01Project 01 of 3
Claude Code is the lead developer. Everything around it exists for two reasons: keep its context, its working memory, for judgment, and turn my working method into steps it cannot skip.
A skill is a written procedure Claude Code loads when a task calls for it, such as /analyze.
searches project memory, surfaces constraints, writes the plan
implements in its own worktree, checking memory before each file
a local judge compares the code changes with the plan
merges in a throwaway worktree, never the shared working copy
build, start, health check
records decisions, mistakes and gotchas
The two locks: no code edit before the plan is written, no merge before the critic’s verdict. Runs over the 30 days to 14 October 2026.
/decision pressure-tests a costly choice: it checks what the project already has, surveys the state of the art, attacks the strongest version of the choice, then gives a go or no-go. /test-ui drives the app like a real user, in a browser, to prove a change end to end. /prompt-work is loaded before editing any text a model reads.
/session-start opens an isolated git worktree, a separate working copy of the repository, through dc, so parallel sessions never collide. /track logs a broken rule so it comes back next session.
A hook is a script Claude Code runs on its own at set moments: at session start, before an edit, before a command.
No code edit before a written plan, nor outside /dev. No edit in the shared working copy: open a worktree. No merge before the critic’s verdict. No edit to a prompt before /prompt-work is loaded. No large read without delegating it or saying why.
At session start, the three rules Claude breaks most. When it is about to start a cloud sub-agent: could a free local worker do the job? Before a risky shell command, such as killing every process of one name: the rule it would break. After the context is compacted: search the vault again, since what it found there is gone.
Skill and sub-agent runs are logged, the plugin version is checked at start, and long answers end on a short TL;DR.
Each worker is pi, an open-source coding agent, on one of the two local Qwen models: the MoE (mixture of experts, fast, for volume) or the dense one (slower, it cites and judges). Only the short answer comes back into Claude’s context · zero API cost.
Outside MRS Memory · Claude Code’s notebook · about 460 notes
Before editing a file, to see what was decided about it. After a context compaction, when a hook tells it to search again. At the end of a batch, when a skill writes the decisions and corrections and pushes them, searchable within seconds. Over three months to July, 73% of coding sessions checked a file’s notes before the first edit.
Exact questions first (this file, this tag, these links), then by meaning: keywords and embeddings merged, reordered by a cross-encoder, with linked notes up to two links away. On 35 test questions in May, a relevant note was in the top five for 76% of them, against 64% with one method alone.
A duplicate check runs before any new note. A note that sets a constraint must carry a proof one can re-run (a command output, a file and line, a commit), and a linter checks it. Each note lists the code it describes and its last check date, so an old one comes back flagged. The note points, the code confirms.
The hooks and the local workers keep logs, and a tracker counts how often Claude breaks each rule it can observe. Two versions are compared by running them alternately in the same time window, so the time of day cannot explain the gap, and a result from a handful of runs is checked on a larger sample. Example: a first test of 8 runs suggested that agents never delegate on their own. A larger one, on a single task, gave 6 of 12 without a reminder and 12 of 12 with it. Fisher’s exact test gives p = 0.014: if the reminder did nothing, a gap this large would appear by chance about once in 70 tries. In everyday sessions the rule was followed far less often (see Limits), so it is now enforced by a hook.
An instruction in prose gets lost as the context fills; a hook acts at the moment of the action, and it works because it blocks. Measured over two months on the same sessions: when a hook blocked an edit in the shared working copy, the command it named followed within ten minutes in 72 cases out of 103; when a local model only added a warning to an action that went ahead anyway, none of its 33 warnings was taken into account. A gate that blocks must offer a cheap way out: opening a worktree takes four seconds, and the plan gate lifts as soon as the plan’s checklist is written.
72 of 103 (70%) · 0 of 33 · June to August 2026
Measured: of 1,669 checkable citations, 129 (7.7%) had an invented line number, hence the re-check described in section D. In the safest setup, a single-target lookup on the dense model, 398 of 398 checkable citations were correct (312 runs).
129 of 1,669 (7.7%) · 398 of 398 · August 2026
The orchestrator’s context is the scarce resource, so I split work by whether a failure would show. Verifiable bulk reading goes to local Qwen workers, at no cost; web collection to a cheaper cloud model (Sonnet), where a wrong answer can be filtered afterwards; judgment that cannot be re-checked stays on the strongest one (Opus), because a weak judgment is invisible.
Each session has its own branch and git worktree. dc land merges in a throwaway worktree, never in the shared working copy. dc deploy refuses to overwrite diverged local work on the server and reports failure when the health check is red.
It runs for one person, on two machines, Windows and Linux, with the same plugin. It is how I work, not a team platform or a hosted service.
“Delegate bulk reading” is the rule Claude breaks most. A reminder after the fact was followed 8.4% of the time (28 of 335, five weeks to August), so a hook now stops the first large read, and the rule comes back at session start.
Claude Code’s notebook, about 460 short notes of corrections and lessons, keeps no history: a note that is overwritten or deleted is gone for good, whereas every change to the vault is kept. This was the first gap found in a comparison of the harness with public practices in September.