aperture-lab-autoresearch

name	aperture_lab_autoresearch
description	Run Aperture Lab F-Stop on imported public trajectories, prepare review artifacts, compare disagreements, and propose bounded semantic-layer improvements under strict replay gates.
metadata	{"openclaw":{"requires":{"bins":["pnpm"]},"os":["linux","darwin"]}}

Aperture Lab F-Stop

Use this skill when the task is to run Aperture's offline semantic improvement loop on a remote worker or long-running harness.

This skill is only for the Lab path.

Do not use it for live runtime behavior or product-surface work.

Load First

Before doing anything else, read:

• packages/lab/research/autoresearch-program.md
• packages/lab/research/autoresearch-config.json

Those files are the source of truth for:

• allowed edit paths
• forbidden edit paths
• evaluation commands
• artifact chain
• expected outcomes
• non-goals

Follow them strictly.

Main Rule

Keep AI out of the hot path.

This loop may:

• import public trajectories
• prepare offline review artifacts
• run reviewer responses into disagreements and recommendation summaries
• promote selected disagreements into a frozen calibration corpus
• evaluate that corpus repeatably
• propose bounded semantic/importer changes
• run replay and release gates

This loop must not:

• change the live decision path to depend on AI
• edit planner or continuity logic
• auto-merge

Core Commands

Use the provider-neutral lab:fstop:* surface as the default operating path:

pnpm lab:fstop:run --provider <provider> --reviewer-provider <provider> --optimizer-provider <provider> --json
APERTURE_OPENCLAW_REVIEW_TIMEOUT=60 pnpm lab:fstop:review --dataset swe-smith --split tool --limit 3 --reviewer-provider <provider> --json
pnpm lab:fstop:propose --reviewer-provider <provider> --optimizer-provider <provider> --json
pnpm lab:fstop:cycle --json
pnpm lab:fstop:optimize --provider <provider> --json
pnpm judgment:battle
pnpm release:check

For the default short OpenClaw agent-run command, use:

pnpm lab:fstop:openclaw

If you need to debug one bundle manually, fall back to:

pnpm trajectory:import --dataset swe-smith --split tool --limit 3
pnpm lab:fstop:prepare --bundle <bundle-path> --json
pnpm lab:fstop:review:run --artifact <artifact-path> --reviewer-command "pnpm lab:fstop:reviewer --provider <provider>" --json

To freeze reviewer-backed disagreements into the repeatable optimization surface, use:

pnpm lab:fstop:promote --report <report-path> --split train --json
pnpm lab:fstop:evaluate --json
pnpm lab:fstop:cycle --json

Use the discovery loop to find candidate problems.

Use the frozen calibration loop to judge whether a patch actually helped.

The canonical reviewer adapter is:

pnpm lab:fstop:reviewer --provider <provider>

Supported providers:

• hermes
• openclaw
• generic

The adapter resolves the actual provider command from:

• APERTURE_HERMES_REVIEWER_COMMAND
• APERTURE_OPENCLAW_REVIEWER_COMMAND
• APERTURE_REVIEWER_COMMAND

For OpenClaw, the adapter can also invoke the local openclaw binary directly when no override command is configured. Use these env vars to tune that path:

• APERTURE_OPENCLAW_BIN
• APERTURE_OPENCLAW_AGENT
• APERTURE_OPENCLAW_REVIEW_SESSION_ID
• APERTURE_OPENCLAW_REVIEW_THINKING
• APERTURE_OPENCLAW_REVIEW_TIMEOUT

By default it uses a fresh OpenClaw session id per review and avoids the shared main agent session unless APERTURE_OPENCLAW_AGENT is explicitly set.

The underlying reviewer command must:

• read the reviewer prompt on stdin
• write valid JSON to stdout
• exit non-zero on failure

If the provider supports skills, use the dedicated reviewer role:

• skills/aperture-lab-reviewer/SKILL.md

For the code-editing phase, use the dedicated optimizer role:

• skills/aperture-lab-optimizer/SKILL.md

The canonical optimizer adapter is:

pnpm lab:fstop:optimizer --provider <provider>

It resolves provider-specific optimizer commands from:

• APERTURE_HERMES_OPTIMIZER_COMMAND
• APERTURE_OPENCLAW_OPTIMIZER_COMMAND
• APERTURE_OPTIMIZER_COMMAND

For OpenClaw, the adapter can also invoke the local openclaw binary directly when no override command is configured. Use these env vars to tune that path:

• APERTURE_OPENCLAW_BIN
• APERTURE_OPENCLAW_OPTIMIZER_AGENT
• APERTURE_OPENCLAW_OPTIMIZER_SESSION_ID
• APERTURE_OPENCLAW_OPTIMIZER_THINKING
• APERTURE_OPENCLAW_OPTIMIZER_TIMEOUT

The unattended optimizer entrypoint is:

pnpm lab:fstop:optimize --provider <provider> --json

It should be run from a clean worktree.

The unattended proposal entrypoint is:

pnpm lab:fstop:propose --reviewer-provider <provider> --optimizer-provider <provider> --json

Prefer this when the worker should go all the way from:

• discovery batch
• to repeated-signal selection
• to candidate calibration promotion
• to a reviewable patch proposal

The unattended top-level runner entrypoint is:

pnpm lab:fstop:run --provider <provider> --reviewer-provider <provider> --optimizer-provider <provider> --json

Prefer this when the provider should manage repeated proposal slices on its own instead of waiting for manual slice selection.

Output Expectations

Good runs should produce:

• review artifacts under packages/lab/results/offline-review/requests
• reviewer-filled artifacts under packages/lab/results/offline-review/responses
• reviewer prompts under packages/lab/results/offline-review/prompts
• raw reviewer outputs under packages/lab/results/offline-review/raw
• disagreement reports under packages/lab/results/offline-review/disagreements
• recommendation summaries under packages/lab/results/offline-review/recommendations
• run summaries under packages/lab/results/offline-review/runs
• calibration cases under packages/lab/calibration
• calibration reports under packages/lab/results/autoresearch/evaluations
• optimization briefs under packages/lab/results/autoresearch/briefs
• optimizer prompts, raw outputs, patches, and run summaries under packages/lab/results/autoresearch/optimizer
• proposal artifacts under packages/lab/results/autoresearch/proposals
• candidate bounded code changes only on the allowed edit surface
• a clear pass/fail result from the gates

Artifact Chain

Execute the loop in this order:

1. input bundle
2. Aperture replay output
3. prepared review artifact
4. reviewer prompt
5. raw reviewer output
6. reviewer-filled artifact
7. disagreement report
8. recommendation summary
9. promoted calibration case
10. frozen calibration evaluation report
11. optimization brief
12. optimizer prompt
13. raw optimizer output
14. proposal artifact
15. optional bounded patch proposal
16. gated evaluation result

Optimization Target

Optimize for:

• better title extraction
• better summary extraction
• better semantic frame reads
• better tool-family reads
• better consequence-band reads

Do not optimize for:

• general product UX
• continuity behavior
• route churn for its own sake

When In Doubt

Prefer:

• narrower edits
• stronger replay safety
• frozen calibration evidence over one-off reviewer disagreements
• better disagreement artifacts

over:

• aggressive autonomous rewriting
• speculative planner changes
• touching unrelated repo areas