feat(webhook): pin local-webhook 0.10.0 — standing watches stop duplicating owned work

[defangdevs]

Companion to defangdevs/local-channels#11 (merged as efa2531), which closes defangdevs/local-channels#10.

Motivation

This box hit the same failure twice on 2026-08-02: five hook-* sessions spawned for CI on PRs a session was already driving — two while #168 went through CI, three more when merging #168 and #169 put CI on master. Three of the four hook slots went with them, from two documentation merges. Nothing was misconfigured: CI-outcome events bypass ignoreSenders by design, and dispatch_event() never asked whether a session peer had just taken the same delivery.

Changes

• services.agent-box.webhook.rev/sha256 → local-channels efa2531 (local-webhook 0.10.0). Two brakes, dispatch path only: a CI event overrides ignoreSenders only when it reports a failure, and no CI event spawns while a live session peer's own filter already claims the topic. Session delivery is untouched, so "merge on green" still works. Non-CI events (new issue, someone else's PR) still spawn regardless of who is subscribed — topics are repo-granular while ownership is object-granular.
• Docs in the two places an agent actually looks: the guide's standing-watch section and agent-box-webhook --help both now state that a watch will not double up on work you own.
• tests/webhook.nix covers the new behaviour end-to-end.

0.10.0 renames peer IPC sockets to <filter key>.<pid>.sock so the ingress owner can resolve a live peer to its subscriptions. That is internal to webhook.py's own state dir; nothing in this module reads those names.

Why the VM test had to change

Its dispatch delivery is a workflow_run failure on defangdevs/agent-box while the peer session is subscribed to that repo (pinned) — exactly the case 0.10.0 now suppresses, so the old assertions would have failed. The test now proves both directions on a real box:

1. Delivery accepted (200), the daemon logs not spawning for workflow_run on defangdevs/agent-box — session agent-main is subscribed to it, and no hook-* session appears.
2. The peer unsubscribes, the same signed delivery is re-sent, and the hook-* session appears with the framed prompt + trusted preamble as before.

The socket-name assertion moved to instances/agent-main.*.sock, which also pins the new naming contract.

User-visible / security effects

Fewer spawned sessions, none of them for work a live session owns. No change to the ingress, HMAC verification, the 401 fail2ban key, socket ownership, or the settings surface. No AWS, IAM, networking, or migration impact.

Operational note: this box's github:defangdevs/* standing watch is currently unsubscribed as a stop-gap. It gets restored after this lands and the box rebuilds onto the new pin.

Checks run

nix build -L .#checks.aarch64-linux.{module-generated-up-to-date,module-single-file,webhook-route,multi-user} — pass. nix eval .#checks.x86_64-linux.webhook.drvPath evaluates.

The webhook VM test is x86_64-only, so the new assertions are unverified locally — CI is their first real run.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PsqGhkpjsb4kKAdVz25xKN