Skip to content

nonhumanoriented.md

Latest commit

 

History

History
95 lines (61 loc) · 2.99 KB

nonhumanoriented.md

File metadata and controls

95 lines (61 loc) · 2.99 KB

Non-Human-First Operation Principles

1. Purpose

To continuously make progress autonomously without constant human monitoring.

Practical definition:

  • No silent deadlocks
  • No infinite loops on the same step with no state change
  • Repeated failures always converted to a different recovery path

Completion policy:

  • Don't assume completion guarantee for all cases including external conditions
  • Guarantee that the system does not intentionally stay stopped
  • When progress degrades, force strategy change via recovery state transition

2. Core Principles

  • Recovery-first over first-attempt success
  • Idempotent control points (lease / run claim / dedupe signature)
  • Backlog-first startup (prioritize consuming existing backlog over new generation)
  • Explicit blocked reasons that are machine-recoverable

3. Stall-Prevention Mechanisms

3.1 Lease and Runtime Lock Discipline

  • Task lease prevents duplicate dispatch
  • Runtime lock prevents duplicate execution
  • Continuously reclaim dangling / expired / orphaned leases

3.2 Judge Idempotency

  • Only unjudged successful runs are processed
  • Claimed runs cannot be double-judged

3.3 Use Recovery State Instead of Halt State

  • awaiting_judge
  • quota_wait
  • needs_rework

Runtime blocked state for planner issue-link ordering:

  • issue_linking

Convert state to recoverable, never abandon it.

3.4 Adaptive Escalation

  • Escalate to rework/autofix on repeated same failure signature
  • On merge conflict after approve, branch to conflict autofix task when possible

3.5 Event-Driven Progress Recovery

Recovery switch is event-driven, not fixed-time triggered:

  • Repeated same failure signature -> needs_rework / rework split
  • Non-approve circuit breaker -> autofix path
  • Quota failure -> quota_wait -> cooldown requeue
  • Missing judgable run -> restore awaiting_judge run context

4. Quota Philosophy

Treat quota pressure as recoverable external pressure, not terminal failure.

  • Single attempt may fail quickly
  • Task waits with explicit reason (quota_wait)
  • Continue cooldown retry until resources recover

5. Observability Requirements

Operators must observe not only trial results but also intended next steps:

  • Run-level failure
  • Task-level next retry reason/time
  • Backlog gate reason returned by preflight

6. Non-Goals

  • Maximizing first-attempt success at cost of recoverability
  • Fixed strict sequential processing when safe concurrency exists
  • Recovery flows that require manual intervention only
  • Recovery design that relies only on fixed-interval watchdog

7. Common Lookup Path (State Vocabulary -> Transition -> Owner -> Implementation)

This page describes design principles; for actual triage, follow state vocabulary -> transition -> owner -> implementation.

  1. state-model (state vocabulary)
  2. flow (transitions and recovery paths)
  3. operations (API procedures and operation shortcuts)
  4. agent/README (owning agent and implementation tracing)