The Four Pillars
AgentiHooks is organised around four pillars: Identity, Guardrails, Context Intelligence, Fleet Command. Each pillar covers one axis of running Claude Code and Codex agents.
The swarm guide covers fleet task execution, including changing the kind of an existing task, ready on a task of any kind, ready handoffs for tasks split across pull requests, labelled halves for tasks without a plan and what the ready merge pass does when a pull request leaves the merge queue, and what it does when a pull request is marked unmergeable.
See recovering a released task after red checks for the handoff and branch continuation, configuration, owner safeguards and measured recovery time.
See proof loop health findings for rerun exclusions, pull request head changes, review counts, cap configuration and verdict examples.
See stalled activity findings for threshold configuration, wrapper loss recovery and investigation of active agents reported as stalled.
See stall watchdog refreshed checks for requeue waits, stale failure handling, fresh GitHub reads and watchdog configuration.
For publication retries, see publishing unchanged files: automatic remote comparison, skipped gates and pushes, output and missing branch handling.
For resumed work, see publishing after a successor prepares a branch: branch selection, restacking, repository configuration and publication retries. See recording a published upstream branch for successor tracking setup, task metadata and pull request head checks.
For an existing pull request, see continuing with one push: retained branch selection, Stop behavior and lease failures.
See Doctor timer pass for what one Doctor timer step reads, its interval setting and how to measure its wall time per pass.
See Inbox escalation sender exclusion for master handoffs, escalation timing, configuration, message history and delivery evidence.
See Inbox receiver read for how the inbox detector resolves every receiver in a fixed number of Redis reads, the check command and the measured result.
See Information only items for which inbox wakes, seat holds, escalations and Doctor findings skip items sent with --fyi, how to send and inspect them, troubleshooting and delivery evidence.
See Doctor measurement task readiness for tune dependencies, application timestamps, configuration examples and releasing a read only seat with its recorded window.
See worker execution lease through a controller rollout for the lease and grant timing that keeps an Anton worker alive through a release, and why a worker exits 70.
See one push per pull request for when a worker may open a pull request, when a push is refused while checks run, gate modes and troubleshooting.
See plan tracing for promotion fixer tasks for usage, gate modes and the limits on run numbers in ledger prose, and plan refusals in the gate log for the row each refused plan writes and how Doctor rates counts it.
See intent verdict timing for why a ready pull request is judged before agent launches, the at and persisted_at verdict fields and how to measure ready to verdict, and tick lock holder record for how a killed tick’s lock is replaced at once.
See hives for joining machines, roles, placement, launches on a hive, failover and fencing, and deployment modes for where the controller, ledger and Redis run.
See hive endpoint limits for the hive endpoint request, handler and Redis time limits, the 503 it answers when Redis does not reply, and what a client does after it.
See headroom for what a hive’s headroom reading holds, how it stays fresh while a launch runs, and why a hive with no fresh reading has no room.
See promotion fixer intake for how one swarm per repository files and closes promotion fixer tasks, hands intake over and shows its owner, and runner outage runs for the alert raised without a task when failed jobs failed on CI infrastructure, with the jobs that failed for want of a lost unit shard’s output in jobs that failed for want of shard output. See red dev workflows outside the promotion gate for the Doctor finding raised when any other workflow fails five dev push runs in a row.
See independent consumer failures beside a lost shard for mixed Sonar and shard check examples, separate fixer tasks and the fixed folding rules.
See agent status comments for when an agent’s status comment amends its earlier one and when it posts a new comment at the end.
See ledger artifacts for what the ledger accepts as an artifact, a master’s plan publish, the two checks and refusal troubleshooting.
See task ownership and master rulings for phase and standalone task owners, ledger write permissions, claimant message refusals and fence configuration.