PROC-0003 — Team coordination across repositories¶

Status: draft. Open for review.

Four people, two repositories, several lanes running at once. PROC-0001 says the topic is the unit of parallel work; PROC-0002 says how an ingestion lane runs. This says how the lanes stay out of each other's way, and — the part that is easy to get wrong — which project root to open.

1. Skill partition: open the repo you are working in¶

Claude Code loads skills from the project root only. Never transitively through a submodule, and never from a parent directory.

This is not a detail. A session rooted at ~/cb — the umbrella workspace — loads ~/cb's skills and none of m-of-n's. Work done that way runs without ingest-reference, summarize, propose-arch or any of the judgement they carry, and nothing warns you.

Working on	Open as project	Skills you get
library records, ingestion, summaries, extraction	`library/`	`ingest-reference`, `summarize`, `extract`, `distill`
architecture, plan, spec, backlog	`mofn/`	`propose-arch`, `close-topic`
both at once, or neither	`~/cb`	the workspace skills only — not these

So the skills are partitioned by what they act on, not gathered in one place:

Skills that act on library records are canonical in library/.claude/skills/.
Skills that act on m-of-n documents live in mofn/.claude/skills/.
Neither is duplicated. Duplicates drift, and a drifted skill is worse than a missing one because it looks authoritative.

The second reason: the submodule checkout is on a detached HEAD¶

mofn/library/ is a submodule checkout, and a submodule checkout sits on a detached HEAD — the pinned commit, on no branch. Commits made there belong to no branch. They survive only until the next git checkout, git submodule update or bin/lib-sync, which silently moves HEAD and orphans them.

This is not hypothetical. On 2026-09-23 a 49-record CBOR/COSE sweep was found uncommitted on a detached HEAD inside mofn/library/ — one checkout from being discarded. It was rescued to a branch and merged as library#13 (library#12 records the incident).

Never edit, commit or run ingestion inside mofn/library/. It is read-only: the pinned copy that reports cite. Do library work in a clone of the library repo, opened as its own project — which is also the only way its skills load (above). Two reasons, one rule.

If you are unsure which root you are in, you are in the wrong one. Check before a long ingestion run, not after.

2. Lane assignment¶

A lane is one person or agent, one topic, one branch, one worktree, in one repository.


Unit	a topic (`library/topics/<id>.yaml`) — one bounded question, one owner
Branch	`topic/<id>` in whichever repo the work lands
Worktree	`bin/wt new topic/<id>` — so two lanes cannot collide on files
Output	one PR per topic, not per record
Concurrency	three to five lanes, the weekly human review budget

Lanes are sized so their records rarely touch the same files. When two lanes keep colliding, the topics were drawn wrong — split them rather than serialising the people.

3. Which repo does a lane land in¶

Work	Repo	Why
Ingestion, records, summaries, distillation	`library`	it owns the records
Topic notes and answers	`library`	topics live with what they own
Research reports (D3)	`mofn` (`research/`)	they cite `library@<commit>`
Architecture, ADRs, mappings	`mofn` (`spec/`)	serialised, sponsor-reviewed
Plan, backlog, process	`mofn` (`project/`)
Reference implementation	`mofn` (`src/`)	empty until DEC-002

A lane that spans both repos is two PRs, and the library one merges first. Then bin/lib-sync --update moves the pin in a separate commit, so it is visible that the bibliography a report was written against has changed.

4. The submodule pin is the coordination point¶

mofn pins library at a commit. That pin is the whole synchronisation mechanism, and it is deliberate:

Merging in library changes nothing in mofn until someone moves the pin. A report citing library@<commit> cannot have its bibliography shift underneath it.
Move the pin in its own commit, never bundled with content changes. A reviewer should be able to see "the library moved" as one line.
bin/lib-sync --check fails a PR where the pin changed without being staged — so an accidental pin move cannot ride along unnoticed.

Version skew is expected, not a bug. The pinned library may predate a tool that mofn now calls. CI and bin/manifest both run only what the pinned library actually provides.

4b. Should `mofn` be only a harness? — not yet, and here is the trigger¶

Asked 2026-09-22. A harness-only mofn would hold the site generator, the publishing manifest, CI and cross-repo coordination; spec/, the application and the library would each live elsewhere.

Not now. The cost is concrete: the submodule pin is the coordination point (§4), and three pins cost three times what one does. The same two people touch spec, app and site this semester, so splitting adds review surface without separating any audience.

The trigger to revisit — any one of these, and the split earns its cost:

someone wants the specification without the site, or cites it independently
the application ships on its own cadence rather than with the semester
a second implementation appears and needs the spec and vectors alone
the site outlives the project and needs contributors who do not touch the spec

Revisit in January, or when one of those happens.

5. Cadence¶

When	What
Start of an increment	lanes assigned, draft PRs opened the same day
Daily	push to the lane branch; a draft PR is the status report
Review, batched	one topic, 6–10 records, one sitting (PROC-0002 §4)
Weekly, 30 min	decisions and unblocking; notes committed same day
End of increment	topics closed or parked with the experiment named

A draft PR opened on day one is worth more than a status update. It gives the reviewer somewhere to comment while comments are still cheap.

6. What must never run in parallel¶

Architecture. One change to spec/ in flight at a time. ARCH-0001 §9.4 forbids parallel essays, and two concurrent branches touching the model produce exactly the divergence it exists to prevent. Check for an open PR touching spec/ before starting.
The record schema. It is the contract every lane depends on; changing it mid-increment invalidates work in flight.
Anything touching the same file. No exceptions, no "small" edits.

7. Regeneration is a team obligation¶

Derived views go stale the moment anyone ingests:

cd library && bin/validate && bin/export && bin/reindex   # index/, exports/
cd mofn    && bin/manifest                                 # nav, PUBLISHING.md, manifest.json

CI gates both. The failure they prevent is the quiet one: the human-visible bibliography and the published site drifting out of date while every individual PR looked fine.

8. Open questions¶

Does ~/cb need a path-scoped rule pointing at this file, so a workspace-rooted session is told to open the sub-repo? It would have caught the problem in §1 immediately.
Three to five lanes is inherited from PROC-0001 and still uncalibrated against a team of four with one reviewer.
Who moves the submodule pin — the lane that merged in library, or whoever next needs it in mofn? The first keeps it current; the second keeps it intentional.
Should topic ownership be exclusive? Two people on one topic is how the worktree discipline gets bypassed.