Durable Provider Throughput Dispatch

Durable Provider Throughput Dispatch

When to Use

Use this when the user explicitly wants to maximize AI-agent usage availability, burn weekly provider capacity, or continue useful overnight/background work even if provider credits may run out near reset.

This is the spend-forward counterpart to conservative quota routing. The goal is not merely to save credits; it is to convert available provider limits into durable artifacts that reduce future bottlenecks.

Core Principle

When the user states that provider credits are not the bottleneck, switch from conservation mode to durable-throughput mode, but do it from fresh capacity telemetry instead of static provider rules:

• refresh provider capacity/usage roughly every 6 hours before planning the next work wave;
• spend available Claude/Codex limits on useful plan prep, adversarial review, execution-readiness, queue synthesis, and approved bounded implementation;
• treat Gemini as a lower-budget account by default: normally reserve it for cross-reviews, adversarial reviews, research/recon, and risk scans, not main implementation/execution;
• if fresh telemetry shows reliable Gemini capacity and an appropriate bounded/review/recon task, use that capacity instead of rigidly excluding it;
• tolerate high utilization or temporary end-of-cycle depletion only when the user has accepted that tradeoff;
• preserve plan gates, legal/IP/privacy gates, and GitHub mutation control.

Required Dispatch Pattern

1. Refresh live state first

   • Check git status/branch/head.
   • Check active local tmux/processes/logs.
   • Check remote worker state before launching duplicate lanes.
   • Check provider CLI availability/auth where relevant.

2. Partition work by gate state

   • status:plan-review or draft issues: planning/review hardening only.
   • status:plan-approved plus required local markers: eligible for implementation lanes.
   • Unknown/stale state: audit-only until verified.

3. Route by provider fit

   • Claude: control-plane synthesis, long-context planning, adversarial review, governance-heavy work.
   • Codex: bounded implementation, tests, refactors, crisp execution-ready issues, adversarial code/readiness review.
   • Gemini: batched research, reconnaissance, standards/source scans, risk enumeration.

4. Route by machine role

   • Primary control surface keeps approvals, decisions, GitHub mutations, and synthesis.
   • Overflow machines run bounded worker lanes only after repo/tool/auth checks.
   • Do not let an overflow machine become a shadow control plane.

5. Write durable artifacts

   • Prompt packs under docs/plans/overnight-prompts/<date-or-wave>/.
   • Unique result file per lane under a results/ directory.
   • Logs under a known log directory.
   • A monitor or handoff artifact that records lane names, expected outputs, and stalled-lane recovery rules.

6. Monitor, reconcile, then feed lanes

   • Use scheduled/local monitors for long runs.
   • Treat zero-byte Claude logs as inconclusive while the process is alive; prefer process liveness and expected output artifacts.
   • If a lane stalls, salvage useful log content and relaunch only the missing bounded deliverable.
   • When the user reports provider credits sitting idle, verify useful provider activity rather than process liveness alone: compare log sizes/mtimes across snapshots, inspect child provider processes, and look for durable output artifacts. Wrappers running with unchanged logs are not enough.
   • Replan provider routing from fresh capacity telemetry about every 6 hours; see references/capacity-aware-provider-routing.md for the current capacity-aware Claude/Codex/Gemini policy.
   • When the machine is already saturated or an autofeed cron is launching overlapping waves, pause the recurring feeder first, then reconcile run directories before any replacement launches. Useful provider throughput requires completed artifacts, not raw process count.
   • Watch for known non-consuming stall signatures: Codex logs stuck at Reading additional input from stdin..., Codex sandbox startup failures such as bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted, Gemini startup validation errors from repo-local .gemini/agents, Gemini capacity failures (429 RESOURCE_EXHAUSTED / No capacity available), Claude stream-json requires --verbose, and Hermes/worker lanes exiting 143 after worktree-command timeouts. See references/provider-stall-recovery-autofeed.md.
   • Do not let monitors stop at passive reporting when the user's goal is continuous throughput. After primary lanes complete, run a safe auto-feed decision: draft -> adversarial review, review MAJOR -> plan patch lane, review APPROVE/MINOR -> status:plan-review command/comment pack, blocked -> blocker-prep packet, approval candidates -> synthesis table. Keep unsafe transitions gated: no status:plan-approved, no outreach, and no unapproved implementation without explicit user approval.
   • Limit each auto-feed tick to a small bounded number of new lanes (for example 1-2), require unique prompt/log/result paths, and write the generated follow-up prompt under the same repo-owned prompt pack (generated/) before launching it.
   • For anti-idle recovery, create both immediate direct provider lanes and a recurring autofeed monitor (for example every 30 minutes for the current work window) so completed/stalled lanes are replaced without a multi-hour gap. For over-saturation recovery, keep the monitor paused until stale/orphan lanes are classified and launcher rules are patched/dry-run verified.
   • After patching an over-saturated autofeed monitor, do not resume recurring cron immediately. Run exactly one controlled live tick while the cron remains paused, then inspect log/result mtimes, result sizes, failure signatures, and child-process duplication. If the live tick produces zero durable useful outputs or only provider-capacity/sandbox/no-output signatures, clean up that tick and switch to provider-health probes instead of restarting the feeder. See references/provider-autofeed-health-recovery.md.
   • If the workspace is dirty and capacity should be used immediately, prefer a read-only Claude plan-mode scout wave with absolute prompt/output paths before creating write-enabled worktrees. Background launches may not honor relative paths as expected, and interrupted git worktree add can leave broken .git/worktrees/* metadata. See references/dirty-workspace-plan-mode-scout-wave.md.

Safety Gates That Do Not Relax

Even in spend-forward mode:

• Do not implement issues before plan approval.
• Do not mutate GitHub labels/comments/closures from worker lanes unless explicitly authorized.
• Do not publish or promote raw/client/standards/GTM artifacts without source, provenance, license, privacy, and IP checks.
• Do not preserve or output secrets; redact credentials.
• Do not count an issue as approval-ready unless current plan/review/legal artifacts support that claim.

Good Deliverables

A strong overnight spend-forward batch produces one or more of:

• approval-readiness package with issue numbers, live labels, plan paths, review paths, legal gate status, blockers, and exact operator commands;
• adversarial review reports from independent providers;
• execution scout report listing already-approved, low-contention issues;
• follow-up prompt pack for the next wave;
• monitor/handoff artifact with lane status and recovery instructions.

Pitfalls

• Conserving credits after the user has explicitly prioritized throughput.
• Burning credits on non-durable chat output instead of repo-tracked artifacts.
• Launching duplicate lanes without checking existing sessions/logs/results.
• Treating status:plan-review as approval-ready without artifact verification.
• Letting provider enthusiasm bypass legal or plan-gate controls.
• Resuming an autofeed cron just because the classifier script was patched; a patched classifier can still reveal that provider launch health is zero. Prove at least one durable provider output in a controlled live tick or minimal health probe before recurring dispatch.
• Counting child provider processes as separate lanes. Deduplicate by result/log path so wrapper, timeout, node, binary, and tee processes do not inflate useful-active counts.
• Letting result-file freshness override explicit failure signatures; capacity/sandbox/stdin/stream-json signatures must classify stale even when a result stub exists.