AI Agent Workflow Collection

Triage a PR's review comments and failing CI (plus merge conflicts) through the /age lens, deciding which claims are worth acting on. Use when the user says "respond to PR comments", "handle review feedback", "affinage the PR", "/affinage ", "fix the failing build", "resolve the conflicts and respond". Fetches inline + review-body comments via gh and CI failures via affinage.pyz pr-status, grades each through the ten age dimensions, and writes a report at .cheese/affinage/pr- .md. Resolves merge conflicts via /melt and routes build/CI failures to fixes through /cure. When invoked standalone (no upstream handoffcontext) it also runs /age over the PR diff and folds fresh findings in; when chained, it skips that pass. By default it auto-applies the recommended set via /cure and posts the drafted replies, gating only on sprawling/structural fixes or conflicts; --safe re-gates. Supports --auto --stake . Before /cure; parallel to /age.

This skill should be used whenever the user asks to research, look up, compare, or investigate something external to the immediate codebase — phrases like "research X", "look up the API for Y", "compare libraries", "what does the doc say about Z", "find examples of how to do W", "is this library maintained", or "before I implement, what's the right approach". Routes the question across library docs (Context7), web research (Tavily), local code patterns (cheez-search), and GitHub examples (gh), then synthesizes with explicit confidence. Use even when the user only mentions a library name without saying "research" — when in doubt, briesearch first so the spec or implementation is informed, not speculative.

Use when the user has an approved spec that decomposes into 5+ independent behavioural curds and wants the whole pipeline run in parallel with reviewable PRs at the end — phrases like "/cheese-factory .cheese/specs/ .md", "send through the factory", "parallelize this spec", "many curds", "fan out the implementation", "cheese-factory this". Runs inline and spawns full-peer general-purpose sub-agents per phase (decomposer, per-curd workers, wiring, post-merge review, PR planner), ending in 1–N reviewable PRs via /pr-stack. Use even when the user mentions /fromagerie — /cheese-factory is the portable harness-agnostic sibling. Supports --hard propagation and --resume to continue a crashed pipeline. Do NOT use for single coherent specs (/cook, /ultracook), fuzzy planning (/mold), review-only work (/age), or specs with fewer than 5 curds (/ultracook).

Use when the user asks to read, view, show, open, or display the contents of a file or directory — phrases like "read src/auth.ts", "show me this file", "what's in this directory", "view lines 44-89", "look at the imports". Replaces cat / head / tail / less / more / bat / ls / tree / eza / find / fd / Read / Glob with AST-aware tilth MCP reading and file listing. Use even when the user says "cat", "less", "bat", "tree", "ls", "find", "fd", or "open the file" — never call host Read, Glob, or any shell file viewer / lister directly. If tilth MCP is unavailable, stop and report rather than fall back. Do NOT use for searching symbols or text (use cheez-search), editing code (use cheez-write), or git/gh operations.

Use when the user asks to find a symbol, definition, caller, import, or text pattern in the codebase — phrases like "where is X defined", "what calls Y", "find all usages of Z", "trace this function", "find the TODO comments", "search for this error string". Replaces grep / rg / ripgrep / ag / ack / find / fd with AST-aware tilth MCP search for name-shaped and text-shaped queries; use ast-grep (sg) only for AST-shape patterns with metavariables tilth cannot express. Use even when the user says "grep", "rg", "ripgrep", "ag", "ack", "fd", or "find" — never call host Grep, Glob, ripgrep, ast-grep, or any shell search directly. If tilth MCP is unavailable, stop and report rather than fall back. Do NOT use for reading whole files (use cheez-read), editing code (use cheez-write), or running tests/builds.

Use when the user asks to edit, replace, modify, update, change, delete, or insert code in a file — phrases like "replace this function", "delete lines 44-89", "update validateToken", "change the implementation", "add this import", "fix this bug" (when fixing requires editing), or apply a cross-cutting structural change (codemod) like "rewrite every X to Y". Replaces sed / awk / perl -i / patch / tee / Edit / Write / shell redirects (>, >>) with hash-anchored tilth MCP edits for one-off block changes; sanctions sg --rewrite (ast-grep) for structural codemods that span many files. Always read first via cheez-read to get hash anchors for anchored edits. Prefer surgical anchored edits over whole-file rewrites. If tilth MCP is unavailable, stop and report rather than fall back. Do NOT use for reading files (cheez-read first to get anchors), searching code (use cheez-search), or running tests/builds.

Use when the user wants to think out loud, rubber-duck a design, walk through trade-offs, or explore an ambiguous problem WITHOUT producing production code or specs — phrases like "let's talk through X", "rubber duck this with me", "I'm trying to decide between A and B", "help me think about Y", "what would happen if we...", "/culture". Output is conversation; the only sanctioned artifact is an opt-in .cheese/notes/ .md handoff slug at session end if the user asks for notes. Culture never writes to production code, never commits, never opens PRs. Use when the user wants shared mental model first; if the dialogue reveals real work to do, recommend /mold (fuzzy → spec) or /cook (clear ask → code) and stop. Before /mold or /cook.

Use when the user has an /age report (or any list of review findings, CI failures, or a "fix these" instruction) and wants the selected items resolved — phrases like "fix these findings", "/cure ", "address the high-stake items", "act on the age report", "fix the failing CI", "apply the cleanup". Loads the report, gates on explicit user selection, applies focused fixes via cheez-write, runs the project's existing test/lint/build gates, and produces a shipping-ready summary. Supports --auto --stake (propagated from /cook --auto) for the autonomous chain (see Auto mode). Use even when the user just says "fix it" if a review report or finding list is in scope. Default selection is empty — never apply everything implicitly. After /age; loops back to /age --scope for re-review or hands off to /gh to ship.

Use when the user wants a metacognitive vibecheck gate before code is shared for review or a pull request is opened — phrases like "/hard-cheese", "/cheese --hard", "gate this before I push", "vibecheck me", "make sure I understand this diff", "epistemic-debt check". Reads the working diff, asks the human author to explain its causal logic in their own words, spawns a fresh-context judge sub-agent to grade the explanation against the SOLO Taxonomy (pass ≥ Multistructural), and either accepts (PASS) or returns Socratic feedback for retry (FAIL, capped at --socratic-cap N). Writes the audit trail to .cheese/hard/ .md. Use standalone before opening a PR, or as the propagated --hard flag on /cook, /press, /age, /cure so the gate fires at the share-for-review handoff. Do NOT use for code review (/age), test hardening (/press), or fix application (/cure) — those review the artifact; hard-cheese reviews the human's understanding of it.

Use when a git merge, rebase, or cherry-pick has produced conflicts and the user wants them resolved — phrases like "melt the conflicts", "fix the merge conflicts", "resolve the rebase conflicts", "what's conflicting after the merge", "/melt", "fix the cherry-pick", or any prompt that surfaces <<<<<<< markers, CONFLICT (...) git output, or a half-finished merge state. First checks for squash-merge residue (rebase doomed by a squash-merged PR — abort and re-cherry-pick), then runs the structural-merge cascade — mergiraf (AST-aware auto-resolve) → git rerere (replay remembered fixes) → kdiff3 (manual fallback) — with helper scripts for batch resolution, ours/theirs picks, and lockfile regeneration. Use even when only one file is conflicting if the user wants the structural pass attempted before manual editing. Do NOT use for general git operations without conflicts. After /cook or /cure if a merge step blocked them; before retrying the gate that surfaced the conflict.

Use when the user has a fuzzy idea, half-formed feature, or design direction and wants to converge on a spec — phrases like "let's design X", "I'm thinking about Y", "what should the API for Z look like", "shape this into a spec", "I want to add a feature that...", "/mold". Runs an iterative dialogue (Explore / Ground / Shape / Sketch / Grill / Diagnose), grounds every critical claim with cheez-search or briesearch, locks public seams in pseudocode, and only writes a spec to .cheese/specs/ .md after an explicit approval gate. Use even when the user is "just thinking out loud" if they want the dialogue to leave behind a written artifact — for pure no-write thinking, route to /culture instead. After /culture (optional); before /cook.

This skill should be used right after /cook produces green changes, when the user wants the test surface hardened before review or shipping — phrases like "press the changes", "harden this", "check coverage", "strengthen the tests", "are the tests good enough", "press before /age", "/press". Reads the spec + cooked diff, maps changed behavior to tests, finds weak assertions and missing boundaries, adds focused hardening tests, writes a press report to .cheese/press/ .md, and prompts /age next. Supports --auto (propagated from /cook --auto) to skip its handoff and chain straight into /age --auto when readiness is ready for /age or follow-up recommended; only blocked halts the auto chain. Use even when the user wants to "tighten things up" before review. Do NOT use to add broad new behavior — only corrective fixes that hardening tests force. After /cook; before /age → /cure.

Use when the user has an approved spec or unambiguous task with a high-blast-radius footprint and wants the whole pipeline (cook → press → age → cure → age → cure → age, all --auto) to run forward in fresh-context isolation per phase — phrases like "/ultracook .cheese/specs/ .md", "ultracook this", "run the full pipeline in isolation", "fresh-context auto chain", "send it through the cave", "pipeline with no contamination". Runs inline and spawns full-peer general-purpose sub-agents (one per phase), each invoking the phase skill with --auto and writing its handoff slug under .cheese/ / .md. Sub-agents inherit the parent's model and full tool / MCP / skill access. Use when the user wants /cook --auto's autonomous chain semantics but with each phase reasoning blind to prior phases. After /mold; ends after the third age spawn (or earlier on halt / early-stop).