developer-relations

The **first developer-relations specialization** in the library. Before it, the only devrel artifact anywhere was a single stub agent (`../sdk-platform-development/agents/devrel/AGENT.md`, marked *implementation pending*) — referenced here as the seed, left in place, never edited. This specialization is additive: its personas are new devrel-* agents defined inline in the flagship process. It carries one product change end-to-end — from API-change intake through executed sample apps, adversarially verified content, policy-gated external publishing, and community engagement — with routed human approval on every action that leaves the org boundary.

Flagship process: devrel-campaign

devrel-campaign.js (@process developer-relations/devrel-campaign) walks one product change through the whole campaign. Every subtask is kind:'agent' (no shell subtasks — repo override); combinators are reused from `../common-utilities/routed-gate-combinators.js`.

**Inputs:** { productChange: {id, title, summary, changedApis[] (non-empty), docsRefs[], sampleAppSpecs?, sourceMaterials?, embargoUntil?} (required), repoRoot?='.', apiSurfaceRef?='docs/', contentChannels?, communityThreadsDir?='artifacts/community/inbox', sampleRepoTarget?, maxFixAttempts?=2, kipEnabled?=true, kipDir?='.a5c/kip', kipModel?='sonnet' }

**Outputs:** { success, intake, sampleApps, sampleAppGate, content, contentGate, gatedActions, publishes, communityTriage, retro, kipFactsAsserted, artifacts, metadata } — success = sampleAppGate.passed && contentGate.passed && every gatedActions record satisfies (!required || approved === executed). A failed sample-app **or** content gate returns success:false before any publish/approval, with all three gated actions recorded as skipped.

Executed-evidence quality gates

Both accuracy gates run through the adversarialGate combinator and demand **executed evidence** — read-only review is a protocol failure:

sample against the current surface, captures its own stdout/exit) + api-currency-critic (re-extracts every API call and matches it against apiSurfaceRef; any deprecated/removed/ hallucinated call fails). Fixer: built-in gateFixerTask edits the offending sample.

every snippet, captures output) + claim-cross-check-critic (builds a claim-trace table; any claim with no evidence source or contradicting the sample run output fails). Fixer: built-in gateFixerTask edits the offending piece.

• **developer-relations.sample-app-accuracy** — sample-execution-critic (RE-RUNS every
• **developer-relations.content-accuracy** — snippet-execution-critic (extracts and runs

Critic independence is enforced by the combinator (fresh parallel instances; producers never review their own work). passed:true with an empty evidence array is coerced to a protocol failure. Each critic prompt pins the exact verdict JSON shape { passed, issues[], evidence[] } at its end (prior insight: 3/12 critics drifted shape). On fix-budget exhaustion the combinator raises developer-relations.sample-app-accuracy.gate-escalation / developer-relations.content-accuracy.gate-escalation (expert owner), pushed to metadata.breakpointsHit.

Policy-gated actions

All approvals go through routedBreakpoint; for the three policy-gated actions the breakpointId **equals** the actionId and tags are ['policy-gated','developer-relations'], strategy 'single'. Fail-closed: the executor task runs **only** on approved === true — a rejection is honored, recorded, and never worked around.

outputs.gatedActions records **every** decision — { actionId, required, approved, autoApproved, response, executed } per action, including non-interactive auto-approvals (recorded raw from the BreakpointResult) and skipped gates ({ required:false, approved:false, autoApproved:false, executed:false } — never omitted). metadata.breakpointsHit logs every raised breakpointId in order.

Composition by name

The content pieces reuse existing library point tasks rather than re-implementing them — dvr.produce-content composes, **by name**:

• `technical-documentation/interactive-tutorials` — the tutorial piece (executable walkthrough).
• `technical-documentation/how-to-guides` — the blog/how-to angle.
• `technical-documentation/api-reference-docs` — the reference sections.
• `communication/content-writer` — the prose-drafting persona for each piece.

The gated-send skeleton follows `communication/multi-audience-announcement-pipeline.js`: the recordGatedAction audit shape, routedBreakpoint(breakpointId = actionId, policy-gated tag), fail-closed executors, a single shared external approval covering multiple outputs, the executed-evidence gate, and the fail-closed early return on a failed gate.

Parallel tracks

• ctx.parallel.all over intake.sampleAppSpecs — concurrent docs-driven sample-app builds (Phase 2).
• adversarialGate fans its critics out concurrently (Phases 3 and 5).
• ctx.parallel.all over CONTENT_PIECES — concurrent content production (Phase 4).
• ctx.parallel.all over CONTENT_PIECES — concurrent publish executors under the single external-content-publish approval (Phase 7).

kip integration

kipRecall at intake (topic: prior devrel content on these APIs, recurring community questions/FAQs, past sample-app patterns for productChange.title; kind developer-relations) and kipAssert at close — content published-to channel with approval provenance, sample-app accuracy results with runOutputPath, the content-gate outcome, new FAQ candidates from triage, and community-thread outcomes with community-reply-send provenance. Gated on kipEnabled; an empty store is a fresh brain, never an error; assert failures are reported by the librarian task, never swallowed. See `../shared/skills/kip-librarian`.

No fallbacks

• Missing productChange / id / title / non-empty changedApis throws at Phase 0.
• Unknown keys in contentChannels throw against CONTENT_PIECES (never silently ignored).
• A derived sampleAppSpecs entry without a runnable build+run command throws before any build.
• expertFor / actionFor / channelFor are throwing lookups — no silent defaults.
• A failed sample-app or content gate ends the run with success:false before any external action; rejected gates are honored, never worked around.

Usage

babysitter run:create \
  --process library/specializations/developer-relations/devrel-campaign.js \
  --inputs '{
    "productChange": {
      "id": "PC-3.4-webhooks",
      "title": "Signed webhook deliveries",
      "summary": "v3.4 adds HMAC signatures + a retry endpoint to the webhooks API.",
      "changedApis": ["webhooks.createEndpoint", "webhooks.verifySignature", "webhooks.retry"],
      "docsRefs": ["docs/webhooks.md"],
      "embargoUntil": "2026-08-01T09:00:00Z"
    },
    "apiSurfaceRef": "docs/",
    "contentChannels": { "blog": "devblog", "tutorial": "docs-site", "video-script": "youtube" },
    "sampleRepoTarget": "org/webhooks-samples",
    "communityThreadsDir": "artifacts/community/inbox",
    "maxFixAttempts": 2
  }'

Files

• `devrel-campaign.js` — the flagship process (8 dvr.* Style-A agent tasks + orchestration).
• Combinators: `../common-utilities/routed-gate-combinators.js` — routedBreakpoint, adversarialGate, kipRecall, kipAssert, gateFixerTask.
• Seed: `../sdk-platform-development/agents/devrel/AGENT.md` — the pre-existing devrel agent stub (referenced, not modified).