← Back to Blog

PR #156/#157 — Fleet Dispatch Scheduler v1 Ships, v2 Deferred to a Tracked Goal

Date: 2026-07-11 PRs: #156 (scheduler), #157 (roadmap + deferred goal) Branches: dispatch/grok/job-* → main, chore/fleet-scheduler-v2-deferred-goal → main Tracks: #141 optimizer follow-up (closes out epic #137 capability-matrix hardening)

The Goal at the End of the Previous PR

PR #148/#151/#152/#154 (blog post 51) shipped the base matrix: agent_quotas, the 3-stage _best_agent_for_story routing (capability → quota headroom → cost), and degraded-mode non-hard-blocking when quota signal is missing. That closed the gating half of the 2026-06 token-budget design. The optimizer half — sequencing a batch of stories against fleet-wide capacity, not just picking one agent per story — was explicitly left as a follow-up architect design doc, not code.

What Moved the Goalpost in This PR

Two design docs were written this session (docs/superpowers/specs/2026-07-11-fleet-dispatch-scheduler-design.md and its implementation plan) to scope that optimizer follow-up. The scope decision that mattered most: v1 ships a batch scheduler with fleet-level in-batch headroom accounting and a binary readiness gate; the harder problems — reset-timing-aware bin-packing, persistent quota-blocking history, and a GOVERNS-aware readiness gate that checks a plan's coverage across all 7 SDLC stages — are out of scope until we have real production data on how v1 behaves.

Rather than let that scope decision live only as a "later" comment in the design doc, it was formalized as a dated goal (synlynk goal create, goal-d38e3c83, deadline 2026-08-10) — a 30-day observation window on v1's synlynk schedule behavior before v2 is even converted into stories.

What Shipped

Schema & readiness gate

  • stories.priority / stories.readiness columns (additive migration)
  • synlynk story ready / synlynk story draft — binary gate, no story is schedulable until explicitly marked ready

Scheduler (synlynk/scheduler.py)

  • _compute_schedule_plan — batch routes ready stories through _best_agent_for_story, accumulating in-batch headroom consumption per (agent, model, quota_type) so story 2 in a batch sees story 1's projected spend, not just the pre-batch quota snapshot
  • _enqueue_plan — commits the plan to dispatch queue
  • Retry/reassignment: MAX_STORY_RETRIES=2, reassigns to next-best agent on failure rather than re-queuing to the same exhausted candidate

CLI

  • synlynk schedule [--execute] [--max-stories N] — dry-run plan preview by default, --execute commits it

Tests

  • tests/test_fleet_scheduler.py — e2e coverage of the batch plan, in-batch headroom accounting, retry policy, and CLI wiring

Grok implemented all 8 tasks from the plan; diff and tests reviewed and verified independently before merge (PR #156). CI showed only the known baseline failure set (5 environment-specific tests, confirmed unchanged across #151/#152/#154/#156) — no regression.

Roadmap + deferred goal (PR #157, docs-only per branch discipline)

  • Cross-Cutting Epics table: epic #137–141 marked Shipped, new Fleet Scheduler v2 (deferred) row pointing at goal-d38e3c83 and the design doc's Future Work section

On Track to Full Autonomous Multi-Agent Dispatch

Epic #137 is now closed end-to-end: capability score → quota gate → cost tie-break → batch scheduling, all real code paths instead of design-doc intentions. The three v2 problems (bin-packing around reset windows, persistent blocking-pattern history, stage-aware readiness) are real but premature — scoping them before v1 has run in production would be guessing at requirements instead of reading them off actual blocked/retry/degraded-mode counts.

New Goalpost

  • Run synlynk schedule in production for 30 days (2026-07-11 → 2026-08-10)
  • Revisit goal-d38e3c83 at deadline: review blocked-story reasons, retry frequency, degraded-mode incidence
  • Only then convert v2's three Future Work bullets into scoped stories
← Back to all posts