git:20260907.eed0760 to git:20260907.65ed017

# AGENTS.md

This file provides guidance to coding agents working in this repository.

## What this repository is

<One paragraph. Project purpose, current baseline, what the artefacts in

this repo represent.>

## Picking up this repo

<!-- Bootstrap GENERATES this list from the recorded artefact root, the

optional layers this repo actually has, and the coordination mode: drop

every line whose file was not written, renumber what remains, and

resolve each path against the artefact root (e.g.

`.docflow/CONVENTIONS.md`). This section is the repo's only read order —

no separate hand-off file is written. -->

Read these, in order, before any tool calls:

1. `AGENTS.md` (this file) — the hard rules. Read in full.

2. `CONVENTIONS.md` — authoring rules, ADR status semantics, the git

contract.

3. `plan/README.md` — how the work queue is used.

4. The newest `plan/done/` entries and

`git log --first-parent --oneline -n 20` — the shipped-work record.

5. `INDEX.md` — the ADR catalogue: an ADR's filename and its dependency

chain. Sorted by number, not by work order.

<!-- Coordination lines: keep 6 in shared-checkout and separate-worktree

modes; keep 7 in shared-checkout mode only; drop both for a single

writer. -->

6. `_agent/ROLES.md` — who writes what.

7. `_agent/LOCKS.md` — which files are currently claimed; claim yours

here before editing.

the ADR(s) it names — both in full.

## Repository structure

- `adr/0000-template.md` — canonical ADR template.

<!-- If two ADR shapes (Q2): also include `adr/0000-template-technology.md` for technology ADRs. -->

- `adr/NNNN-<kebab-slug>.md` — one ADR per decision, contiguous

numbering, no gaps.

- `INDEX.md` — table regenerated from every ADR's metadata block.

- `CONVENTIONS.md` — authoring rules (read before editing anything).

<!-- If plan folder skipped (Q4a): drop the two plan/ lines and the

"Plan folder" section below. -->

- `plan/todo/NNNN-<slug>.md` — pending work, lower numbers run first.

- `plan/done/<YYYY-MM-DD>-<slug>.md` — shipped work, chronological.

<!-- `_agent/` line — list the members the coordination mode (Q5) writes,

and drop the line entirely when no `_agent/` exists (a single writer

without both a verify gate and a plan queue):

single writer: `prompts/autonomous.md`

shared checkout: `ROLES.md`, `LOCKS.md`, `prompts/autonomous.md`

separate worktrees: `ROLES.md`, `prompts/autonomous.md`

`prompts/autonomous.md` appears only where a verify gate was recorded

(Q8) AND the plan queue it walks exists (Q4a); drop it from the list

otherwise. -->

- `_agent/` — the agent operating contract: who writes what, the one

real mutex, and how an unattended run behaves. Here: `ROLES.md`,

`LOCKS.md`, `prompts/autonomous.md`. It holds nothing git already

records.

<!-- If GLOSSARY.md (Q7): also include `GLOSSARY.md` — shared terms. -->

<!-- If domains/ (Q7): also include `domains/<slug>/README.md`. -->

## Hard rules when editing ADRs

These come from `CONVENTIONS.md` and override default behaviour:

- **One decision per ADR.** Splits become new ADRs that supersede;

never expand scope inside an existing one.

- **Status lifecycle:** `<from Q3>`.

- **Capability ADR section order:** metadata → Context → Capability

statement → User stories / scenarios → Acceptance criteria → Out of

scope → Open questions → References → Revision History → Approvals.

<!-- If two ADR shapes (Q2): keep the NEXT TWO bullets. A single-shape repo

drops both — it writes no `shape:` field at all. -->

- **Technology ADR section order:** metadata → Context → Decision →

Rationale → Consequences → Acceptance criteria → Out of scope → Open

questions → References → Revision History → Approvals. Rationale must

name alternatives considered with specific rejection reasons.

- **Shape is declared, never numbered.** Each ADR's `shape:` field says

which order applies (`capability` — the default when the field is

absent — or `technology`). Every ADR takes the next contiguous number

whatever its shape.

<!-- End of the two-shape bullets. -->

- **Acceptance criteria are testable and numbered.**

- **ADRs are internal artefacts — never user-visible.** ADR numbers,

ADR titles, and the existence of the ADR catalogue must NEVER appear

in any string the product emits to users: UI copy, API response

bodies, error messages, customer-visible log lines, public

documentation, release notes, marketing copy, or support

communications. The catalogue is a builder's tool, not a user-facing

surface. References ARE allowed in: code comments

(`// see adr/0042-foo.md`), commit messages, PR descriptions,

internal docs, `AGENTS.md`, `CONVENTIONS.md`, `INDEX.md`, and the

`plan/` queue. Rule of thumb: if a non-builder could ever see the

string, the ADR reference comes out.

<!-- Insert any Q10 domain-specific hard rules here as additional bullets. -->

## Implementation work

- Start from the ADRs. Identify which ADRs a code change implements or

affects before changing behaviour.

- If implementation reveals a capability gap or changed decision, update

the relevant ADR rather than silently diverging.

- Add or update tests for implemented behaviour. Map tests back to ADR

acceptance criteria where practical.

- **Do not leak ADR identifiers into user-visible surfaces.** When

writing error messages, UI copy, API responses, log lines that ship

to customers, public docs, or release notes, refer to the behaviour

by its product-level name — never by ADR number, ADR title, or

phrases like "per the ADR catalogue". The ADR link belongs in the

commit message and (optionally) an inline code comment, not in the

string the user reads.

## Audit trail and revision discipline

- Substantive ADR changes append a row to the Revision History table.

Editorial changes (typos, formatting, link fixes) are excluded but

flagged `editorial` in the commit message.

- Approvals table populates when an ADR is Accepted and updates on each

later substantive revision.

- Regenerate `INDEX.md` from ADR metadata after any ADR status change

or new ADR.

## Multi-agent workflow

<!-- Keep the ONE block matching the coordination mode (Q5); delete the

others. `_agent/` holds contracts only — no mode writes a log, a

dashboard, or a snapshot of what git already records. -->

<!-- Single writer. Keep this block. -->

A single writer owns this repo — one human/agent integrates at a time,

so there is nothing to serialise. What happened is in git history and

`plan/done/`; what is in flight is the branch you are on and any open

pull request.

<!-- If a verify gate was recorded (Q8) AND the plan queue exists (Q4a),

add: The one coordination file is `_agent/prompts/autonomous.md`, the

contract for an unattended run. Without both — the prompt walks

`plan/todo/` and ships through the gate — there is no `_agent/`

directory at all. -->

<!-- Several writers, one shared checkout. Replace the block above with: -->

<!--

Named actors answer for areas (see `_agent/ROLES.md`); work is assigned

by claim.

Coordination rules:

- One working tree means no branch per item: claim a queue item **on the

item itself** — the actor and the date, in its own status section where

it has one — before you start it.

- Before editing a file, claim it in `_agent/LOCKS.md` by appending

`<agent-id> | <path> | <ISO-8601 timestamp>`. Remove the line on

commit. LOCKS is the one real mutex — it prevents simultaneous

writes to the same file in this shared checkout.

- The shipped record is git history and `plan/done/`.

-->

<!-- Several writers, separate worktrees / PR branches. Replace the

block above with: -->

<!--

Named actors answer for areas (see `_agent/ROLES.md`); work is assigned

by claim. Each writer works in its own git worktree or PR branch.

Coordination rules:

- **The claim on a queue item is the branch `claim/<item-key>` pushed

for it** — the queue file name without its extension, e.g.

`claim/0007-rate-limit` — plus its draft pull request where

integration is PR-based. Commit, then push: the push is what excludes

the second writer, because one item maps to one ref. There is no lock

ledger: worktrees cannot collide on the filesystem, and an advisory

ledger nobody can rely on is noise. What has to be guarded against is

duplicated work and contradictory merges — see the guardrails below.

- **State what you own** in the pull-request description, or the

branch's first commit message under direct-to-main: the identifier

block reserved for you and the artefacts you are the single writer of.

- What is in flight is **derived** — `git worktree list`, the remote

`claim/*` branches, and the open draft pull requests — not read from a

file; what shipped is git history and `plan/done/`.

-->

<!-- Concurrency guardrails hard rule — bootstrap INCLUDES this bullet

ONLY for several-writer (shared-checkout / worktree) OR PR-based repos

(omit for a single writer on direct-to-main). See CONVENTIONS.md

§Concurrency Guardrails.

- **Before integrating, check for number collisions (G2/G3).** Sync onto

the current `main` and run the audit skill; if your ADR or `plan/todo`

number now clashes with what landed on `main`, renumber locally before

integrating. The single-threaded merge gate rejects a duplicate as the

backstop. Numbers are immutable once merged. (G1 — landing the ADR

before implementation — is recommended guidance, in CONVENTIONS.md.)

- **Claim a queued item before implementing it (G4).** In separate

worktrees / PR branches, commit and then push `claim/<item-key>` — the

queue file name without its extension — and open a draft PR from it

where integration is PR-based; the push is the exclusion, so a rejected

push means someone else holds the item. In a shared checkout, claim the

item on the item itself. A single writer claims nothing. Never start a

queued item you have not claimed.

-->

## Plan folder

- A pending item gets a `plan/todo/NNNN-<slug>.md` file BEFORE work

starts, naming the owning ADR(s), scope, and exit criteria.

- The completion event is: `<from Q4 — e.g. merge to main, deploy +

smoke passes, release tag>`. Under direct-to-main integration, the

completion commit moves the file to

`plan/done/<YYYY-MM-DD>-<slug>.md` with a footer naming the HEAD SHA

and any artefact id, then pushes. Under pull-request integration,

the move, any plan-item Status removal, ADR advance, and INDEX

regeneration are the final branch commit before the pull request is

marked ready; the footer names the pull request, and the merge is

the completion event.

- The owning ADR(s) advance `Accepted → Implemented` on the same

commit. Regenerate `INDEX.md`.

## Git contract

- Commit messages follow **Conventional Commits**.

- Mandatory `Rationale:` footer on any commit touching an ADR.

- <Signed commits: yes/no per Q6.>

- <ADR-revision tags `adr-NNNN-rN`: yes/no per Q6.>

- <Co-Authored-By trailer: yes/no per Q6 — default no.>

- Cross-references between ADRs use relative paths (`adr/NNNN-*.md`).

<!-- Integration model per Q4b — keep one block. -->

<!-- Direct-to-main (if no verify gate was recorded (Q8), drop the

gate sentence):

- **Integration:** direct-to-main, **fast-forward only**. No merge

commits on `main`. The verify gate (`<command from Q8>`) runs

locally and must pass before push. Completion event:

fast-forwarded to `main` + remote push succeeded.

-->

<!-- PR-based:

- **Integration:** every change ships via a pull request. CI must be

green before merge — the verify gate (`<command from Q8>`) runs in

CI on the PR, not (only) locally. Merge strategy:

<squash | merge | rebase>. Completion event: PR merged to `main`

with CI green. Completion changes are committed on the pull-request

branch before it is marked ready; no follow-up commit is made on the

integration branch after merge.

-->