v0.7.0

Added

• Per-prompt reasoning-level control via --jev-thinking / PI_JEV_THINKING=1 / /jev thinking [on|off]. Prompt intent escalates (high/xhigh) for planning, debugging, security, and review tasks and de-escalates (minimal) for short mechanical prompts. The model is never changed, so the prompt-cache identity stays fixed (#22).
• /jev status and /jev usage list the new thinking mode (#22).
• modelSwitchCostUsd, modelStayCostUsd, hasKnownCost, and evaluateModel are exported, and ModelRouteResult reports fit, switchCostUsd, and stayCostUsd so the decision is inspectable (#24).

Changed

• Auto-model scoring is now cost-aware. The scorer previously had a single cost reference, (model.cost?.input ?? 0) * -0.01, which mixed arbitrary capability points with dollars and left cost at roughly 0.3% of the score; four of the six profiles had no cost term at all. Candidates are now ranked by capability fit, and cost only chooses between models within FIT_TOLERANCE of the best fit. Decision granularity is unchanged: still one decision per prompt, no new flags (#24).
• Cost can no longer downgrade a capability class. An additive cost term against a bounded fit scale would let a long conversation's cache miss outweigh the entire capability range — a reasoning task could be routed to a non-reasoning model purely because the reasoning one was pricier. Keeping cost as a tie-breaker means price decides only between near-equivalent models (#24).
• Capabilities are scored only when the profile needs them. Image and URL points were credited in every profile, so a multimodal model won text-only tasks while being the most expensive option — for debug this failing test, a $15/Mtok vision model beat an equivalent $3/Mtok text model. Hard input requirements are now filtered before scoring (#24).
• Cache misses are priced with the model's real rates (cost.input, cost.cacheRead, cost.cacheWrite) and the live prefix size from ctx.getContextUsage().tokens: switch cost is prefixTokens * (input + cacheWrite), stay cost is prefixTokens * cacheRead. Pricing tiers are applied, as pi's own cost calculation does — the bundled catalog tiers 24 models, doubling flagship rates above ~272k tokens. An omitted cost is treated as unknown rather than free (it is charged the current model's rates), while an explicit zero cost remains known and free (#24).

Fixed

• Skill matching checks requested activity and product scope rather than related topics. The skill router no longer recommends setup-pstack for an exhausted-model routing bug in the live regression corpus, while retaining genuine pstack configuration matches. The shared 0.65 cutoff is unchanged. Added an opt-in live replay and the investigation/evidence in docs/skill-routing-investigation.md (#21).
• Jev API key source precedence aligns origin with the active key. Non-empty runtime keys set in-session properly override environment keys while preserving source attribution, and empty in-session strings safely fall back to the environment key (#20).
• URL prompt model routing detects URLs in raw prompt text. Model routing automatically routes requests containing raw URLs to URL-capable models using promptHasUrl regex matching (#19).

Notes

• Budget-based Anthropic thinking (models without compat.forceAdaptiveThinking) is skipped: Pi derives budget_tokens from max_tokens, which is itself derived from the level, and Anthropic keys the message cache on budget_tokens. That claim comes from Pi's own source comment and is not verified against Anthropic's public documentation.
• FIT_TOLERANCE is calibrated against the fit scale: wide enough for price to act, narrower than the reasoning (+30) and image (+40) gaps so cost can never overturn a real capability difference.
• When Jev is configured, a switch is additionally gated by one Noul judgment, thresholded by requiredConfidence(switchCost). Jev supplies the probability and code owns the threshold (P > cost/value), so no dollar value is assigned to a correct answer.
• The model-switch question is batched with the tool/skill questions when /jev auto is also on, so a prompt still costs one Jev request rather than two. Auto-model alone also costs one; without Jev it stays fully local and only breaks ties on cost.

Acknowledgements

Thanks to the contributors who filed issues and opened pull requests for this release:

• @armstrys — per-prompt reasoning-level control (#22) and cost-aware auto-model scoring (#24).
• @leftytennis — skill-routing false-positive fix (#21).

v0.6.0

Added

• Custom Jev-compatible endpoint support via PI_JEV_BASE_URL or TYPESAFE_BASE_URL, including unauthenticated local servers such as Laya laya-serve.
• /jev status now shows the active Jev endpoint.
• OpenCode Zen URL-aware model routing.

Fixed

• pi-jev-gate runtime loading now works from package installs without dev dependencies in the caller cwd.
• /jev status reports the real API key origin and counts SDK input_tokens/output_tokens usage.
• --jev-auto-model and --jev-agents take effect independently of /jev auto.
• /jev auto now sends one Jev request per prompt as documented.

v0.5.0

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[0.5.0] - 2026-09-20

Added

• Tool Guard: Opt-in tool call validation and anti-hallucination interceptor (--jev-tool-guard, PI_JEV_TOOL_GUARD=1, /jev tool-guard [on|off]). Evaluates tool parameters with Jev System One to block hallucinated paths/flags and enhances error output with targeted recovery hints.

Fixed

• Safe Fallback: When Jev is unreachable or unconfigured, tool router no longer auto-activates tools blindly and reports 0 probability rather than false certainty (1.0). Skill router only surfaces keyword matches with 0 probability (closes #1: "A failed request activates three tools and reports them at probability 1.0").
• Removed outdated reference to nonexistent /jev login in jev_evaluate error message.
• Documentation clarifies that heuristic routing (/jev auto-model, topology fallback) executes locally without spending Jev requests.

[0.4.0] - 2026-09-18

Added

• Jev Gate CLI binary (bin/jev-gate.js, exposed as pi-jev-gate and jev-gate) for subagent post-run gate checks and CI/CD validation. Evaluates git diff, stdin, or files against acceptance criteria with fast System One noul probability.
• Typed Jev Subagent (agent: "jev" / agentType: "jev") handler in pi-subagents RPC for sub-second, zero-LLM-overhead choice, score, and probability decisions inside workflows.

Fixed

• /jev agents <task> now directly constructs multi-agent workflowScript topologies delegating to builtin agents (scout, worker, reviewer, researcher, evidence-auditor), replacing single delegate subagent calls.

[0.3.0] - 2026-09-17

Added

• Opt-in automatic model routing via --jev-auto-model, PI_JEV_AUTO_MODEL=1, and /jev auto-model [on|off].
• Model profiles for fast, balanced, reasoning, long-context, and vision tasks. Selection respects scoped models and attached images.
• Provider-limit handling: quota, rate-limit, timeout, unavailable, auth, and context-limit errors are classified; retry-prone models are temporarily avoided on later prompts without loops or silent truncation.
• Opt-in Jev-guided /compact via --jev-compact, PI_JEV_COMPACT=1, or /jev compact on. Important tool history is retained in a custom compaction summary, with Pi's built-in summary as fail-open fallback.
• Explicit agent orchestration via /jev agents <task> and opt-in automatic orchestration via --jev-agents, PI_JEV_AGENTS=1, or /jev auto-agents on, using the installed pi-subagents RPC.

[0.2.1] - 2026-09-17

Documentation

• Add secret store key resolution option (~/.pi/agent/secrets/typesafe_api_key) to Setup section in README.

[0.2.0] - 2026-09-17

Added

• Dynamic evaluation command: /jev test <prompt> (aliases /jev eval, /jev evaluate) asks the session's active model to design the Jev question schema from the user's prompt, then runs it on TypeSafe Jev. /jev test alone still runs the fixed smoke test.
• Automatic mode: --jev-auto flag / PI_JEV_AUTO=1 env var and /jev auto [on|off] command run one Jev routing pass before each prompt, activating tools and surfacing matching skills.

Changed

• Single activation threshold JEV_THRESHOLD (0.65) in src/skills.ts, used by the router, both tools, /jev skills, and auto mode. /jev skills previously used 0.6, so manual skill search could show matches auto mode hid.

Fixed

• Router no longer offers pi-jev's own tools as routing candidates. After /jev disable, automatic routing used to re-activate jev_find_skill and jev_evaluate.
• /jev subcommands now match exactly, so /jev autofoo on and /jev skillsfoo report an error instead of silently toggling or searching.
• /jev skills discloses local heuristic fallback instead of presenting 1.00 probabilities as Jev judgments.
• /jev status reports where the API key came from ($TYPESAFE_API_KEY vs ~/.pi/agent/secrets/typesafe_api_key) and counts only genuinely routable tools.
• /jev help lists usage at info level instead of warn-as-unknown-command.
• Router and skill fallback tests no longer depend on the machine being unconfigured.
• npm run smoke passes -ne so it no longer collides with an already-installed pi-jev copy.

[0.1.1] - 2026-09-17

Added

• jev_find_skill tool and /jev skills [query] command for semantic skill discovery and recommendation.

[0.1.0] - 2026-09-17

Added

• Initial public release of pi-jev package for the Pi coding agent.
• jev_find_tools tool for semantic candidate shortlisting and additive tool activation.
• jev_evaluate tool exposing typed TypeSafe Jev decisions (Choice, Noul, Score).
• /jev slash commands (status, enable, disable, test).
• Bounded TypeSafe client integration with safe error handling and usage accounting.
• Comprehensive unit test suite and CI workflows.

v0.4.0: Jev gate CLI, subagent runner & workflow delegation

Full Changelog: v0.3.0...v0.4.0

v0.3.0: Agent orchestration & Jev-guided compaction

Full Changelog: v0.2.1...v0.3.0

v0.2.1: Secret store support & bug fixes

Full Changelog: v0.2.0...v0.2.1

v0.2.0: Dynamic /jev test & opt-in auto mode

Full Changelog: v0.1.1...v0.2.0

v0.1.1: Skill Discovery Support

Added\n- jev_find_skill tool and /jev skills [query] command for semantic skill discovery and recommendation.

v0.1.0: Initial Release

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[0.1.0] - 2026-09-17

Added

• Initial public release of pi-jev package for the Pi coding agent.
• jev_find_tools tool for semantic candidate shortlisting and additive tool activation.
• jev_evaluate tool exposing typed TypeSafe Jev decisions (Choice, Noul, Score).
• /jev slash commands (status, enable, disable, test).
• Bounded TypeSafe client integration with safe error handling and usage accounting.
• Comprehensive unit test suite and CI workflows.