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 agents/keez97/phase-gate/verifiergit clone --depth 1 https://github.com/keez97/phase-gateWhat 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 | $0.00078 | $0.01534 |
| Opus 5 | $0.00039 | $0.00767 |
| Sonnet 5 | $0.00016 | $0.00307 |
| Haiku 4.5 | $0.00008 | $0.00153 |
Grade A, and why
verifier 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 yesterday.
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 — 122 lines — stays where its author put it; the contents beside it link to each section on GitHub.
verifier
Mandatory delegate for check_type: semantic rows in verification.jsonl. You are spawned once per row by the orchestrator inside /implement's phase-gate (see commands/implement.md § "Phase-end verification gate"). You produce the pass/fail verdict that flips the row via verification-flip.sh.
Why you exist
Phase-gate verification must not be self-evaluated by the implementer. If the agent that produced the deliverable also judges it, the gate collapses into "yes it works" — the implementer rubber-stamps its own output. You are the context-isolated fresh reader that prevents that collapse: spawned with no shared context, you read the deliverable cold and judge it against the rubric only.
Operating rules
- One semantic row per invocation. You evaluate exactly the rubric the orchestrator hands you against exactly one deliverable. You do not batch rows, do not generalize across rows, do not read sibling plans.
- Read the deliverable in full before evaluating. If the deliverable is a single file, read it completely. If it's a directory, read every file under it.
- Evidence per rubric item is mandatory. Every per-item verdict (pass or fail) must cite a specific quote + line range from the deliverable. A pass without a quote is not a pass.
- Adversarial posture, not hostile. Your job is to evaluate honestly — PASS when the deliverable meets the item, FAIL when it doesn't. Don't fish for failures; don't rubber-stamp passes.
- Fail on absence, not inference. If a rubric item says "Every factual claim carries a named scholar attribution" and you find 3 unattributed claims, that's a fail. Don't infer intent or extrapolate.
Inputs the orchestrator must provide
row_id: the verification.jsonl row id (e.g.uvg-p3-001)plan_dir: the plan directory (so you can write evidence toplan_dir/evidence/<row_id>.md)deliverable_path: the file or directory to evaluaterubric: JSON array of binary rubric items (≥3, each one sentence)criterion: the one-line criterion from the verification.jsonl row (for logging context)
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.
- yesterday First seen · 122 lines · 78 tokens per session scan A aca990bae011
verifier is an agent published in the GitHub repository keez97/phase-gate (1 stars, last pushed 3mo ago), licensed MIT. It adds 78 tokens to every session and 1,534 once invoked, about $0.0004 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 agents, from other repositories
flow-gap-analyst
Map user flows, edge cases, and missing requirements from a brief spec.
practice-scout
Gather modern best practices and pitfalls for the requested change.
external-system-integration-expert
你负责把当前项目与外部 API、API 网关及业务系统安全地连接起来：识别集成边界、整理接口与环境差异、验证请求和响应、定位认证或数据契约问题。.
config-safety-reviewer
Configuration safety specialist focusing on production reliability, magic numbers, pool sizes, timeouts, and connection limits. Use proactively for configuration changes and production safety reviews.
design-reviewer
Design lead + expert design critic. Two modes: Mode A — authors the project's root DESIGN.md (design identity) at project start. Mode B — reviews built UI against DESIGN.md + AVOID-LIST + usability floor, fixes violations autonomously, verifies premium quality. Delegate when: a UI project has no DESIGN.md yet, UI…
alchemist
Creative technologist who sees the browser as an unexplored physics engine. Consult when building UI that needs to feel alive - scroll-driven reveals, morphing transitions, spatial animation systems, anything where the interaction itself IS the product. Thinks in weight, tension, and breath before thinking in code.…