archive/skills/agent-dispatch/SKILL.md
Reusable inter-agent dispatch contract; responsible for delivering structured handoff requests from MARS to other sub-agents (Andy, Cooper) via the documented transport path.
npx skillsauth add maximebaudette/agent-skills agent-dispatchInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
This SKILL documents the contract for a reusable agent-dispatch skill used to hand off tasks from MARS to other sub-agents. It defines required inputs, supported targets, success/failure semantics, and expected behaviour when the transport path fails. Scope: delivery of dispatch messages only — this skill does NOT perform the dispatched work.
The agent-dispatch skill MUST accept these inputs:
andy, coopersource-system)artifact_123456)The context input is intentionally lightweight. It MAY be either a plain string summary or a small JSON object. Do not embed large artifacts — include a reference_id so the target agent can fetch the full item.
Keep context aggressively short. Do not include score-by-score examples, long link lists, or copied message bodies.
Recommended minimal object shape (informational, not required):
This keeps the contract simple while making it clear how to use context in practice.
Implementations MUST return a deterministic, machine-readable envelope on success or error. This makes downstream automation and auditing reliable.
Success envelope (example fields):
JSON example (success):
{
"status": "delivered",
"timestamp": "2026-04-15T12:34:56Z",
"transport": { "name": "telegram", "details": { "chat_id": -1003989798620 } },
"transport_message_id": "1234567890",
"target": "andy",
"task_source": "source-system",
"reference_id": "artifact_1234567890"
}
Error envelope (example fields):
invalid_input, transport_error, auth_error, not_found)JSON example (error):
{
"status": "error",
"code": "transport_error",
"message": "Transport API returned 401: authentication failed",
"retryable": false,
"details": { "http_status": 401 },
"target": "cooper",
"task_source": "source-system",
"reference_id": "artifact_987654321"
}
target, task_source, context, reference_id. Missing required inputs → immediate error (use code: "invalid_input").target to one of the supported targets or return validation error.This is the live runtime transport for this host.
Use Hermes profile queries, not a conceptual no-op and not a Telegram group handoff.
Target mapping:
andy -> profile career-managercooper -> profile health-coachRequired runtime behavior:
target to the corresponding Hermes profile.task_sourcereference_idcontextreference_id when neededemail-triage, prefer wording like "Maxime replied with scoring feedback on the weekly job hunt email"hermes chat -p <profile> -q "<delegation prompt>"
Use a generous command timeout for Hermes profile queries. For email-triage, the minimum recommended timeout is 180 seconds.
0, andRecommended prompt shape:
Delegated task from MARS.
target=<target>
task_source=<task_source>
reference_id=<reference_id>
context=<short summary or compact JSON>
Fetch the full artifact using reference_id if needed, process it using your own workflow, and reply with a concise summary.
For email-triage, keep context to one or two sentences and rely on reference_id for the full message. Do not include score-by-score examples, long link lists, or copied message bodies.
Recommended success envelope for this runtime transport:
{
"status": "delivered",
"timestamp": "2026-04-17T01:00:00Z",
"transport": {
"name": "hermes-profile-query",
"details": { "profile": "career-manager" }
},
"transport_message_id": null,
"target": "andy",
"task_source": "email-triage",
"reference_id": "artifact_123456"
}
If the Hermes query exits non-zero, produces empty output, or otherwise cannot confirm delivery, return an error envelope and do not claim the task was delivered.
retryable: true in the error envelope so callers can decide to retry.retryable: false and an appropriate code.YAML example (conceptual):
- skill: agent-dispatch
inputs:
target: andy
task_source: source-system
context: "Short summary: candidate message; Subject: Senior Engineer; Deadline: 2026-04-20"
reference_id: "artifact_123456"
JSON example (programmatic):
{
"skill": "agent-dispatch",
"inputs": {
"target": "cooper",
"task_source": "external-system",
"context": {"snippet": "Lab results attached, appointment 2026-04-21"},
"reference_id": "artifact_987654"
}
}
development
Compile LaTeX documents using the Tectonic engine. Use this skill whenever the user wants to create, edit, or compile a LaTeX document — letters, articles, CVs, reports. Tectonic is a self-contained compiler that auto-downloads only the packages it needs.
tools
Maintain living stack documentation for the MARS host in the Prime Radiant KB. Use this skill whenever: updating the mars host page after infrastructure changes, appending archive entries for deprecated components, documenting a new service/skill/plugin being added or removed, or when the user asks to 'update stack docs', 'document this change', 'archive this setup', 'what's in the stack', or 'generate stack summary'. Also use when you've just finished installing or removing any service, plugin, skill, tool, or cron — proactively suggest documenting it.
tools
Run autonomous dev tasks using Gemini CLI (gemini-cli) with Maxime's free Google quota. Use for programming projects where MARS needs to delegate agentic coding work — file edits, code gen, refactoring, analysis — to Gemini in a project directory.
testing
Weekly or on-demand audit of all agent workspaces (Cooper, Andy, MARS). Checks for stale script references, wrong skill paths, snapshot count drift, config vs file drift, and contradictory instructions. Auto-fixes simple unambiguous issues. Emails Maxime only for items requiring judgment. Silent on clean runs. MARS-only — do not invoke as Cooper or Andy.