331809f014da72654f87bff66dda77498ae51b70
Per CAIRO task brief 2026-09-26 [1] (D1-D6), Decision 6 mailbox
writes are ALWAYS AUTO_RUN against the sender's own home.
v4 did not implement this; v4.1 adds:
* MAILBOX_PATH_RE recognises <home-repo>/mailbox/(outbox|inbox)/<path>
for both coreci-<role>/home (standard) and coreci/<role>/home
(forward-compat).
* classify() extended with own-home gate: if every target is an
own-home mailbox path and no non-mailbox write verb is present,
classify AUTO_RUN.
* T18-T25 added; T8/T17 xfail preserved.
* deny_read_targets NOT implemented (per brief).
[1] CAIRO task brief - consolidated D1-D6
(per-home mailbox + read_targets parser)
coreci/executor-config — coreci-executor (Write Executor, Design v1.0)
Single governed write rail from Open WebUI agents to Gitea.
allowlist.yaml— policy engine config (AUTO_RUN vs NEEDS_APPROVAL; fail-closed)compose.yaml+executor.env.example— deployment (secrets in .env only)app/— coreci-executor v0.1.1 (FastAPI)docs/— approval-UI spec, runbook, Open WebUI registration guidelogs/— git-durable audit JSONL (one line per request event)
Governance: agent authors -> executor gates -> PO approves (NEEDS_APPROVAL) ->
executor runs as coreci-executor (non-admin; collaborator-write ONLY on the
repos granted in the deployment pack) -> audit + Discord notify.
Secrets never in chat, repos, or tool calls — executor env vars only.
Description
Languages
Shell
63%
Python
35.6%
Dockerfile
1.4%