A command that advances a long-term project goal by exactly one planned work item.

Selecting the next goal increment, implementing it, running the project’s verification checks, and recording blocked work.

It prevents accidentally starting multiple increments and requires each item to be started and verified separately.

step — Command by zebbern

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods

npx agentmods add commands/zebbern/agent-collab/step

Clone the repo

git clone --depth 1 https://github.com/zebbern/agent-collab

Per session 8 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.

When invoked 711 The whole file, excluding the scripts and references it only reads on demand.

Security scan A 0 findings. Scan, not verified.

Origin original No closer match found in the catalogue.

Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

Model	Per session	Once invoked
Fable 5	$0.00008	$0.00711
Opus 5	$0.00004	$0.00356
Sonnet 5	$0.00002	$0.00142
Haiku 4.5	$0.00001	$0.00071

Measured 2d ago against content hash d418b5a12348, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

step scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

plugins/goal/commands/step.md · 51 lines

What it actually says

Advance the goal one increment. One increment per invocation — when step 7 completes, stop. Do not start another increment; repetition is the user's call (or /loop while they watch).

Run node "${CLAUDE_PLUGIN_ROOT}/scripts/goal-companion.mjs" next [slug] --json. If it refuses (goal blocked/done, an item already in progress, nothing todo), surface the reason verbatim and stop.
Announce the increment in one line — what it is and whether you intend to delegate it.
Run node "${CLAUDE_PLUGIN_ROOT}/scripts/goal-companion.mjs" start <slug> <itemId>.
Execute with judgment: trivial work stays local; otherwise delegate via the codex-delegation / cursor-delegation skills. Analysis and implementation are separate delegations when both are needed, within the goal's budget.perStepDelegations (advisory: say so if you exceed it).
Verify through the project's own gates (this repo: npm run verify) and land the change as a PR; the user merges.
If delegated work fails verification: refine the brief with the failure evidence and re-delegate once (it counts against the step budget). If it fails again, record the item as blocked: node "${CLAUDE_PLUGIN_ROOT}/scripts/goal-companion.mjs" record <slug> <itemId> --disposition blocked --notes "<evidence>"
Record the real disposition (merged with --pr <n> and --delegate, or discarded/blocked with --notes), show the one-line output of status, and stop.

Unattended (scheduled) operation

The same one-increment choreography can be driven without a human watching — by /loop or a Claude Code scheduled agent. Only the trigger changes; the architecture does not. Extra rails for that mode:

Branch naming. Do the increment's work on a branch named goal/<slug>/<itemId>, so a later wake can find its PR mechanically.
An unattended step never merges PRs. It opens the PR, leaves the item in-progress, and stops. Merging stays human.
Reconcile before stepping. On wake, if an item is in-progress, find its PR by head branch (gh pr list --head goal/<slug>/<itemId>) before doing anything else:
- merged → record <slug> <itemId> --disposition merged --pr <n>
- closed unmerged → record <slug> <itemId> --disposition discarded --notes "<why>"
- still open → stop and wait; do not start another increment.
Blocked halts the loop. A blocked goal stops every subsequent wake until a human resolves it — exactly as in attended mode.
Budget stays advisory, but every delegation is still announced into the session log, so the wake's transcript carries the same disclosure a watching user would have seen.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

2d ago First seen · 51 lines · 8 tokens per session scan A d418b5a12348

Subscribe to this mod's changes

step is a command published in the GitHub repository zebbern/agent-collab (32 stars, last pushed 16d ago), licensed Apache-2.0. It adds 8 tokens to every session and 711 once invoked, about $0.0000 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

agents-team ai ai-agents ai-agents-cli ai-agents-framework ai-tools

step

Getting it into your agent

What it costs to keep this loaded

Grade A, and why

Nothing flagged

What it actually says

Unattended (scheduled) operation

What this file has done since we first saw it

Other commands, from other repositories

design-review

setup-pm-skills

statusbar-style

fest-show

config

scan