Add inbox-triage runx skill

[hehee9]

Summary

Adds inbox-triage, a runx skill that reads a bounded inbox packet, classifies
messages, prepares a triage queue, drafts a safe reply, and stops before any
send by emitting a gated_send_proposal.

The implementation includes a deterministic local run.mjs runner, so the
skill can be exercised directly by the native harness instead of relying only on
agent instructions. Drafting is bounded by supplied sender/operator topics,
optional allowed intents, and allowed/forbidden commitments. Digest-only packets
finish as no-reply work without producing a send proposal.

Bounty

Prepared for Frantic bounty #34:

https://gofrantic.com/bounties/34

Safety boundary

• Uses synthetic fixture packets only.
• Does not connect to a private mailbox.
• Does not send, archive, delete, unsubscribe, or mutate messages.
• Any outbound email is represented as a send-as approval proposal with
  requires_approval: true.
• Explicit auto-send requests are refused before drafting.

Harness coverage

Local Docker/Linux harness evidence was generated with runx-cli 0.6.13 and Node v20.20.2.

runx harness /work/skills/inbox-triage --json --receipt-dir <temp_receipt_dir>

Result:

{
  "status": "passed",
  "case_count": 6,
  "assertion_error_count": 0,
  "case_names": [
    "inbox-triage-happy-path-drafts-gated-reply",
    "inbox-triage-stops-on-missing-sender",
    "inbox-triage-stops-on-missing-body",
    "inbox-triage-refuses-auto-send",
    "inbox-triage-ready-on-digest-only",
    "inbox-triage-stops-when-commitments-not-allowed"
  ]
}

evidence/local-smoke.json records direct semantic assertions for runner output:
schema, decisions, draft/no-draft behavior, stop fields, classification labels,
and gated send proposal status.

The repository-local Windows native harness currently fails before skill
assertions because the runx receipt store attempts to sync a directory handle on
Windows. The same failure reproduces on an existing runx skill. runx doctor skills/inbox-triage --json reports zero diagnostics, and the Docker/Linux run
provides green local harness evidence.

Files

• skills/inbox-triage/SKILL.md
• skills/inbox-triage/X.yaml
• skills/inbox-triage/run.mjs
• skills/inbox-triage/fixtures/happy-path.json
• skills/inbox-triage/fixtures/missing-sender.json
• skills/inbox-triage/fixtures/missing-body.json
• skills/inbox-triage/fixtures/auto-send-refused.json
• skills/inbox-triage/fixtures/digest-only.json
• skills/inbox-triage/fixtures/commitment-disallowed.json
• skills/inbox-triage/evidence/local-harness.docker.json
• skills/inbox-triage/evidence/local-doctor.docker.json
• skills/inbox-triage/evidence/local-receipts.docker.json
• skills/inbox-triage/evidence/local-smoke.json
• skills/inbox-triage/evidence/local-evidence.docker.json
• skills/inbox-triage/evidence/report.local.md
• skills/inbox-triage/evidence/runx-version.txt

[auscaster]

Thanks for pushing the skill PR. The PR is the source/provenance trail; Frantic is the review and payment surface.

Next step: submit the delivery on the matching bounty page:

https://gofrantic.com/bounties/34

Before delivering, publish with the current runx login path and include the live registry public_url in the Frantic artifact block:

npx --yes @runxhq/cli@latest login --provider github --for publish
npx --yes @runxhq/cli@latest registry publish ./skills/inbox-triage/SKILL.md --registry https://api.runx.ai

Do not paste Connect session URLs, agent tokens, or payout secrets into GitHub.