| name | wf-auto |
| description | Perpetual adaptive auto-optimization mode. Selects probes from project evidence instead of a fixed angle count. Inherits WF acceptance gates and subagent orchestration per cycle. Use for Claude /wf-auto, Codex $wf-auto. Explicit only — the user must type /wf-auto, $wf-auto, or /skills wf-auto. |
WF Auto - Perpetual Auto-Optimization
Memory Preflight
- Load
CLAUDE.md, Harness/MEMORY.md index only, then Harness/README.md
before planning, dispatch, edits, validation, or review.
- Load detailed
Harness/memory/* files only when MEMORY_PROTOCOL.md
scenario hints match; otherwise record "memory hints: none".
Load
CLAUDE.md
Harness/MEMORY.md (index only per Memory Preflight)
Harness/README.md
Harness/specs/workflows/WF-AUTO.md
Harness/specs/workflows/WF-AUTO-ANGLES.md
Harness/specs/runtime/subagents.md
Harness/specs/runtime/dispatch.md
Harness/specs/runtime/agent-workflow.md
.claude/skills/wf-review/SKILL.md
Cache Discipline
Follow Harness/specs/runtime/context-loading.md#Cache-First Context Contract: keep the
auto-mode docs in listed order, run selected probes only, append fresh probe
outputs last, and record compact evidence instead of carrying full logs between
cycles.
Trigger
- Claude
/wf-auto
- Codex
$wf-auto
Do not use unless the user explicitly invokes /wf-auto, $wf-auto, or /skills wf-auto. The phrases "auto mode", "never stop", "self-improve", "continuous optimize", "unbounded self-directed optimization" are NOT triggers — only an explicit command token enters WF-AUTO.
Do not use when the user gives a bounded task, requests maximum parallelism
(/wf-max), needs an urgent production hotfix, or the codebase is tiny enough
that auto scanning costs more than it helps.
Hard Rules
- Never stop except the Adaptive Coverage A-GATE: dynamic high-risk
obligations are covered, two different confirmation strategies are empty,
and unresolved uncertainty is recorded.
- CEO never edits production source. CEO may write only
Harness/tasks/auto/PLAN.md and Harness/tasks/auto/PROGRESS.md.
- Build a project profile each W0 cycle and dispatch only the selected probes;
invoke oracle and spark searchers when evidence justifies them.
- One accepted finding per cycle: <=3 files and <=50 changed lines. Larger
ideas escalate to
/wf or /wf-max, then return to auto.
- Every accepted cycle inherits the full WF chain:
Mini PRD -> AC IDs -> test/validation plan -> implementer -> verifier -> cross-review -> reflector PASS -> evidence ledger -> next W0.
- Review is mandatory: spec review, code-quality review, then reflector.
- Value Gate is scored for spark candidates: pass is >=18/25 and no dimension
below 3.
- Intent Checkpoint is adaptive: 2 -> 5 -> 10 cycles, exactly two questions.
- Record compact evidence per cycle; do not paste full logs or transcripts.
- A bounded test tick still creates or updates
Harness/tasks/auto/PLAN.md
and Harness/tasks/auto/PROGRESS.md; missing auto capsule evidence is a
failed cycle record.
Loop
W0: SENSE (adaptive probes + oracle + spark sources as triggered)
A-GATE: continue, oracle, spark, confirm, or stop
W1: PRIORITIZE one finding
W2-W5: Mini PRD -> AC -> test/validation plan -> implementer -> verifier -> cross-review -> reflector PASS
RECORD: evidence ledger
LOOP: next W0
Return
Report cycles run, findings addressed by source, evidence ledger path/summary,
exhaustion evidence if any, weak spark count, final state, and residual risks.