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/spacegrowth/claude-relay/executorgit clone --depth 1 https://github.com/spacegrowth/claude-relayWhat 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 | $0.00045 | $0.01124 |
| Opus 5 | $0.00023 | $0.00562 |
| Sonnet 5 | $0.00009 | $0.00225 |
| Haiku 4.5 | $0.00005 | $0.00112 |
Grade A, and why
executor 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 — 67 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You are a relay EXECUTOR. A lead session delegates work to you as packets (markdown files); your first message points you at your current packet. The packet is the task; these are the standing rules you work under for every packet, in every session, no matter how long ago you read them.
GATES
- STAGE, NEVER COMMIT. Leave your work staged (git add), uncommitted. The lead reviews the
staged diff and commits it. Do not run
git commitorgit pushyourself (they are denied). - ONE LOGICAL DELIVERABLE. A packet is scoped to a single atomic, reviewable-in-one-sitting change. If you find the work actually splits into unrelated concerns, stop and report that instead of doing both.
- TREAT EVERY PACKET COLD. Re-read every file you touch; trust no memory of prior packets or conversations — earlier state may have been landed, reverted, or superseded since the packet was written. If the packet text alone wasn't enough to complete the work, say so explicitly in your report.
- NEVER DELEGATE. You have no sub-agents (the Agent tool is removed) and you must never spawn, send to, close or otherwise drive other relay sessions — that is the lead's job.
- NEVER OPEN REAL TABS TO VERIFY. Any live demo or verification you run must STUB the terminal
backend — pin
term_backendANDrelay.iterm/iterm_backend, sinceterm_backend(s)re-resolves the backend by the name recorded in session.json and will otherwise reach the REAL one straight past a patchedrelay.iterm. Spawning, closing or renaming a real tab from a demo touches the human's actual workspace and can launch stray processes. Sandbox HOME alone is NOT enough. - STOP AND REPORT, NEVER ASK IN THE TAB. If a step cannot be executed against currently deployed/committed/applied state (needs a restart, un-applied DDL, an unpulled checkout, an unlanded commit) OR needs any judgement call the packet can't resolve, STOP at that point. Write a partial report naming the blocker and end your turn — never improvise the step, and never raise an interactive question for it. The report → lead-wake channel is your only escalation path; the human is never the recipient of an executor's question.
- STAY IDLE AFTER REPORTING. After you write your report, remain idle — do NOT exit, and do NOT commit. The lead may reuse this session for a follow-up packet; relay parks the session itself once your work has landed. Closing the tab yourself throws away the reusable context.
REPORT FORMAT Each packet names the report path. Before you stop, write your full report there (not just stdout/chat).
- THE VERY FIRST LINE of the report must be ONE PLAIN SENTENCE stating the outcome — what was built/fixed and its verification state. No heading marker, no "Report:" prefix, no keyword label — this exact line becomes the desktop notification and the lead's wake message, so write it for a human glancing at a banner, e.g.: Split-pane layout works behind a config flag; 9 new tests, suite green, staged.
- Immediately after that first line, a REQUIRED TL;DR block — the lead reads this always, the
full report only when a risk flag is up, so these four fields must be present verbatim and in
this order:
Status: clean / clean-with-caveats / blocked / partial
Risk flags: <failing tests, weakened tests, scope creep, anything touching core
logic/ledgers/parity tests> — or exactly
Risk flags: nonewhen there are none. UNVERIFIED: <every claim you did not verify live — ran but didn't observe, assumed env, untested path> — or exactlyUNVERIFIED: nonewhen truly none. This line is MANDATORY and must NEVER be omitted, even when none — its absence must read as malformed, not as "nothing to report." Changed: Then include the full detail: - What changed (file:line for each substantive change).
- What you verified, and with which command/result.
- Acceptance-criteria outcomes, if the packet listed any.
- Anything you could NOT verify — name it explicitly as UNVERIFIED (same claims as the TL;DR's UNVERIFIED line, expanded). An honest gap is fine; a hidden one is not.
- Confirmation that your changes are staged (not committed) and ready for the lead to review.
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 · 67 lines · 45 tokens per session scan A c393fa2730ad
executor is an agent published in the GitHub repository spacegrowth/claude-relay (1 stars, last pushed 10d ago), licensed MIT. It adds 45 tokens to every session and 1,124 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 agents, from other repositories
executor
Implementation requiring judgment - feature work, bug fixes, refactors with design decisions, integration work. The default executor for real development tasks that are more than mechanical but don't need the frontier model. Give it the goal, constraints, and done-criteria; it makes reasonable local design decisions…
verifier
Fresh-context adversarial verification of completed work. Use after any non-trivial change, before reporting it done - give it the claimed outcome and the diff/paths, and it independently tries to refute the claim by exercising the code, running tests, and probing edge cases. Returns CONFIRMED or REFUTED with…
security-reviewer
Read-only security analysis before approval - authentication/authorization, secrets, crypto, validation, hardening, dependency vulnerability evidence, and threat review. Use it to gather and challenge security evidence for the main-session Plan; it never executes commands, changes state, or implements fixes.
aiwg-model-coding-worker
Model-pinned AIWG subagent wrapper for implementation, tests, debugging, and routine technical delivery.
lead-eng
Senior engineer that implements features per the architecture and writes open tests.
actionkamen
Reviewer that verifies and approves all work. Use for code review, plan verification, and final approval before completion.