Don't pay a frontier model to be a shell script.

A Claude Code session in Continuum Engine starts by loading .claude/CLAUDE.md, everything it imports with @, and every file in .claude/rules/ without a paths: key. On 28 September that was 13 files and 300,198 bytes: the rules, the project map, the local model protocol and every model's settings, the system as built, why it exists, the queue of what's next, and the reference repos on disk. At about four bytes a token that is 75,000 tokens, in the context window on every turn of every session.

Prompt caching keeps the rereading cheap. Claude Code writes the set to a one-hour cache on the first request, at twice the input price, and each later turn reads it back at a twentieth of the input price on Opus 5.5. Caching gives none of the room back: the tokens are in the window on every turn, they bring the next compaction closer, and the project's CLAUDE.md is loaded again after it.

The set is paid for on purpose. CLAUDE.md says what it buys: "You should never need to go looking for what _ref/ holds, what is built, or what to work on next." The model settings are always on because "a measurement taken under different settings is a different measurement, and the drift is invisible." Each file is held to one test, whether every session needs it or only the sessions that touch its area, and only the first kind stays always on.

The engine's rules call it "any manual step an agent runs to keep Continuum working." On 26 September Claude stopped and started the AppHost by process id through WMI, queried Postgres, Redis and Temporal through docker exec, and messaged another session about build windows and process ids. The next morning LM Studio had restarted overnight with nothing loaded, and the profile went back on by hand. Each step cost tokens every time, and none of them happened while no session was running.

The rule that came out of it, Glue Becomes Code, retires each step on the cheapest rung that can do it:

1. Plain code or a Terminal verb, for anything deterministic.
2. An MCP tool, which answers in one compact reply where docker exec … psql returned pages.
3. A local model, for routine judgement, once a battery has measured it on that job.
4. A paid model, only with the person's confirmation, under a cap of $1 a day across every paid provider.

The build list written that day estimates that about half its steps were rung-one work done by a model. The WMI went first: continuum stack start|stop|status, proven the same day. Hand-run polls of Temporal's HTTP API became continuum ops queues, and on 28 September the hand-typed terminates became continuum ops terminate, proven live after the day's reset. The test for each one is whether it still works with no session running.

• Rules scoped to paths. Five of the engine's nine rules load only when a session touches their files. This site's 67 KB animation guide went behind one on 28 September, loading when a session opens an animation script, component or skill, and its always-on set fell from 161 KB to 94 KB.
• A README on touch. A hook hands a session a project's README the first time it reads a file in that project, once per session. STRUCTURE.md was rewritten on 22 September, when its file tree had reached 45,898 characters under one heading, 70 times the median section, and each project's detail moved into the README that arrives with its code.
• Hooks that say nothing until there's something to act on. A hook that asks at each prompt whether anything is waiting prints nothing when nothing is, so it costs no tokens while things are quiet.
• The catch-up at session start, under a budget. The last session's open questions arrive with the new one, under 8,000 bytes, for reasons the next entry tells.
• Compaction rescue. Before a compaction a hook saves the person's messages the journal hasn't recorded, and hands them back after it.
• Subagents by tier. On 23 September the first architecture survey launched 13 agents at once, each costing about 250,000 to 450,000 tokens. From then on, at most four ran at a time, with Sonnet reading and Opus judging, and from 27 September one or two.
• No polling. A session waits for the notification, since every look is a turn.
• No command that stops for approval. On 28 September a delete on a path built from a variable sat on an approval prompt for 40 minutes while Tristan was away from the desk. Runs now write to a new folder instead of deleting to make room.

Summarising long output, classifying a line, triaging a failure and writing a first draft go to the RTX 5090 once a model has been measured on that exact job. Anything not yet measured stays on the rung above. The engine's LOCAL_LLM.md holds the rules every integration has hit:

• Send every sampler, every request. An omitted one is the runtime's chat default, and LM Studio's repeat penalty of 1.1 punishes JSON, which repeats by construction.
• Read the request on the wire. On 22 September three samplers were missing from the wire while the configuration looked right. On 28 September the same client library turned a schema's maxLength into description text, removing the bound that stops a model looping inside a string.
• Check the response's model. LM Studio answers a request for a model that isn't loaded with one that is. On 24 September that recorded 46 claims as the extractor's, and the extractor hadn't answered any of them.
• A big model loads alone. Loaded beside the chat and embedding models, the extractor warmed up at 4.1 tokens a second, with its layers on the CPU.

The receptionist that routes Continuum's messages is on the same path: a routing battery decides which local model makes the choice.

Its chapter on eliminating toil defines it as work "that tends to be manual, repetitive, automatable, tactical, devoid of enduring value, and that scales linearly as a service grows," and keeps it "below 50% of each SRE's time." Glue is toil done by a model: agent toil. It is paid in tokens, and it comes back every session, since a session starts with nothing but what was written down.

Mitchell Hashimoto's "Engineer the Harness" (February 2026) says what to do about a mistake: "you take the time to engineer a solution such that the agent never makes that mistake again." Every's compound engineering (Dan Shipper and Kieran Klaassen, December 2025) expects "each feature to make the next feature easier to build." Toyota's jidoka is a line that stops itself when something is wrong, and the lease gate is one.

What this work adds is the ladder, whose third rung is a local model measured on the job before it is trusted; a hook that notices the agent's own glue while the session is still running; and the test that a retired step works with no session running.