Development record
A cheap watcher wakes an agent only when something needs action
Press Release
A stuck pull request or a failed queue entry wakes an agent within minutes, and nothing runs a model while all is well. A cheap script follows one component's events and wakes the coordinating agent only on an actionable event. The agent reports and proposes, and acts only through the session that owns the change.
"A pull request sat behind its base for an afternoon with every check green," said Frank, an SRE. "Now the watcher wakes the session that owns it, which updates the branch, and no model sits refreshing charts in between."
Why This Matters
Survived, reframed, from an ideate run of 2026-10-14 (private verdict component-watcher-agents) on the engineering write-up "Redesigning source control at Anthropic" (Anthropic, 2026-10-13, https://www.anthropic.com/engineering/redesigning-source-control), whose "superfan" agents each watch one component's telemetry and logs. The write-up gives no measurement of their effect, and a standing model per component spends tokens and sub-agent slots while idle; both reviewers of the run reframed it to an event-triggered cheap watcher, the shape an earlier private verdict (coordinator-crew, 2026-10-09) had also reached. The real cost it targets is on the record: pull requests wedged behind their base (iss-172).
Mechanism
We expect faster recovery at almost no cost because the watcher is a script, not a model, and the agent is woken only on an event that needs action, so the expense scales with incidents, not with time.
Scope Conditions
- The watcher acts on nothing itself; changes go through the owning session, under the concurrency rules (scan, announce, the cutting session's say).
Acceptance Criteria
- Given a pull request behind its base with all checks green, when the watcher sees it, then the owning session is woken with the event and a proposed action, and no change is made by the watcher.
- Given no actionable event, when time passes, then no model runs.
Open Questions
- Which components are watched first, and how the owning session is found (technical facilitator).
Audit Notes
Empty. Populated by intent-auditor when intent moves to shipped/.
Frontmatter
| id | itd-2610141646469491 |
|---|---|
| type | intent |
| lifecycle | drafts |
| kind | standalone |
| severity | minor |
.abcd/development/intents/drafts/itd-2610141646469491-a-cheap-event-watcher-follows-one.md
open on GitHub ↗ · commit history ↗
mentions1
- iss-172abcd-managed repos need PR-queue hygiene when branch protection requires up-to-date branches: a queued auto-merge PR whose base moved sits BEHIND forever because GitHub auto-merge never updates branches. Two delivery rungs per basics-built-in/SOTA-delegated: (1) protocol-level now — after arming auto-merge, poll mergeStateStatus and run gh pr update-branch on BEHIND (works with the agent's own token, CI triggers normally); (2) scaffolded CI later — abcd scaffolds an update-branch workflow into managed repos, which requires a PAT/GitHub-App secret because the default GITHUB_TOKEN's pushes do not trigger workflows (the recursion suppression trap), or points at a merge queue where the plan allows. Same scaffolding family as itd-93's release gate; same merged-work-hygiene family as itd-95.