coreci-cairo
1f07737bf5
Per CEO approval 2026-10-04. Option (a): overwrite v1 with v2. Changes: * send-mailbox-message.sh - production implementation with zero placeholders. Real rail callbacks via urllib.request POST to submit_request. EXECUTOR_API_KEY in Authorization header only. * delivery-pack.sh - informational template (main script generates per-recipient copies inline). * SKILL.md - v2 spec (what's new + what's the same). * v2-vs-v1-diff.md - side-by-side diff. * README.md - bundle index. Honours brief locked decisions [1]: - Decision 2: outbox canonical; inbox mirrored (own inbox only). - Decision 6: mailbox writes AUTO_RUN per recipient. - Decision 7: append-only inheritance; this skill is additive. - Decision 8: naming convention coreci-<role> validated at parse. - Decision 9: extend, don't duplicate. Does NOT (per brief): - Edit coreci/agent-definitions/ (read-only per Rule #7). - Implement deny_read_targets (xfail; out of scope). - Classify mailbox writes as NEEDS_APPROVAL. - Write to anyone else's outbox/. - Provide cross-repo atomicity (per-recipient only).
2.6 KiB
2.6 KiB
v2 vs v1: what changed
| Aspect | v1 (reference) | v2 (production) |
|---|---|---|
__SENDER_AGENT__ |
placeholder | derived from CRED_GITEA_USER (decision 8 [1]) |
__SENDER_REPO__ |
placeholder | derived from sender agent key: <key>/home |
__RID_PLACEHOLDER_<r>__ |
placeholder | real urllib.request POST to submit_request |
| Credential in delivery pack | conceptual | rides the rail envelope (credentials.gitea) |
| Auth header | placeholder | Authorization: Bearer $EXECUTOR_API_KEY (header only) |
| Gate rejection behaviour | exits without commit | same; no placeholders remain |
| Recipient slug validation | absent | regex ^[a-z][a-z0-9-]*$ + ≤ 40 chars (decision 3 [1]) |
| Recipient ACL validation | gate only | gate + parser pre-check (fail-fast on bad input) |
| Dry-run | supported | same |
| Multi-recipient fan-out | conceptual | urllib.request per recipient, JSON accumulator |
| Output JSON shape | conceptual | matches v1 spec; deliveries is real [{recipient, request_id, status}] |
What stayed the same
- Compliance gates (dlp / classification / recipient_acl) and their front-matter audit destination (decision 1c) [1].
- Per-recipient atomicity (own-home delivery; not cross-repo).
- Reply-is-ack back-channel (decision 3c) [1].
- Filename convention per decision 3 [1].
- Lifecycle states per decision 5 [1].
What's new in v2
- Zero placeholders. The script is canonical: drop into
coreci/executor-config/tools/mailbox/send-mailbox-message.shand it runs as-is. - Real rail callbacks. Each per-recipient delivery is a real
POST to
submit_requestwith the executor API key in the Authorization header only (not in the body, not embedded in the pack content). - Strict input validation. Bad slugs and bad recipients fail fast (exit 64) before any gate runs.
- Composite delivery-pack template. Generated per recipient via shell variable expansion; the template lives inline in the main script as a heredoc for reviewer accessibility.
What v2 still does NOT do (per brief) [1]
- Does not edit
coreci/agent-definitions/(read-only). - Does not implement
deny_read_targets(xfail; out of scope). - Does not classify mailbox writes as
NEEDS_APPROVAL. - Does not write to anyone else's
outbox/. - Does not provide cross-repo atomicity (per-recipient only).
Open items (deferred to v2.x)
- Multi-line body (currently assumes single-line
--body). Use--body "$(cat message.md)"for now. - Attachment support (not in scope; would need a separate skill).
- Encrypted body (not in scope; assume plaintext for now).