Triages test failure traceback, compilation error, or runtime exception using Jev System One non-autoregressive decision classification (70-300ms, zero LLM generation). Detects missing environment packages, flaky transient glitches, or deep logic defects.

Use when: An agent encounters a test failure, traceback, compiler error (e.g. TS2307, ModuleNotFoundError, E0463), or needs to decide whether to invoke a frontier LLM. Do NOT use when: Tests pass, or for general code review or feature generation.

Returns: JSON object with 'category' (env_missing, flaky_transient, deep_logic, no_failure), 'skip_llm' (boolean: true if resolvable deterministically without frontier LLM), and 'action_recommendation' (string with concrete recovery action).

Guards against doom loops, repetitive circular retries, dead-ends, and destructive refactoring before burning frontier reasoning tokens. Evaluates the agent's proposed plan against recent attempt history.

Use when: An agent is about to retry a failed step, execute a code edit after previous failed attempts, or before embarking on a potentially circular fix. Do NOT use when: Making the first attempt on a fresh task with no prior failure history.

Returns: JSON object with 'should_abort' (boolean: true if the trajectory is stuck in a circular loop), 'abort_probability' (number: 0.0 to 1.0), 'action' (string: PROCEED, HALT, PIVOT), and 'reasoning_summary' (string explaining the decision).

Routes a programming task to the minimal sufficient model tier (deterministic script, lightweight flash model, or heavy frontier reasoning model) to minimize latency and token expenditure.

Use when: Starting a new task, refactoring step, or bug fix to choose between lightweight models and expensive reasoning frontier models. Do NOT use when: Diagnosing test execution tracebacks (use jev_triage_test_failure instead).

Returns: JSON object with 'selected_tier' (string: deterministic, lightweight, heavy), 'complexity_score' (number: 1.0 to 5.0), 'recommended_model' (string), and 'rationale' (string).

Calibrates step completion against acceptance criteria using typed rubric scoring. Evaluates whether produced evidence proves the task is finished without running redundant review loops.

Use when: An agent believes a task or milestone is complete and wants to verify acceptance criteria before concluding. Do NOT use when: Work is still underway or tests are actively failing.

Returns: JSON object with 'is_verified' (boolean: true if acceptance criteria are satisfied with proof), 'satisfaction_probability' (number: 0.0 to 1.0), 'rigor_score' (number: 1.0 to 4.0), 'needs_rework' (boolean), and 'confidence' (number: 0.0 to 1.0).

Dynamically modulates reasoning effort (low, medium, high, etc.) and generation stability lease steps for the immediate LLM call. Maps provider-specific parameters for OpenAI (GPT-6 Astra/o3), DeepSeek (V4.1-Flash/R1), Qwen (3.8 Max), Anthropic (Claude Fable 5.1), and Gemini (3.8 Thinking) to prevent reasoning token waste.

Use when: Preparing a prompt or tool call for a reasoning-capable LLM to calibrate thinking effort according to task complexity. Do NOT use when: Calling standard non-reasoning models or local deterministic scripts.

Returns: JSON object with 'effort' (string), 'provider' (string), 'provider_params' (object with provider-native kwargs), 'is_reasoning_supported' (boolean), 'cache_safe_recommendation' (string), and 'lease_steps' (integer).

Evaluates whether an autonomous agent paused prematurely with unfinished work or unverified changes (covering workflow phases: research, ask, plan, execute, verify, complete). Vetoes nudges when waiting on user input or when repeated nudges make no progress.

Use when: A background worker or agent loop stops and you need to determine if it should be nudged to continue autonomously. Do NOT use when: The agent explicitly requested user confirmation or required credentials.

Returns: JSON object with 'should_nudge' (boolean), 'workflow_phase' (string: research, ask, plan, execute, verify, complete), 'action' (string: NUDGE, WAIT, STOP), and 'confidence' (number: 0.0 to 1.0).