Anthropic Routines

Quick reference

Routines are triggers, not durable executors. Inngest owns durability. Always include budget + routine_id_env_key in frontmatter.

When to use

• Authoring a new Routine .md file under .claude/agents/_routines/
• Refining an existing Routine's frontmatter or operating procedure
• Debugging why a Routine did not fire on its cron schedule
• Understanding how a Routine receives and validates a trust spec

When NOT to use

• For interactive C-suite agents (they use the standard agent schema, not Routine schema)
• For worker agents (workers are spawned via Task, not via /fire)

What a Routine is

A Routine is an agent that runs on a cron schedule OR is fired on-demand via Anthropic's /fire endpoint. It starts a fresh main thread each invocation. It has no persistent state between runs — state lives in Mem0, DECISIONS.md, and Supabase.

Routines are triggers + orchestrators. They are NOT durable executors. Inngest owns durability.

Required frontmatter

---
name: cto-daily-plan
description: "Reads open Linear tickets + overnight code activity, drafts a prioritized engineering plan, posts to war-room project. Fires daily at 10:30 UTC."
when_to_use: |
  - Fires automatically at 10:30 UTC via cron schedule
  - Can be manually fired via /fire for an on-demand engineering plan
  - Triggered by Inngest timeout-watcher if previous run timed out
schedule: "30 10 * * *"           # cron expression (UTC)
routine_id_env_key: ROUTINE_CTO_DAILY_PLAN_ID
routine_token_env_key: ROUTINE_CTO_DAILY_PLAN_TOKEN
model: claude-sonnet-4-6           # or claude-opus-4-7 for depth Routines
tools: [Read, Write, Edit, Bash, Glob, Grep, Task]
maxTurns: 25
color: blue                        # matches CTO color
isolation: none                    # Routines share no worktrees
mcpServers:
  - linear
  - github
  - supabase
  - mem0
skills:
  - war-room-orchestration
  - linear-mvp-recipe
  - writing-plans
budget:
  max_cost_usd: 2.0
  max_runtime_minutes: 20
  max_tool_calls: 100
delivery:
  channel: linear-comment
  project: war-room
  ticket_prefix: BMX-W              # war-room tickets use W prefix
trust_spec_parsing: true            # this Routine accepts inbound trust specs
escalates_to: ceo                   # on BLOCKED or over-budget
---

Fields reference

Field	Required	Notes
schedule	Yes (for cron Routines)	Standard cron expression, UTC
routine_id_env_key	Yes	Env var name storing the Routine ID. Value set via Wrangler secrets.
routine_token_env_key	Yes	Env var name storing the per-Routine bearer token.
model	Yes	claude-sonnet-4-6 default. claude-opus-4-7 for advisor/synthesis Routines.
budget.max_cost_usd	Yes	Hard ceiling. Inngest runaway-watcher enforces this.
budget.max_runtime_minutes	Yes	Inngest timeout-watcher fires Auto-Unblock if exceeded.
trust_spec_parsing	Conditional	Set true if this Routine can receive an inbound trust spec via /fire. Must validate HMAC.
delivery.channel	Yes	Where the Routine posts its output: linear-comment, slack, telegram.

Routine body structure

All Routines follow the same 6-section body:

# <Routine Name> — <Tagline>

## Identity & mission
## Schedule & triggers
## Pre-flight reads      (files to read at the start of every run)
## Operating procedure   (numbered steps)
## Trust spec handling   (if trust_spec_parsing: true)
## Return format         (what the Routine posts and where)

Trust spec handling section (mandatory if trust_spec_parsing: true)

## Trust spec handling

When fired via `/fire` with a trust spec in the payload:

1. Extract spec from `---BEAMIX-SPEC-V1-START---` / `---BEAMIX-SPEC-V1-END---` sentinels
2. Verify HMAC signature against `BRIDGE_HMAC_SECRET` env var
3. Verify `nonce` not seen before (read KV or fail-open with audit_log row)
4. Verify `expires_at > now()`
5. Verify `issued_by.linear_user_id` ∈ `ALLOWED_ISSUERS` env list
6. Write `audit_log` row: status=accepted
7. Execute the scoped task from `scope.intent` + `scope.constraints`
8. Write `audit_log` row: status=complete (or blocked)

If any verification step fails: write audit_log row status=rejected, post Linear comment with reason, exit.

Standing Routines roster

Routine	Schedule (UTC)	Model
advisor-daily-thinking	05:30 daily	Opus
morning-digest	05:30 daily	Sonnet
competitor-pulse	05:30 daily	Sonnet
geo-algorithm-signal	05:30 Sundays	Opus
cto-daily-plan	10:30 daily	Sonnet (Opus for plan depth)
content-idea-generator	10:30 daily	Sonnet
monday-standup	10:30 Mondays	Sonnet
friday-retro	15:30 Fridays	Sonnet
eod-sync	20:30 daily	Sonnet
auto-unblock	Event-triggered (timeout)	Sonnet
synthesizer	Event-triggered (board R3)	Opus

See also

• war-room-orchestration — [[war-room-orchestration]]
• trust-spec-contracts — [[trust-spec-contracts]]
• linear-mvp-recipe — [[linear-mvp-recipe]]
• board-meeting-protocol — [[board-meeting-protocol]]

Anti-patterns

• Hardcoding the Routine ID or token in the .md file (use env var name, not value)
• Omitting budget fields — Inngest runaway-watcher cannot enforce without them
• Setting trust_spec_parsing: true without implementing the HMAC verification steps
• Posting output to main product Linear project (Routines post to war-room or strategy projects)
• Using schedule in local timezone — always UTC
• Awaiting a downstream Routine's completion inline (fire-and-forget only; completion via Inngest fan-in)