القائمة
Run Aperture Lab F-Stop on imported public trajectories, prepare review artifacts, compare disagreements, and propose bounded semantic-layer improvements under strict replay gates.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
Apply bounded semantic-layer improvements for Aperture Lab F-Stop's frozen calibration corpus. Use when acting as the optimizer model inside the VPS F-Stop loop; edit only the allowed semantic/importer files, run the required gates, and return a structured JSON outcome.
Run Aperture Lab F-Stop's provider-neutral runtime loop. Use when operating the self-hosted reviewer/optimizer harness against OpenClaw, Hermes, or another compatible runtime through the canonical lab:fstop:* commands.
Run Aperture Lab F-Stop, the offline discovery, calibration, and proposal subsystem for Aperture. Use when operating the product-facing lab:fstop:* command surface across OpenClaw, Hermes, or another compatible runtime.
Review Aperture Lab F-Stop offline review prompts and return JSON-only findings about title, summary, status, intentFrame, toolFamily, and consequence. Use when acting as the reviewer model inside the Aperture Lab F-Stop loop; do not edit code and do not return prose outside the required JSON.
استنادا إلى تصنيف SOC المهني
| 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"]}} |
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.
Before doing anything else, read:
packages/lab/research/autoresearch-program.mdpackages/lab/research/autoresearch-config.jsonThose files are the source of truth for:
Follow them strictly.
Keep AI out of the hot path.
This loop may:
This loop must not:
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:
hermesopenclawgenericThe adapter resolves the actual provider command from:
APERTURE_HERMES_REVIEWER_COMMANDAPERTURE_OPENCLAW_REVIEWER_COMMANDAPERTURE_REVIEWER_COMMANDFor 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_BINAPERTURE_OPENCLAW_AGENTAPERTURE_OPENCLAW_REVIEW_SESSION_IDAPERTURE_OPENCLAW_REVIEW_THINKINGAPERTURE_OPENCLAW_REVIEW_TIMEOUTBy 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:
If the provider supports skills, use the dedicated reviewer role:
skills/aperture-lab-reviewer/SKILL.mdFor the code-editing phase, use the dedicated optimizer role:
skills/aperture-lab-optimizer/SKILL.mdThe canonical optimizer adapter is:
pnpm lab:fstop:optimizer --provider <provider>
It resolves provider-specific optimizer commands from:
APERTURE_HERMES_OPTIMIZER_COMMANDAPERTURE_OPENCLAW_OPTIMIZER_COMMANDAPERTURE_OPTIMIZER_COMMANDFor 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_BINAPERTURE_OPENCLAW_OPTIMIZER_AGENTAPERTURE_OPENCLAW_OPTIMIZER_SESSION_IDAPERTURE_OPENCLAW_OPTIMIZER_THINKINGAPERTURE_OPENCLAW_OPTIMIZER_TIMEOUTThe 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:
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.
Good runs should produce:
packages/lab/results/offline-review/requestspackages/lab/results/offline-review/responsespackages/lab/results/offline-review/promptspackages/lab/results/offline-review/rawpackages/lab/results/offline-review/disagreementspackages/lab/results/offline-review/recommendationspackages/lab/results/offline-review/runspackages/lab/calibrationpackages/lab/results/autoresearch/evaluationspackages/lab/results/autoresearch/briefspackages/lab/results/autoresearch/optimizerpackages/lab/results/autoresearch/proposalsExecute the loop in this order:
Optimize for:
Do not optimize for:
Prefer:
over: