gradient

name	gradient
description	Use Gradient itself: discover install state, initialize or upgrade repositories, operate the Work/Fleet/Context/Evidence/Policy loop, and know when repository work should flow through Gradient. Trigger: /gradient.
argument-hint	[status|init|work|capture|eval|close|upgrade]

/gradient

Gradient is the local agent control plane. In any initialized repository, use Gradient as the first-class interface for governed agent work instead of inventing parallel backlog, evidence, harness, or eval flows.

When To Use

Use this skill when you need to:

• check whether Gradient is installed and active;
• initialize a repository with Gradient;
• update an already initialized repository from the Gradient core checkout;
• inspect or claim local work;
• capture evidence for completed work;
• run Gradient evals and validation gates;
• close work into backlog.d/_done;
• explain the Gradient lifecycle to another agent or operator.

Lifecycle

Gradient preserves one lifecycle:

Intent -> Work Graph -> Fleet Run -> Evidence -> Policy/Eval -> Feedback

Every command should strengthen one part of that lifecycle. Do not bypass it by manually moving work, inventing one-off run logs, or treating test output as a substitute for an evidence packet.

Command Map

Discovery:

gradient status
gradient status --check
gradient config

Initialize or update a repository:

gradient init --profile solo-frontier /path/to/repo
gradient upgrade --dry-run /path/to/repo
gradient upgrade --apply /path/to/repo

Resolve and verify the active harness:

gradient resolve
gradient validate
gradient eval

Operate local work:

gradient work list --status all
gradient work next
gradient work show <id|path>
gradient work claim <id|path> <owner>

Capture and close work:

gradient capture backlog.d/<work-item>.md
gradient report --latest
gradient close backlog.d/<work-item>.md

Feedback intake:

gradient feedback report \
  --module Work \
  --classification work-adapter \
  --severity high \
  --summary "short public-safe summary" \
  --expected "expected behavior" \
  --actual "actual behavior" \
  --evidence "public-safe evidence" \
  --route backlog

Operating Rules

• In initialized repositories, prefer gradient over repo-local ad hoc harness, backlog, evidence, trace, or eval commands when Gradient has an equivalent command.
• Keep repo-owned state repo-owned: gradient.yaml, backlog.d, private source config, evidence, policy, feedback, and run artifacts are not blindly overwritten by upgrades.
• Keep Gradient-managed state managed: scripts, schemas, profiles, standards, evals, Go module files, and native harness primitives should be refreshed through gradient upgrade.
• Run gradient validate after changing profiles, schemas, harness primitives, work items, evidence, policies, or module contracts.
• Run gradient eval before claiming a Gradient behavior change is ready.
• Capture evidence with gradient capture before closing work.

Harness Notes

The native HARNESS makes this skill available everywhere Gradient initializes:

• canonical skill root: .agents/skills/gradient;
• Claude bridge: .claude/skills/gradient;
• Codex bridge: .codex/skills/gradient;
• Pi bridge: .pi/skills/gradient.

If an agent can read the repository harness, it should be able to discover that Gradient exists, which commands to run, and how Gradient expects work to move through the lifecycle.