It directly addresses the core problem of stochastic agent behavior by making the vast majority of actions deterministic and verifiable, with measured reductions in model output and hidden LLM-mediated fetches.
One question: of everything an agent does, how much can be made reliably deterministic?
Detent is a strictly passive, deterministic leverage layer for coding agents. The LLM's irreducible remainder is reasoning — deciding, and emitting specs. Everything else — every read, write, retrieval, verification, and presentation of fact — runs through deterministic machinery instead of the default token transport: pure functions of the event and disk state, no model call anywhere, ever. Generation is stochastic exactly once; the moment output exists it is a content-addressed artifact, and every later use is a deterministic copy, never a re-emission. Denial is never the point — it is only the redirect at the boundary, used exactly where the default path is provably identical to a free deterministic one.
Read LAW.md (the admission test any move must pass). The primitive layer itself —
5 stations, 20 flow cells, complete by construction, every cell SERVED,
PARTIAL-with-named-machinery, or VOID-with-reason; silence is not a state — ships as code:
detent/cells.py, the punchcard.
| Piece | What it is |
|---|---|
detent/contract.py | the bound: Claude Code's documented hook schema as data, plus the three envelope types (rewrite dict / Deny / Block / advisory str) |
detent/dispatch.py | the pivot: one stdin→stdout hook entrypoint; (event, tool) table lookup; picks the envelope by return type, never by content |
detent/moves.py | the arms: every move is one composition m = α∘φ∘π — total projections, exact guards, typed envelope actions; 29 moves across three tiers, enforcement (pure), capture (write only to the store, atomically), and display (render-by-address at the screen boundary), each declaring its own station flows; coverage is a quotient, not an enumeration — wildcard rows bound EVERY tool's oversized output and gate the whole →WORLD class, including tools that don't exist yet |
detent/store.py | the substrate: content-addressed artifact store — put/get/materialize/slice + an append-only firing ledger; identical bytes coincide, so N concurrent writers need zero coordination |
detent/cells.py | the punchcard: the 20-cell coverage topology as data; the move rows are derived from each move's own flow declaration, reconciled against importable reality by tests — coverage drift is a red |
Similar projects matched by category, topics, and programming language.
Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.
A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.
mimic captures traffic from any iOS or web app and automatically generates a Python client library that lets you call the app's API like a regular library.
