Skill reference¶
Start with Usage tiers to choose a workflow. This page is a lookup for individual tools; you do not need to learn them all.
Commands below use Claude Code syntax. In Codex, replace /mycelium-extra:
with $mycelium-extra:. In OpenCode, use the same /mycelium-extra: commands
(@grill also works) and send approve plan <hash> the same way; the plugin
adapter handles it. See Host compatibility.
Common prompts¶
Need |
Command |
|---|---|
Make a plan |
|
Enable execution checks |
|
Check a completed run |
|
Create an analysis folder |
|
Resume in a fresh session |
|
Park a passing idea |
|
Choose the next idea |
|
grill¶
Reads project evidence and turns your question into a sourced, numbered plan.
It asks only consequential questions you own, at most five after the opening
question. It calls decision-status and data-contract-check when needed.
It writes nothing and runs no analysis.
The brief ends with READY, READY_WITH_ASSUMPTIONS, or DECISION_REQUIRED.
Its Inputs: line identifies files the gate will pin. With the gate enabled,
use the displayed approve plan <hash> message to authorize covered runs.
See Approving a plan.
init¶
Turns on the approval gate for the current repository. Writes
.mycelium-extra/gate.json and adds .mycelium-extra/ to .gitignore;
leaves an existing gate unchanged. Default paths are analysis/** and nbs/**.
See Gate settings to change coverage.
verify¶
Checks an approved plan against receipts, pinned inputs, outputs, lint, and
explicit claims. It never re-runs the analysis. Requires the approval gate
for a current plan’s report; verify stale, status and runs can read saved
provenance without it. Other subcommands (list, diff, multiplicity,
explore, and --require-reportable) are described in
Verification reference; stale, status, diff and runs are in
Advanced features.
What the report shows¶
Planned scripts and their run evidence, changed inputs, output attribution, lint findings, and whether declared claims match saved output cells. See Report details.
Status¶
CONFORMS means the checked record agrees; CONFORMS_WITH_GAPS means evidence
is missing; DOES_NOT_CONFORM means a blocking mismatch or failure was found.
A missing check never counts as clean. See Status details.
What it writes¶
After confirmation, saves the frozen plan, receipts, output inventory, lint,
environment records where available, and report under <analysis>/provenance/.
See Provenance files.
new-analysis¶
Creates a new folder with numbered step stubs, a Snakefile, run.sh, an analysis
doc, one PLAN.md, one TRACKER.md, and data/output/log/report locations.
Refuses non-empty folders and missing link targets; stubs stop until implemented.
Uses <NAME>.md in Mycelium projects and README.md elsewhere. Writes only the
new folder; Mycelium’s analyze handles manifest and memory updates.
See Full project workflow.
plan-review¶
Challenges a grill draft before approval. In Claude Code or OpenCode, the host
assembles a minimal packet for separate Codex engineering and Biomni biomedical critiques.
Before anything leaves the machine it runs a pre-send check and asks you to type send packet <digest>; Biomni uses cloud credits. The check and the consent mechanics are in Advanced features. Missing reviewers are reported unavailable.
Writes nothing and approves nothing (apart from the prepared packet folder the send route needs). Material revisions need a new approval. In Codex, this skill supplies the engineering critique only. See Host compatibility for reviewer requirements.
decision-status¶
Resolves which past Mycelium decision applies. Explicit supersession wins,
then matching scope; contested choices go to you. After confirmation, appends
a resolution to .living/decisions.md without editing old entries.
Usually called by grill when a plan depends on prior decisions.
data-contract-check¶
Compares a plan’s assumptions with CSV/TSV sample tables or H5AD obs.
Checks columns, cohort levels/counts, unique units, complete pairs, and batch
versus contrast nesting. Writes nothing. H5AD needs h5py; a missing dependency
is a gap. Never loosens the contract to make a failed check pass without agreement.
Usually called by grill when sample assumptions matter.
handoff¶
Writes a short root HANDOFF.md with the goal, current state, locked decisions,
dead ends, code pointers, and one exact next action. Replaces the previous
handoff and ends with a resume prompt. A grill plan still waiting for approval is
carried word for word: HANDOFF.md points to the gate’s stored copy (or holds the
whole brief when there is none), and the new session prints it unchanged for a
fresh approval card. Does not commit or change .living/.
idea¶
Parks a passing idea without a question, so the main work keeps its flow. Writes
todo/<slug>.md and one todo/TODO_REGISTRY.md row (priority idea, status
open) in Mycelium’s todo format, then replies with one line. Uses your words as
written. Writes only under todo/: no code runs, and a plan waiting for approval
stays approvable. Mycelium’s own todo-idea mode asks up to six questions first;
idea is the no-question path into the same registry.
idea-triage¶
Ranks the open ideas in todo/ so you can choose the next one. Each idea gets a
pitch, two impact ratings, the number of things that must go right, its riskiest
assumption (1 to 5) with months to test, a search of existing findings, and red
flags; the top idea gets a cheapest go/no-go test. It shows a ranked table with a
park all option and waits. After you pick, it records the triage in the item file
and marks the pick in-progress; it offers grill but does not start it. Based on
Fischbach and Walsh (2024), Cell.
harden¶
Turns one Mycelium learning’s concrete mitigation candidate into a test in the repository’s existing test setup. Demonstrates failure on a minimal reproduction and a pass on current code. After confirmation, marks the learning’s mitigation as structural and records the test path. Leaves analysis code unchanged.