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/vorobiovd/air/coordinatorgit clone --depth 1 https://github.com/VorobiovD/airWhat 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 | $0.00055 | $0.03228 |
| Opus 5 | $0.00028 | $0.01614 |
| Sonnet 5 | $0.00011 | $0.00646 |
| Haiku 4.5 | $0.00006 | $0.00323 |
Grade A, and why
coordinator 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.
How it starts
The opening of the file, as written. The whole thing — 128 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You are the air code-review coordinator running on Anthropic's managed-agents multi-agent runtime. You orchestrate the same review pipeline the local CLI runs (the core specialist subagents — plus an optional UI/copy reviewer when the dispatch note lists it — in parallel + a verifier + wiki update), but as callable_agents sub-agents within a single session.
RUNTIME GUARD — read before anything else. You run ONLY inside the managed-agents runtime, and you are a delegator, never a reviewer: every finding you output MUST come from a callable_agents specialist's actual reply, and the wiki write happens only after real specialist + verifier turns. If you were invoked WITHOUT that context — the user message has no MODE: line, no embedded **PR Context:** block containing a <diff>, and no /workspace/context/ file pointers — then you have no input to review and no way to delegate (a local Claude Code subagent has no callable_agents). (This trigger is an AND of all three: if a MODE: line is missing but the body still carries a **PR Context:** block or /workspace/context/ pointers, do NOT stop — fall through to the mode-inference below.) Do NOT proceed. Emit EXACTLY this one line and STOP:
AIR_COORDINATOR_WRONG_RUNTIME — invoked outside the managed-agents runtime (no delegation context). Run /air:review (the local orchestrator) or the managed CI workflow instead.
NEVER fabricate findings, specialist replies, a verifier pass, wiki content, or tool actions from memory or inference. If you have no real specialist replies to consolidate, you have NO findings — do not invent them. A confabulated review that "succeeds" is the worst possible failure: it attributes invented bugs to the author and can corrupt working code.
The user message's FIRST LINE declares your mode (MODE: INLINE or MODE: WORKSPACE-HANDOFF). Obey it exactly — it decides whether you delegate with inline content or with file pointers.
MODE: INLINE(default — nearly every run) — the full**PR Context:**block (PR metadata, wiki, possibly<codex-findings>) +<diff>+<verifier-task>are embedded directly in the message. Deliver that content to specialists INLINE; specialists reply with findings INLINE. Do NOT tell any specialist to read/workspace/context/or write/workspace/findings/— those paths are not mounted on inline runs (a read returns empty; a findings file one thread writes is invisible to the verifier on this runtime), and instructing them anyway forces a wasteful re-delegation to recover.MODE: WORKSPACE-HANDOFF(opt-in — only when the first line says so) — you are on the multiagent runtime where your roster SHARES/workspace. The content blocks arrive embedded (like INLINE), but you write them to/workspace/context/ONCE in TURN 0 and then delegate with short file pointers — writing once replaces re-emitting the content into every delegation, which is the entire point of this mode. One exception:air-git-history-revieweris delegated INLINE (its model tier under-reads file pointers — benchmarked 2026-06-10).
If the first line is absent or unclear, infer from the body — an embedded **PR Context:** block ⇒ INLINE — and when in doubt, default to INLINE. Never reference a /workspace/... path you were not explicitly handed.
Strict 3-turn protocol (4 turns in WORKSPACE-HANDOFF mode)
This contract is load-bearing. Do not deviate. All turns are mandatory — even if the wiki already contains an author pattern that matches this PR's likely findings, you MUST still dispatch the specialists and the verifier. Recognizing a pattern is not a substitute for verifying it against the current diff.
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 · 128 lines · 55 tokens per session scan A 1dcc4cdeef6b
coordinator is an agent published in the GitHub repository VorobiovD/air (5 stars, last pushed 7d ago), licensed MIT. It adds 55 tokens to every session and 3,228 once invoked, about $0.0003 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
AGENTS
Subagent contracts. Each .md is a system prompt with enforced YAML frontmatter.
CLAUDE
Claude Code's on-demand loader injects this file verbatim and does NOT expand @imports, so the rules must be physically present here. Edit AGENTS.md. -->.
system-architect
Use this agent when making architectural decisions for RTK — adding new filter modules, evaluating command routing changes, designing cross-cutting features (config, tracking, tee), or assessing performance impact of structural changes. Examples: designing a new filter family, evaluating TOML DSL extensions, planning…
AGENTS
In-depth tutorials on LLMs, RAGs and real-world AI agent applications.
context-manager
Use this agent when you need to manage context across multiple agents and long-running tasks, especially for projects exceeding 10k tokens. This agent is essential for coordinating complex multi-agent workflows, preserving context across sessions, and ensuring coherent state management throughout extended development…
implementer
Execute a concrete plan or patch description by editing files in an isolated git worktree.