agentleFS
kody-wkody-w/rappterbook/.github/skills/rappterbook-gauntlet/SKILL.md

rappterbook-gauntlet

Skill18 starsChanged 6 months agoMITView on GitHub

Improve Rappterbook through a bounded, supervised reproduce-fix-challenge loop. Use when authorized to run an improvement gauntlet or supervised worker cycle.

The file

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 9 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈9

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Inherit evidence: Read the latest lab notebook and the previous cycle's actual outcome and next decision.

    line 19
  2. Distinctive:

    If none exists, establish a dated baseline.

    line 21
  3. Distinctive:

    Prefer newcomer friction, incorrect credit, incomplete readback, or a broken useful contribution over cosmetic churn.

    line 24
  4. Distinctive:

    Reproduce first: Run the relevant baseline tests and write the smallest failing regression.

    line 26
  5. Distinctive:

    Publish and feed back: Only the supervisor may publish authorized changes through normal PR gates.

    line 37
  6. Distinctive:

    Default to two nonoverlapping workers only when the operator requests subagents.

    line 44
  7. Distinctive:

    Keep shared files such as the lab notebook and supervisor workflow owned by the supervisor.

    line 53
  8. Very specific:

    Use scripts/rappterbook_gauntlet.py for the repeatable offline gate suite:

    line 73
  9. Distinctive:

    Stop a lane after a supported fix passes, two repair attempts fail, the objective requires new authority/private sources, or no worthwhile …

    line 96
  10. Distinctive:

    Never create an additional schedule merely by loading this skill.

    line 110

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Rappterbook improvement gauntlet

Success is a demonstrated improvement to useful participation, not a larger diff, more posts, or a worker's confidence. This workflow does not grant publication permission, start a fleet, or replace repository instructions.

Resolve the authorized kody-w/rappterbook checkout first. Commands below run from that checkout, not the personal skill directory. If the repository or existing test tooling is unavailable, report the concrete blocker rather than inventing a substitute runner or claiming a completed gate.

Loop

  1. Inherit evidence. Read the latest lab notebook and the previous cycle's actual outcome and next decision. State which evidence changes this run. If none exists, establish a dated baseline; do not invent prior feedback.
  2. Select a bounded challenge. Name an observable failure, its affected user, and the exact acceptance condition before changing implementation. Prefer newcomer friction, incorrect credit, incomplete readback, or a broken useful contribution over cosmetic churn.
  3. Reproduce first. Run the relevant baseline tests and write the smallest failing regression. A missing tool, unavailable credential or incomplete corpus is a blocker, not proof of a product bug.
  4. Repair one cause. Make the smallest complete fix in an isolated worktree. Preserve the counterexample and existing intended behavior.
  5. Challenge the repair. Try the neighboring failure cases and rerun the original regression plus the relevant existing suite. Do not weaken an assertion, delete a test or relax a safety boundary to obtain green output.
  6. Supervisor decision. Review the evidence and diff, integrate accepted candidates, and run the combined gates. Record accept, revise, reject or blocked with a reason. Passing tests alone is not independent adoption.
  7. Publish and feed back. Only the supervisor may publish authorized changes through normal PR gates. Verify submission, merge and live effects separately. Record the outcome and the next falsifiable challenge before beginning another cycle.

Worker contract

Default to two nonoverlapping workers only when the operator requests subagents. Otherwise run one lane directly. Use the host's approved agent configuration; do not create another model runtime.

  • Pin every worker to the same baseline commit in a separate worktree.
  • Assign exact owned paths and a specific user-facing objective.
  • Allow one issue per lane and at most two repair attempts in a cycle.
  • Workers do not spawn workers, push, merge, publish social content, change account identity or credentials, or operate private integration sources.
  • Keep shared files such as the lab notebook and supervisor workflow owned by the supervisor. Workers return their entry text rather than editing them.
  • A worker returns its hypothesis, baseline command/result, counterexample, changed files, regression result, neighboring cases, remaining uncertainty, and suggested next challenge. "No supported issue found" is a valid result.
  • The supervisor reuses the same worker for one focused revision rather than launching a replacement to get a more favorable answer.

Acceptance gates

GateRequired evidence
NeedA concrete affected workflow and reproducible failure
ImprovementThe original regression fails before and passes after
PreservationRelevant existing tests and neighboring cases pass
IntegrityNo invented authorship, double-counted edits, lost receipts or weakened consent
ScopeOnly owned, relevant files; no shared canonical-state mutation
SupervisionExplicit supervisor diff review and combined gate result
DeliveryActual PR/merge/readback evidence, not a local success claim

Use scripts/rappterbook_gauntlet.py for the repeatable offline gate suite:

python3 scripts/rappterbook_gauntlet.py --list
python3 scripts/rappterbook_gauntlet.py --lane onboarding --lane credit
python3 scripts/rappterbook_gauntlet.py --repo /absolute/worker/worktree
make gauntlet

The runner unions selected suites into one pytest invocation, excludes live tests and inherited test-selection options, and requires nonzero completed coverage with no failures or skips. Exit codes are 0 (passed), 1 (failed) and 2 (blocked). The default timeout is 300 seconds. It strips common credential environment variables; this is not an OS sandbox, so inspect candidate test code before running it. It does not compare baselines or assess usefulness: the supervisor must do those parts of the loop.

Its JSON report is test evidence, not an automatic approval or publication action. Store run reports in the host's session/artifact area, never canonical state/. Do not run make all or make clean as a validation shortcut.

Stop conditions

Stop a lane after a supported fix passes, two repair attempts fail, the objective requires new authority/private sources, or no worthwhile issue is supported by evidence. Do not widen the task to justify keeping workers busy. Return the strongest evidence and precise blocker; do not retry writes with a new key or empty ledger after an ambiguous remote effect.

Live social tests are not part of this gauntlet. Read-only checks may use authorized public sources, but offline fixtures must exercise writes. Existing posts, contributor work, service identities and private installations remain untouched. Distinguish direct-link visibility from API verification and feed distribution; preserve contradictory observations rather than forging success.

For a recurring run, carry forward the previous cycle's report and next decision. A scheduler is optional and host-dependent, not an always-on service. Never create an additional schedule merely by loading this skill.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.