Repository navigation
Implement the session conductor: merge openchamber-dispatcher + openchamber-babysitter into openchamber-conductor #2153
Description
Implementation follow-on from the #2132 triage (queue task 3163, 2026-10-29). Parent design: #2132; babysitter design origin: #2021; fleet epic: #2055.
Build the merged session conductor designed on #2132:
- New job
openchamber-conductor(20m cadence, owner goal:keep-the-four-repos-healthy) with one instruction file merged from the current dispatcher and babysitter instructions, preserving both rule sets verbatim where they do not conflict. One run = one capacity inventory + one quota snapshot (measurement gate: no dispatch or reroute without a quota picture ≤60 min old covering all pools), then:- SWEEP phase first: one pass over all live sessions — answer questions, nudge/kick stalls within the babysitter's existing limits, apply the escalation ladder: (a) reroute pool (Workhorse → Specialist; Summit only for synthesis-class work), (b) move off an exhausted pool, (c) stop the session and hand its task back to the queue parked with a reason — never leave a stopped session silent.
- DISPATCH phase second: launch from the queue head into capacity the sweep confirmed or freed; Workhorse default; verified launch with session↔task binding; fire-and-forget (no babysitting inside dispatch).
- Cutover: enable the conductor, disable
openchamber-dispatcherandopenchamber-babysitter(disable, do not delete — rollback is re-enabling the pair). Update the fleet-watchdog roster to the merged job id in the same pass. - Measurement: for 48h after cutover, the daily token check-in reports the conductor's input per cycle against the pair's check-in Bump build/buck/preludes/buck2-prelude from
3a02ea3to370cd4d#5 baseline (149M + 124M per ~15h window); success = ≤60% of the pair at equal sessions served. Report the first comparison on Design: session conductor — merge OpenChamber dispatcher + babysitter into one job (implements #2021 as the merged job) #2132.
Acceptance:
- Conductor job exists, enabled, running on a 20m cadence with the merged instruction (both phases, escalation ladder, measurement gate)
- Dispatcher and babysitter disabled (not deleted); fleet-watchdog roster names the conductor
- First 48h cost comparison posted on Design: session conductor — merge OpenChamber dispatcher + babysitter into one job (implements #2021 as the merged job) #2132 against the ≤60% target
- design: scheduled OpenCode preliminary babysitter — sweep sessions, fix easy stalls, escalate the rest #2021's open escalation question is answered by the ladder above; link design: scheduled OpenCode preliminary babysitter — sweep sessions, fix easy stalls, escalate the rest #2021 from the cutover report
Activity
developing-today commentedon Oct 29, 2026
CODE-1048
dezren39 commentedon Oct 30, 2026
Triage verdict — cutover plan CONFIRMED against live definitions, implementation packet enqueued
Queue task 3173 (triage,
repo-worker-code, OpenChamber scheduled lane), 2026-10-30T04:4xZ. Read-only: no job was created, enabled, disabled or edited by this triage.Live state re-verified (source of truth, not the packet)
openchamber-dispatcher~/hatch-personal/workspace/goals/keep-the-four-repos-healthy/crons/minutely/openchamber-dispatcher__interval@15m.mdtruehidden_files/openchamber-dispatcher-instruction.mdopenchamber-babysitter.../crons/minutely/openchamber-babysitter__interval@20m.mdtruehidden_files/openchamber-babysitter-instruction.mdBoth carry
owner: goal:keep-the-four-repos-healthyand deliver to the same chata15e5512-51ce-49c1-ad0f-46beaea69b04, each running its own PRELUDE and its own full inventory. The duplicated fixed cost the merge removes is confirmed in the definitions themselves.openchamber-conductoris free: no cron file, no branch inhatch-personal, no open PR mentions it. #2132's triage verdict (task 3163) and this re-check agree; no premise correction needed.Cutover plan — confirmed, with the exact touch points
openchamber-conductor__interval@20m.mdin the samecrons/minutely/directory,enabled: true,owner: goal:keep-the-four-repos-healthy,every: 20m,timeout_secs: 2700(the dispatcher's, per Design: session conductor — merge OpenChamber dispatcher + babysitter into one job (implements #2021 as the merged job) #2132 decision 1), delivery chata15e5512-…unchanged, PRELUDE block copied verbatim.hidden_files/openchamber-conductor-instruction.md: dispatcher + babysitter rule sets preserved verbatim where they do not conflict; one capacity inventory + one quota snapshot per cycle shared by both phases; SWEEP first then DISPATCH; escalation ladder (nudge/kick within existing limits → reroute pool Workhorse→Specialist, Summit for synthesis only → move off an exhausted pool → stop and hand the task back parked with a reason); measurement gate binding on dispatch and reroute.enabled: falsein both files above (definitions are versioned inhatch-personal, so this is a one-line-per-file, fully reversible change).crons/minutely/fleet-watchdog__interval@15m.mdMETHOD Query Ajob_id IN (…)list;job_id IN (…)list;hidden_files/fleet-watchdog-instruction.md§Fleet roster (line 18), the §Stuck-run threshold table (line 38, currently "dispatcher 60m, babysitter 45m" → one conductor threshold), and the §Watcher gaps rule (line 46, "openchamber-babysitter or repo-fleet-monitor" →openchamber-conductor). Missing the instruction-file edits leaves the watchdog alerting on two jobs that no longer run.Two gaps the implementation must close (not blockers, but silent-failure traps)
3a02ea3to370cd4d#5) comes from the daily token check-in, not fromdispatcher-log.md/babysitter-log.md, which record dispatches and rounds but no token input. The 48h comparison must read the same check-in source for the conductor, and the conductor's log must additionally record sessions swept + dispatched per window, otherwise "at equal sessions served" is unmeasurable. Ask for this in the implementation acceptance, do not leave it to prose.timeout_secs: 2700on a 20m cadence means runs can overlap pending occurrences; the scheduler coalesces them ascancelled, which the watchdog correctly does not count as failure. Same behaviour the pair has today (2700s on 15m), so it is a documentation note for the conductor instruction, not a design change.Routing
Implementation packet enqueued on
repo-health-code-pilotfor this issue, kindsmall_fix, flaggedlive-fleet-change, carrying the four acceptance criteria above plus the watchdog three-place roster edit and the measurement-source requirement. Dedupe keyrepo-health:v1:fix:developing-today/code:issue:2153. #2021's open escalation question is answered by the ladder (step 2); the cutover report must link #2021 and #2055.dezren39 commentedon Oct 30, 2026
🤖 Agent note · picked up
dezren39 commentedon Oct 30, 2026
🤖 Agent note · parked
dezren39 commentedon Oct 30, 2026
🤖 Agent note · picked up
openchamber-conductor(20m, 2700s), merged instruction from dispatcher+babysitter, disable (not delete) the pair, update all three fleet-watchdog roster touch points, cutover report. Note: the real scheduler change goes through the cron tools — thecrons/minutely/*.mdfiles are read-only copies of saved definitions.26 remaining items
dezren39 commentedon Nov 3, 2026
🤖 Agent note · parked — one acceptance item remains
hidden_files/conductor-log.mdheader): baseline dispatcher+babysitter = 273M input per ~15h window (daily token check-in Bump build/buck/preludes/buck2-prelude from3a02ea3to370cd4d#5); read the conductor's input for a named 48h window from the same daily token check-in source; normalise by sessions swept + sessions dispatched summed from the conductor-log cycle entries in that window; verdict against the ≤60%-of-baseline target at equal sessions served; post on Design: session conductor — merge OpenChamber dispatcher + babysitter into one job (implements #2021 as the merged job) #2132 and link it here.dezren39 commentedon Nov 4, 2026
agent-note:v1
event: picked_up
task: 8468
lane: code-i
execution_lane: direct
summary: Task 8468 (triage, acceptance remainder of #2153) claimed by lane code-i on the direct lane. Acceptance item 2 (#2021 link) already done per prior progress; resuming item 1: compute the conductor 48h cost comparison and post it on #2132.
next: assemble figures from the token check-in source and conductor-log.md, post on #2132, complete task 8468
dezren39 commentedon Nov 4, 2026
agent-note:v1
event: done
task: 8468
lane: code-i
execution_lane: direct
summary: Both #2153 acceptance items are now met. Item 2 (#2021 linked from the cutover report) was done earlier (6614026909 / 6614030429). Item 1 (48h cost comparison) is posted on #2132 (comment 6640808470): conductor first 48h (2026-10-30T22:13Z -> 2026-11-01T22:13Z) = 398.35M input over 139 runs (2.87M/run, 8.30M/h, 14 sessions dispatched) vs the check-in #5 dispatcher+babysitter baseline of 273.6M per ~14.7h (18.68M/h) = 44% per hour / 46% per 20-min slot — PASS against the <=60% target. Attribution method cross-validated against check-in #8's published conductor figure (96.39M measured vs 96M published). Closing #2153: implementation live since 2026-10-30 and every acceptance item now has posted evidence.
next: none — task 8468 completing as triaged