loops/skills: orchestrator context economy — subagent return-payload contracts + GitHub-externalized state

[kyle-sexton]

Category: operational · Source: live lane operation — long-running loop orchestrators accumulate subagent return payloads until auto-compact fires; quality degrades well before the window fills.

Problem

Loop skills (babysit, triage, work) spawn subagents whose full final reports return into the orchestrator's context. At day-scale (dozens-hundreds of dispatches) this alone walks the orchestrator into the degradation zone and forces auto-compaction mid-shift.

Design principles to codify in the loop-running skills

1. Return-payload contract: every subagent brief specifies exactly what comes back — identifiers + verdict + parked-payload only (e.g. work lane: PR URL, CI status, ≤3-line summary; babysit worker: PR#, action, outcome one-liner). The 10x lever is WHAT returns, not its format; markup choice is marginal.
2. Detail lives on GitHub, never in orchestrator context: PR bodies, issue comments, labels are the durable record. A subagent explains itself in the PR body, not to its parent.
3. Context is disposable by design: after any compaction (or restart), the orchestrator re-orients from gh queries (queue labels, open PRs, own status comments) — never from conversational memory. Lane state must survive a cold start with zero context.
4. Restart cadence: daily refresh restart already standard; add guidance to restart when the statusline crosses ~50% context rather than riding into auto-compact.

Relation

Extends #477/#478/#479 (skill absorption umbrellas) with a cross-cutting principle; complements #451 (dispatch posture).

[kyle-sexton]

This was generated by AI during triage. · Category: operational (pipeline throughput) · Priority: priority: high (week 2nd-priority: the pipeline improving itself; orchestrator context economy directly bounds how long a loop lane runs before degradation). Routed status: ready.

Plan

Codify four context-economy principles into the loop-running skills (babysit, triage, work) — all established best practice, low controversy:

1. Return-payload contract per lane. Every subagent brief specifies exactly what returns: identifiers + verdict + parked-payload only. Recommended per-lane defaults:
   • work: PR URL, CI status, ≤3-line summary.
   • babysit worker: PR#, action taken, one-line outcome.
   • triage planner: target issue #, ready/needs-human verdict, the plan comment already posted to the issue (not re-narrated to the parent).
2. Detail lives on GitHub, never in orchestrator context. PR bodies, issue comments, labels are the durable record; a subagent explains itself in the PR/issue, not to its parent.
3. Context is disposable by design. After any compaction or restart, the orchestrator re-orients purely from gh queries (queue labels, open PRs, own status comments) — never conversational memory. Lane state must survive a cold start with zero context.
4. Restart cadence. Keep the daily refresh; add: restart when the statusline crosses ~50% context rather than riding into auto-compact.

Cross-cutting principle that extends the #477/#478/#479 absorption umbrellas and complements #451 (dispatch posture) — land it as shared loop-running guidance the three lane skills reference, not three copies.

Acceptance

• Each loop skill states its subagent return-payload contract (identifiers + verdict + parked-payload), and its briefs enforce it.
• Each loop skill documents GitHub-externalized state + cold-start re-orientation from gh queries.
• Restart-at-~50%-context guidance is present.
• The guidance lives once (shared) and is referenced by babysit/triage/work, not duplicated.

Decision defaulted: codify the return-payload contract as identifiers+verdict+parked-payload with the per-lane defaults above, and the restart trigger at ~50% context — veto/tune before merge

[kyle-sexton]

Gap-hunt input (control tower 7/19): extension for the return-payload contract — artifact/reference passing. Subagents write full outputs to filesystem/GitHub and return lightweight REFERENCES up the chain, not payload summaries (anthropic.com/engineering/effective-context-engineering-for-ai-agents, effective-harnesses-for-long-running-agents). Also: resumable checkpoints — resume a failed dispatch from its failure point rather than restarting the cycle; distinct from the daily context-reset restart.