi-Agent Workspace Structure
Each agent gets their own git clone of a project, enabling parallel development without conflicts:
{ROBOCO_WORKSPACES_ROOT}/ # Default: /data/workspaces
+-- {project-slug}/
+-- {team}/
+-- {agent-slug}/
+-- [git repository]
Example:
/data/workspaces/
+-- roboco/
+-- backend/
| +-- be-dev-1/ # be-dev-1's workspace
| +-- be-dev-2/ # be-dev-2's workspace
+-- frontend/
+-- fe-dev-1/
+-- fe-dev-2/
Note: the Next.js control panel now lives at roboco/panel/ inside this repo (no longer a separate roboco-panel project or workspace).
Key Configuration (roboco/config.py):
ROBOCO_WORKSPACES_ROOT: Root directory for workspaces (default:/data/workspaces)ROBOCO_WORKSPACE_AUTO_CLONE: Auto-clone repos on first access (default:true)ROBOCO_WORKSPACE_CLONE_TIMEOUT: Clone timeout in seconds (default:300)
On a Python workspace, WorkspaceService runs uv sync --extra dev (not plain uv sync) so the clone's .venv carries the full gate toolchain (ruff/mypy/xenon/pytest) — the lint/type/complexity tools live in the dev extra, which plain uv sync skips. Without it an agent's make quality fails on ruff: command not found and the agent can't gate its own work.
Because the clone is shared across a dev's tasks, a fresh claim git-resets the workspace to a clean tree (git reset --hard) before checking out the new task's branch — discarding abandoned uncommitted cruft from a finished task while preserving all commits and the gitignored .venv. A resume short-circuits before this, so committed work is never reset.
Git Workflow
Branch Naming Convention
Branch names follow the pattern: {type}/{team}/{task-hierarchy}
Types: feature, bug, chore, docs, hotfix
Task Hierarchy: Uses -- separator (not /) to avoid git ref conflicts.
Examples:
- Root task:
feature/backend/ABC12345 - Subtask:
feature/backend/ABC12345--DEF67890 - Sub-subtask:
feature/backend/ABC12345--DEF67890--GHI11111
Commit Format
Commits are automatically prefixed with the task ID:
[{task-id[:8]}] {message}
Example:
[ABC12345] Add user authentication endpoint
Work Sessions
When a developer claims a task, a WorkSession is created that tracks:
- Branch name and base/target branches
- All commits made during the session
- Files modified
- PR number/URL when created
- Merge status and who merged
A task has at most one active WorkSession: re-claiming a task (pool release, reaper unclaim, escalation redirect) supersedes any prior agent's stale active session, enforced both at the service layer and by a DB partial-unique index (migration 047). Without it, duplicate active sessions made the one-row active lookup raise and crashed the claim/plan flow into a respawn loop.
A developer's clone is shared across all their tasks, so push and PR-head operate on the task's recorded branch by name, independent of the clone's current checkout — fixing the BRANCH_MISMATCH / "No commits between" failures when the clone was parked on a later task's branch. A missing local task-branch ref is first recovered from origin/<branch> before the push-by-name.
Git Credentials
Git authentication is managed per-project through encrypted GitHub PATs:
- Each project stores its own git token - no global fallback
- Tokens are encrypted at rest using Fernet symmetric encryption
- API never exposes tokens - only returns
has_git_token: boolean - Self-service via UI - users set/update tokens in project settings
Project fields:
| Field | Description |
|---|---|
git_token_encrypted |
Fernet-encrypted GitHub PAT (DB column) |
has_git_token |
Boolean indicator for API responses |
Token flow:
- User creates project in UI, enters GitHub PAT
- Token encrypted and stored in
projects.git_token_encrypted - WorkspaceService decrypts token when cloning repos
- GitService decrypts token for PR operations (gh CLI)
HTTPS URLs require tokens - attempting to clone without a token will raise WorkspaceError.
Task Lifecycle
Task States
The complete task lifecycle is defined in roboco/foundation/policy/lifecycle.py (roboco/enforcement/task_lifecycle.py is a backwards-compat shim over it):
backlog -> pending -> claimed -> in_progress -> [blocked|paused] -> verifying
| |
v v
awaiting_qa <------------------+ awaiting_documentation
| (needs_revision) | |
v | v
awaiting_documentation --------+ awaiting_pm_review
| |
v v
awaiting_pm_review awaiting_ceo_approval
| |
v v
completed completed
In-path PR-review gate (awaiting_pr_review): each assembled PR is reviewed before the PM merges. The cell PM's submit_up opens the cell→root PR and the Main PM's submit_root opens the root→master PR; both enter awaiting_pr_review, where a reviewer pr_passes it on to awaiting_pm_review or pr_fails it back to needs_revision — the merge-level reject the PM otherwise lacks. Leaf dev tasks and branchless coordination roots skip the gate.
States:
| State | Description |
|---|---|
backlog |
PM setup phase - dependencies or session setup needed |
pending |
Ready for work - orchestrator can spawn agents |
claimed |
Agent has locked the task |
in_progress |
Active development |
blocked |
External dependency blocking progress |
paused |
Temporarily stopped (can resume) |
verifying |
Self-verification by developer |
needs_revision |
QA or CEO requested changes |
awaiting_qa |
Submitted for QA review — PR must already exist |
awaiting_documentation |
Documentation phase — PR already open from pre-QA; doc writes docs |
awaiting_pr_review |
In-path PR-review gate: a reviewer checks the assembled cell→root / root→master PR before the PM merges (assembled, PR-bearing tasks only) |
awaiting_pm_review |
Docs complete, PM reviews + merges |
awaiting_ceo_approval |
Major tasks escalated for CEO final approval |
completed |
Terminal state - work done and merged |
cancelled |
Terminal state - work cancelled |
Role-Based Transitions
All status transitions are validated through the enforcement layer. Key restrictions:
| Transition | Allowed Roles |
|---|---|
backlog → pending (activate) |
PM roles only |
pending → claimed (claim) |
Role must match task type (QA for awaiting_qa, etc.) |
claimed → pending (unclaim) |
Assignee or PM |
awaiting_qa → awaiting_documentation (pass) |
QA only |
awaiting_qa → needs_revision (fail) |
QA only |
awaiting_documentation → awaiting_pm_review |
Documenter or Developer (parallel completion) |
in_progress → awaiting_pr_review (submit_up / submit_root) |
PM roles (opens the assembled cell→root / root→master PR) |
awaiting_pr_review → awaiting_pm_review (pr_pass) |
PR reviewer only |
awaiting_pr_review → needs_revision (pr_fail) |
PR reviewer only |
awaiting_pm_review → completed |
PM roles only |
awaiting_pm_review → awaiting_ceo_approval |
PM roles only |
awaiting_ceo_approval → completed/needs_revision/cancelled |
CEO only |
Any → cancelled |
PM roles only |
Unclaim Operation: Agents can release claimed tasks back to the pool using unclaim(). This transitions claimed → pending and optionally reassigns to another agent.
Board never owns a coordination root: a Board role (Product Owner / Head of Marketing) is never assigned a Main-PM coordination root (delivery root or MegaTask root-subtask) via escalation or reassignment — Board roles have no unblock verb, so such a hand-off would deadlock. The transition is diverted to the pool for a role-matched Main-PM reclaim.
Git Integration Requirements
All tasks follow git workflow. PR is created BEFORE QA review (not after) so QA can review the real PR diff on GitHub and downstream PM/CEO approval chain off a PR that already exists:
- claimed -> in_progress:
branch_nameis auto-set on claim (hierarchical branches) - verifying -> awaiting_qa (submit-qa): Requires
self_verified,commits,pr_number(PR open), and at least oneprogress_updatesentry - awaiting_qa -> awaiting_documentation (pass-qa): Requires
pr_numberand substantive QA notes - awaiting_documentation -> awaiting_pm_review: Requires
docs_complete=True(PR already exists from step 2 above) - awaiting_pm_review -> awaiting_ceo_approval: Must have
pr_numberset and all subtasks in a terminal state
CEO Approval Workflow
Major tasks are escalated to CEO for final approval:
- PM reviews and approves, escalates to
awaiting_ceo_approval - CEO can:
- Approve: Merges PR, task ->
completed - Request changes: Task ->
needs_revision - Cancel: Task ->
cancelled
- Approve: Merges PR, task ->
Data Models
Core Models (roboco/models/)
| Model | Purpose |
|---|---|
Task |
Atomic unit of work with acceptance criteria |
Project |
Git repository configuration and CI/CD commands |
WorkSession |
Links agent work to task, tracks branch/commits/PR |
Agent |
AI agent with role, team, capabilities |
Session |
Communication session with messages |
Channel |
Team communication channel |
Message |
Extracted message from agent streams |
Notification |
Formal notification requiring acknowledgment |
Journal |
Agent personal log for reflections/learnings |
Task Model Key Fields
# Git configuration (all tasks follow git workflow)
task_type: TaskType # code, documentation, research, planning, design, administrative
project_id: UUID # Project this task works on (required)
branch_name: str # Branch for this task (auto-created on claim)
work_session_id: UUID # Active work session
# PR tracking (parallel execution in awaiting_documentation)
pr_number: int # GitHub/GitLab PR number
pr_url: str # Full URL to PR
docs_complete: bool # Documenter has finished
pr_created: bool # Developer has created PR
# Commits linked to task
commits: list[CommitRef] # All commits made for this task
Communication Model
Communication = constant stream (always flowing, logged, observed) Notifications = formal signals (require acknowledgment, sent by PMs/Board only)
Channel Structure
- Cell channels:
#backend-cell,#frontend-cell,#uxui-cell - Cross-cell:
#dev-all,#qa-all,#pm-all,#doc-all - Management:
#main-pm-board,#board-private - Special:
#announcements(read-only except Board/Main PM),#all-hands
The Auditor has silent read access to ALL channels.
Agent learnings (note scope='learning') broadcast as knowledge-share notifications only to other agents — the human / human-driven roles (CEO, prompter, secretary) are excluded, since agent knowledge-sharing is noise in a human's inbox.
Key Principles
- Everything is a task - All work is tracked and documented
- No work without a task - Create task record first
- No task without acceptance criteria - How do we know it's done?
- No closure without documentation - Future agents need context
- Communication is constant - Stream reasoning, log everything
- State is sacred - If interrupted, state must be recoverable
- The Auditor sees all - Quality monitored silently
- Commits linked to tasks - Every commit references its task ID
- CEO approves major changes - Escalation path for important work
Agent Gateway
Agents do not call the API or per-domain MCP tools directly. They go through two thin MCP servers (roboco-flow, roboco-do) backed by the server-side Choreographer in roboco/services/gateway/. The Choreographer composes the existing services (TaskService, JournalService, GitService, etc.) into intent-verb sequences. Tracing, claim-locking, evidence assembly, and remediation hints are all centralized there.
Each agent gets a spawn manifest at /app/tool-manifest.json listing the verbs its role is allowed to call. The orchestrator builds the manifest from roboco/services/gateway/role_config.py and mounts it read-only into the agent container.
Verb surface (canonical source: lifecycle.intents_for_role; every role also gets i_am_idle)
| Role | Flow verbs (beyond i_am_idle) |
|---|---|
| developer | give_me_work, i_will_work_on, open_pr, i_am_done, i_am_blocked, resume, sync_branch, unclaim |
| qa | give_me_work, claim_review, pass_review, fail_review, i_am_blocked, resume, unclaim |
| documenter | give_me_work, claim_doc_task, i_documented, i_am_blocked, resume, unclaim |
| cell_pm | give_me_work, i_will_plan, delegate, complete, submit_up, triage, unblock, escalate_up, reassign, resume, unclaim |
| main_pm | give_me_work, i_will_plan, delegate, complete, submit_root, triage, triage_all, unblock, escalate_up, escalate_to_ceo, resume, unclaim |
| pr_reviewer | give_me_work, claim_pr_review, post_pr_review (inbound external/fork PRs), claim_gate_review, pr_pass, pr_fail (in-path assembled-PR gate), unclaim |
| product_owner | triage, escalate_to_ceo |
| head_marketing | triage, escalate_to_ceo |
| auditor | triage (read-only — no say/dm) |
| prompter | (none beyond i_am_idle — not a delivery-lifecycle role; intake interviewer, human-only) |
| secretary | (none beyond i_am_idle — human-only chief-of-staff; reads company state + runs gated CEO directives) |
Content tools (do_server) — most roles: commit, note, say, dm, evidence. Delivery roles (developer / qa / documenter / cell_pm / main_pm) also get draft_playbook (draft a curated playbook for the KB). Auditor is restricted to note (scope=reflect) + evidence, plus the playbook-curation verbs approve_playbook / reject_playbook / archive_playbook (a bounded, deliberate expansion — KB curation, not agent comms, so its no-say/no-dm restriction holds). The pr_reviewer posts its change-request on the PR itself (no agent comms). The prompter (intake) and secretary are restricted to note + evidence — human-only, no say/dm/notify. The note/journal write returns as soon as the entry is persisted; RAG indexing (Ollama embedding) runs fire-and-forget, so the tool no longer times out under concurrent load.
MCP servers running per agent container
| Server | Purpose |
|---|---|
roboco-flow |
Intent verbs (give_me_work, i_am_done, claim_review, complete, ...) |
roboco-do |
Content tools (commit, note, say, dm, evidence) |
roboco-git-readonly |
Read-only git: status, log, diff, branches |
roboco-optimal |
RAG: roboco_ask_mentor, roboco_kb_search |
roboco-docs |
Project docs file management (selected roles) |
Every verb returns a standardized Envelope:
- ok:
{status, task_id, next, evidence?, context_briefing} - error:
{error, message, remediate, missing}
The next field tells the agent what to call next; the remediate field on errors tells them exactly how to fix and retry. Agents should not guess state — trust the response. The verb runner re-checks the task after each composed atomic action and, on a concurrent mid-verb state change, fails fast with a clean INVALID_STATE (re-fetch + re-issue) rather than crashing on a None dereference.
Agent Providers
Agent backends are pluggable. roboco/llm/providers/ defines an AgentProvider lifecycle ABC (base.py) and a ProviderRegistry keyed by ModelProvider (registry.py), with ClaudeCodeProvider (default) and GrokCliProvider. The orchestrator resolves a provider at spawn from the agent's ModelProvider; when no dedicated provider is registered it falls back to the built-in Claude Code spawn. ModelProvider (roboco/models/base.py) is ANTHROPIC (default), GROK, LOCAL, OLLAMA_CLOUD, OPENAI (reserved). The seam is additive: only GROK routes through GrokCliProvider; Anthropic / Ollama Cloud / self-hosted spawns are unchanged, and every provider gets the same MCP gateway + tool-manifest wiring by construction.
Grok runtime. GROK agents run xAI's official grok CLI (model grok-build) authenticated by a SuperGrok subscription, not a metered API key — so a Grok workforce can't stall mid-task on out-of-credits. The host ~/.grok/auth.json is mounted read-only into each agent (GrokCliProvider._append_grok_auth_mount; ROBOCO_HOST_GROK_DIR is the host mount source, set up once with grok login). It reaches parity with the Claude path by construction: same MCP gateway + manifest, per-role tool-removal and git-operation deny rules, a prompt-injection guard on the task prompt, headless tool auto-approval, and per-agent token/cost capture from the grok session store. It covers both one-shot delivery roles and the interactive Intake (Prompter) and Secretary chats (per-turn grok -p with session resume).
Token auto-refresh. The grok access token has a fixed ~6h server-set TTL and the CLI cannot refresh it headlessly — on an expired token it hangs forever at an interactive login prompt. The orchestrator mints a fresh token from the offline-access refresh token (xAI's OIDC refresh_token grant) before expiry and rewrites the shared auth.json in place (roboco/llm/providers/grok_auth.py refresh_if_stale, run once per dispatch tick; the orchestrator's ~/.grok mount is read-write so it can rewrite it). As a backstop the agent entrypoint runs python -m roboco.llm.providers.grok_auth --check and refuses to start (exit 78) on a missing/expired token instead of hanging.
Self-Healing & Feature Flags
Self-healing CI loop (default-off). RoboCo can watch its own repository's CI (a single named workflow) and, on a detected regression, open a fix task that is held out of dispatch until the CEO approves it (it terminates at awaiting_ceo_approval), then dispatch it through the normal delivery flow. It is dormant by default and armed by ROBOCO_SELF_HEAL_ENABLED plus a second opt-in ROBOCO_SELF_HEAL_ORIGINATE_ENABLED; origination is bounded by ROBOCO_SELF_HEAL_MAX_OPEN_TASKS / _MAX_PER_CYCLE so it can't flood the backlog. It never auto-merges or self-deploys (roboco/services/self_heal_engine.py).
Multi-repo CI-watch (default-off). The fan-out generalization of self-heal: instead of RoboCo's single own repo, it watches every project the operator opts into (projects.ci_watch_enabled, migration 048) and, on a red CI conclusion on that project's default branch, opens one fix task into that project's lifecycle that rides the normal delivery flow (+ PR-review gate) and never auto-merges. It reuses the exact hardened per-project GitService.get_latest_ci_conclusion (a missing signal is "unknown", never a false green; per-project errors are isolated and never abort the sweep), and is bounded + deduped per repo by git_url (a monorepo's cell-projects share one fix task) with per-cycle / rolling caps. Armed by ROBOCO_CI_WATCH_ENABLED (+ _INTERVAL_SECONDS / _MAX_OPEN_TASKS / _MAX_PER_CYCLE / _DEFAULT_WORKFLOW) and per-project ci_watch_enabled / ci_watch_workflow; MultiProjectCITelemetrySource (roboco/services/telemetry/source.py) + CiWatchEngine (roboco/services/ci_watch_engine.py) + a dedicated orchestrator _ci_watch_loop. The single-repo self-heal loop is untouched.
Dependency-update bot (default-off). A per-project engine mirroring the self-heal/CI-watch shape: weekly (default) it probes whether a dependency upgrade would change a project's lockfiles and, if so, opens one "update dependencies" task that rides the normal delivery flow (+ PR-review gate) and never auto-merges. Detection is read-only — WorkspaceService.dry_upgrade_changes_lockfile runs the project's dep_update_command (e.g. uv lock --upgrade) in a throwaway clone of the read clone and diffs the lockfile paths (dep_update_paths, or inferred uv.lock/pnpm-lock.yaml); the read clone is never mutated, nothing is committed/pushed, and a null/failing command originates nothing (fail-safe). A project participates only when projects.dep_update_command is set (migration 049); bounded + deduped per git_url with per-cycle/rolling caps. Armed by ROBOCO_DEP_UPDATE_ENABLED (+ _INTERVAL_SECONDS default 604800 / _MAX_OPEN_TASKS / _MAX_PER_CYCLE); DepUpdateEngine (roboco/services/dep_update_engine.py) + a dedicated _dep_update_loop.
Gated release manager (default-off). The autonomy that automates cutting a release up to the decision. A default-off background loop (ReleaseManagerEngine + _release_manager_loop) runs the deterministic readiness sweep (ReleaseReadinessService.assess, roboco/services/release_readiness.py) — diff-since-tag → conventional-commit classification → semver bump → version-reference completeness (the missed-ref guard) → CHANGELOG completeness → docs-drift (agent count) → migration single-head → gate state — and, past a threshold (ROBOCO_RELEASE_MIN_COMMITS, or any feat/security) with a green gate, originates ONE release proposal held for the CEO. The proposal is a source='release_manager' task owned by the Secretary, HELD (confirmed_by_human=False) and skipped by every dispatcher — acted on only by the CEO-gated routes, never delivered. The CEO approves or rejects-with-changes in the panel (release-proposal-card.tsx; GET/POST /api/release/proposal{,/approve,/reject}, CEO-only); approval runs the fail-closed ReleaseExecutor (roboco/services/release_executor.py): write the bumps across the canonical set (derived from the previous chore(release): commit) + the CHANGELOG entry, run make quality (abort before commit on red), commit chore(release): X.Y.Z (signed) + push, wait for green release-commit CI (abort before publish on red), then gh release create vX.Y.Z. Idempotent (an already-published version is a no-op) and never publishes without the CEO. Correctness is deterministic code, not agent judgment; the only generative step is the CHANGELOG prose, which the CEO reviews. Armed by ROBOCO_RELEASE_MANAGER_ENABLED (+ ROBOCO_RELEASE_MIN_COMMITS / _INTERVAL_SECONDS). Auto-deploy stays out of scope — publishing builds images; deploying to the NAS is the CEO's manual step.
Organizational memory loop (default-off). Closes the learn→reuse loop so agents stop cold-respawning blind. Three parts, all gated by ROBOCO_ORG_MEMORY_ENABLED: ① capture — at task completion TaskService._completion_learnings_for distills ONE high-signal lesson (Problem→Approach→Gotcha, ≤120 words) via the local model (MemoryDistiller, roboco/services/memory_distiller.py) instead of the noisy raw-notes/duration capture (flag-off keeps the legacy capture); journal indexing excludes is_private reflections from the shared corpus. ② retrieve (keystone) — on claim, _briefing_for injects context_briefing["institutional_memory"]: top-K (ROBOCO_ORG_MEMORY_TOP_K) relevance-floored (ROBOCO_ORG_MEMORY_MIN_SCORE) lessons + approved playbooks from a role-shaped query (EvidenceRepo.similar_memory over the LEARNINGS + PLAYBOOKS pgvector indexes); below the floor nothing is injected (no briefing bloat). ③ playbooks — a first-class curated procedure store: PlaybookTable (migration 050), the PLAYBOOKS OptimalService index, the draft_playbook content verb (delivery roles), Auditor approve_playbook/reject_playbook/archive_playbook curation (approval indexes it), and the panel review queue (playbook-review-queue.tsx; /api/playbooks Auditor/CEO routes). Distillation runs on the local model only — never a cloud LLM in the hot path; every step is best-effort (a failure never blocks completion or the briefing).
Feature flags / company-in-a-box. Env-gated, default-off subsystems toggle from the panel's Settings → Feature Flags card (panel/src/components/settings/feature-flags-card.tsx) instead of hand-editing env: web research (ROBOCO_RESEARCH_ENABLED), the strategy engine (ROBOCO_STRATEGY_ENGINE_ENABLED), pitch provisioning (ROBOCO_PROVISIONING_*), external / internal PR review, the agent-runtime toolchain match (ROBOCO_TOOLCHAIN_MATCH_ENABLED), the architectural-conventions standard (ROBOCO_CONVENTIONS_ENABLED), gateway-health recovery (ROBOCO_GATEWAY_HEALTH_ENABLED), multi-repo CI-watch (ROBOCO_CI_WATCH_ENABLED), the dependency-update bot (ROBOCO_DEP_UPDATE_ENABLED), the gated release manager (ROBOCO_RELEASE_MANAGER_ENABLED), the organizational memory loop (ROBOCO_ORG_MEMORY_ENABLED), and the self-heal flags above. A toggle persists in the settings store and takes effect on the next backend restart; an unset flag falls back to its environment / config default.
Architectural Conventions Standard
Per-project architectural standard (default-off). Beyond the make-style gates (which check syntax/types/tests, not where code lives), each project can carry a repo-canonical .roboco/conventions.yml — an architecture map (which definition kinds belong in which modules), a toggleable rule set, custom regex rules, and waivers — so an agent cannot land a Pydantic model defined inside a router or a # noqa / # type: ignore. Placement of a helper (any top-level function) only warns — too blunt to hard-block; thin_routes doesn't count an explicit db.commit(); and a small allowlist of unavoidable framework suppressions (ruff TC001–TC003, pydantic prop-decorator) is exempt. Gated by ROBOCO_CONVENTIONS_ENABLED; fully inert when off. RoboCo itself ships a canonical .roboco/conventions.yml.
Effective map. Consumers read the effective map — auto-derived defaults (from a repo scan + BUILTIN_RULES, excluding tests//docs/ trees) overlaid by the committed file — so behaviour is identical whether the file is present, absent, or partial. ConventionsService (roboco/services/conventions.py) builds it, caches it per (project, HEAD sha) in project_conventions_cache (migration 043), renders the per-task baseline constraints + the ambient prompt block, and scaffolds/restores the file via a PR (GitService.open_conventions_pr). The committed file + scan are read from a dedicated project-level read clone the service ensures on demand (WorkspaceService.ensure_read_clone, pinned to the default branch's HEAD) — the backfill that makes the standard resolve even for a project created before it existed, with no manual workspace_path. The schema lives in roboco/foundation/policy/conventions/ (pure).
Validator. A single Python CLI, python -m roboco.conventions check --root <repo> --files <a> <b> ... (roboco/conventions/), uses tree-sitter (Python + TypeScript grammars, shipped in the agent image) to classify each changed definition and flag forbidden placements + hygiene + custom-rule matches as JSONL findings, after waiver filtering. Precision over recall (it abstains when uncertain so a block gate can't false-positive-strand a task) and fail-loud (a validator that cannot run exits 3 so the gate blocks, never silently passes).
Threading + enforcement. The standard reaches the work two ways: an ambient "Architectural Standard" block injected at spawn (compose_prompt) and an auto-attached ## Constraints section on every project task (TaskService.create). Enforcement is deterministic: a block-level finding refuses i_am_done (dev pre-submit) and pr_pass (the in-path PR gate) with the offending file:line + fix hint; findings also surface in QA's claim_review evidence (convention_findings). A false positive is relieved by a waiver the dev commits in their branch — accountable, reviewed in the PR. The panel's per-project Conventions tab (in the edit-project dialog) shows the map + health and offers Save / Restore.
MegaTask (sequenced batch intake)
MegaTask lets the CEO describe several tasks in one Intake chat and ship them as one collision-aware, sequenced batch — even across projects that don't share a codebase (the motivating case: a SaaS app + its OSS core engine + a framework adapter). It is a core capability, not a feature flag (additive + opt-in by nature: proposed only when the CEO asks for several tasks; single-task intake is byte-for-byte unchanged), branded "MegaTask" on every user-facing surface while internal names stay technical (batch_id, SequencingService).
The umbrella model. A MegaTask's identity is a real umbrella task — branchless, no PR of its own — over N root-subtasks, each a real Main-PM coordination root with its own project_id, branch, and PR. Hierarchy: Umbrella (Main PM) → N Root-subtasks (Main PM) → Cell tasks (cell PMs) → Dev subtasks. One extra Main-PM layer on top of the normal model. The umbrella is the single board-review / CEO-approve / Main-PM-coordinate unit, so the batch plugs into the existing coordination-root flow for free (task tree, progress rollup, CEO queue).
Identity predicate (single source of truth). roboco/foundation/policy/batch.py: is_batch_umbrella (batch_id set AND parent_task_id None), is_batch_root_subtask (batch_id set AND parented), is_branchless_coordination ((no-project AND product) OR umbrella). Every git-exemption site consults it so the umbrella's exemptions can't drift: the orchestrator's _is_coordination_task, the claim→in_progress branch gate (GitContext.is_coordination), _ensure_branch_for_task (returns "" for an umbrella), and the CEO-reject routing. submit_root hard-rejects an umbrella (it assembles no PR); umbrella completion reuses the existing branchless path (all_subtasks_terminal, PR waived → escalate to CEO).
Sequencing. The pure SequencingService.analyze(surfaces, cell_of, cell_capacity) (roboco/services/sequencing.py; schema in roboco/foundation/policy/sequencing/) turns each draft's collision surface — intends_to_touch (globs), adds_migration, touches_shared — into a dependency DAG + Kahn-layered waves: file-overlap serializes (more-important first by (priority, idx)), migration-adders chain serially, a shared-surface edit runs after each non-shared task it overlaps (file-overlap-conditioned), independent tasks run in parallel; cell-contention only warns. Correctness lives in code, not agent judgment. The columns tasks.batch_id + intends_to_touch / adds_migration / touches_shared are migration 046.
Intake + create path. The intake chat can be scoped to a MegaTask (a multi-project picker → StartLiveRequest.project_ids); the orchestrator clones each repo (_clone_intake_scope / _slugs_for_project_ids, the multi-repo machinery products already used). The intake agent proposes the whole batch with one propose_batch tool call — wired on both runtimes (the Claude SDK driver emits one batch stream chunk; the grok intake_server POSTs a batch relay event). The panel's third intake scope accumulates it into a Review-MegaTask card → POST /prompter/live/{session}/confirm-batch. PrompterService.confirm_live_batch builds the umbrella + N root-subtasks (via create_task_from_draft + a BatchPlacement) and wires the analyzer edges through add_dependency. The Board route holds the root-subtasks in BACKLOG until approve_and_start releases them (_activate_batch_root_subtasks); the Main-PM route dispatches wave 0 at once. The Product Owner + Head of Marketing review the whole batch (their identity prompts carry a MegaTask section).
Services
Core services in roboco/services/:
| Service | Purpose |
|---|---|
TaskService |
Task CRUD and state transitions |
WorkSessionService |
Git session management, PR lifecycle |
WorkspaceService |
Multi-agent workspace resolution and cloning |
ProjectService |
Project/repository management |
MessagingService |
Channels, sessions, messages |
NotificationService |
Formal notifications |
JournalService |
Agent journals and entries |
OptimalService |
RAG queries (in-house pgvector engine) |
PermissionsService |
Role-based access control |
Configuration
Key settings in roboco/config.py (env prefix: ROBOCO_):
# Database
ROBOCO_DATABASE_HOST=localhost
ROBOCO_DATABASE_PORT=5432
ROBOCO_DATABASE_USER=roboco
ROBOCO_DATABASE_PASSWORD=roboco
ROBOCO_DATABASE_NAME=roboco
# Redis
ROBOCO_REDIS_HOST=localhost
ROBOCO_REDIS_PORT=6379
# Security (REQUIRED)
# Generate with: python -c 'from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())'
ROBOCO_ENCRYPTION_KEY=<your-fernet-key>
# Workspaces
ROBOCO_WORKSPACES_ROOT=/data/workspaces
ROBOCO_WORKSPACE_AUTO_CLONE=true
ROBOCO_WORKSPACE_CLONE_TIMEOUT=300
# RAG (in-house pgvector engine)
ROBOCO_RAG_CHUNK_STRATEGY=fixed
ROBOCO_RAG_CHUNK_SIZE=512
ROBOCO_RAG_USE_HYDE=true
ROBOCO_RAG_USE_HYBRID_SEARCH=true
# AI/LLM
ROBOCO_DEFAULT_EMBEDDING_MODEL=qwen3-embedding:0.6b
ROBOCO_LOCAL_LLM_MODEL=glm-5.2:cloud
ROBOCO_LOCAL_LLM_BASE_URL=http://roboco-ollama:11434/v1
ROBOCO_OLLAMA_BASE_URL=http://roboco-ollama:11434
Docker Deployment
Container Architecture
The system runs as Docker Compose services. All Dockerfiles live under docker/ at the project root; every service uses context: . plus dockerfile: docker/<name>.Dockerfile.
| Service | Purpose | Healthcheck |
|---|---|---|
postgres |
PostgreSQL + pgvector | pg_isready |
redis |
Cache, sessions, event bus | redis-cli ping |
ollama |
Local LLM + embeddings | ollama list |
ollama-init |
Pulls models on startup | One-shot |
agent-base-image / agent-*-image |
Pre-built images spawned per agent | One-shot |
orchestrator |
API + agent spawner | Depends on all above |
panel |
Next.js control panel (internal, port 3000) | — |
nginx |
Reverse proxy fronting panel + orchestrator | — |
Single Entry Point
nginx is the only externally-exposed service. It listens on localhost:3000 and routes:
/api/*and/ws/*→orchestrator:8000- everything else →
panel:3000
This avoids CORS since the browser sees one origin. The Next.js code uses relative URLs (/api, /ws) and lets nginx do the dispatch.
WebSocket streams
The orchestrator exposes WebSocket endpoints under /ws (router in roboco/api/websocket.py, ConnectionManager + broadcast_* helpers):
| Endpoint | Purpose |
|---|---|
/ws/channels/{id}, /ws/agents/{id}, /ws/sessions/{id}, /ws/notifications/{id} |
Per-resource live streams — /ws/channels + /ws/sessions carry live message.new frames (from EventType.MESSAGE_SENT), so a session transcript updates without a manual refresh |
/ws/system |
Operator/system-wide stream (no per-agent keying) — the rate-limit lifecycle (RATE_LIMIT_HIT / RATE_LIMIT_LIFTED) and live usage (USAGE_SNAPSHOT, pushed to the usage dashboard) |
Server-side events reach these sockets through roboco/api/websocket_bridge.py, which subscribes to the StreamEventBus and forwards each event to the matching connections. To add a new live event: define an EventType (dotted value), publish it to the bus, add a _handle_* forwarder in websocket_bridge, and consume it on the panel via the useWebSocket("/<endpoint>", …) hook — do not stand up a parallel endpoint or client stack. MESSAGE_SENT is the worked example: send_message publishes it, _handle_message_event fans it out to /ws/sessions/{id} + /ws/channels/{id} as a message.new frame, and the panel's useSessionStream consumes it.
Rate limiting & usage
- Provider rate limits are tracked in Redis (
RateLimitStateTracker,roboco/services/gateway/). On a provider 429 an agent callsi_am_blocked(reason="rate_limited"); the spawn gate then queues (never drops) further work for that provider, and a background probe-and-resume loop in the orchestrator clears the limit and revives parked agents when it lifts. - Provider overloads reuse the same park-and-probe break. A persistent model-API overload (HTTP 529 / 500 / 503 — the SDK already retries transient ones) parks the provider exactly like a 429 instead of crash-retrying the agent straight back into the overload and burning tokens; the overload is detected orchestrator-side from the dead container's log markers, and the background loop revives the parked work when it recovers. The same break also catches the Claude session-limit 429 (the org's 5-hour usage window): an agent exiting with a 0-token session-limit rejection parks the provider and is auto-revived when the window resets, instead of fleet-wide crash-respawning straight back into the limit. Gated by
ROBOCO_OVERLOAD_BREAK_ENABLED(default-on). - Gateway-health recovery closes a blind spot in the stale-claim reaper: the heartbeat is bumped only by gateway verbs, so a broken-but-alive agent (a corrupted
/app/.venvso no gateway tool imports) goes heartbeat-stale yet keeps its container up, and the reaper's live-skip would protect it forever. On a stale-heartbeat live container the reaper now probes the gateway out-of-band (_probe_gateway_health→docker execthe gateway venv imports) and, once broken pastROBOCO_GATEWAY_HEALTH_GRACE_SECONDS(a transient probe miss is tolerated), kills + evicts it (_maybe_recover_broken_gateway) so it falls through to release + respawn; healthy or inconclusive probes spare it. Gated byROBOCO_GATEWAY_HEALTH_ENABLED(default-on). It is the third leg beside the shipped bash-guard/appblock (prevents the self-corruption) and the reaper Docker-liveness fallback (stops over-reaping live containers). - PM coordinator concurrency. A Main / Cell PM plans and delegates many root tasks in parallel — the actual work then runs in the delegated children/cells, not in the PM's own hands. The claim-time concurrency guards that keep a developer to one task at a time (
already_active/paused, inroboco/services/gateway/claim_guards.py) are therefore skipped for the coordinator PM roles (_COORDINATOR_ROLES = {main_pm, cell_pm}, consulted in_run_claim_guards); only a genuine upstream sequence dependency (unmet_dependency, which parks the task back topending) holds a PM's root back. Without this a single PM that claimed one root could never plan a second — it thrashed between its claimed roots and respawned forever, burning tokens for zero progress (the livei_am_idle-auto-paused-umbrella deadlock). Thepausedguard also excludes the target task itself, so a PM re-entering its own paused umbrella never self-blocks. - Orchestrator runtime-state durability. The PM-respawn loop breaker (
_pm_respawn_tracker, the(agent_slug, task_id) → strike-countcircuit breaker) is DB-durable via therespawn_trackertable (migration 051): each gate mutation write-throughs fire-and-forget on the_bg_tasksset (_schedule_respawn_persist→_persist_respawn_record), andrestore_respawn_tracker()repopulates it atstart(), validating each row against live tasks (terminal/missing rows are evicted). Kept only in memory it reset tocount=1on every restart and re-burned the whole strike threshold (4 spawns) against a still-wedged task. It mirrors theWaitingRecordTable/restore_waiting_recordspattern: best-effort (a DB hiccup degrades to in-memory-only — it can only ever suppress a spawn, never manufacture one) and inert when the table is empty. The companion_instancesregistry is reconciled-from-Docker (not persisted) at startup via_readopt_running_agents, so the reaper's liveness path and the spawn gate's_is_agent_activecheck see surviving containers immediately after a restart. - Token usage is captured per agent session from the Claude Code transcript via the SDK server's
/usage/sync(hook → orchestrator finalize →agent_spawn_sessions→daily_usage_rollups→ dashboard). Cost uses provider-aware pricing inroboco/billing/pricing.py(Anthropic priced; local/Ollama intentionally$0). The token sweep also publishesUSAGE_SNAPSHOTto/ws/system, so the dashboard's "Token Usage & Cost" panel updates live and falls back to HTTP polling when the stream is down. - Delivery observability (the panel's Metrics → "Delivery" tab) shows how work flows, computed by
MetricsServicefrom data already captured — no new feature flag. Per-stage cycle time and the bottleneck distribution are reconstructed from theaudit_logtransition journey (each generictask.<status>event marks entry into a status; the namedtask.qa_fail/task.pr_failevents are excluded from the reconstruction). Rework rate readstasks.revision_count— incremented once per transition intoneeds_revisionat the single chokepointTaskService._emit_status_transition_audit— and attributes each bounce to the QA / PR-reviewer via those named audit events; rework cost joinsagent_spawn_sessions.task_id. Read-only endpoints:/dashboard/metrics/{cycle-time,bottlenecks,rework,scorecard/agent/{id},scorecard/team/{team}}.
Startup Sequence
The startup order is critical due to dependencies:
postgres ──┐
redis ─────┼──> ollama ──> ollama-init ──> orchestrator ──> panel ──> nginx
│ │ │
│ │ └── Pulls qwen3-embedding:0.6b, glm-5.2:cloud
│ └── Healthcheck: ollama list
└── Healthcheck: pg_isready, redis-cli ping
Important timing notes:
ollama-initpulls models (~30s for embedding model, ~2min for LLM)- Orchestrator waits for models before starting
- FastAPI lifespan indexes documents using Ollama (~30-60s)
- Orchestrator polls
/healthuntil API is ready before starting dispatcher - After orchestrator is up,
panel(Next.js) builds/starts, thennginx
Database migrations
Schema changes ship as Alembic migrations under alembic/versions/. Run:
docker compose exec orchestrator alembic upgrade head
after pulling any change that adds a new migration.
Ollama Configuration
Ollama provides two APIs:
/v1/*- OpenAI-compatible API (for LLM chat/completion)/api/*- Native Ollama API (for embeddings, model management)
The embedder uses /api/embed endpoint with the qwen3-embedding:0.6b model.
Environment variables for Docker:
ROBOCO_LOCAL_LLM_BASE_URL=http://roboco-ollama:11434/v1 # OpenAI-compat
ROBOCO_OLLAMA_BASE_URL=http://roboco-ollama:11434 # Native API
Common Issues
| Symptom | Cause | Fix |
|---|---|---|
404 /api/embed |
Model not pulled | Check docker logs roboco-ollama-init |
All connection attempts failed |
API not ready | Orchestrator starts before FastAPI lifespan completes |
| Healthcheck failing | Wrong endpoint | Use ollama list not curl |
Blueprint Reference
The organizational structure, communication matrix, role descriptions, and access-control model are documented inline above and in the user-facing documentation site (MkDocs Material; source under docs/, built by mkdocs.yml, deployed by .github/workflows/docs.yml via GitHub Pages Actions and served at rennf93.github.io/roboco). docs/rag/ remains the agent-facing RAG corpus (excluded from the published site); the old root usage.md / deployment.md are now redirect stubs into the site.