Agentic software factory: inception

Inventory of the MCP Inspector's agentic software factory, a verdict on how each part carries over to this repo, and the ordered sub-issues that build our own. Tracker: #4858. This doc: #4859.

Sources, as read for this doc

Source	Revision
inspector#2498, "MCP Inspector: Our AI Software Factory" (docs/ai-software-factory.md)	102344f5 (open PR)
Inspector v2/main: AGENTS.md, .claude/skills/*, scripts/*, .github/*, docs/quality-gate.md, docs/skill-authoring.md	64a50d6f
This repo v2/main (identical to main at the time of writing)	f46d9578

The Inspector moves fast. Where a sub-issue below says "copy X", re-read X on the Inspector's v2/main at the time, not this doc's summary of it.

Contents

1. What the factory is
2. Inventory and verdicts
3. What only this repo needs
4. Retiring CLAUDE.md
5. Decisions carried over from #4473
6. Reusable templates
7. Existing issues to reconcile
8. Findings along the way
9. Proposed sub-issues
10. Maintainer decisions

1. What the factory is

The Inspector's factory is not a separate tool. It is the repo's contribution process, written precisely enough for a coding agent to run it for long stretches without a human confirming each step. It has four parts:

• Rules (AGENTS.md). Conventions a reviewer cites against a diff. It is loaded in full on every turn, so it survives auto-compaction and must be complete enough to work from on its own.
• Procedures (.claude/skills/*/SKILL.md). Multi-step recipes with live commands, IDs and gotchas. They load on demand, either by description match or by /name. Claude Code and the Copilot CLI both read the same directory.
• A gate strict enough to trust. npm run local:gate runs every check CI runs, plus one. A green local gate is the best available predictor of green CI, and that is what makes unattended runs reasonable.
• Issue-driven tracking. Every PR closes an issue, every issue is a card on the board, and anything automated (dependency bumps, SDK releases) files an issue instead of opening a PR.

A session starts with a maintainer's /goal create a PR for #N. It then runs branch → gate → PR → Copilot review loop → merge → close-out, with the rules and skills supplying every step.

The split between rules and skills follows how each one fails. A skill can drop out of context in a long session, and its description may simply not fire. So anything that must never silently disappear goes in AGENTS.md, and anything that can safely be re-read from disk goes in a skill.

2. Inventory and verdicts

Verdicts:

• Transfer: copy it, changing only names, IDs and paths.
• Adapt: keep the intent but change the mechanics. The adaptation is named.
• N/A: doesn't apply here, with the reason.

The Sub-issue column points into §9.

2.1 Rules: AGENTS.md sections

Section	What it does	Verdict	Adaptation / reason	Sub-issue
Header: rules vs procedures	States that AGENTS.md holds the rules and skills hold the recipes	Transfer	—	S1
Skills index	Table of every skill, what it covers, how it loads	Transfer	Our skill list (§9)	S1, then each skill PR
Project Structure	Annotated tree of the repo	Adapt	7 servers × 2 languages; package name and registry for each server	S1
Every file carries a purpose header	Each source file opens with a comment stating its purpose and rationale, so AGENTS.md never duplicates it	Adapt	Our source files don't follow this today (e.g. src/filesystem/index.ts, src/time/src/mcp_server_time/server.py). The adaptation is no bulk migration: new files, and files a PR substantially rewrites, get a header: the first comment in the file, after any shebang (#!/usr/bin/env node must stay on line 1 of an executable entry point). It spreads as the refactor in #4857 touches each server	S1
Development setup	Root npm install, build, dev loop	Adapt	npm workspaces for TS, uv sync per Python server. Node 22, Python ≥ 3.10	S1
Dependency placement (+ its rationale in local-dev)	Rules for a non-workspace multi-install repo: root-only runtime deps, bundler externals, vitest pin trio, lockstep	N/A	This repo is an npm workspace for its four TS servers, each with its own package.json. The three Python servers use pyproject.toml, and every server publishes independently. The few general rules survive as S1 rules: pin transitive deps with overrides, never npm audit fix; one version of a shared devDependency across workspaces	S1 (the survivors), S11 (local-dev)
Dependency updates are issue-driven	Dependabot PRs off; scheduled sweeps file issues	Adapt	npm and uv/PyPI and Actions ecosystems. Dependabot security-fix PRs are currently on here (§8)	S13
Action pinning (#2484)	SHA-pin actions in credentialed jobs, enforced by verify:action-pins	Transfer	release.yml holds id-token: write in publish-npm / publish-pypi, and claude.yml holds id-token: write + ANTHROPIC_API_KEY. Lands with the release split, before the first milestone release	S12
SDK watch (third sweep)	Nightly issue per MCP SDK release we're behind; a hardened LLM-in-CI analyze job	Adapt	Two SDKs, two registries (npm @modelcontextprotocol/*, PyPI mcp). The security posture carries over unchanged	S13
Contributing	External contributors file issues, not PRs, including org members with write access	Transfer	Decided (§10): outside PRs are turned off, as in the Inspector. External contributors, including org members with write access, file issues; maintainers open PRs. The existing backlog (316 open outside PRs) is handled per S8	S8
Issue forms	Bug and feature forms, blank issues off, security routed to a private advisory	Adapt	Needs a server dropdown (7 servers) and a spec-era/client field. We have no forms today	S8
Every PR references an issue	Closes #N first line; no issue-less PRs	Transfer	Until S13 retires them, Dependabot still opens issue-less PRs. So S1 states the rule with an explicit temporary exception for Dependabot PRs, and S13 removes that exception	S1, S6, S13
Project Status and Direction	Branch table: v2/main develop, main release, v1/main maintenance	Adapt	No v1/main line here. v2/main develops and main releases (the default branch, and what users see)	S1
Maintenance rules	Keep READMEs and AGENTS.md in sync; procedures change in their skill	Transfer	Plus per-server READMEs and RELEASING.md (#4473)	S1
Maintaining the skills	verify:skills, disable-model-invocation explicit and defaulting to false, eval cases, listing budget, paths only when a skill is useless outside the matched files (with the trade-off stated in the PR)	Transfer	—	S2 (rules land with the harness)
Issue-driven Work Style	Board invariants: real issues only, labels, milestones, Priority, Incoming ⇔ unmilestoned, Done = shipped, branch naming, Copilot loop, manual close on v2/main	Adapt	One board (#43), not two. The version label is always v2. Type labels need chore (§8). Server-scope labels already exist	S1 (rules), S5–S7 (recipes)
Responding to Code Reviews	Judge against the issue, decline scope creep, reply in each thread, then a PR-level summary	Transfer	—	S1
Always test new or modified code	Per-file ≥ 90 on all four dimensions, justified v8 ignore, test placement	Adapt	The TS half (all four dimensions) comes from #4854. The Python half (per-file lines and branches, the dimensions coverage.py measures) comes from #4855: coverage.py, justified # pragma: no cover, per-file script	S10
Test-gate timeouts	Budgets in one place, no retry, no fixed sleeps, timeout-minutes per CI job	Adapt	Keep no retry, no scaled sleeps and timeout-minutes on every job. The shared-budget machinery is sized for 6 Vitest projects and a browser, so defer it	S10
Mandatory pre-push gate	npm run format, then npm run local:gate; validate is not a substitute; gate lease	Adapt	A two-language gate: TS workspaces plus uv per server	S3, S4, S10
Waiting on long-running work	Arm a notifier; never poll per turn	Transfer	—	S1
Build output is never a gate target	Lint/format/typecheck read only first-party source	Transfer	src/*/dist, src/*/coverage, Python dist/ and .venv/	S3, S4
Lint has no warning tier	--max-warnings 0; no-floating-promises at error	Transfer	TS only. For Python, ruff has no warning tier, so a finding is a finding	S3
TypeScript instructions	Never any, no config-level suppression, avoid double casts, no floating promises	Transfer	Plus the server idioms from CLAUDE.md (§4)	S1
Web source layout (lib vs utils)	Web-client directory rule	N/A	No web client	—
React instructions	Mantine, hooks, state stores	N/A	No UI	—
Web backend auth token	The Inspector's proxy auth	N/A	No backend proxy	—

2.2 Procedures: skills

Skill	What it does	Verdict	Adaptation / reason	Sub-issue
board-ops	gh project recipes for two boards; ID tables; resolving option IDs by name; the option-deletion hazard and its recovery	Adapt	One board, #43. Its fields: Status (with Incoming) and Priority. Hazard and recovery copy unchanged	S5
issue-create	Five-step create flow: version label, type label, milestone, card, Status + Priority; duplicate check across all states	Adapt	Version label is always v2. Add a server-scope label step (server-<name>). Milestone = nearest due v2.x	S5
issue-triage	Two-pass sweep (board as Incoming, then human approval), priority rubric with a score comment, 12-check board audit	Adapt	The highest-leverage skill here given the inflow (§3.4). Add spam/registry-redirect classes, server-scope labelling, and outside-PR triage	S7
pr-flow	Assign + In Progress, branch naming, DCO signoff, screenshots, Closes #N, addCloseIssueReferences, In Review, Copilot loop to exhaustion, per-thread replies, manual close-out	Adapt	Board #43; branch v2/<type>/<N>-<slug>. DCO: on, as in the Inspector (decided, §10). S6 installs the DCO app and requires git commit -s. Screenshots → client evidence: for a server-facing change, Inspector and LLM-client transcripts in both spec eras; otherwise a targeted probe (§3.2). The Copilot loop copies unchanged	S6
pre-push-gate	Running local:gate; diagnosing each stage	Adapt	Rewrite around our stages: TS workspaces, Python per server, per-file coverage both sides	S10
project-structure	Where a file goes; who owns what	Adapt	Per-server layouts (e.g. everything's tools/, resources/, prompts/, transports/)	S11
local-dev	Install/run each client; dependency-placement reasoning	Adapt	Workspaces + uv; running each server over the transports it implements (stdio for all seven; everything also serves SSE and Streamable HTTP); npx/uvx local builds	S11
testing	Test placement, commands, tiers, coverage gate, renderWithMantine	Adapt	In-process protocol harness (Client ↔ server over in-memory transport; ClientSession for Python), per #4854/#4855	S11
test-servers	Picking and running the Inspector's fixture MCP servers	Adapt	Inverted: here the servers are the product. The equivalent is driving a server with a client: Inspector V2 (web/CLI) and an LLM client, in both spec eras (#4857). Proposed name: client-smoke	S11
release	Name-only. Two PRs (audit + bump on v2/main; milestone merge to main), a ledger artifact, then a human-published GitHub Release	Adapt	Two registries, per-package versions, CalVer (Py) vs semver/changesets (TS, #4472), release environment approvals	S12
security-advisory	Private advisory flow: draft [GHSA-…] card, ownership check, accept, private fork, publish, public tracking. Accept and publish are human-gated	Adapt	One release line. 61 advisories are in triage, and SECURITY.md says the repo is ineligible for reports (§8)	S9

2.3 Scripts and gates

Element	What it does	Verdict	Adaptation / reason	Sub-issue
validate (+ per-client check/validate)	Fast inner loop: guards, format:check, lint, typecheck, build, test	Adapt	Root validate over TS workspaces (#4473 design) plus a Python equivalent per server	S3, S4
local:gate / local:gate:stages / local:validate	Mandatory pre-push command; every CI check plus local-only extras, run once instrumented	Adapt	Chains the TS and Python validate, per-file coverage (both languages), skills CLI check	S10
coverage / coverage:*	Per-client test:coverage at 90/90/90/90 per file	Adapt	TS from #4854; Python from #4855 (coverage.py fail_under is global, so a per-file check script is needed)	S10
format / format:check:*	Prettier across every scope	Adapt	Root Prettier for TS; ruff format for Python	S3, S4
lint:* (--max-warnings 0)	ESLint flat config, type-aware	Adapt	Root flat config across workspaces; ruff check for Python (not run in CI today)	S3, S4
gate-lease.mjs	Machine-wide FIFO lease so concurrent sessions' gates queue	Transfer	Our gates are cheaper, but concurrent sessions still contend. Also needed if any stage binds a fixed port (HTTP transport tests)	S10
lib/workflow-gate.mjs (+ tests)	Keeps the local gate out of CI: no workflow may invoke a local:* script, and local:gate stays exactly the lease wrapper	Adapt	Same invariant for our local:* namespace; the Inspector's engine-pass specifics drop out	S10
test:scripts	Runs every scripts/**/*.test.mjs under node --test, inside validate:guards, so a regression in a guard script fails the gate	Transfer	Keep the exact *.test.mjs naming: node --test silently skips a file its glob misses	S2 (adds it), S3 (root-guards job)
verify:skills / verify:skills:cli / lib/skill-manifest.mjs	Frontmatter parse, explicit invocation mode, eval cases, listing budget; claude plugin validate at a pinned CLI	Transfer	—	S2
skills:eval (skill-eval.mjs, lib/claude-cli.mjs)	Runs each skill's eval cases headless (Claude or Copilot); trigger rate, chains, negatives	Transfer	—	S2
verify:format-coverage	Every first-party file is format-gated	Adapt	Workspace globs; Python via ruff config	S3
verify:typecheck-coverage	Every tracked TS file gets a tsc pass	Adapt	Per-workspace tsconfig (tests are excluded in some servers today)	S3
verify:action-pins	Credentialed jobs use SHA pins with # vX.Y.Z	Transfer	—	S12
verify:test-timeouts	Resolves every Vitest project's budgets; asserts no retry	Adapt	Keep the no-retry assertion only. Defer the budgets machinery until a timeout problem shows up	S10
verify:dep-lockstep	One version per install-crossing dependency across 5 installs	Adapt	A single workspace lockfile can still resolve different versions per workspace, and the manifests already declare different ranges (e.g. typescript ^5.6.2 / ^5.8.2 / ^5.3.3). The adaptation is a smaller guard: every shared TS devDependency (typescript, vitest, @vitest/coverage-v8, prettier, @types/node) is declared with one range across workspaces, or hoisted to the root. Python servers stay independent by design	S3
verify:install-fresh	node_modules matches its lockfile	Adapt	Root workspace install only. CI's npm ci is always fresh, but a long-lived local checkout can have a node_modules older than package-lock.json, so local:gate would test stale dependencies that CI doesn't. It runs in local:gate (and is harmless in CI)	S10
verify:bundle-externals / verify:build-gate	Bundler guards for tsup/Vite output	N/A	Servers compile with plain tsc	—
smoke:* (launcher/cli/tui/web/engines), local:storybook	Built-artifact smokes of the three clients	Adapt	Inverted: a boot smoke per server over each transport it implements (stdio for all; SSE and Streamable HTTP for everything), from the built dist/ (TS) and console script (Py): connect, list, call one tool. Only a thin spawn test, per #4854/#4855	S10, S11
pack:verify (pack-and-verify.mjs)	Installs the exact publish tarball into a throwaway consumer and runs the bin	Adapt	Per package: npm pack → install → npx boot; uv build → install wheel → console-script boot	S12
install-clients.mjs, install-smoke-browser.mjs, run-engine-smokes.mjs, docker-healthcheck.mjs	Inspector-specific install/browser/Docker helpers	N/A	No non-workspace installs, browsers or Docker healthcheck. Our Dockerfiles aren't published by CI	—

2.4 CI and release workflows

Element	What it does	Verdict	Adaptation / reason	Sub-issue
main.yml → build	validate, skills CLI, build guards, smokes, Storybook	Adapt	typescript.yml / python.yml run the new validate (and ruff), keeping the per-package matrix	S3, S4
main.yml → coverage (parallel job, #2159)	CI enforces the per-file gate	Adapt	See §7: #4854/#4855 say "coverage stays local, as in the Inspector", but the Inspector now enforces it in CI	S10
main.yml → package / publish split (#2483)	Build and pack:verify in a job with no id-token; publish downloads the tarball only	Adapt	release.yml today builds and publishes in the same job as id-token: write	S12
main.yml → GHCR image	Publishes a container	N/A	We don't publish images	—
timeout-minutes on every job	Hung-job guard sized from observed runs	Transfer	None of our jobs declare one	S10
dependency-refresh.yml	Monthly: npm outdated + action majors → one tracking issue	Adapt	npm workspaces + uv lock --upgrade --dry-run-style check per Python server + actions	S13
dependabot-alerts.yml	Daily: alerts → one issue per bump, re-checked against v2/main	Adapt	npm + pip ecosystems	S13
sdk-watch.yml	Nightly SDK-release issues + hardened LLM analysis (3 jobs split by permission, no Bash, scanned artifact)	Adapt	Two SDK groups (TS @modelcontextprotocol/sdk → v2 packages; Python mcp)	S13

2.5 Board, labels, milestones, docs, session practice

Element	What it does	Verdict	Adaptation / reason	Sub-issue
Two boards (v2 #28, v1 #11)	One board per release line	Adapt	One board: Servers V2 (#43)	S5
Status: Incoming → Todo → In Progress → In Review → Done	Approval-aware lifecycle	Transfer	#43 already has all five options	S5
Priority field + rubric	Scored, with a posted comment	Transfer	#43 has Urgent/High/Medium/Low	S7
Version labels v1/v2	Line routing	Adapt	Only v2; it marks work tracked by the factory	S5
Type labels (5)	Exactly one per issue	Adapt	bug/enhancement/documentation/question exist; chore is missing	S5
Milestones = release buckets	Incoming ⇔ unmilestoned	Transfer	v2.0.0, v2.1.0 exist	S5, S7
docs/ai-software-factory.md	The overview for humans	Adapt	Write ours once the pieces exist, after Wave 6's automation lands	S13 (closing doc task)
docs/quality-gate.md	Canonical CI-vs-local split	Adapt	Two languages	S10
docs/skill-authoring.md	How to write a description that fires; eval-case design	Transfer	—	S2
.claude/settings.json	Enables the Playwright plugin	Adapt	Not for our own code, which has no UI, but client-smoke drives the Inspector V2 web client, which needs browser automation. S11 enables the plugin, or documents the Inspector CLI as the scripted path and the web client as the hand-driven one	S11
/goal session start	Persistent sessions, one per issue	Transfer	Practice, not a file. Documented in the closing factory doc	—
Copilot review loop	Request via requestReviews (bot id BOT_kgDOCnlnWA), wait, answer, repeat until one clean round	Transfer	—	S6
Co-Authored-By trailer	Attributes agent-authored commits	Transfer	—	S6

Board #43 also had a Size field (XS–XL) with no Inspector counterpart. It held no values and has been deleted (decided, §10).

3. What only this repo needs

3.1 Seven servers, two languages

• Gates at two levels. Every gate runs per package (CI already matrixes over packages) and across the workspace (one root command runs all of them). A Python server has no npm workspace, so the root gate orchestrates uv per server. local:gate is the one command for everything.

• Python equivalents of every TS gate:

  TS gate	Python equivalent
  prettier --check	ruff format --check
  eslint --max-warnings 0	ruff check (not in CI today)
  tsc	pyright (in CI)
  vitest	pytest (+ pytest-asyncio)
  vitest --coverage with per-file 90	pytest --cov + branch coverage + a per-file check script (#4855)
  verify:format-coverage	ruff include/exclude reviewed so no first-party file drops out

• Per-file coverage in Python. coverage.py's fail_under is global. #4855 owns the check script (coverage json → fail any file below threshold, same semantics as the TS gate). The factory wires it into local:gate and CI.

3.2 Server-oriented testing

• Protocol-level client harnesses, in-process: an SDK Client (TS) or ClientSession (Py) over an in-memory transport, asserting on the wire. This is the design of #4854/#4855, and the testing skill (S11) documents it.
• Client smoke tests in both spec eras, per #4857: every change to server behavior is checked against a 2026-07-28 client and a 2025-11-25 client, using both the Inspector V2 and an LLM client. This replaces the Inspector's screenshot rule. Instead of images, a server-facing PR carries client evidence: what each client was asked to do, and what it returned. A change with no client-observable surface (docs, skills, workflows, gate tooling) instead carries a targeted probe, as the Inspector's ledger allows: the thing that proves it, such as a guard made to fire or a before/after run. That gives the client-smoke skill (S11) and the pr-flow evidence step (S6).
• Interface-diff CI (#4860) gives interface-level evidence that a change is transparent. The gate sub-issue (S10) wires it in once #4860 lands.

3.3 Release and publish

• v2/main → main milestone merges. Same shape as the Inspector: bump on v2/main first, a pure merge PR, then the human release step.
• Two registries, both on OIDC trusted publishing. npm is bound to release.yml + environment release (#4463); PyPI uses pypa/gh-action-pypi-publish with skip-existing.
• Per-package versions. The Inspector has one version; we have seven. Today everything is CalVer, stamped at release time by scripts/release.py. #4472 moves TS to semver via changesets and keeps Python on CalVer, stamped by a prepare-release PR. The milestone-release flow (S12) is written against #4472's end state, so #4472 lands first.
• Only changed packages publish. Change detection by file extension since the last tag, today. Under #4472 it becomes a registry diff.

3.4 Triage under heavy community inflow

At the time of writing there are 245 open issues and 316 open PRs, nearly all from outside contributors. There are several distinct spam and misdirected classes:

• "Add my server to the README / ADDITIONAL.md" PRs. readme-pr-check.yml already labels and redirects README-only PRs.
• New server implementations, which go to the Registry.
• Renames and no-op PRs (e.g. "Rename README.md to README.md").
• Duplicate fixes: several outside PRs often race for the same bug (e.g. #4809 and #4810).

issue-triage (S7) needs a class and a canned response for each, plus server-scope labelling. The Inspector has neither, because it has no public PR inflow. Outside PRs are being turned off (S8, §10), so the triage recipe's PR half is short: close an outside PR with a pointer to the issue flow, after harvesting anything worth doing into an issue.

3.5 Security advisories for servers with real reach

filesystem, git and fetch read and write the local disk, run git, and make outbound requests. Advisories are real and frequent here (§8). The security-advisory skill (S9) has to cover:

• Ownership: is it this repo's server, or the SDK underneath (route it to the SDK repo)?
• Reach classes: path traversal, symlink escape and Roots bypass (filesystem, git); SSRF and robots bypass (fetch).
• The reference-implementation caveat. SECURITY.md currently tells reporters this repo is ineligible, which contradicts the enabled private reporting and the 61-advisory backlog.

4. Retiring CLAUDE.md

CLAUDE.md will be deleted, with no pointer file left behind. Claude Code reads AGENTS.md directly. S1 verifies this in a fresh session before deleting. Every section goes somewhere:

CLAUDE.md section	Destination	Sub-issue
Project Overview	AGENTS.md intro (one paragraph: what this repo is, the rules-vs-skills split)	S1
Monorepo Structure (7 servers, package names, registries)	AGENTS.md Project Structure: annotated src/ tree with package name + registry per server. The fuller per-server map goes to the project-structure skill	S1, S11
Build & Test Commands (TS)	AGENTS.md Development setup (the short form). The rest goes to the local-dev skill. validate / local:gate rules are added when S3 / S10 land	S1, S3, S10, S11
Build & Test Commands (Python)	Same as TS: uv sync --frozen --all-extras --dev, uv run pytest / pyright / ruff check .. Hatchling / uv build go to local-dev	S1, S4, S11
Code Style: TypeScript	AGENTS.md TypeScript instructions: the Inspector's rules plus our server idioms (ESM .js suffixes, Zod input schemas, naming, verb-first kebab-case tool names, import grouping). 2-space / trailing commas become Prettier config and drop out of prose	S1, S3
Code Style: Python	AGENTS.md Python instructions: pyright-clean type hints, ruff, async/await + pytest-asyncio, per-server module layout	S1
Contributing Guidelines (accepted / selective / not accepted)	AGENTS.md Contributing, linking CONTRIBUTING.md rather than duplicating it. Rewritten for the issues-only policy (§10)	S1, S8
CI/CD Pipeline (dynamic package detection, test → build → publish)	Dropped from AGENTS.md as derivable: the workflows describe themselves. The CI-vs-local split goes to docs/quality-gate.md; release goes to RELEASING.md + the release skill. (The "publish on release events" line is already stale: release.yml is dispatch-only, #4466)	S10, S12
MCP Protocol Reference (.mcp.json docs server, schema repo)	AGENTS.md: a two-line rule to look protocol questions up via the mcp-docs server, with a link to the schema repo	S1
Key Patterns: registerTools/registerResources/registerPrompts	AGENTS.md TS instructions (the rule). Where each server keeps them goes to project-structure	S1, S11
Key Patterns: tool annotations	AGENTS.md (rule: set readOnlyHint, idempotentHint, destructiveHint on every tool)	S1
Key Patterns: transports	AGENTS.md: stdio default, Streamable HTTP; HTTP+SSE is deprecated (still deprecated, not removed, in the 2026-07-28 spec, #4857)	S1
Key Patterns: PR template checklist	AGENTS.md Contributing (MCP docs read, security practice, tested with an LLM client). The evidence step goes to pr-flow	S1, S6

src/everything/AGENTS.md (a per-server guide) also exists. S1 decides its fate: move its generic style rules into the root file, keep its extension-point guidance next to the server, and check whether a nested AGENTS.md is picked up at all.

5. Decisions carried over from #4473

#4473 is closed and superseded by #4859. Each decision it recorded is mapped here.

#4473 decision	Where it goes	Note
Reference is the Inspector's v2/main AGENTS.md, not main	S1	—
Tool-agnostic rules, one document	S1, S2	S2's eval runs against both Claude and Copilot (AGENT=copilot)
Project Structure: annotated src/ tree with package name + registry	S1	—
Development setup / build & test commands	S1, S11	—
Repository & board: repo, base branch, single board #43	S1	Base branch is v2/main, not main (#4473 predates the v2/main flow)
gh recipes + stable-ID table	S5 (board-ops), not AGENTS.md	The Inspector keeps IDs in exactly one place, the skill, and resolves option IDs by name at run time, because option IDs change whenever the option list is edited. The #4473 table is also incomplete: it lacks Incoming (9f267269) and Priority
Issue-driven work style (created = labelled + boarded + Status; issues only; no drafts; dedupe; assign; status flow; Closes #N first line; new work → new issues)	S1 (rules), S5, S6 (recipes)	On v2/main, Closes #N does not auto-close. Close by hand, move to Done, and link with addCloseIssueReferences
Maintenance rules (READMEs, per-server READMEs, RELEASING.md, AGENTS.md; link, don't duplicate)	S1	—
Always test new or modified code	S1 (baseline), S10 (the per-file 90 rule)	#4473's "no 90% gate on day one" is superseded by #4854/#4855
Responding to code reviews (verbatim etiquette)	S1, S6	—
Root Prettier + ESLint flat config; root validate = format:check → lint → build → test; format before commit, validate before push; CI runs validate	S3	validate becomes the inner loop, and the push rule later becomes local:gate (S10)
Python equivalent documented per server	S4	Upgraded from "documented" to "run in CI" (ruff isn't run in CI today)
TypeScript instructions + server idioms	S1	no-floating-promises at error lands with S3
Python instructions	S1	—
Contribution boundaries + PR checklist	S1, S8	—
Omit React/Mantine and web-auth-token sections	—	Confirmed N/A (§2.1)

6. Reusable templates

Inspector files that can be copied in as starting points (paths on its v2/main), and the edits each needs.

Inspector file	Copy to	Edits needed	Sub-issue
AGENTS.md	AGENTS.md	Keep: header, Skills index, Maintenance rules, Maintaining the skills, Issue-driven Work Style, Responding to Code Reviews, Waiting on long-running work, Build output is never a gate target, Lint has no warning tier, TypeScript instructions. Rewrite: Project Structure, Development setup, Project Status (drop v1/main), Contributing. Drop: Dependency placement (keep the overrides rule), web layout, React, auth token, SDK-watch internals (they belong in the workflow's own comments). Add: Python instructions, MCP server idioms, protocol lookup	S1
.claude/skills/board-ops/SKILL.md	same	Board #43 only; Status/Priority; drop #11 and the dual-Priority-field section; keep the option-deletion hazard and recovery verbatim	S5
.claude/skills/issue-create/SKILL.md	same	--repo modelcontextprotocol/servers; no v1 rows; add a server-scope label step; chore type	S5
.claude/skills/issue-triage/SKILL.md	same	One board; the rubric's severity axis reworded for servers ("reports something false about the protocol", "escapes an allowed root"); add spam/registry/duplicate-PR classes; audit checks for one board. Update the total-issue-count --limit (this repo has far more issues than 884)	S7
.claude/skills/pr-flow/SKILL.md	same	Repo, board 43, branch naming; keep DCO (adopted, §10); drop screenshots; add client evidence; the Copilot loop copies as is	S6
.claude/skills/pre-push-gate/SKILL.md	same	Rewrite the stage list for our gate; keep "verify by exit code, not by grepping" and "waiting on the lease"	S10
.claude/skills/release/SKILL.md	same	Two registries; per-package versions; changesets / CalVer; the release environment approvals; keep the shape of preparation PRs then one pure merge PR (here: audit, TS "Version Packages" and Python CalVer PRs, then the merge; §9 S12), "bump on v2/main first", "never back-merge main" and the ledger	S12
.claude/skills/security-advisory/SKILL.md	same	One line (no v1 path); server reach classes; SDK routing	S9
.claude/skills/*/evals/evals.json	same	Rewrite the prompts in our terms; keep ≥ 5 positives + negatives per model-invoked skill	S2 and each skill PR
scripts/verify-skills.mjs, scripts/verify-skills-cli.mjs, scripts/skill-eval.mjs, scripts/lib/skill-manifest.mjs, scripts/lib/claude-cli.mjs, and the helpers they import, scripts/lib/npm-scripts.mjs and scripts/lib/win-shell-args.mjs (+ all their *.test.mjs)	scripts/	Paths and skill list; a budget recomputed for our skill set. S3 and S10 reuse the two helpers	S2
scripts/gate-lease.mjs (+ test)	scripts/	Env var rename (SERVERS_SKIP_GATE_LEASE). Imports lib/win-shell-args.mjs, which S2 already ports	S10
scripts/lib/workflow-gate.mjs (+ test)	scripts/lib/	Our workflow list and local:* scripts; drop the browser-engine rationale and its lib/headless-browser.mjs import	S10
scripts/verify-format-coverage.mjs, scripts/verify-typecheck-coverage.mjs, and scripts/lib/tsc-program.mjs (+ tests), which the latter imports	scripts/	Workspace globs instead of clients/*. Both also import S2's lib/npm-scripts.mjs	S3
scripts/verify-action-pins.mjs (+ verify-action-pins.test.mjs, covering credential detection, transitive artifact producers and fail-closed workflow parsing)	scripts/	Workflow list. It imports SHA_REF from dependency-refresh.mjs, which doesn't land until S13, so extract the SHA matcher into scripts/lib/action-refs.mjs in S12; S13's dependency-refresh then imports it from there	S12
scripts/dependency-refresh.mjs, scripts/dependabot-alerts.mjs, scripts/sdk-watch.mjs + workflows	scripts/, .github/workflows/	Add the uv/PyPI ecosystem; SDK groups for TS and Python; board #43; labels	S13
docs/skill-authoring.md	docs/	Paths only	S2
docs/quality-gate.md	docs/	Rewrite for two languages; keep the structure (tiers table, local-only steps, lease)	S10
.github/ISSUE_TEMPLATE/*	same	Server dropdown, spec-era/client fields, registry redirect in config.yml	S8
.github/pull_request_template.md	same	Depends on the contribution-model decision	S8
.github/workflows/main.yml (coverage job, package/publish split, timeout-minutes)	typescript.yml, python.yml, release.yml	Patterns only, not the file	S10, S12

7. Existing issues to reconcile

Issue	Decision
#4472: release Phase 2, changesets (TS) + GitHub-Release-triggered publishing	Fold in as a sub-issue of #4858, unchanged in scope, in Wave 5. It is the versioning and publish half of the release flow; the milestone-merge half is new (S12) and depends on it. Two notes to add to #4472: it lands on v2/main like everything else, and its release: [published] trigger must fire from main after a milestone merge.
#4854 / #4855: per-file 90% coverage, TS / Python	Stay under #4857 (they're the refactor's regression net). The factory depends on them and doesn't duplicate them: S10 wires their coverage commands into local:gate and CI and writes the AGENTS.md coverage rule (their carry-over task). One correction to feed back: both said the coverage gate stays local "following the Inspector", but the Inspector's CI now runs coverage as a parallel job (#2159). Decided (§10): CI enforces coverage here too, and #4854/#4855 are amended to match. Amended when S10 was built: the wiring moved to #4854/#4855 themselves, so S10 does not wait on them; see the note under S10 in §9.
#4857: 2026-07-28 spec refactor tracker	Unchanged. Its "verify against both eras with the Inspector and an LLM client" rule becomes the client-smoke skill (S11) and the pr-flow evidence step (S6).
#4860: interface-diff CI for everything	Unchanged. S10 includes it in the gate once it lands.
#4473: AGENTS.md plan	Closed, superseded by #4859. Every decision is mapped in §5.

8. Findings along the way

Facts discovered while writing this doc. Each is owned by a sub-issue.

1. Dependabot security-fix PRs are enabled (automated-security-fixes: enabled), and .github/dependabot.yml opens weekly Actions PRs. Both are issue-less PRs, the exception the Inspector removed. → S13.
2. Private vulnerability reporting is on, with 61 advisories in triage (6 published, 2 closed). Meanwhile SECURITY.md tells reporters the repo is "not eligible for security vulnerability reporting". → S9.
3. ruff is a dev dependency of every Python server but is not run in CI. → S4.
4. No chore label. The five-type taxonomy needs it. → S5.
5. No issue forms. Blank issues are the only path. → S8.
6. No DCO app is installed, so the Inspector's signoff rule has nothing to enforce it. DCO is adopted (§10). → S6.
7. No CI job declares timeout-minutes. → S10.
8. release.yml builds, installs and publishes in the job holding id-token: write. The Inspector split these after #2483. → S12.
9. CLAUDE.md's CI/CD section describes publish-on-release, which is gone. release.yml is dispatch-only since #4466. → S1 (dropped section).

9. Proposed sub-issues

The sub-issues now exist under #4858. The S-ids used throughout this doc map to them as follows. #4472 sits in Wave 5, between S11 and S12, as Part 13.

Id	Issue	Title prefix
S1	#4862	Part 2
S2	#4863	Part 3
S3	#4864	Part 4
S4	#4865	Part 5
S5	#4866	Part 6
S6	#4867	Part 7
S7	#4868	Part 8
S8	#4869	Part 9
S9	#4870	Part 10
S10	#4871	Part 11
S11	#4872	Part 12
S12	#4873	Part 14
S13	#4874	Part 15

Each targets v2/main, carries the v2 label and a milestone, and sits on the Servers V2 board (#43). "After" means the listed issue must merge first.

W1  #4859 inception (this doc)
W2  S1 AGENTS.md · S2 skills harness · S4 Py validate · then S3 TS validate (after S2)
W3  S5 board-ops + issue-create · S8 contribution model · then S6 pr-flow (after S5), S9 security-advisory (after S5) and S7 issue-triage (after S5, S8)
W4  S10 local:gate + pre-push-gate (after S2, S3, S4; coverage moved to #4854, #4855) · S11 knowledge skills (after S2)
W5  #4472 changesets + Release-triggered publish → S12 milestone release flow + release skill (after #4472, S5, S10, S11)
W6  S13 dependency & SDK sweeps replace Dependabot PRs; closing factory overview (after S5, S12)

Wave 2: rules and scaffolding (S1, S2 and S4 in parallel; S3 after S2)

S1 (#4862). AGENTS.md: the absolute rules; delete CLAUDE.md

• Scope: write AGENTS.md from the Inspector's template (§6), holding only rules that are true on the day it merges. A rule whose machinery doesn't exist yet (validate, local:gate, per-file coverage) is added by the sub-issue that builds it. Map every CLAUDE.md section per §4 and every #4473 decision per §5. Settle src/everything/AGENTS.md. Delete CLAUDE.md.
• Acceptance:
  • AGENTS.md exists at the root; CLAUDE.md is deleted and no pointer file remains.
  • Every row of §4 and §5 marked S1 is present.
  • A fresh Claude Code session in the repo demonstrably follows an AGENTS.md-only rule, recorded in the PR.
  • The skills index lists only skills that exist; later skill PRs add their own rows.

S2 (#4863). Skills infrastructure: .claude/skills/, verify:skills, skills:eval

• Scope: port verify-skills, verify-skills-cli, skill-eval and their libs and tests (§6); root npm scripts; docs/skill-authoring.md. Add the "Maintaining the skills" rules to AGENTS.md (or to S1, if S1 hasn't merged).
• Acceptance:
  • The ported verifier keeps its "no skills found" failure. Because S2 lands before any skill, it ships with an explicit, temporary bootstrap allowance for an empty .claude/skills/, and the first skill PR removes it: whichever of the skill-adding issues (S5, S6, S9, S10, S11) lands first. That removal is an acceptance criterion of each of them.
  • npm run verify:skills fails on a fixture with malformed frontmatter or a missing disable-model-invocation.
  • npm run skills:eval runs against Claude, and against Copilot with AGENT=copilot.
  • verify:skills runs in CI, in a new root-guards job that S2 creates (in typescript.yml, or a small repo-level workflow; this repo has no main.yml). S3 extends that job. verify:skills:cli runs there too, on every PR.
  • The verifier's wiring checks (checkWiring) are ported against this repo's real files, not the Inspector's main.yml. Two of the links they assert don't exist yet, so each gets an explicit, temporary allowance that a later issue removes: root validate reaching verify:format-coverage (S3 removes it) and local:gate reaching verify:skills:cli (S10 removes it).

S3 (#4864). TypeScript workspace gate: Prettier, ESLint, root validate, CI (after S2, whose root-guards job it extends and whose wiring allowance it removes)

• Scope: the #4473 design. Root Prettier config and format / format:check; a root ESLint flat config, type-aware, --max-warnings 0, no-floating-promises at error, build output ignored. A per-workspace validate script in each TS server (format:check → lint → build → test, for that package only). Today every TS workspace's test runs vitest run --coverage, so S3 splits it: test becomes the fast run, and a separate coverage script keeps the instrumented one (#4854 then adds its per-file thresholds to that script). A root validate that aggregates them (npm run validate --workspaces) plus any root-only guards. verify:format-coverage and verify:typecheck-coverage adapted. typescript.yml keeps its per-package matrix, and each leg runs only its own package's validate rather than the whole monorepo. Removes S2's temporary verify:format-coverage wiring allowance. The separate root-guards CI job (created by S2) also runs what no package leg covers: the root format and lint of root files, verify:format-coverage, verify:typecheck-coverage, test:scripts (the guard scripts' own unit tests), and the shared-devDependency version guard adapted from verify:dep-lockstep (§2.3). Add the format/lint/validate rules to AGENTS.md. If S1 hasn't merged yet, hand these rules to S1 instead, the same fallback as S2.
• Acceptance:
  • npm run validate passes on a clean checkout, and so does npm run validate -w <package> for each server.
  • Each CI matrix leg gates only its own package, and the root-guards job gates the rest.
  • CI fails a PR with a formatting or lint finding in a package or a root file, or with a divergent shared devDependency range.
  • everything's per-package Prettier setup is folded into the root one.

S4 (#4865). Python gate parity

• Scope: for each of fetch, git, time: ruff check, ruff format --check, pyright, pytest and uv build (CI already gates the build), with a single per-server validate entry (a uv run chain, or a scripts/ helper called from the root). A root npm run validate:py (or equivalent) runs all three. python.yml runs ruff. Add the Python rules to AGENTS.md. If S1 hasn't merged yet, hand them to S1 instead, the same fallback as S2.
• Acceptance:
  • One root command gates all Python servers.
  • CI fails on a ruff or format finding.
  • Existing findings are fixed, not suppressed through config.

Wave 3: work-tracking and security skills (after S1, S2)

Whichever skill-adding issue merges first (S5, S6 or S9 here, or S10 or S11 in Wave 4) also removes S2's empty-skills bootstrap allowance. That is part of each one's acceptance.

S5 (#4866). board-ops and issue-create skills; label taxonomy

• Scope: adapt both skills (§6). Create the chore label. Board #43's Size field is already deleted (§10), so the create flow sets only Status and Priority. Server-scope labels (server-<name>) are part of create where the issue concerns one server (repo-wide issues carry none). Board #43's IDs live only in board-ops, and option IDs are resolved by name.
• Acceptance:
  • Filing an issue through the skill yields labels (v2 + type, plus a scope label when one applies), milestone, card, Status and Priority, verified by a query in the PR.
  • Eval cases pass the threshold.
  • The skills index is updated.

S6 (#4867). pr-flow skill (after S5, whose board-ops it uses)

• Scope: adapt §6. Branch v2/<type>/<N>-<slug> from origin/v2/main; assign and move to In Progress; the gate; Closes #N on the first line; addCloseIssueReferences; In Review; the Copilot review loop to exhaustion; per-thread replies plus a PR summary; manual close and Done on merge. A client-evidence step replaces screenshots (§3.2). DCO is on (§10): install the DCO app, sign off every commit with git commit -s, and add the signoff rule to AGENTS.md once the app enforces it.
• Acceptance:
  • The DCO app is installed and fails a PR with an unsigned commit.
  • A PR taken end to end through the skill.
  • The loop's exits (clean round / out-of-scope only / two silent rounds / timeout) documented.
  • Eval cases pass the threshold.

S7 (#4868). issue-triage skill and board audit, for community inflow (after S5 and S8)

• Scope: adapt §6. Two-pass sweep (Incoming → approval), rubric with a posted score comment, the board audit. Add triage classes for server submissions, README/ADDITIONAL.md listing PRs, new-server implementations, duplicate racing fixes and no-op PRs, each with a canned response and close/label action. With outside PRs off (§10), every outside PR gets the same close with a pointer to the issue flow. Fold in readme-pr-check.yml's behavior.
• Acceptance:
  • A triage pass over the current open backlog runs, and the audit prints all zeros afterwards.
  • Each spam class has a documented response.
  • Eval cases pass the threshold.

S8 (#4869). Contribution model: outside PRs, CONTRIBUTING.md, templates, issue forms

• Scope: decided (§10): outside PRs are turned off, as in the Inspector. External contributors, org members with write access included, file detailed issues (sharing the prompt they used, not a diff), and maintainers open every PR. Make CONTRIBUTING.md, AGENTS.md's Contributing section, the PR template (the Inspector's "issues, not PRs" banner) and new issue forms (bug / feature, with a server dropdown and spec-era field; security → private advisory; new-server → Registry) say so. Decide how the open outside-PR backlog is handled (for example, close each with a pointer to the issue flow, filing an issue for any fix worth keeping); S7's triage pass carries it out.
• Acceptance:
  • The decision is recorded on the issue.
  • The docs and templates match it.
  • Forms are validated against GitHub's schema (they only go live after the next milestone merge to main).

S9 (#4870). security-advisory skill; reconcile SECURITY.md and the advisory backlog (after S5: the flow uses board-ops for the draft card and issue-create for public tracking)

• Scope: adapt §6: draft [GHSA-…] card, ownership check (this server vs the SDK), accept/reject, private fork, fix, publish, public tracking. Accept and publish stay human-only. Add the two narrow advisory exceptions to AGENTS.md's board rules, as the Inspector does: a [GHSA-…] draft card is the one allowed non-issue card, and accepting an advisory moves its (necessarily unmilestoned) draft from Incoming to Todo. Give the board audit the matching [GHSA- carve-out; whichever of S7 and S9 lands second adds it. Rewrite SECURITY.md so it matches the enabled private reporting. Plan how the 61-advisory triage backlog is worked (the plan only, not the triage itself).
• Acceptance:
  • The skill merged with eval cases.
  • SECURITY.md is consistent with repo settings.
  • A backlog-triage issue is filed.

Wave 4: quality gate and knowledge skills

S10 (#4871). local:gate, per-file coverage in CI, pre-push-gate skill (after S2, S3, S4, #4854, #4855)

Amended when S10 was built (#4871). The plan below put the whole gate behind #4854 and #4855, for the sake of one piece: per-file coverage cannot be enforced until their test suites exist. Nothing else in S10 depends on them, so that piece moved out. S10 shipped local:gate, the lease, the guards, the boot smoke, docs/quality-gate.md, the pre-push-gate skill and the pre-push rule, with no coverage stage and no coverage rule. #4854 (TypeScript) and #4855 (Python) each add their thresholds, their parallel CI coverage job, their stage in local:gate, and their half of the AGENTS.md coverage rule. The scope and acceptance lines below are the original plan, kept for the record; where they mention coverage, read #4854/#4855.

• Scope: root local:gate (under gate-lease) chaining the TS and Python validate, verify:skills:cli, per-file coverage for both languages, a thin per-server boot smoke over each transport the server implements (stdio for all seven; SSE and Streamable HTTP for everything), and #4860's interface diff. #4860 is independent, so whichever of S10 and #4860 lands second wires the interface diff into local:gate, keeping "every check CI runs" true. Removes S2's temporary verify:skills:cli wiring allowance. verify:install-fresh (root workspace install only) runs first, so the gate never tests a stale node_modules. workflow-gate ported, so no workflow can invoke a local:* script. CI runs coverage as a parallel job (§7). timeout-minutes on every job. No test retries (asserted). docs/quality-gate.md. The pre-push-gate skill. The AGENTS.md rules: mandatory pre-push gate, and the per-file ≥ 90 coverage rule with justified ignores (the carry-over from #4854/#4855). For TypeScript that is all four Vitest dimensions (lines, statements, functions, branches). For Python it is the per-file metrics coverage.py measures, lines and branches: coverage.py has no native function dimension, so the rule doesn't invent one.
• Acceptance:
  • npm run local:gate runs every check CI runs.
  • A PR dropping any file below 90 fails CI.
  • Concurrent gates queue.
  • Skill eval cases pass.

S11 (#4872). Knowledge skills: project-structure, local-dev, testing, client-smoke

• Scope: adapt §6. testing documents the in-process harnesses from #4854/#4855. client-smoke drives a server with Inspector V2 and an LLM client in both spec eras (#4857). Driving the Inspector's web client needs browser automation, so this issue also enables the Playwright plugin in .claude/settings.json, or documents the Inspector CLI as the scripted path and the web client as the hand-driven one.
• Acceptance:
  • Four skills merged with eval cases.
  • A full skills:eval re-run shows no regression in the skills that already exist.

Wave 5: release

#4472. changesets (TS) + GitHub-Release-triggered publishing: folded in with its scope unchanged (§7). Its two bump PRs are why S12's preparation step is more than one PR.

S12 (#4873). v2/main → main milestone release flow and release skill (after #4472, S5, S10, S11; the release issue is filed through S5's issue-create, and the ledger uses S11's client-smoke)

• Scope:
  • A release issue per milestone (Release vX.Y.Z), filed through issue-create. Every preparation PR (on v2/main, where closing keywords don't fire) opens with Closes #N. The merge PR targets main, the default branch, where Closes #N would auto-close the issue before the Release is published, so it uses a non-closing reference (Part of #N) instead. The issue closes by hand once the Release is published. This is what keeps release PRs inside "every PR references an issue" after S13 removes the Dependabot exception. The one bot-authored PR, changesets' "Version Packages", gets the reference added to its body after the bot opens it. If the action overwrites the body on each update, AGENTS.md names it as the one standing release-automation exception.
  • The preparation PRs, all on v2/main: the audit report (npm and uv/pip) with any fixes it forces, plus the bumps. #4472 makes the bumps two separate PRs, the changesets "Version Packages" PR for TS and the prepare-release CalVer PR for Python, so a milestone touching both ecosystems has up to three preparation PRs rather than the Inspector's one. All of them merge before the merge PR opens.
  • The merge PR: a pure v2/main → main merge whose tree hash matches origin/v2/main, with a release ledger artifact: local:gate, per-package pack:verify, each server-facing milestone issue exercised via client-smoke, and a targeted probe for each issue with no client surface (§3.2).
  • The maintainer publishes the GitHub Release.
  • Split release.yml so build and verify run without id-token.
  • Pin actions in every credentialed job before the first release through this flow, with verify:action-pins enforcing it: release.yml's publish jobs, every job whose artifact a credentialed job downloads (the build/pack jobs the split introduces), and claude.yml (id-token: write, ANTHROPIC_API_KEY). Wire verify:action-pins into the continuously run chain: root validate's guards, S2's root-guards CI job, and local:gate. Extract the SHA matcher into scripts/lib/action-refs.mjs (§6), so S12 doesn't depend on S13. Add the AGENTS.md SHA-pinning rule.
  • The release skill (name-only).
  • RELEASING.md rewritten for the merged state.
• Acceptance:
  • verify:action-pins passes, and fails on a tag-pinned action in a credentialed job or in a job whose artifact a credentialed job downloads. Unrelated artifact uploads (e.g. python.yml's CI dist upload) stay out of scope, as in the Inspector's guard.
  • One milestone released end to end through the skill.
  • The ledger is linked from the merge PR.

Wave 6: automation

S13 (#4874). Replace Dependabot PRs with issue-filing sweeps; SDK watch; the factory overview (after S5, and S12, whose scripts/lib/action-refs.mjs its dependency-refresh imports)

• Scope:
  • Turn off automated security-fix PRs (a repo setting) and delete dependabot.yml; keep alerts on.
  • Add dependency-refresh (monthly; npm workspaces, each uv.lock, Actions) and dependabot-alerts (daily; npm + pip, re-checked against v2/main).
  • Add sdk-watch (nightly; TS SDK packages and Python mcp), including the hardened analysis job's properties unchanged.
  • Keep S12's SHA pins current: dependency-refresh ranks each pin by its # vX.Y.Z comment, as the Inspector's sweep does.
  • Add the AGENTS.md "dependency updates are issue-driven" rules, and remove S1's temporary Dependabot exception to "every PR references an issue".
  • Work down the existing Dependabot PR backlog (six open at the time of writing): convert each still-needed bump into an issue for the sweep flow, and close the PR with a pointer to it.
• Acceptance:
  • No new Dependabot PRs open after merge.
  • Every Dependabot PR that was open at merge time is closed, with a pointer to its replacement issue or a reason.
  • Each sweep's dry run writes nothing and prints the issue payload it would file, with correct labels and milestone. Live filing is exercised by script tests with a mocked gh, not against the real tracker.
  • Script tests pass.
  • A closing docs/ai-software-factory.md for this repo, written once the whole factory, this automation included, has landed.

10. Maintainer decisions

These were open questions in the first draft of this doc; the maintainers answered them on PR #4861.

1. Outside PRs (S8): turned off, as in the Inspector. External contributors file issues; maintainers open PRs. S8 rewrites the policy docs and templates and plans the open backlog; S7 carries it out.
2. DCO (S6): on, as in the Inspector. S6 installs the DCO app and requires git commit -s on every commit.
3. Coverage in CI (S10): yes. CI enforces the per-file gate in a parallel job, as the Inspector does. #4854 and #4855 are amended where they said the gate stays local.
4. Size field (S5): removed. Board #43's Size field held no values and has been deleted.
5. Milestones: kept as is. Every sub-issue stays in v2.0.0.