Claude Code · the machinery around the model

The harness that runs this foundry.

The model does the reasoning. Everything else — hooks, guards, routing rules, recorders — decides what it may do, how work gets dispatched, and what survives the session. Drawn below as the machine hall it behaves like.

Walk the line ↓Help me get better ↗

Audited by its own machinery

79 / 100↑ from 73

Independent adversarial re-score, not self-asserted. Honest gaps below — see what's still weak.

intake the askguards deny the catastrophicrouting mode · tier · effortfurnaces fable/opus · sonnet · haikucasting commitsrecords session logsreturn pipe lessons become rules

01 · The switchyard

Same yard, different trains.

Every task is routed twice: an orchestration mode picks the shape of the work, then each dispatch sets two knobs — model tier and reasoning effort. Pick a scenario; watch the line light up.

Workflow — plan → approval → phases land serially/fix BUG-004 — “barcode lookup silently overwrites the name”

plan authored at planner tier · fable/opus · xhigh — then WAITS for approval

implementation is TDD at executor · sonnet · med–high

digests, checklists, sweeps at chore · haiku · low

every phase: compile → whole CubbyTests → reviewer → push → docs: checkpoint

The rule underneath: pick the cheapest mode that fits, escalate effort one step before tier, and never dispatch with both knobs implicit — an unset knob silently inherits the orchestrator's expensive defaults. Calibration point: one 26-agent review left on inherited defaults cost ~1.8M tokens; the same shape on the matrix runs at a fraction.

02 · The conveyor

Hooks the model can't reason around.

Five lifecycle events, each with shell scripts the runtime executes — not the model. Conventions become mechanisms at the exact moment they could be violated.

The honest line: three guards are genuinely hard — bash-guard, test-scope, generated-files (it fails closed). The two checkpoint gates are block-once reminders: they stop the first push, and pass the second identical one. That gap is named in the audit, not smoothed over.

03 · The ledger — shipped this week

Two loops, opposite cost profiles.

Six new slash commands turn "found a bug" into a committed ledger row in seconds — and keep all research out of capture. Resolution is the expensive loop, run with full wave ceremony. The rows below are the real backlog, logged from one device pass.

Capture — chore tier, no research

/issue /feature /task — dedup against the ledger, assign an ID, append, commit. A screenshot token grabs the live simulator frame. Reporting a bug never turns into an uninvited debugging session.

Resolve — full ceremony

/fix BUG-004 groups a wave, recommends a /goal, authors a plan at planner tier, then stops for approval. Only after the yes: TDD, per-phase review, gates, device-verify, archive.

Learn — suggest-only

/tracker-learn mines items that took ≥2 attempts and asks why. It emits a prose report naming a target — a skill, an AGENTS.md amendment, a hook — and drafts nothing. The user decides what becomes machinery.

04 · The paper

One source of truth. Literally one.

CLAUDE.md is eight lines — an @import shim. Everything binding lives in AGENTS.md, where every rule traces to a named incident, commit hash, or measured token count.

cubby/CLAUDE.mdverbatim · all 8 lines

# CLAUDE.md

@AGENTS.md

The import above loads AGENTS.md — Cubby's canonical
conventions, the single source of truth across tools
(Claude Code, Cursor, Codex, …) — into every Claude Code
session. Do not duplicate rules here; amend AGENTS.md instead.

Two copies of a rule is how rules rot. The shim guarantees there's only ever one.

cubby/AGENTS.md — the register51 KB · 305 lines

Part 1 — Cubby engineeringLocked stack decisions
Roadmap amendments
Persistence rules
Protocol boundaries
Scan-quality rules (12)
Deep-link contract
Swift 6 concurrency playbook
Enforcement map
Testing · Build status

Part 2 — portable processAgent behavior
Orchestration modes
Wave workflow
Model routing & plans
Effort & dispatch cost
Session logs
Compaction protocol

Amendment rule: a decision that constrains future sessions must fold back into this file — a decision living only in a session log is session-local by definition.

"Current tier mapping (edit me, not the rules): planner/orchestrator = Fable/Opus · executor = Sonnet · chore = Haiku."
— the one line to edit when the lineup changes

"Challenge by default, then commit. Push back with substance — yours or the user's — before the call locks. Once made, execute; don't re-litigate."
— agent behavior, evidenced by sessions 0025/0027

"Never dispatch with both knobs implicit — an unset knob silently inherits the orchestrator's expensive values."
— effort & dispatch cost, after a measured 1.8M-token waste

There's more here than the outline. The page source greets any agent pointed at it with a fuller digest of the register — or open the console and type AGENTS().

05 · The arsenal

What's actually on the racks.

The enabled surface, not the installed one — and the distinction matters: ~360 skills are installed; curation is the open problem the audit keeps naming.

Plugins

13 on · 4 retired 07-12

superpowersios-devecccartooncavemancontext7visual-explainercore-workflowcommit-commandsclaude-md-managementskill-creatorplugin-devswift-lsp

cartoon wraps noisy build output into TOON (~70% fewer input tokens); caveman strips output filler (~65%). Compression is what makes the exhaustive default affordable. Retired as safe-defaults: security-guidance, claude-code-setup, ralph-loop, frontend-design.

Skills that earn their keep

114 enabled · ~360 installed

Process — run first, always

brainstorming · writing-plans · test-driven-development · systematic-debugging · verification-before-completion

Mined from real Cubby incidents (~51)

avfoundation-capture-delivery-watchdog · swift6-mainactor-compile-fixes · xcodegen-test-targets · release-inapp-vs-asc-whatsnew-surfaces

Project-local

session-log · tracker new

Precedence when several match: project-mined > AGENTS.md playbook > global rules > generic catalog. Declared this month; before that, first-match won.

Slash commands

the daily levers

Tracker new this week

/issue/feature/task/backlog/fix/tracker-learn

Session & quality

/goal/code-review/release testflight/compact

Learning

/learn/evolve/promote/instinct-status

/goal arms a Stop hook that blocks the session from ending until the stated condition holds. /compact is user-typed only — the agent's ceiling is suggesting a safe compaction point.

06 · The audit

79/100, and what that number hides.

Six agents read the actual config; an independent planner-tier pass scored it and refuted two of its own draft findings. Up from 73 — driven by real repo-scoped enforcement — and stale by its own rules until the next re-score.

Dimension scores — independent re-score, 2026-07-12

Workflow & orchestration17 / 20

strongest layer — still prose-enforced

Hooks & enforcement16 / 18

bash-guard hardened, 20/20 fixtures live

Recorders, memory & checkpoints12 / 17

instinct pile uncurated and growing

Skills ecosystem10 / 15

precedence declared; sprawl unchanged

Safety, cost & token control11 / 15

allowlist narrowing deferred by choice

Documentation13 / 15

every rule traces to an incident

Still weak — on the record

272 banked instincts, uncurated — the observer that minted them is paused after ~92 failed runs; pruning found zero safe deletions.
~360 installed skills across four populations that were never designed together; Swift 6 concurrency alone has 5–6 overlapping sources.
The push and stop-log gates verify that a checkpoint ritual happened, not that tests ran or a review was real.
The permission allowlist still auto-approves a few destructive shapes; narrowing it was deliberately deferred.

Full method + scorecard: cubby/docs/harness/EVALUATION.md · architecture: cubby/docs/harness/ARCHITECTURE.md (private repo — not published here)

Open workshop

Steal what's useful. Help me get better.

This is a working setup, not a finished product — the audit above is the proof it isn't done. If you run a harness of your own, I'd genuinely like to compare notes.

Mail feedback ↗LinkedIn ↗

Foundry · Abhijit Bansal · exploration 05 — the harnessfigures drawn from the live config · 2026-07-13