workbench

Foundation audit mode. Walk a codebase as an adversarial building inspector - stress-test load-bearing logic to bank-grade, limit-test feature flows by actually running them, and file a fragility report backed by runnable repros. Maintains an AUDIT.md ledger at repo root across runs. Use when the user asks to audit the foundations, check whether things are foundationally strong or flaky, run a bank-logic check, limit-test or stress-test features after a heavy build sprint, question the codebase from first principles, or harden what the audit found ("report and fix"). Two modes - report (default) and report-then-fix.

Turn competing options and available evidence into a reasoned recommendation, with the tradeoffs, unresolved assumptions, and conditions for revisiting it. Use when the user asks which direction to take, what to prioritize, whether something is worth doing, to weigh alternatives, or to decide between opportunities. Works for product, architecture, operations, and other consequential choices. Conversational by default; a recommendation alone does not authorize implementation.

Enter codebase mastery mode before implementation, or refresh an existing map after changes. Use when the user asks to DEVOUR a repo, deeply onboard, map architecture, trace runtime flows, understand blast radius, refresh MAP.md, or establish what changed since the last study. Persists its atlas to MAP.md with revision, scope, and verification provenance so later runs can reuse valid understanding without claiming stale areas were rechecked.

Run a realistic user journey end to end and sweep-test the real subsystems behind it. Triage the whole journey, then test the risky stations with varied inputs, volume, and independent verification in an isolated reversible environment. Retain useful findings and reproductions; separate observed behavior from inferred user experience. Use when the user says gauntlet this flow, sweep-test the pipeline, simulate a user and stress the machinery, is X sound front to back, or run the gauntlet on a journey.

Reason about a whole system by mapping it onto a mature, structurally-similar domain that already paid for its mistakes, then read that domain's laws, invariants, and blindspots back onto the system. Use when the user wants the big-picture / first-principles view of an architecture, codebase, product, or vision; asks "what is this really like", "find the analogy", "what are we blind to", "think about this systemically / from first principles", "is this like X", or asks mid-conversation to "isomorph this" because the talk got too technical; or wants to pressure-test a design by inheriting a proven domain's failure modes. A thinking mode, not a file-producing workflow - with one exception - when the user adopts a twin as the project's design bible, offer to record it into VISION.md as the system shape.

Turn a vague idea, intuition, frustration, goal, screenshot, rough request, or incomplete brief into a grounded and execution-ready prompt for the right agent or tool. Use across software, debugging, architecture, infrastructure, operations, research, decisions, writing, product/UI, and creative work when the user knows what they are reaching for but has not yet expressed the complete or technically correct shape. Adapts its reasoning and output to the situation instead of forcing every request into a UI or coding template.

Capture a conversation's essence into the active host's memory directory, and recall it back like a colleague - not a database. Use at the tail end of a substantive session ("log this", "memory-scribe this"), progressively during one ("scribe as we go", "checkpoint this"), or in reverse when the user asks "where did we leave off", "what do you remember about this project", "pick up the thread". Supports Codex with a Workbench project-memory convention and Claude Code project memory without assuming the two plugin runtimes share state.

See what a codebase wants to become and which opportunities deserve attention. Ground possibilities in existing structure, assess who benefits, ongoing burden, opportunity cost, and smaller alternatives. Two modes - open (what could this become) and wish (whether and how the structure can grant a wish). Use when the user asks what is latent, what features could emerge, whether a capability is being used fully, or says run potential. Conversational; never implements or writes files within this mode.

Resolve a consequential uncertainty with the smallest useful experiment before committing to a larger direction. Use when the user asks to test an assumption, check feasibility, run a bounded spike, or find out whether an approach works before building it. Define success, failure, limits, and cleanup before execution; distinguish observed results from inference. Runs authorized reversible experiments and returns evidence to the decision, without expanding into full implementation.

Run a deep research dive on any topic and either produce a polished Quarto HTML report (official mode — files written to research/<umbrella>/<title>/) or reply with structured findings inline (scan mode — no files). Use when the user invokes $workbench:research-report, /workbench:research, /workbench:scan, or says "research X for me", "do a deep dive on X", "give me a quick read on X", "what's the state of X", "scan X for me", "write up findings on X". Covers maximum surface area across web, Reddit, GitHub, ProductHunt, docs, papers, and any other available sources. Always converges output to exactly notes.md + report.html (official mode) or a structured chat reply (scan mode). Never proliferates files.

The reality-check hotline. Mid-conversation, leave the closed room and go to the internet to ground what was just said against how the world actually does it. Two faces - verify (we concluded something confidently; pull the load-bearing claims and check each against real fetched sources - confirmed / wrong / oversimplified / outdated / no-consensus) and discover (we're stuck or starting fresh; find the dominant pattern and the gotchas). Use when the user says "scour this", "are we on the right page?", "is this actually how people do it?", "go check the internet", "make sure we're not confidently wrong", or asks how something is normally done. Fast and conversational like isomorph - never writes files, never answers from memory.

Recommend which of the user's ALREADY-INSTALLED skills best fits a task or query. Use when the user asks "what skill should I use for X", "which of my skills fits this", "do I have a skill for Y", "what's at my disposal for Z", "recommend a skill from my toolkit", or when you need to pick the right skill from their available toolkit before starting work. This skill is read-only — it advises, it does not install. For installing new skills, use the separate skill-management workflow available in the current agent environment.

Distill a successful chat, project workflow, repeated agent behavior, or a way of thinking into a reusable Codex or Claude skill. Use when the user wants to "skillify" a process, capture a reasoning/voice mode that just worked, create a new skill from what just worked, generalize a workflow without overfitting one project, or update a skillpack/plugin with reusable instructions.

The exhaustive-surface cartographer. Map the COMPLETE surface of any object — a resume, a domain, a dataset, a market, a decision space, a codebase's concerns — with PROVABLE coverage instead of vibes. Anchor against official exhaustive taxonomies, rotate enumeration generators, expand every example into its full basket recursively (the Orange Rule), hunt the underrepresented niche, and SHOW the whole tree with an audit. Use when the user says "/totality of X", "the totality of X", "map the full surface area", "what am I blindspotted on", "everything X encompasses", "give me ALL of it, exhaustively", or keeps saying "but what else is missing" after a normal answer. Companion: scour fetches unknown anchors; research-report carries totality-shaped deep dives.

Keep a project's vision written down and alive. Creates and maintains VISION.md at the repo root - the statement of what the building is trying to be, which bedrock, potential, devour, and any implementing agent read before judging, dreaming, or building. Three moments - birth (carry a warroom-style exploration's distilled intent into a new project repo), backfill (a project already alive with no vision artifact - read the building, interview the visionary, write it), and refresh (the artifact has drifted from current intent - reconcile and update). Use when the user says write the vision, vision this project, this repo has no vision doc, the vision is stale, refresh the vision, backfill the vision, or wants to hand a project off from exploration into development.