Repository navigation
epic(fleet): pool & routing intelligence — lanes everywhere, live dispatch + babysitting, quota stewardship, reviewer oversight, AA + measured-speed model #1334
Description
Purpose
Drew (2026-10-27, with the Muse usage upgrade): the fleet has token
headroom now — spend it on throughput and on intelligence about the
fleet's own routing. This epic scales the lane pattern to every queue
and every producer, puts the OpenChamber dispatcher and babysitter
live, and builds the oversight stack that keeps the model pools
honest: a quota/config steward that makes bounded routing decisions
and logs every change, a meta-reviewer that grades the reviewers
and updates their guidelines, and a pool-intelligence function that
combines Artificial Analysis data with our own measured session
speeds to flag models to add or retire, with controlled tests.
Scope (children)
- Lane scale-out: A/B shifts for the opencode, openchamber, and nix
queues (implements Queue workers: give the non-code queues lane-style long-budget shifts (claim evaporation is a run-budget problem) #1130's pattern for all queues). - Producer lanes: second offset lanes for the issue creator and the
PR diff reviewer. - Dispatcher v1 live (extends OpenChamber dispatch v1.1: admission control, hand-backs, stall nudges, harvesting, Workhorse-first routing #1187): launches OpenCode/OpenChamber
sessions for queued issues under admission control. - Babysitter live (extends OpenChamber babysitter v1: one worker, session rounds, quota + progress report (report-only) #1185): rounds over ALL OpenCode sessions
on an interval, with verdicts, quota, and pool attribution. - Quota & config steward: monitors every pool's quota, makes bounded
ad-hoc routing decisions (config changes land ONLY via PR to the
canonical repo config), and keeps an append-only change-event
ledger with before/after quota snapshots and measured impact. - Meta-reviewer: reviews the reviewers' (steward, lanes, dispatcher,
babysitter) activity and outcomes; updates the reviewer guidelines
from evidence — when model-up/effort-down wins, when max effort is
waste, which pools earn their place. - Pool intelligence: Artificial Analysis monitoring + our own
measured session speed per provider/model/effort; an interpretive
model combining AA scores with our history; add/retire flags with
quantified tradeoffs (score vs speed vs cost per pool), validated
by controlled tests before a change is recommended.
Execution mode
Scheduled fleet (cron lanes + steward/reviewer/intelligence jobs
under the keep-the-four-repos-healthy goal), reporting to the token /
fleet side chat. Config authority stays canonical: repo PRs only,
never hand-edited deployed files. Credentials by name only.
Dependencies
- epic(openchamber): session dispatch + babysitting — OpenChamber as the fleet's workhorse executor #1234 (OpenChamber dispatch + babysitter design/pilots) — this
epic operationalizes it. - Babysitter access: enroll a paired OpenChamber client for sandbox → laptop, and document the credential home #1271 (paired-client enrollment) gates remote babysitter/
dispatcher reach; early runs measure around it from the laptop. - Routing policy alignment: map OpenChamber routing categories to the fleet's Workhorse-first tiers + POOL MAP gates #1272 (routing-policy alignment) feeds the steward's decision
bounds; the steward starts conservative until it lands.
Acceptance criteria
- All children filed, labelled, and linked as sub-issues.
- Every job named in the children exists as a live cron with a
written instruction file, verified by cron list/status. - The change-event ledger exists and every config change in the
window has a matching event with before/after quota snapshots. - Roadmap rows added for this epic (child issue tracks it).
Filed by Muse (operator) 2026-10-27 on Drew's direction.
Activity
developing-today commentedon Oct 27, 2026
CODE-438
50 remaining items
dezren39 commentedon Nov 4, 2026
🤖 Agent note · handed off
dezren39 commentedon Nov 4, 2026
🤖 Agent note · handed off
dezren39 commentedon Nov 4, 2026
🤖 Agent note · picked up
dezren39 commentedon Nov 4, 2026
🤖 Agent note · triaged (current-main revalidation)
18f0aa1a68d83d3df127cca8523f578f4adb33a3; issue epic(fleet): pool & routing intelligence — lanes everywhere, live dispatch + babysitting, quota stewardship, reviewer oversight, AA + measured-speed model #1334 remains OPEN. GitHub sub-issues API lists all 14 linked children: 6 closed and 8 open.Child map / tracker state (fresh labels and states):
queue/blocked; its latest note says verification needs a read-only path to the authoritative Muse scheduler); Dispatcher v1 live: scheduled launches of OpenCode/OpenChamber sessions for queued issues #1337 dispatcher (queue/blocked; old dispatcher path is superseded by the claimant cutover, with only the by-id claim/bind remainder tracked by task 6669, whose packet has a source-repo defect).user/escalated): the function moved intoopenchamber-conductor, but the successor is disabled; the issue already records the choice to re-enable at 20m/hourly or retire it.status/triaged; real per-pool/per-effort data remains thin, with telemetry adoption tracked by fleet telemetry adoption: workers complete tasks without pool/model/effort/cost fields (1 of 846 post-#1582 completions) #4919); Reviewer scope gap: diff review is static — no runtime verification lane #1361 runtime review (status/triaged; remaining hatch-personal PR Bump checkmarx/ast-github-action from 2.0.11 to 2.0.19 #14 + live cron sync is outside this repo).Epic acceptance / dependencies: all children are filed, linked, and classified; the ledger and roadmap child trackers (#1339, #1343) are closed. The live-cron criterion is not yet established: #1335/#1337 remain blocked and #1338 records a disabled successor. Dependency #1272 is now CLOSED (the older note calling it queued is stale); #1271 remains OPEN/blocked on paired-client enrollment. Existing child trackers already own the remaining paths, so this pass creates no duplicate epic-level task. Keep #1334 OPEN until the live-job criterion and child blockers are resolved; only #1338's already-recorded re-enable/retire choice needs Drew.