feat(seo): add agent collaboration patterns guide

[lilyshen0722]

Summary

• publish the static, light-shell guide at /guides/ai-agent-collaboration-patterns/
• add sitemap/canonical Article metadata, hub entry, and four approved reciprocal links
• encode the approved lead, reviewer, handoff, child-lane, and research-race patterns without enforcement claims

Verification

• node --test scripts/generate-seo-pages.test.mjs
• npx jest --runInBand src/v2/__tests__/V2Login.test.tsx
• npm run typecheck
• npm run build
• npm test -- --watch=false (86 suites, 487 tests)

[lilyshen0722]

Gated at ece56007. PASS. The one hard technical claim in this guide is implemented — I checked it rather than trusting the hedge — and there is a refinement worth making before it goes out.

Verified true: "when the same source reference already has a task, the documented API returns the existing task rather than creating another one."

backend/routes/tasksApi.ts:255 does exactly this: on create, if sourceRef is present it runs Task.findOne({ podId, sourceRef }) and returns { task: existing, alreadyExists: true } without creating. It is backed at the storage layer too — models/Task.ts:94 declares a unique partial index podId_1_sourceRef_1_partial over { sourceRef: { $type: 'string' } }, and the create path catches the 11000 collision as a fallback. So this is a genuine three-layer guarantee, not a documented intention. The guide's hedge "the documented API" is more cautious than it needs to be.

The refinement: the dedup path is not side-effect-free, and this guide is exactly the place that should say so.

If the existing task's status is done, the same call does not simply return it — it reopens it (tasksApi.ts:258–269): status back to pending, assignee overwritten with whatever the repeat call passed, claimedAt and claimExpiresAt cleared, notes replaced with "Reopened — the same source is active again", an update appended, and emitTaskUpdated + notifyAgents fired. The response is { reopened: true }.

So repeating a create against a completed source-ref clears a claim and wakes the pod. That is deliberate and reasonable behaviour, but it means "safer to repeat" holds for pending/claimed tasks and has a real consequence for done ones. Given this guide's whole thesis is don't mistake coordination for enforcement, and it already carefully hedges "not a blanket guarantee that an agent's external actions are idempotent", one clause on the reopen case would be consistent with its own standard — something like "unless the task is already done, in which case the same call reopens it."

Not blocking; it makes a true statement more precise rather than correcting a false one.

Also verified: "A claim is a coordination signal … It does not lock files, branches, deployments, or external systems." Correct and correctly scoped — ADR-018 claims are an advisory CAS lease over messages and tasks, with no filesystem or deployment reach.

Mechanical checks: rendered all 26 pages through renderStaticPage in a scratch tree — 0 hits for [object Object] or stray undefined; node --test scripts/generate-seo-pages.test.mjs 2/2 on Node 22; base is current main (0 behind); pages.length 25→26 and guidePages.length 15→16 are anchored; the four reciprocal-link assertions loop over the linking guides; assert.doesNotMatch(/cm_agent_[A-Za-z0-9]{8,}/) carries the token-leak guard forward. Test & Coverage pending, other 10 green.

Unlike #1388, this guide names no API identifiers at all — I swept the added content for commonly_* tools, /api/* routes, dotted event types and SCREAMING_CASE env vars and found none. That is why the defect class I filed on #1388 cannot recur here.