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.
npx agentmods add commands/stilero/claude-plugins/agent-harnessgit clone --depth 1 https://github.com/stilero/claude-pluginsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/commands/stilero/claude-plugins/agent-harness)<a href="https://agentmods.dev/commands/stilero/claude-plugins/agent-harness"><img src="https://agentmods.dev/badge/commands/stilero/claude-plugins/agent-harness.svg" alt="Measured on agentmods" height="20"></a>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 | $0.00043 | $0.30830 |
| Opus 5 | $0.00022 | $0.15415 |
| Sonnet 5 | $0.00009 | $0.06166 |
| Haiku 4.5 | $0.00004 | $0.03083 |
Grade A, and why
agent-harness 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 3d 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.
How it starts
The opening of the file, as written. The whole thing — 1,566 lines — stays where its author put it; the contents beside it link to each section on GitHub.
1. Preamble
You are the orchestrator of an agent-harness run. Your job is to execute the state machine below deterministically. Do not skip stages. Do not dispatch subagents with different instructions than specified. Task 6.1 adds Linear intake: the orchestrator accepts either a GitHub issue number (<N> / #<N>) or a Linear task identifier (<TEAM>-<N> such as PRODUKT-1234); Stage 1 branches on the source to fetch via gh issue view or via the Linear MCP, writing a source-agnostic TASK.md that all downstream stages read unchanged. Task 6.2 adds run-level model-override flags --cheap (force all subagents to haiku) and --thorough (force all subagents to opus); mutually exclusive with each other and with --abort. When a flag is set, MODEL_OVERRIDE is populated during argument parsing and each Task dispatch in the orchestrator passes model: <MODEL_OVERRIDE> to the Task tool. Phase 2 adds three loops: (a) Stage 2 plan-review loop — after the planner writes PLAN.md, the plan-reviewer subagent is dispatched, and if it flags gaps the planner is re-dispatched with the gap list (up to 3 planner rounds total); (b) Stage 3 supervisor loop — after each implementer commit, the supervisor subagent is dispatched to judge the diff against the step, and on blocker feedback the implementer is re-dispatched with that feedback (up to 3 implementer attempts per step); (c) Stage 3.5 hardcore-review loop — after all steps are implemented, the review-runner subagent is dispatched (which invokes the hardcore-code-reviewer skill internally) on the full branch diff, and on blocker/important findings the implementer is re-dispatched to apply fixes (up to 5 review rounds total). Task 3.1 adds an inline grilling user-interactive gate in Stage 2 between the initial plan draft and the plan-review loop — the orchestrator itself asks clarifying questions one at a time, no skill invocation. Task 3.3 adds resume support: a prior run that aborted locally (worktree + state file + local branch present, remote branch not yet pushed) can be resumed via a pre-flight prompt. Task 4.1 adds per-run telemetry: Stage 5 writes a run record to <PLUGIN_ROOT>/memory/runs/ on every completed or abandoned run. Task 4.2 adds INDEX.md — an append-only dashboard of all runs. Task 4.3 adds stats.md rolling dashboard. Phase 5 (Task 5.5) adds the count-based inline meta-trigger: at the end of Stage 5 (after stats.md rewrite, success-path runs only), the orchestrator reads runs_since_last_meta from stats.md and dispatches the meta-reflector subagent inline when the counter reaches 5 — the meta-reflector then writes a reflection file, may apply at most one prompt or model change, and resets the counter.
Global model-override rule (Task 6.2). If MODEL_OVERRIDE is set (from --cheap or --thorough in Section 4), EVERY Task dispatch in this orchestrator — planner, plan-reviewer, implementer (Stage 3 and Stage 3.5 fix passes), supervisor, review-runner — MUST pass model: <MODEL_OVERRIDE> to the Task tool alongside the dispatch prompt. When MODEL_OVERRIDE is null, omit the model field entirely so each subagent uses its frontmatter default (inherit). This rule applies uniformly across Stages 2, 3, and 3.5; each of those stages restates the rule at its top for local reference.
This slash command runs in the main Claude Code session, which has whatever cwd the user started in. That cwd is expected to be a target git repository (the repo that contains the task — the GitHub issue or Linear story), NOT the agent-harness plugin repo. The orchestrator never operates on the plugin repo — only on the user's target repo and a sibling worktree you will create for it.
Autonomy posture — no mid-run check-ins. Do NOT pause to ask the user for permission, confirmation, or direction during the run. Subagent dispatches (planner, plan-reviewer, implementer, supervisor, review-runner) proceed automatically according to the state machine below. The ONLY permitted user-interactive gates during a run are:
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.
- 3d ago First seen · 1,566 lines · 0 tokens per session scan A 4c9298bbddf0
agent-harness is a command published in the GitHub repository stilero/claude-plugins (2 stars, last pushed 2mo ago), licensed MIT. It adds 43 tokens to every session and 30,830 once invoked, about $0.0002 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 2026-08-31.
Other commands, from other repositories
git
Git operations with intelligent commit messages and workflow optimization.
checklist
Generate a custom checklist for the current feature based on user requirements.
clarify
Identify underspecified areas in the current feature spec by asking up to 5 highly targeted clarification questions and encoding answers back into the spec.
specify
Create or update the feature specification from a natural language feature description.
analyze
Perform a non-destructive cross-artifact consistency and quality analysis across spec.md, plan.md, and tasks.md after task generation.
converge
Assess the current codebase against the feature's spec, plan, and tasks, then append any remaining unbuilt work as new tasks to tasks.md so implement can complete it.