AutoSkillit

Analyze issue dependencies and produce a dispatch execution map for parallel orchestration

Execute recipe sessions batch-by-batch for triaged GitHub issues. Reads the triage-issues output manifest, processes each batch sequentially, and launches the appropriate recipe for each issue. Use when user says "process issues", "run issues", or "execute pipeline for issues".

Validate review-decisions audit findings with mandatory intent analysis and seven evidence-gathering rules. Adds docstring-as-contract recognition, deliberate-change detection, test-as-intent-signal, consumer-impact verification, architectural feasibility checks, behavioral simulation, and symmetry-as-design recognition to the standard validation workflow. Use when validating reports from audit-review-decisions specifically.

Generate YAML recipes for .autoskillit/recipes/. Use when user says "make script skill", "generate script", "script a workflow", "write a script", "create a script", "new recipe", "write a pipeline", or when loaded by other skills for script formatting.

Validate audit findings from audit-arch, audit-tests, or audit-cohesion against actual code, git history, and design intent using 9–10 parallel subagents. Removes contested findings, documents exceptions, adjusts severities. Use when user says "validate audit", "validate findings", "validate report", or "check audit results".

Validate audit findings from audit-arch, audit-tests, audit-cohesion, audit-feature-gates, audit-docs, or audit-review-decisions against actual code, git history, and design intent using 9–10 parallel subagents. Removes contested findings, documents exceptions, adjusts severities. Use when user says "validate audit", "validate findings", "validate report", or "check audit results".

Validate test audit findings with test-domain semantic rules and intent analysis. Adds import-path-as-contract recognition, precondition-as-assertion detection, provenance verification, split-era lifecycle awareness, and deletion-vs-improvement distinction to the standard validation workflow. Use when validating reports from audit-tests specifically.

Synthesize experiment results into a structured research report in the research/ folder. Supports --inconclusive flag.

Merge a single PR into the integration branch. For simple PRs, uses gh pr merge --squash --auto to enforce GitHub's required status checks. For needs_check PRs, re-assesses complexity and returns needs_plan=true with a conflict report when conflicts are detected. Use inside the merge-prs loop.

Orchestrates 2–4 vis-lens skills in parallel to produce a figure inventory (visualization-plan.md) and a report-placement outline (report-plan.md). Runs after design review GO, before worktree creation.

Goal-aware resolution of rebase conflicts when merging a conflict-resolution worktree back into the integration branch. Analyzes the intent of each side of a conflict, resolves it in-place when confidence is HIGH or MEDIUM, and escalates when LOW.

Validate an experiment plan before execution using triage-first, fail-fast dimensional analysis with an adversarial red-team. Emits verdict (GO/REVISE/STOP), experiment_type, evaluation_dashboard, and revision_guidance.

Create Methodology Norms visualization planning spec showing ML sub-area mandatory figures, community conventions, and coverage gaps. Methodology-Normative lens answering "Which domain-specific figures are expected by reviewers?"

Composition executor for pull requests. ALWAYS invoke this skill when instructed to compose a PR. Do not read prep files or create PRs directly — use this skill first to load the composition workflow.

Diagnostic executor for CI failures. ALWAYS invoke this skill when instructed to diagnose CI failures. Do not fetch CI logs directly — use this skill first to load the diagnosis workflow.

Deploy experiment artifacts in an isolated git worktree following an approved experiment plan, with per-phase commits.

Implementation executor. ALWAYS invoke this skill when instructed to implement a plan in a worktree. Do not read the plan or edit files directly — use this skill first to load the full implementation workflow.

Elaborate all work packages for a target phase via parallel L0 subagents. One invocation per phase; spawns one L0 per WP concurrently. (Pass 3 loop body)

Preparation executor for pull-request metadata. ALWAYS invoke this skill when instructed to prepare PR metadata. Do not read plans or classify files directly — use this skill first to load the preparation workflow.

Worktree retry executor. ALWAYS invoke this skill when instructed to continue or retry an implementation in an existing worktree. Do not resume editing files directly — use this skill first to load the retry workflow.

Execute a designed experiment in a worktree and collect structured results. Supports --adjust retry mode.

Plan validation executor. ALWAYS invoke this skill when instructed to validate or dry-walkthrough a plan. Do not read the plan or trace changes directly — use this skill first to load the validation workflow.

Worktree implementation executor. ALWAYS invoke this skill when instructed to implement a plan in a worktree with testing and merging. Do not read the plan or edit files directly — use this skill first to load the full implementation workflow.

Planning executor. ALWAYS invoke this skill when instructed to create, devise, or write an implementation plan. Do not explore the codebase or draft a plan directly — use this skill first to load the planning workflow.

Failure resolution executor. ALWAYS invoke this skill when instructed to fix test failures in a worktree. Do not read test output or edit code directly — use this skill first to load the failure resolution workflow.

Fetch PR review comments, run intent validation (ACCEPT/REJECT/DISCUSS) before applying fixes, and post inline replies. MCP-only — used exclusively by recipe orchestration via run_skill after review_pr reports changes_requested or needs_human verdict.

Automated diff-scoped PR code review using parallel audit subagents. Posts inline GitHub review comments and submits a summary verdict. Use after a PR is opened to gate CI on review approval.

Pre-flight environment gate for the research recipe. Reads the experiment plan, detects the required environment type, builds a Docker image or creates a host micromamba environment, and emits an env_mode verdict consumed by downstream steps.

Pre-flight resource gate for the research recipe. Reads the experiment plan's data_manifest, checks disk space and network connectivity for external/gitignored entries, creates data directory structure, and emits a PASS/WARN/FAIL feasibility verdict.

Analyze WP complexity per phase and emit consolidation group manifests for trivial WP merging (L1+L0 pattern)

Elaborate all assignments for a target phase via parallel L0 subagents. One invocation per phase; spawns one L0 per assignment concurrently. (Pass 2 loop body)

Refine elaborated work packages with cross-phase visibility via per-phase L0 subagents (L1+L0 pattern)

Audit the test suite for useless tests, consolidation opportunities, over-mocking, weak assertions, placement/organization issues, xdist safety violations, test path filter integrity, and other test quality issues. Use when user says "audit tests", "audit test suite", "review tests", or "test quality check". Generates an improvement plan in {{AUTOSKILLIT_TEMP}}/ with explanations for each proposed change.

Assess each work package for review-approach benefit before implementation. Writes review_approach_assessment.json; does NOT invoke review-approach.

Elaborate a single phase into a full result, parallel-safe — receives plan snapshot + target phase ID

Extract domain knowledge and naming conventions for planning context

Generate high-level phases from project analysis (Pass 1 entry point)

Refine elaborated assignments for a single phase via parallel L0 subagents (L1+L0 pattern), using per-phase context file with peer_summaries for cross-phase visibility

Refine elaborated phases with cross-phase visibility via parallel L0 subagents (L1+L0 pattern)

Targeted fix of validate_plan findings — re-elaboration, duplicate resolution, dependency corrections