OMP Code Model

omp-code-model adds same-conversation coding phases to OMP. The main model defines requirements and interfaces, hands implementation to a configured coding model, and returns for final review with the full history and tool state intact.

Watch the complete accelerated MP4 recording.

Flow

main model
  → code-model start
  → configured coding model and effort
  → implementation and targeted checks
  → code-model finish
  → original model review

Recorded cost example

The current demo records one complete task launched with omp. GPT-6-Astra plans the work, Gemini 3.8 Flash implements it in the coding phase, and GPT-6-Astra returns to inspect the diff and correct three issues.

Model and role	Input	Output	Cache read	Recorded cost
GPT-6-Astra · planning and final review	136,367	12,085	1,627,136	$3.595056
Gemini 3.8 Flash · coding phase	360,267	42,471	4,200,137	$0.744477
Complete conversation	496,634	54,556	5,827,273	$4.339533

Applying GPT-6-Astra's listed rates to every observed token produces a price-normalised estimate of $13.521413. The recorded model split is $9.181880 lower, a 67.9% reduction for the complete conversation. The coding phase alone costs 92.5% less at the recorded Gemini rates. This comparison holds token counts constant; an all-GPT run can follow a different execution path and produce a different token count.

The recording and its machine-readable usage data remain in demo/session-cost.json.

The extension provides:

• an essential code-model tool with recommend, start, finish and status actions;
• optional TypeSafe Jev routing across the current main model, the configured coding model and an explicitly authorised subagent;
• opt-in before_agent_start routing modes: off, observe and enforce;
• quota-aware recommendations sourced from omp usage, with configurable main_agent or code_model fallback;
• graphical /code-model settings for Provider → Model → Effort → Save and a /code-model routing selector for mode and fallback;
• persistent phase, recommendation and automatic-routing audit state in the session tree;
• restoration after completion, cancellation, retry fallback, session navigation and shutdown;
• English and Simplified Chinese menus selected from the current locale.

Install

Install directly from GitHub:

omp plugin install github:cyriusweng/omp-code-model

Restart OMP after installation. Local development can use a link:

git clone https://github.com/cyriusweng/omp-code-model.git
omp plugin link ./omp-code-model

OMP 18.2.0 is the compatibility target.

Configure

Open the graphical selectors:

/code-model
/code-model routing

The first menu presents providers and models from the active OMP catalogue. The routing menu configures Jev preflight behaviour and fallback. Both selections are stored at:

~/.omp/agent/code-model.json

PI_CODING_AGENT_DIR changes the active agent directory. OMP_CODE_MODEL_CONFIG can name a dedicated configuration file.

Useful commands:

/code-model show
/code-model models
/code-model recommend implement the requested change and run targeted tests
/code-model routing
/code-model routing observe main_agent
/code-model routing enforce main_agent
/code-model routing off
/code-model status
/code-model start
/code-model finish

The main model usually calls the code-model tool itself. recommend is advisory and sends its task summary plus sanitised model and quota facts to TypeSafe when /login typesafe or TYPESAFE_API_KEY provides a credential. Its result identifies the judgment backend, confidence, quota state and fallback reason. start is a standalone tool call before implementation; its optional effort applies the recommendation for that coding phase, while omission uses the saved profile. finish is a standalone tool call after targeted checks. The active OMP approval policy continues to govern tool use.

Automatic routing defaults to off. observe asks Jev before every user-containing prompt and records a receipt while preserving the current model. enforce applies eligible routing recommendations whose finite confidence meets routing.minConfidence, which defaults to 0.5 and accepts values from 0 to 1 in the configuration file. This threshold applies to both code_model and main_agent recommendations. Low-confidence or missing-confidence recommendations preserve the actual active phase. Interactive routing changes preserve a saved custom threshold.

The configured fallback runs when the TypeSafe credential or request is unavailable. main_agent retains or restores the main phase according to phase ownership; code_model enters the configured coding phase when its model and quota remain eligible. This explicit fallback policy operates independently of the TypeSafe confidence threshold. Receipts record the requested route, effective route, confidence, threshold and decision reason. Automatically owned phases follow the existing completion and retry lifecycle, while manually started phases retain their explicit ownership.

Working with OMP Jev Gate

OMP Jev Gate complements Code Model with a global judgment policy and a first-tool checkpoint. Code Model's Jev preflight selects the main agent or configured coding model and the coding effort. Jev Gate's separate preflight supplies decision mode, reasoning depth, verification depth and later-checkpoint likelihood, then governs edit, write and bash. OMP serialises both before_agent_start hooks before the provider request and preserves system-prompt amendments, so an automatically selected coding model receives Jev Gate's policy from its first response.

Phase controls and guarded work retain separate scopes. code-model start can enter a coding phase while a Jev Gate turn is pending; the selected model then calls jev-judge when the policy's four trigger conditions hold, and the resulting disposition unlocks guarded tools for that task turn. code-model finish restores the original model with the same conversation history and current turn disposition. When agent completion starts a fresh review turn, automatic routing and Jev Gate each produce a new preflight and receipt.

Each plugin normally issues its own focused TypeSafe request. Their fallback settings compose. With Code Model routing set to enforce main_agent and Jev Gate set to enforce continue, an unavailable routing judgment retains or restores the main executor, while an unavailable gate judgment records degraded_continue and lets the agent proceed with audited reasoning. Jev Gate's enforce block setting gives the affected turn fail-closed availability.

Reliability behaviour

The extension records the original model and configured effort before entering a coding phase. A model or effort selected through another OMP control remains active when the phase ends. Retry fallback entries remain phase-owned so a repeated start keeps the original restoration target.

An interrupted switch stores a recoverable phase marker. Session lifecycle events restore recorded state, navigation waits for restoration, and the configuration writer uses an atomic rename with concurrent-update detection.

Development

Requires Node.js 22 or later.

npm test
npm run check

The test suite covers configuration integrity, interactive menu flows, manual and automatic Jev routing, prompt-hook idempotency, fallback behavior, phase transitions, recovery, retry fallback, configured auto, extension registration and lifecycle completion.

Licence

MIT