feat(dispatcher): auto-dispatch + limits (Phase 8)#132
Merged
Conversation
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (10)
✅ Files skipped from review due to trivial changes (1)
📝 WalkthroughWalkthroughAdds Phase 8: slot accounting, a debounced per-repo auto-dispatch loop, per-repo pause/resume and mm config CLI, complexity pause propagation from adapters to dispatcher with GitHub comment formatting, rate-limit observer, dispatch source metadata, implementation workflow prompts, and tests/docs. ChangesPhase 8 Auto-dispatch, Limits, and Pause/Resume
Estimated code review effort🎯 4 (Complex) | ⏱️ ~75 minutes Possibly related issues
Possibly related PRs
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. Warning There were issues while running some tools. Please review the errors and either fix the tool's configuration or disable the tool if it's a critical failure. 🔧 ESLint
ESLint skipped: no ESLint configuration detected in root package.json. To enable, add Comment |
This was referenced May 25, 2026
Add packages/dispatcher/src/slots.ts: getSlotState derives the three gating dimensions (per-adapter, per-repo, global) from live workflows rows + merged config; hasFreeSlot is the enqueue guard; reserveSlot is the auto-dispatch loop's pure local decrement. The recommender's row is excluded from slot counts (its dedicated slot). Covers at-capacity, free-slot, and per-adapter-vs-global-vs-repo cases.
Add packages/dispatcher/src/auto-dispatch.ts: walk readyToDispatch, skip rate-limited adapters and exhausted per-adapter slots, stop on a full repo/global, decrementing a local SlotState as it enqueues. No pre-dispatch complexity gate. Wire the four triggers in main.ts behind a debounced, re-entrancy-guarded scheduler: workflow terminal-state transitions (broadcastWorkflow), rate-limit changes (new rate-limit observer), recommender-run completion (threaded triggerAutoDispatch), and manual mm dispatch (ControlPlane.afterDispatch). The loop's enqueue goes through the daemon's collision-guarded startDispatch directly, so its own enqueues never re-trigger the loop.
repo-config.ts: setPausedUntil/clearPaused/isPaused/getPausedUntil over the repo_config.paused_until column (lazy row create, future-dated pause auto-expires). The auto-dispatch loop's enable-check now composes the [recommender] auto_dispatch toggle (default false) with !isPaused. CLI: mm pause / mm resume (DB-keyed by the git-remote slug, shared deriveRepoSlug in paths.ts) and mm config <repo> <key> <value> (a formatting-preserving, section-scoped TOML edit; v1 supports auto_dispatch). Tests cover toggle off/on, paused-until-in-future, comment/sibling preservation, and unknown-key/invalid-value rejection.
A complexity pause is a kind-tagged asked-question: BlockedSentinel
gains kind ('question'|'complexity'), the claude adapter parses it, and
the implementation workflow passes the kind to postQuestion on park
(already routing to waiting-human). The default gh-backed poster frames
a complexity pause with the literal 'complexity pause' label so the
recommender classifies it under Needs human input. The dispatch brief
now carries the repo's complexity_ceiling (resolved per repo, default 3)
as the agent's fork budget and instructs it to tag a complexity overrun.
No pre-dispatch ceiling gate in the loop (pinned by a test).
#53) isEpicApproved defaults to reading the Epic's 'approved' GitHub label (new github.getIssueLabels); when present the dispatch brief authorizes proceeding past a complexity overrun with a best-judgment call instead of pausing. Manual mm dispatch now respects slot limits via a route-level ControlPlane.slotAvailable gate (429 when full), distinct from the 409 collision guard; the loop bypasses the route so its accounting is unaffected. Dispatch origin recorded as meta_json.source ('manual' vs 'auto'), threaded through ImplementationInput and carried by continuations. Tests: approved-proceeds brief, manual 429/200 slot gate, source round-trip.
A flaky gh label read (isEpicApproved) or per-repo config load (resolveComplexityCeiling) must not fail the whole dispatch — fall back to safe defaults (ceiling 3, not approved) and log. Also gate the resolution on the brief not already existing, so a resume (which writes its own brief) skips the gh round-trip entirely.
thejustinwalsh
force-pushed
the
middle-issue-48
branch
from
ec570a7 to
ad5ab28
Compare
…nfig key Self-review follow-ups: - slotAvailable now receives the full ControlDispatchInput and resolves slot caps from the request's own repoPath, so manual mm dispatch respects max_concurrent even for a repo the daemon hasn't dispatched yet this lifetime (previously a cold repoPaths returned 'slot free'). - setTomlKey escapes the key before building its regex (SETTABLE is the extension point; a future key with regex metacharacters must match literally).
thejustinwalsh
commented
May 25, 2026
thejustinwalsh
left a comment
thejustinwalsh
left a comment
Owner
Author
There was a problem hiding this comment.
Decision-log rationale, distilled inline (the why behind the load-bearing choices). Source: planning/issues/48/decisions.md.
| * distinct (a repo's per-repo cap must not be charged for another repo's agents). | ||
| * The recommender's row is excluded by {@link countActiveImplementationSlots}. | ||
| */ | ||
| export function getSlotState(db: Database, repo: string, limits: SlotLimits): SlotState { |
Owner
Author
There was a problem hiding this comment.
Why a separate authority over re-counting: countActiveImplementationSlots already counts non-terminal implementation rows (recommender excluded) per-repo and globally; slots.ts only adds the limit/availability layer on top. Keeping the guard a pure function of (SlotState, adapter) makes at-capacity / free-slot / per-adapter-vs-global unit-testable without the engine, and lets the loop and manual dispatch share one consultation. I deliberately did not overload the daemon's startDispatch string|null (whose null means the 409 collision reservation) with a slots-full signal — that would conflate "already running" with "queue full".
| const autoDispatchTimers = new Map<string, ReturnType<typeof setTimeout>>(); | ||
| const autoDispatchRunning = new Set<string>(); | ||
| const autoDispatchRerun = new Set<string>(); | ||
| function scheduleAutoDispatch(repo: string): void { |
Owner
Author
There was a problem hiding this comment.
Why a debounced, re-entrancy-guarded scheduler: all four trigger events (recommender-complete, terminal transition, rate-limit change, manual dispatch) funnel here so a burst coalesces into one pass per repo, and a trigger arriving mid-pass re-runs exactly once after. Crucially the loop's own enqueue calls startDispatchImpl directly (not the HTTP route), so it never fires afterDispatch and can't recursively re-trigger itself — afterDispatch is manual-dispatch-only.
| * hasn't dispatched yet this lifetime (cold `repoPaths`). Conservative: an | ||
| * unreadable config reports a free slot rather than blocking a manual dispatch. | ||
| */ | ||
| function slotAvailable(input: ControlDispatchInput): boolean { |
Owner
Author
There was a problem hiding this comment.
Why the slot gate is route-level (not inside startDispatchImpl): the auto-dispatch loop already does its own local slot accounting and enqueues through startDispatchImpl directly; a DB-slot check inside the shared path would double-gate the loop inconsistently (rows are created async, so a mid-pass DB read disagrees with the loop's local view). Resolving caps from the request's own repoPath keeps the gate effective even for a repo the daemon hasn't dispatched yet this lifetime. A distinct 429 (vs the collision 409) tells mm dispatch why it was refused.
| * timestamp are touched on conflict, so a later config sync isn't clobbered). | ||
| * `until` is unix-ms; omit it to pause indefinitely (`mm pause`). | ||
| */ | ||
| export function setPausedUntil( |
Owner
Author
There was a problem hiding this comment.
Why keyed by the git-remote slug + indefinite default: the auto-dispatch loop reads pause state by repo = the git-remote owner/name (the same key a dispatch/recommender resolves), so mm pause must derive the slug identically or it writes a row the loop never reads. mm pause with no duration is indefinite (MAX_SAFE_INTEGER); isPaused honors the timestamp so a future-dated pause auto-expires with no cleanup.
| if (outcome.kind === "asked-question" && deps.postQuestion) { | ||
| // The sentinel's `kind` distinguishes a complexity pause (surfaced under the | ||
| // `complexity pause` state-issue label) from a plain question. | ||
| const kind = outcome.sentinel?.kind === "complexity" ? "complexity" : "question"; |
Owner
Author
There was a problem hiding this comment.
Why a complexity pause reuses the asked-question spine: distinguishing it by a kind field on the sentinel (rather than a new park kind) means the waitFor/resume plumbing (Phase 5) needs no change — a complexity pause resumes the same way once the human clarifies or applies approved. The dispatcher can't write the state issue's "Needs human input" section (recommender-owned), so the surface is a GitHub comment the recommender keys off for the complexity pause label.
| }), | ||
| resolveComplexityCeiling: args.resolveComplexityCeiling, | ||
| // Default: the Epic is approved iff it carries the `approved` label (#53). | ||
| isEpicApproved: |
Owner
Author
There was a problem hiding this comment.
Why approved is delivered via the brief: it lands alongside complexity_ceiling — one place the agent reads its fork policy — rather than a separate channel. The default reads the Epic's labels via gh; a flaky read is made failure-safe in driveOnce (falls back to not-approved, never fails the dispatch).
Owner
Author
Reviewer's brief — Phase 8 (auto-dispatch + limits), PR #132What shipped: all five sub-issues (#49–#53). Slot accounting + an auto-dispatch loop (four triggers behind one debounced scheduler), per-repo How to run itbun install
bun run typecheck # tsc --noEmit
bun run lint # oxlint --fix --deny-warnings
bun test packages # 540 passWhat to verify (and what "correct" looks like)
Fragile / extra eyes
How to reviewDecision rationale is distilled into inline review comments on the PR (source: |
thejustinwalsh
added
the
ready-for-review
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 8
🧹 Nitpick comments (1)
packages/dispatcher/src/build-deps.ts (1)
22-24: ⚡ Quick winUse
APPROVED_LABELin pause text to avoid label drift.The complexity comment text hardcodes
approvedwhile approval checks useAPPROVED_LABEL; keeping one source of truth prevents mismatch later.Proposed fix
- A human resolves this by **scope reduction or clarification** — or applies the \`approved\` label to authorize a best-judgment call within the ceiling on resume.`; + A human resolves this by **scope reduction or clarification** — or applies the \`${APPROVED_LABEL}\` label to authorize a best-judgment call within the ceiling on resume.`;Also applies to: 40-45
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/dispatcher/src/build-deps.ts` around lines 22 - 24, Replace the hardcoded "approved" string in the complexity pause/comment text with the APPROVED_LABEL constant to keep a single source of truth; update every occurrence in this file (the pause text variable(s) around the complexity pause and the other instance noted between lines 40-45) so they reference APPROVED_LABEL instead of the literal "approved".
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@packages/adapters/claude/test/adapter.test.ts`:
- Around line 252-267: The test that verifies fallback for an unrecognized kind
should assert the classification explicitly instead of guarding with an if;
update the test around claudeAdapter.classifyStop to assert that result.kind ===
"asked-question" (e.g., using expect(result.kind).toBe("asked-question")) and
then assert the sentinel content (result.sentinel) equals { question: "Q" }, so
the test fails if classification regresses rather than silently passing when the
guarded block is skipped.
In `@packages/cli/src/commands/config.ts`:
- Around line 4-7: Add a top-level TSDoc comment for the exported ConfigOptions
type describing its purpose and fields; update the file by adding a JSDoc/TSDoc
block immediately above the exported type ConfigOptions explaining that it
represents CLI configuration overrides and documenting the optional configFile
property (what it overrides and the default path behavior). Ensure the comment
uses proper TSDoc tags/format (brief summary and param/field description for
configFile) so the export complies with the project's “every public export must
carry a TSDoc/JSDoc comment” guideline.
- Around line 30-48: The header detection currently uses headerRe which only
matches bare “[section]” and thus misses headers with trailing comments; update
headerRe (and the similar use around lines 50-55) to accept optional trailing
comments/whitespace (e.g., allow a "#" and anything after the closing bracket)
so headers like "[recommender] # note" are recognized; keep the captured group
as the section name (m[1]) unchanged, leave keyRe, sectionStart, assignment,
lines and source logic intact, and run the same adjusted regex wherever headerRe
is used to avoid appending duplicate sections.
In `@packages/cli/src/commands/pause.ts`:
- Around line 45-47: The resolve(...) and openAndMigrate(...) calls in the pause
command run outside the try/catch so thrown errors can bypass the intended
exit-code handling; move or duplicate those calls inside the existing try block
(or wrap each call in its own try/catch) so any exception from resolve("pause",
repoPath, opts) or openAndMigrate(resolved.dbPath) is caught and the function
returns the documented exit code (1 on failure, 0 on success), and ensure the
same pattern is applied to the other occurrence around lines 66–69.
- Around line 8-15: Add a top-level TSDoc/JSDoc block immediately above the
exported type declaration export type PauseResumeOptions that briefly describes
the purpose of this options object (e.g., options for pause/resume CLI
commands), noting that it overrides config and db paths and can supply a
resolveSlug function; keep it concise (one summary sentence and optional
`@remarks` or `@example` if useful) so the public export carries documentation per
guidelines while keeping the existing per-field comments intact.
In `@packages/dispatcher/src/hook-server.ts`:
- Around line 398-400: The best-effort post-dispatch callback
control.afterDispatch(normalizedRepo) can throw and turn a successful manual
dispatch into a 500; wrap the call in a try/catch to isolate errors so they
don't affect the HTTP response. Locate the invocation of control.afterDispatch
in hook-server.ts (the lines calling afterDispatch with normalizedRepo) and
surround it with a try block, catch any error, and log it (or swallow it)
without rethrowing so startDispatch's success path still returns a successful
response.
In `@packages/dispatcher/src/main.ts`:
- Around line 254-256: scheduleAutoDispatch reads the variable shuttingDown
before it's declared, risking a TDZ ReferenceError when hookServer.start
triggers afterDispatch; move the declaration/initialization of shuttingDown (and
any related state like autoDispatchTimers if applicable) above any
functions/calls that may reference it (specifically ensure let shuttingDown =
false; is defined before scheduleAutoDispatch, hookServer.start, and any
registration of afterDispatch handlers), or alternatively convert shuttingDown
to a const/var initialized earlier so scheduleAutoDispatch and afterDispatch can
safely read it at runtime.
In `@planning/issues/48/plan.md`:
- Line 7: Update the scoped statement on line containing the phrase "the
recommender runs on cron" to remove the claim that cron scheduling is delivered
in this Epic; reword it to indicate the recommender can be scheduled later or
that cron scheduling is out of scope (referencing follow-up `#135`) — keep the
rest of the sentence about ready Epics auto-dispatching within slot limits and
the approved-label override unchanged.
---
Nitpick comments:
In `@packages/dispatcher/src/build-deps.ts`:
- Around line 22-24: Replace the hardcoded "approved" string in the complexity
pause/comment text with the APPROVED_LABEL constant to keep a single source of
truth; update every occurrence in this file (the pause text variable(s) around
the complexity pause and the other instance noted between lines 40-45) so they
reference APPROVED_LABEL instead of the literal "approved".
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: 7566d9b1-7b1d-41a2-b235-4938a3a4bdd7
📒 Files selected for processing (31)
packages/adapters/claude/src/classify.tspackages/adapters/claude/test/adapter.test.tspackages/cli/src/commands/config.tspackages/cli/src/commands/pause.tspackages/cli/src/index.tspackages/cli/src/paths.tspackages/cli/test/config.test.tspackages/cli/test/pause-resume.test.tspackages/core/src/adapter.tspackages/dispatcher/src/auto-dispatch.tspackages/dispatcher/src/build-deps.tspackages/dispatcher/src/github.tspackages/dispatcher/src/hook-server.tspackages/dispatcher/src/index.tspackages/dispatcher/src/main.tspackages/dispatcher/src/rate-limits.tspackages/dispatcher/src/recommender-run.tspackages/dispatcher/src/repo-config.tspackages/dispatcher/src/slots.tspackages/dispatcher/src/workflow-record.tspackages/dispatcher/src/workflows/implementation.tspackages/dispatcher/test/auto-dispatch.test.tspackages/dispatcher/test/build-deps.test.tspackages/dispatcher/test/control-routes.test.tspackages/dispatcher/test/implementation-workflow.test.tspackages/dispatcher/test/recommender-run.test.tspackages/dispatcher/test/repo-config.test.tspackages/dispatcher/test/slots.test.tspackages/dispatcher/test/workflow-record.test.tsplanning/issues/48/decisions.mdplanning/issues/48/plan.md
A bare `[section]` header match missed valid TOML like `[recommender] # note` and `[ recommender ]`, so `mm config` appended a duplicate section instead of editing in place. Match via a `headerName` helper (trailing-comment- and whitespace-tolerant, trimmed name) used by both the section-find loop and the key-scan break. Also adds the missing TSDoc on the exported `ConfigOptions`.
`resolve()` and `openAndMigrate()` ran before the try/catch, so a throw (e.g. a failed git-remote slug derivation) rejected the promise instead of returning the documented exit code 1. Move both inside the try; `db` is now nullable with `db?.close()` in finally. Also adds the missing TSDoc on `PauseResumeOptions`.
…response A throwing `afterDispatch` turned a manual dispatch that already succeeded into a 500. Wrap the callback so a failure logs and is swallowed; the route still returns the workflow id.
…atch `scheduleAutoDispatch` reads `shuttingDown`, but the `let` was declared after `hookServer.start`; a slot-freeing broadcast firing that path early threw a TDZ ReferenceError. Declare it up front, before any reader.
The pause comment hardcoded `approved` while approval checks use APPROVED_LABEL; interpolate the constant so the label has one source of truth.
Without asserting `result.kind`, the test passed even if classification regressed and the guarded block never ran. Also reword plan.md so it no longer claims cron scheduling shipped in this Epic (out of scope, tracked as #135).
Owner
Author
Review round 1 — addressedBatched all 8 actionable findings + the nitpick, one fix (with a test) per class, pushed once (
Verification: |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Closes #48
Phase 8 makes dispatch autonomous within limits. Slot accounting gates the enqueue path (per-adapter / per-repo / global), an auto-dispatch loop walks the recommender's ranked plan and dispatches ready Epics — triggered on all four events (recommender-run completes, workflow terminal transition, rate-limit change, manual
mm dispatch) behind one debounced scheduler. Auto-dispatch is opt-in per repo ([recommender] auto_dispatch, default off) and pausable (mm pause/mm resume). There is no pre-dispatch complexity gate: a complexity overrun is a runtime pause on a sub-issue that routes the workflow towaiting-humanand surfaces under the state issue'scomplexity pauselabel — unless the Epic carries theapprovedlabel, which the dispatch brief uses to authorize a best-judgment call.Status
auto_dispatchtoggle + pause/resumeapproved-label handling for complexity pausesAcceptance criteria (Epic #48)
completed)approvedlabel — no pre-dispatch ceiling gate (Route sub-issue complexity overruns to waiting-human #52), complexity pause routes towaiting-human(Route sub-issue complexity overruns to waiting-human #52),approvedauthorizes proceeding (Implement approved-label handling for complexity pauses #53). See Route sub-issue complexity overruns to waiting-human #52, Implement approved-label handling for complexity pauses #53.What changed
packages/dispatcher/src/slots.ts(new) —getSlotState/hasFreeSlot/reserveSlot: the slot authority over the three gating dimensions, derived from liveworkflowsrows + merged config (recommender row excluded).packages/dispatcher/src/auto-dispatch.ts(new) — the loop: walkreadyToDispatch, skip rate-limited adapters + exhausted per-adapter slots, stop on a full repo/global, decrement a local view as it enqueues. No complexity gate.packages/dispatcher/src/main.ts— wires the four triggers behind a 250 ms debounced, re-entrancy-guarded scheduler;startDispatchImplrecordssource;slotAvailablegates manual dispatch;resolveComplexityCeilingfeeds the brief.packages/dispatcher/src/repo-config.ts(new) —paused_untilhelpers (mm pause/mm resume); future-dated pause auto-expires.packages/dispatcher/src/rate-limits.ts— a rate-limit-change observer (a reset re-runs auto-dispatch for every known repo).packages/dispatcher/src/hook-server.ts—ControlPlane.slotAvailable(manual dispatch → 429 when full, distinct from the 409 collision) +afterDispatch(manual-dispatch trigger).packages/dispatcher/src/workflows/implementation.ts— complexity-pause routing (akind-taggedasked-question),complexity_ceiling+approvedinjected into the dispatch brief,sourcetracking; brief-context resolution is failure-safe.packages/dispatcher/src/build-deps.ts— defaultpostQuestion(gh comment,complexity pauseframing) + defaultisEpicApproved(reads theapprovedlabel) +formatPauseComment.packages/dispatcher/src/github.ts—getIssueLabels.workflow-record.ts—sourceinmeta_json+getWorkflowSource.packages/core/src/adapter.ts—BlockedSentinel.kind.packages/adapters/claude/src/classify.ts— parse it.packages/cli/—mm pause/mm resume(DB-keyed by the git-remote slug) andmm config <repo> <key> <value>(formatting-preserving, section-scoped TOML edit; v1auto_dispatch); sharedderiveRepoSluginpaths.ts.Why these changes
The slot guard is a pure function of
(SlotState, adapter)so the loop and manual dispatch share one consultation and the cases are unit-testable without the engine — built on the existingcountActiveImplementationSlots. The loop enqueues through the daemon's collision-guardedstartDispatchImpldirectly (not the HTTP route), so its own enqueues never re-fire the manual-dispatch trigger — no recursion. A complexity pause reuses the existingasked-questionpark spine (distinguished by a sentinelkind), so the Phase-5 waitFor/resume plumbing needs no change; since the dispatcher doesn't own the state issue's "Needs human input" section, it surfaces the pause as a recognizable Epic comment the recommender classifies.approvedrides in the dispatch brief alongsidecomplexity_ceiling— one place the agent reads its fork policy. Full rationale:planning/issues/48/decisions.md, distilled into inline review comments on this PR.Verification
bun run typecheck,bun run lint, andbun test packagesall green (540 tests). Per phase:slots.test.ts(10): at-capacity, free-slot, per-adapter-vs-repo-vs-global, recommender-excluded, clamp,reserveSlot.auto-dispatch.test.ts(10): normal pass, rate-limited-adapter skip, per-adapter/repo/global exhaustion, local-decrement mid-pass, collision no-op, no-complexity-gate;recommender-run.test.tstrigger-wiring.repo-config.test.ts(6) +pause-resume.test.ts+config.test.ts: toggle off/on, paused-until-in-future, TOML preservation, bad-key/value rejection.implementation-workflow.test.tscomplexity-pause routing + brief ceiling + in-ceiling-no-pause + failure-safe;build-deps.test.tsformatPauseComment;adapter.test.tssentinel-kindparse.control-routes.test.ts429/200 slot gate,workflow-record.test.tssource round-trip.Stumbling points
waiting-human/rate-limited) workflow still counts as a used slot under the existingcountActiveImplementationSlotssemantics; I kept that convention rather than redefine it, so slot-freeing triggers fire only on the four terminal states.repoPathsis populated lazily), which is why a true recommender cron is a follow-up (Periodic recommender scheduling (cron) + daemon repo registry #135) rather than a small addition here.Suggested CLAUDE.md updates
setRateLimitObserver(mirrorssetUpdateWorkflowObserver: process-scoped, reset to null on shutdown) so a future change preserves the shutdown reset.Follow-up issues
Known limitations (for reviewer attention)
mm dispatchslot-limiting is a soft cap checked just before enqueue; two concurrent manual force-dispatches of different Epics could over-subscribe by one (the collision guard only de-dupes the same Epic). Acceptable for an explicit force-dispatch; flagged rather than fixed with heavier atomic reservation.inFlightEpicsreleases the dispatch reservation via the first workflow broadcast; ifcreateWorkflowRecorditself threw afterengine.startsucceeded (a narrow DB-insert failure), the reservation would leak until restart. Pre-existing behavior, not introduced here.Out of scope
implementing-github-issuesskill) — Route sub-issue complexity overruns to waiting-human #52 defines only the sentinelkindcontract the dispatcher reads, delivered to the agent via the brief.Mergeability
Rebased onto
main(clean,rerere-assisted) and re-verified;mergeablereadsMERGEABLE.Summary by CodeRabbit
New Features
Tests
Documentation