* fix(board): LEARN decisions name the item, not its per-cycle index
A cycle's reject reasons are rendered into the NEXT cycle's exploration
prompt, but the ref recorded alongside each reason was the item's stored
id (item-0/item-1) — a per-cycle index that means something different
every cycle and appears nowhere the explorer can resolve. The reason
survived the loop; what it was about did not.
Record the item's title instead, via a shared learn_ref() helper (falls
back to the id when title-less, and reads target_task_title for Scales,
whose items name the live task they mutate).
* chore(lint): satisfy ruff 0.16 — keyword-only signatures and markdown formatting
The dev toolchain resolved ruff 0.16.0, which stabilises PLR0917 (too many
positional arguments) and formats python code blocks inside markdown. Both
fired repo-wide and neither had anything to do with the code they flagged.
- 36 signatures gain a `*` so their tail arguments are keyword-only, and
the 104 call sites that passed them positionally are converted. mypy was
the safety net for the static ones; the full suite caught nine more that
only bind at runtime (the MCP tool functions, whose real callers already
pass named JSON arguments).
- 28 markdown files reformatted by 0.16's code-block formatter.
- One RUF036 (`None` mid-union) autofixed in the GitLab provider.
* fix(gateway): log the reason when a verb rejects
A rejected envelope rides an HTTP 200, its body is never logged, and there
is no trace table — so in the access log a verb an agent could not satisfy
looks identical to one that worked. On 2026-07-25 four Board Programs
(Periscope, Sentinel, Scales, Barfly) each POSTed their propose verb three
or four times, persisted nothing, and left their exploration tasks PENDING;
the reason was unrecoverable afterwards, from the logs or from the agents'
own transcripts.
Log error/message/remediate/missing plus the calling agent at
envelope_to_response — the one chokepoint every v1 flow and do route
returns through. Success envelopes stay silent.
---------
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
* fix(rag): per-index chunk floors — journals and learnings were never indexed
The global 200-char garbage floor (sized for code/doc chunks) discarded
every templated journal note and most distilled org-memory lessons,
silently: ingest returned success with zero chunks, so agent journals
and learnings were never retrievable via RAG. IndexConfig now carries a
per-type min_chunk_length (journals 40, learnings 80, others unchanged).
* fix(mcp): git-readonly tools default project_slug from the container env
Agents 404ed /api/git/status with 'Project not found: roboco' — the
tools made the LLM supply the slug and six doc examples taught a slug
that matches no registered project. The tools now fall back to the
ROBOCO_PROJECT_SLUG the orchestrator already injects, and the stale
examples are corrected.
* feat(rag): startup backfill re-ingests zero-chunk journals and learnings
Before the per-index chunk-floor fix, ingest() returned success with
chunk_count=0 for undersized content: every historical journal entry and
distilled learning below the (then-global) 200-char floor was durably
recorded in journal_entries but silently never got a chunks_journals /
chunks_learnings row, and no exception meant the existing dead-letter
(rag_index_failures) never saw it either.
Extends the startup reconcile (roboco/api/app.py _reconcile_rag_indexes)
with a new pass: backfill_unindexed_journals (roboco/services/
rag_index_failures.py) queries journal_entries for rows missing from each
vector table and re-ingests them through the same live code paths
(_reindex_journal_entry / record_learning).
Journals and learnings are backfilled independently since a LEARNING entry
can clear the (lower) JOURNALS floor while still failing the (higher)
LEARNINGS floor — a learning's doc_source is a content hash, not the entry
id, so presence there is checked by hashing each candidate the same way
LearningsIndexPlugin.record_learning does and batch-querying chunks_learnings
for those exact sources.
Bounded to 200 rows per pass per boot (converges over restarts on a larger
backlog) and best-effort per row (one failure never aborts the pass). Rows
still under the current floor are excluded by a length filter in the SELECT
so they are never retried forever, and private entries are excluded from
the JOURNALS pass exactly like the live indexing path.
* test(rag): scope backfill assertions to their own rows
---------
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
Markdown and editors soft-wrap on their own, so the manual ~75-char line
breaks across the docs added nothing but noise. Join wrapped prose, list
items, and paragraphs into single lines across 67 docs — README, CLAUDE.md,
deployment, usage, the RAG knowledge base, and the agent role prompts.
Whitespace-only: code fences, tables, and blockquote alerts are byte-identical
and the change is token-verified (no content altered). Applied with a
deterministic reflow tool (committed separately).
Also lands two doc edits that were awaiting commit: the measured under-load
resource numbers in usage.md and the pr_reviewer additions to the
org-structure RAG doc.
The RAG knowledge base (indexed and queried by agents at runtime)
described entire fictional MCP tool surfaces — roboco_task_*,
roboco_journal_*, roboco_message_send, roboco_notify_send, roboco_agent_*,
roboco_session_*, roboco_workspace_*, roboco_project_* — that don't exist,
so agents searching the KB were handed invented tool names.
Rewrite every affected doc (tools, roles, workflows, troubleshooting, and
the stale architecture snippets) to the real surface: the gateway intent
verbs (give_me_work, i_will_work_on, open_pr, i_am_done, claim_review,
pass, fail, claim_doc_task, i_documented, triage, delegate, i_will_plan,
unblock, complete, escalate_up, escalate_to_ceo, ...) and content tools
(commit, note(scope=...), say, dm, evidence, notify*, open_session,
channels). Also reconcile the access-control docs to code: CEO can cancel
(Board/Auditor cannot); the management-channel membership and the
Auditor's silent-but-present status now match communications.py.
- Add roboco_project_* tools (list, get, create, update) with CEO bypass
- Add roboco_workspace_* tools (ensure, status, list) for workspace management
- Add project_slug and requires_git fields to TaskCreateInput schema
- Validate project exists and cell matches when creating git-enabled tasks
- Register project MCP server in orchestrator with proper permissions
Workspace permissions by role:
- Developer: Write to own workspace only
- QA: Read-only access to all cell workspaces
- Documenter: Write to all cell workspaces (add docs to dev branches)
- Cell PM: Write to own workspace, project_update for own cell
- Main PM: Full project access (create, update all, workspace_list all)
- CEO: Full bypass on all permission checks
Also includes:
- Git templates for commits, branches, PRs (separation of concerns)
- Updated blueprints with project/workspace tools documentation
- Updated RAG docs with project tools reference
