Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
8 changes: 4 additions & 4 deletions .agents/skills/arena/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ Every candidate gets the same prompt, so the prompt is the contract.

1. State the artifact each candidate is producing.
2. Derive the rubric, 3-6 concrete gradeable criteria for what success looks like on *this* task. It is the picker's tool in Phase D. Candidates only see the task.
3. Pick the runners. Use the `arena runners` line in `~/.agents/pstack-models.md`. If the rule or that line is missing, default to one each on `claude-opus-5-5-xhigh` and `grok-4.7-xhigh-fast`. An `auto` or `inherit-parent` entry in this line or the cross-judge line means the parent model, so omit `model` for it. If the Task tool rejects a configured entry, run that seat on its family's default and say so. Families go by prefix: `claude-*`, and `grok-*` with `cursor-grok-*` and `composer-*`. A refused `composer-*` entry runs on `grok-4.7-xhigh-fast`. With no family match, use `claude-opus-5-5-xhigh`. If it rejects a default, use the closest valid slug of the same family from its error message. Spawn more runners when the arena covers multiple design directions. Use the same model N times when the work is generation-bound rather than judgment-sensitive. Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section of the **poteto-mode** skill.
3. Pick the runners. Use the `arena runners` line in `~/.agents/pstack-models.md`. If the rule or that line is missing, default to one each on `claude-opus-5-5-xhigh` and `grok-4.7-xhigh-fast`. An `auto` or `inherit-parent` entry in this line means the parent model, so omit `model` for it. If the Task tool rejects a configured entry, run that seat on its family's default and say so. Families go by prefix: `claude-*`, and `grok-*` with `cursor-grok-*` and `composer-*`. A refused `composer-*` entry runs on `grok-4.7-xhigh-fast`. If the family has no default, or the Task tool rejects that default, use the closest valid slug of the same family from its error message, or drop the seat and say so when there is none. If dropped seats leave fewer than two runners, add runners until there are two, or one when the line has one entry. Run each added runner on a remaining runner's model, or omit `model` so it runs on the parent model when none remain, and say so. Spawn more runners when the arena covers multiple design directions. Use the same model N times when the work is generation-bound rather than judgment-sensitive. Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section of the **poteto-mode** skill.
4. Give each candidate its own output location (a git worktree where possible, otherwise `/tmp/arena-<slug>/candidate-<n>/`), per the **separate-before-serializing-shared-state** principle skill.

## Phase B: Fan out
Expand All @@ -26,11 +26,11 @@ If a candidate produces no output, proceed with N-1 and note the dropout.

## Phase C: Cross-judge

After all Phase B candidates complete, choose one model from the `arena cross-judge pool` line in `~/.agents/pstack-models.md`. If the rule or that line is missing, choose from `claude-opus-5-5-xhigh` and `grok-4.7-xhigh-fast`. Prefer a different model family from the parent's. Spawn one readonly judge subagent on that model. It sees the rubric and the candidates by path label, scores each criterion, and recommends a base with rationale. It runs in parallel with your reading in Phase D. Don't spawn it while candidates are still writing.
After all Phase B candidates complete, choose one model from the `arena cross-judge pool` line in `~/.agents/pstack-models.md`. If the rule or that line is missing, choose from `claude-opus-5-5-xhigh` and `grok-4.7-xhigh-fast`. Spawn one readonly judge subagent. The judge is never from the parent's family, as Phase A step 3 defines families. Try only the entries from another family, in pool order. An `auto` or `inherit-parent` entry is the parent model, so it never judges. With the default pool, a Claude parent gets `grok-4.7-xhigh-fast`. A parent from any other family gets `claude-opus-5-5-xhigh` first. If the Task tool rejects an entry, use a valid slug of the same model at the closest level or speed from the error message, or move to the next entry when there is none. If no entry from another family runs, try the default pool's entries from another family the same way. If none of those runs, skip the judge, pick the base alone in Phase D, and say so. If you do not know the parent's model, try the pool's real slugs in pool order and say that the family check did not run. The judge sees the rubric and the candidates by path label, scores each criterion, and recommends a base with rationale. It runs in parallel with your reading in Phase D. Don't spawn it while candidates are still writing.

## Phase D: Pick a base

Read every candidate end to end before picking. Score each against the rubric criterion by criterion, not on holistic feel, and compare with the cross-judge. Agreement on the base confirms the pick. Disagreement means one of you is biased or the rubric was ambiguous. Read both rationales before deciding.
Read every candidate end to end before picking. Score each against the rubric criterion by criterion, not on holistic feel, and compare with the cross-judge when Phase C ran one. Agreement on the base confirms the pick. Disagreement means one of you is biased or the rubric was ambiguous. Read both rationales before deciding.

Pick the base a future maintainer can extend most easily without breaking invariants. When two feel tied, prefer the cleaner boundary or smaller API, per the Laziness Protocol.

Expand All @@ -46,4 +46,4 @@ Verify the synthesized artifact like any other output, per the **prove-it-works*

## Outputs

One synthesized artifact, with a short synthesis note alongside naming the base and why, the cross-judge's verdict, each graft with its source candidate, what was rejected and why, any convergence or dropouts, and the verification result.
One synthesized artifact, with a short synthesis note alongside naming the base and why, the cross-judge's verdict or that Phase C skipped the judge, each graft with its source candidate, what was rejected and why, any convergence or dropouts, and the verification result.
2 changes: 1 addition & 1 deletion .agents/skills/create-formal-verification-skill/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ Pick the tool that answers the question at the lowest cost: a model checker, a p

Write a project-local skill with frontmatter that names the area and says when to use it. Without frontmatter, the skill never registers. From the skill alone, the next agent must be able to tell whether a change touches the verified code, run the check, tell a real bug from a modeling error, and update the verification when a change is intended. Put the coverage record and the sync commit in the skill, and tell the next agent to move the sync commit forward only after it confirms that the verification describes the code at the new commit. Use this repo's real commands, paths, and nouns, and leave no placeholders.

If the check is a TLA+ model, tell the next agent to read every result with the pstack [`/statechart`](../statechart/SKILL.md) skill. It shows each invariant violation the checker finds as its shortest path in plain words, and which of those violations a change fixes. When the check passes, `/statechart` draws a chart of how the system moves between states. Refer to `/statechart` by name. Its commands, options, and file paths change between releases and machines, so copy none of them into the generated skill or its scripts.
If the check is a TLA+ model, tell the next agent to read every result with the pstack [`/statechart`](../statechart/SKILL.md) skill. It shows each invariant violation the checker finds as its shortest path in plain words. When the check passes, `/statechart` draws a chart of how the system moves between states. Refer to `/statechart` by name. Its commands, options, and file paths change between releases and machines, so copy none of them into the generated skill or its scripts.

## Prove it before you hand off

Expand Down
71 changes: 14 additions & 57 deletions .agents/skills/interrogate/SKILL.md
Original file line number Diff line number Diff line change
@@ -1,75 +1,32 @@
---
name: interrogate
description: "Use for \"interrogate\", \"adversarial review\", \"multi-model review\", \"challenge this\", \"stress test this code\", \"find blind spots\", or \"tear this apart\". Multiple LLM reviewers challenge changes from independent angles."
description: "Use for \"interrogate\", \"adversarial review\", \"multi-model review\", \"challenge this\", \"stress test this code\", \"find blind spots\", or \"tear this apart\". Adversaries try to break a change or design, and only findings proven by a run survive."
---

Before following this workflow, read the [Codex adapter](../poteto-mode/references/codex.md) once per session. Its tool, model, and host mappings apply to the upstream terminology below.

# Interrogate

Spawn one reviewer per configured model to adversarially review code changes. The adversarial signal comes from model diversity, not assigned personas. The deliverable is a synthesized verdict. Do NOT auto-apply changes.
Adversaries try to break a change or a design. A finding survives only when a run proves it. Report the findings. Do not apply fixes.

## Step 1, Scope
## Frame

Identify what to review from context:
The scope is what the user names. Otherwise it is the branch diff from its merge base with the base branch, through the working tree. State the intent with the **in-your-own-words** skill, from the user's message, the commits, and the PR description. Ask the user when the intent is unclear.

- If the user points at specific files or a diff, use that.
- If on a feature branch, run `git diff main...HEAD` (or the appropriate base branch) for the full changeset.
- If the user's message references recent work, gather the relevant files.
## Fan out

Package the diff or file contents with any surrounding context files the reviewers need to understand the code.
Run Phase A steps 3 and 4 and Phase B of the **arena** skill, with four changes. Read the `interrogate reviewers` line in place of `arena runners`. Copy the scope's uncommitted changes into each worktree. Spawn each adversary as `generalPurpose`. In place of the arena task and rationale, give each adversary the scope, the intent, and this brief:

## Step 2, Intent
> Break this change. Prove each finding with a run against the real code, such as a failing test, a repro script, or command output that shows the wrong behavior. For a design or plan, run a prototype or probe the current tree. Keep repros in your own worktree. Report each finding with a title, the exact command, the output it printed, and the output the intent requires. Leave out anything you did not run. That covers style, naming, structure, and risks that need a future change to matter. No findings is a valid result.

Before spawning reviewers, state the intent in one clear paragraph from the user's message, commit messages, the PR description if one exists, and the code. If you're unsure about the intent, ask the user before proceeding.
## Verify

## Step 3, Spawn Reviewers
Verify each finding per arena Phase F by rerunning its command. Keep a finding when the output shows the failure and the failure breaks the intent. Drop the rest. Merge findings that share a root cause.

Launch all reviewers in a single message with the Task tool, one per entry on the `interrogate reviewers` line in `~/.agents/pstack-models.md`, labeled Reviewer A, B, and so on to match the entry count. If the rule or that line is missing, use the table defaults.
Skip arena Phases C, D, and E. Adversaries find different failures, so the verdict keeps every verified finding, not one adversary's list.

| Subagent | Default model |
|----------|---------------|
| Reviewer A | `claude-opus-5-5-xhigh` |
| Reviewer B | `grok-4.7-xhigh-fast` |
## Output

Each reviewer gets `subagent_type: generalPurpose`, `readonly: true`, and `model` set to its configured entry, or its table default when there is no configured line. Omit `model` for an `auto` or `inherit-parent` entry, so that reviewer runs on the parent model.

If the Task tool rejects a configured entry, run that reviewer on its family's table default and say so. Families go by prefix: `claude-*`, and `grok-*` with `cursor-grok-*` and `composer-*`. A refused `composer-*` entry runs on Reviewer B's default. With no family match, use Reviewer A's default. If it rejects a table default, spawn with the closest equivalent from the valid slugs in its error message (prefer the same family and reasoning tier), and open a separate PR to update the default table. Do not block the review on the slug issue. Never treat an alias entry as a rejected slug or apply either fallback to it.

Read `references/reviewer-prompt.md` and fill in the same template for every reviewer with the stated intent, the diff or file contents, the rubric from `references/rubric.md`, and the code-quality lens from `references/code-quality-review.md`. Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section of the **poteto-mode** skill.

## Step 4, Synthesize

As results come back, build a unified picture. Parse every reviewer's findings. Merge findings that describe the same issue differently, and note which models raised each one. Findings two or more models raised independently are the highest signal. Read a lone model's finding, but weight it accordingly. Note disagreements. If one model flags something and another explicitly says the opposite, that is context for the verdict.

## Step 5, Lead Judgment

You are the lead reviewer, a pragmatic senior engineer, not a neutral aggregator. Read `references/lead-judgment.md` for the full framework. Put every finding in one of four buckets, with the model(s) that raised it and a one-line rationale.

- **Act on**. Real issues affecting correctness, security, or maintainability given the actual goals. They would block a real PR.
- **Consider**. Legitimate, but you're not sure they outweigh the cost of addressing them now. Worth the user's attention.
- **Noted**. Technically valid but not actionable. Context-dependent, premature optimization, or low-impact at this stage.
- **Dismissed**. Wrong, nitpicky, or missing context, with a brief reason.

## Output Format

### Intent
> [The stated intent paragraph from Step 2]

### Reviewers
- Reviewer [label]: [model name], [N findings] (one bullet per reviewer)

### Act On
[Each: description, which models raised it, why it matters.]

### Consider
[Each: description, which models raised it, the tradeoff.]

### Noted
[Brief list.]

### Dismissed
[Each with a brief rationale.]

### Agreement Map
[Where models agreed and diverged, and what that pattern tells us.]
- **Intent**: the restatement.
- **Findings**: for each verified finding, the command, its output, how it breaks the intent, and the adversaries that found it. Say "no verified findings" when none survive.
- **Dropped**: for each adversary, the count of dropped findings and why.
47 changes: 0 additions & 47 deletions .agents/skills/interrogate/references/code-quality-review.md

This file was deleted.

Loading
Loading