Widen babysitter intake to routed pull requests

.agentworkforce/agents/factory-feature-guardian/agent.test.ts

@@ -683,7 +683,7 @@ describe('factory-feature-guardian runtime paths', () => {
 
     const catalog = await loadFactoryGuardianCatalog(production.ctx);
 
-    expect(catalog.features).toHaveLength(310);
+    expect(catalog.features).toHaveLength(314);
     expect(catalog.features.flatMap((feature) => feature.locations)).toEqual(
       expect.arrayContaining(['src/environments/', 'src/triage/', 'src/fleet/', 'src/writeback/'])
     );

.agentworkforce/features/manifest.yaml

@@ -1,14 +1,14 @@
 version: '1.1'
-updated: '2026-08-05'
+updated: '2026-08-07'
 catalog:
   category_count: 25
-  feature_count: 310
+  feature_count: 314
   tier_counts:
     1: 48
-    2: 124
+    2: 125
     3: 11
-    4: 52
-    5: 60
+    4: 53
+    5: 62
     6: 15
 
 # Every user-facing feature in @agent-relay/factory, categorized and scored.
@@ -821,6 +821,27 @@ categories:
         location: src/orchestrator/factory.ts, src/triage/heuristic.ts, src/dispatch/templates.ts
         verify_tier: 6
 
+      - id: pr-routed-babysitter-discovery
+        name: Routed Repository PR Babysitter Discovery
+        api: babysitter.mode=routed-open-prs
+        description: Discover open non-draft same-repository PRs only from repos.names, fall back to PR title and body, and account for every excluded, deferred, unreadable, or failed candidate
+        location: src/github/routed-pr-babysitter.ts, src/orchestrator/factory.ts
+        verify_tier: 5
+
+      - id: pr-routed-babysitter-ownership
+        name: Routed PR Work-Unit Ownership
+        api: StateStore.claimRoutedPrBabysitter()
+        description: Atomically claim exact normalized repo and PR work units across routed and issue-created intake before spawn, with durable restart recovery and bounded admission
+        location: src/ports/state.ts, src/state/in-memory-state-store.ts, src/state/file-state-store.ts, src/orchestrator/factory.ts
+        verify_tier: 4
+
+      - id: pr-routed-babysitter-opt-out
+        name: Routed PR Pre-Write Opt-Out
+        api: babysitter.excludeLabels / babysitter.excludePullRequests
+        description: Reject configured PR identities or labeled PRs before spawn and require automatic workers to re-check labels before their first provider write
+        location: src/config/schema.ts, src/github/routed-pr-babysitter.ts, src/dispatch/templates.ts
+        verify_tier: 5
+
       - id: pr-standalone-babysitter-validation
         name: Standalone Babysitter PR Validation
         api: readStandalonePullRequest()
@@ -2142,6 +2163,13 @@ categories:
         location: src/config/schema.ts, src/orchestrator/factory.ts
         verify_tier: 2
 
+      - id: config-babysitter-routed-intake
+        name: Routed Babysitter Intake Controls
+        api: factory.config.json#babysitter.mode
+        description: Keep factory-created intake as the default while explicitly configuring routed discovery, exclusions, and notifications-off behavior
+        location: src/config/schema.ts, src/orchestrator/factory.ts
+        verify_tier: 2
+
       - id: config-merge-policy
         name: Merge Policy
         api: factory.config.json#mergePolicy

.agentworkforce/features/verify/procedures.md

@@ -351,6 +351,19 @@ Verify the PR contains the issue key, commits, tests, and review activity; no dr
 
 Enable `babysitter.enabled`, open a non-draft PR through the issue-driven path, and confirm only one babysitter starts. Make CI fail or add a review request, confirm only the exact repo/PR owner wakes, then enter the documented destructive critical section and prove Factory ACKs only after the no-submit fence persists. Confirm coalesced activity is delivered once after exit, the babysitter fixes the current PR branch, then send/observe its readiness signal. A mismatched signal or draft/closed PR must not advance the issue.
 
+For widened intake, additionally set `babysitter.mode` to
+`routed-open-prs` in a disposable workspace. Provide one eligible PR, one PR
+with `factory:skip-babysitter`, one configured `excludePullRequests` identity,
+one draft, and one same-number PR in a repository absent from `repos.names`.
+Run a dry/read-only sweep and verify the exact discovery counters account for
+every candidate without spawning or writing. Then use a fake fleet/state test
+to prove the eligible PR is atomically claimed before spawn, two owners cannot
+claim it, the opt-outs spawn nothing, only one new routed PR is admitted per
+sweep, capacity deferrals are logged, restart restores the exact repo/PR owner,
+and the automatic prompt disables notifications and checks labels before its
+first provider write. Do not run a provider-writing widened sweep until the
+GitHub App identity path is independently validated.
+
 Do not attempt a cross-host active/active control-plane test: the supported ownership topology is multiple Factory processes sharing one same-host `FileStateStore`. Remote relay execution nodes are supported and do not need that directory.
 
 ### Standalone babysitter

README.md

@@ -303,6 +303,46 @@ existing PR branch, and always leaves the final review and merge to a human.
 The command prints a spawn receipt and returns; the PR-keyed task-exit worker
 continues on the relay broker and reports completion or access blockers there.
 
+### Automatic routed-PR babysitting
+
+`babysitter.enabled` continues to cover Factory-created PRs by default. To
+admit existing PRs independently of issue dispatch, set:
+
+```json
+{
+  "babysitter": {
+    "enabled": true,
+    "mode": "routed-open-prs",
+    "excludeLabels": ["factory:skip-babysitter"],
+    "excludePullRequests": [],
+    "notifyHumans": false
+  }
+}
+```
+
+The widened mode scans only repositories named by `repos.names`; other routing
+fallbacks do not silently expand the sweep. It admits open, non-draft,
+same-repository PRs and uses their title/body as the standalone definition of
+done when no Factory issue lifecycle exists. Such PRs never gain an issue state
+transition or automatic merge. The exact normalized `owner/repo#number` work
+unit is claimed durably before spawn, so issue-driven and routed discovery
+cannot start two babysitters for one PR.
+
+Opt-out labels and `excludePullRequests` are checked before spawn. Automatic
+workers re-check the labels before their first and every later provider write. Human-facing
+status comments, mentions, and escalation stay disabled unless
+`notifyHumans` is explicitly enabled. Admission uses `batchSize` as its active
+ceiling and starts no more than one new routed PR per sweep. Every sweep logs
+exact scanned, eligible, excluded, incomplete, already-owned, admitted,
+capacity-deferred, unchanged, and failed counts; deferred work is retried, not
+silently truncated. This bounded intake is deliberate: issue #222 recorded a
+GitHub API measurement of 302 open PRs, 248 ready for review, across 41 routed
+repositories on 2026-08-07.
+
+Do not enable `routed-open-prs` until the GitHub App identity path from #221 is
+deployed and validated. The code can land first because the default mode remains
+`factory-created`.
+
 ### Scheduled sync-fidelity canary
 
 `factory canary` is the regression detector for upstream sync drift: if a synced

src/cli/fleet.ts

@@ -884,7 +884,10 @@ async function runStandaloneBabysitCommand(
       crossRepository: pr.crossRepository,
       maintainerCanModify: pr.maintainerCanModify,
     },
-    standaloneBabysitter: { specSource },
+    // The explicit CLI command is itself the human opt-in and preserves its
+    // existing conversational behavior. Automatic routed sweeps pass the
+    // safer notification flag from config instead.
+    standaloneBabysitter: { specSource, notifyHumans: true },
     integrationsMountRoot: resolveIntegrationsMountRoot(mount),
     testGuidance,
   })

src/config/schema.test.ts

@@ -36,7 +36,13 @@ describe('FactoryConfigSchema', () => {
       reviewer: 'spawn:claude',
       babysitter: 'spawn:claude',
     })
-    expect(parsed.babysitter).toEqual({ enabled: false })
+    expect(parsed.babysitter).toEqual({
+      enabled: false,
+      mode: 'factory-created',
+      excludeLabels: ['factory:skip-babysitter'],
+      excludePullRequests: [],
+      notifyHumans: false,
+    })
     expect(parsed.terminalState).toBe('human-review')
     expect(parsed.stateIds.humanReview).toBeUndefined()
     expect(parsed.loop.registryPath).toBe('/tmp/factory-run/factory-loop-registry.json')
@@ -314,6 +320,35 @@ describe('FactoryConfigSchema', () => {
     expect(parsed.repos.names).toEqual(['pear', 'cloud', 'agentswarm'])
   })
 
+  it('requires explicit routed repositories for routed PR babysitting', () => {
+    expect(() => FactoryConfigSchema.parse({
+      babysitter: { enabled: true, mode: 'routed-open-prs' },
+      repos: {},
+    })).toThrow(/repos\.names must contain at least one repository/u)
+
+    expect(FactoryConfigSchema.parse({
+      babysitter: { enabled: true, mode: 'factory-created' },
+      repos: {},
+    }).babysitter.mode).toBe('factory-created')
+
+    expect(() => FactoryConfigSchema.parse({
+      babysitter: { enabled: true, mode: 'routed-open-prs' },
+      repos: { names: ['pear'] },
+    })).toThrow(/must resolve at least one owner\/repository route/u)
+  })
+
+  it('accepts one-character repository opt-out identities', () => {
+    const parsed = FactoryConfigSchema.parse({
+      babysitter: {
+        enabled: true,
+        excludePullRequests: ['owner/r#1'],
+      },
+      repos: {},
+    })
+
+    expect(parsed.babysitter.excludePullRequests).toEqual(['owner/r#1'])
+  })
+
   it('lets explicit byLabel/clonePaths/labels override the derived ones', () => {
     const parsed = FactoryConfigSchema.parse({
       subscription: { labels: ['pear'] },

src/config/schema.ts

@@ -121,6 +121,22 @@ const slackSchema = z.object({
 
 const babysitterSchema = z.object({
   enabled: z.boolean().default(false),
+  // Preserve the historical issue-created intake unless an operator
+  // deliberately opts into the wider routed-repository sweep. Keeping the
+  // rollout switch separate from `enabled` lets the identity fix in #221 land
+  // before automated writes are attributed across existing human PRs.
+  mode: z.enum(['factory-created', 'routed-open-prs']).default('factory-created'),
+  // This label is an author-controlled hard stop. Discovery reads it before
+  // spawning an agent, and routed babysitters are instructed to re-check it
+  // before their first provider write.
+  excludeLabels: z.array(z.string().trim().min(1)).default(['factory:skip-babysitter']),
+  excludePullRequests: z.array(z.string().regex(
+    /^[A-Za-z0-9](?:[A-Za-z0-9_.-]{0,99})\/[A-Za-z0-9_.-]{0,99}#[1-9]\d*$/u,
+    'expected owner/repo#number',
+  )).default([]),
+  // Routed intake stays quiet until an operator explicitly approves the
+  // notification shape observed in a read-only sweep.
+  notifyHumans: z.boolean().default(false),
 }).default({})
 
 const reportingSchema = z.object({
@@ -307,9 +323,40 @@ const NodeConfigObjectSchema = z.object({
 
 const FactoryConfigObjectSchema = WorkspaceConfigObjectSchema.merge(NodeConfigObjectSchema)
 
-export const WorkspaceConfigSchema = WorkspaceConfigObjectSchema.transform((cfg) => normalizeWorkspaceConfig(cfg))
+const requireRoutedBabysitterRepos = (
+  cfg: z.infer<typeof WorkspaceConfigObjectSchema>,
+  ctx: z.RefinementCtx,
+): void => {
+  if (cfg.babysitter.mode === 'routed-open-prs' && (cfg.repos.names?.length ?? 0) === 0) {
+    ctx.addIssue({
+      code: z.ZodIssueCode.custom,
+      path: ['repos', 'names'],
+      message: 'repos.names must contain at least one repository when babysitter.mode is routed-open-prs',
+    })
+    return
+  }
+  if (cfg.babysitter.mode === 'routed-open-prs') {
+    const routedRepos = (cfg.repos.names ?? []).map((name) =>
+      cfg.repos.byLabel[name] ?? cfg.repos.overrides[name] ??
+      (cfg.repos.org ? `${cfg.repos.org}/${name}` : name)
+    )
+    if (!routedRepos.some((repo) => /^[^/]+\/[^/]+$/u.test(repo))) {
+      ctx.addIssue({
+        code: z.ZodIssueCode.custom,
+        path: ['repos', 'names'],
+        message: 'repos.names must resolve at least one owner/repository route when babysitter.mode is routed-open-prs',
+      })
+    }
+  }
+}
+
+export const WorkspaceConfigSchema = WorkspaceConfigObjectSchema
+  .superRefine(requireRoutedBabysitterRepos)
+  .transform((cfg) => normalizeWorkspaceConfig(cfg))
 export const NodeConfigSchema = NodeConfigObjectSchema.transform((cfg) => normalizeNodeConfig(cfg))
-export const FactoryConfigSchema = FactoryConfigObjectSchema.transform((cfg) => normalizeFactoryConfig(cfg))
+export const FactoryConfigSchema = FactoryConfigObjectSchema
+  .superRefine(requireRoutedBabysitterRepos)
+  .transform((cfg) => normalizeFactoryConfig(cfg))
 
 function normalizeWorkspaceConfig(cfg: z.infer<typeof WorkspaceConfigObjectSchema>) {
   const resolved = resolveRepos(cfg.repos, cfg.repos.cloneRoot)

src/dispatch/templates.test.ts

@@ -186,7 +186,11 @@ describe('renderAgentTask', () => {
         baseRef: 'main',
         headRepo: 'AgentWorkforce/hoopsheet',
       },
-      standaloneBabysitter: { specSource: 'pull-request' },
+      standaloneBabysitter: {
+        specSource: 'pull-request',
+        excludeLabels: ['factory:skip-babysitter'],
+        notifyHumans: false,
+      },
       integrationsMountRoot: '/workspace/.integrations',
     })
 
@@ -210,6 +214,10 @@ describe('renderAgentTask', () => {
     expect(task).toContain('push the same PR head')
     expect(task).toContain('re-read the live merge state and fresh checks')
     expect(task).toContain('never merge it yourself')
+    expect(task).toContain('Before your first provider write')
+    expect(task).toContain('factory:skip-babysitter')
+    expect(task).toContain('Do not post status comments, mention humans, send notifications, or escalate')
+    expect(task).not.toContain('proactively offer to discuss')
     expect(task).toContain('Never search for, read, or substitute credentials or tokens')
     expect(task).toContain('output `/exit` on its own line')
     expect(task).not.toContain('DM `broker`')

src/dispatch/templates.ts

@@ -47,6 +47,10 @@ export interface RenderAgentTaskInput {
    */
   standaloneBabysitter?: {
     specSource: 'pull-request' | 'linked-issue'
+    /** Labels which abort this run before its first provider write. */
+    excludeLabels?: string[]
+    /** Human-facing comments, mentions, and escalation are opt-in for sweeps. */
+    notifyHumans?: boolean
   }
   slackDispatchThread?: {
     channel: string
@@ -222,13 +226,21 @@ export function renderAgentTask(input: RenderAgentTaskInput): string {
       const standaloneMergePolicy = input.config.mergePolicy === 'on-green-with-review'
         ? 'Merge policy: on-green-with-review. This standalone run has no guarded merge executor, so never merge the PR yourself; leave the final merge to a human.'
         : 'Merge policy: never - leave the PR open for human review and approval; never merge it yourself.'
+      const optOutLine = input.standaloneBabysitter.excludeLabels?.length
+        ? `Before your first provider write and again before every later provider write, re-read the live PR labels. If any label in JSON ${JSON.stringify(input.standaloneBabysitter.excludeLabels)} is present, make no further provider writes, report the opt-out, and exit.`
+        : undefined
+      const notificationLine = input.standaloneBabysitter.notifyHumans
+        ? 'You may notify or mention humans when a concrete decision is required.'
+        : 'Do not post status comments, mention humans, send notifications, or escalate this run. Only write the code, commits, pushes, and direct review-thread replies required to shepherd the PR.'
       return [
         `GitHub repo: ${repo}`,
         cloneInstruction,
         ...specHeader,
         '',
         `You are the standalone PR babysitter for ${prRef}.`,
         'Your job: drive this PR to genuinely green and correct against the definition of done above, then hand it to a human. Do NOT merge it yourself.',
+        ...(optOutLine ? [optOutLine] : []),
+        notificationLine,
         'Fix things directly and aggressively: inspect the existing implementation, make substantive corrections, and keep the PR scope anchored to the definition of done.',
         ...(branchLine ? [branchLine] : []),
         checkoutLine,
@@ -242,7 +254,9 @@ export function renderAgentTask(input: RenderAgentTaskInput): string {
         'After every push, wait for the checks on the newly pushed head commit. Never reuse green results from an older commit when declaring the PR ready.',
         'Commit and push fixes only to the existing PR head branch. Use a normal push when possible; if rebasing requires rewriting the PR head, use `--force-with-lease`, never an unconditional force push.',
         'If the push is denied, stop and report the access blocker. Never search for, read, or substitute credentials or tokens, and never modify Git/GitHub authentication configuration.',
-        'If a human can be reached, proactively offer to discuss the PR status, trade-offs, and open questions.',
+        ...(input.standaloneBabysitter.notifyHumans
+          ? ['If a human can be reached, proactively offer to discuss the PR status, trade-offs, and open questions.']
+          : []),
         'When the PR is green — no failing CI, no merge conflicts, and every review comment addressed — report a concise completion summary and output `/exit` on its own line so the Agent Relay task-exit lifecycle closes cleanly.',
         standaloneFinishLine,
         standaloneMergePolicy,

src/github/index.ts

@@ -13,6 +13,11 @@ export {
   readStandalonePullRequest,
   standaloneBabysitterAgentName,
 } from './standalone-babysitter'
+export {
+  discoverRoutedPullRequests,
+  routedPrIdentity,
+  routedPrRepos,
+} from './routed-pr-babysitter'
 export type {
   GhRunner,
   GhRunResult,
@@ -30,3 +35,4 @@ export type {
   StandaloneBabysitTarget,
   StandalonePullRequest,
 } from './standalone-babysitter'
+export type { RoutedPrCandidate, RoutedPrDiscoveryReport } from './routed-pr-babysitter'