# Changelog

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

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

## [0.3.0] - 2026-06-12

### Added
- TypeScript typechecking: `tsconfig.json`, dev dependencies on the pi packages, and a `typecheck` script wired into `npm run check`.
- Unit tests for stage detection, tool-result summarization, exit-code parsing, and verification-command matching (`tests/advisor-signals.test.mjs`).
- Edit/write summaries in advisor context now include `(+added/-removed)` line stats derived from the edit tool's unified patch details.

### Changed
- Advisor system prompt now states it sees a curated subset of the executor's context (not "what the executor sees") and instructs the advisor to request missing evidence as its first action item instead of guessing when the transcript is too thin to judge.
- The final-check stage directive points the advisor at the original user request retained at the top of the transcript, so sign-off is checked against the user's requirements rather than the executor's interpretation.
- Context policy wording fixed to match actual behavior: tool results are summarized (not "not replayed"), truncation omits middle messages (task framing is retained), and the transcript sits above the closing context block.
- Removed the duplicated `<advisor-tool>` system prompt injection: usage guidance now lives solely in the tool's `promptGuidelines`, which pi injects natively when the tool is active. The executor system prompt no longer carries the advisor instructions twice.
- Advisor context (stage, directive, executor signals, recent tool activity) moved from the head of the transcript to the closing user message, next to where model attention is strongest; the advisor system prompt is now stage-agnostic and stable across calls.
- Advisor model calls now pass the pi session ID for provider-side cache affinity across repeated consultations.
- Extracted pure signal logic (`detectStage`, `summarizeToolResult`, `buildExecutorSignals`, `isVerificationCommand`, `shouldNudge`) into `src/advisor-signals.ts` so it is unit-testable without pi imports.
- Tool activity tracking now listens to the `tool_result` event, which carries tool input and typed details directly — the `tool_execution_start`/`tool_execution_end` bookkeeping and the per-call args map are gone.
- `isVerificationCommand` now matches the leading token of each pipeline segment instead of substrings anywhere in the command, so paths like `cat tests/foo.test.ts` no longer count as verification runs.
- Default advisor model is now `anthropic/claude-fable-5` (was `claude-opus-4-6`). Existing `advisor.json` files keep their configured model.
- Default `maxTokens` raised from `8192` to `16384` — adaptive-thinking models (Opus 4.6+, Fable 5) count thinking tokens against the output cap, and 8k risked truncated advice at `reasoning=high`.
- `/advisor on provider/model` now splits on the first slash only, so model IDs containing slashes (e.g. OpenRouter) work.

### Fixed
- Bash exit codes are now actually extracted from tool output: the parser looked for `exit code: N` while Pi's bash tool reports `Command exited with code N`, so the `(exit N)` suffix in advisor context never appeared.
- `/advisor on provider/model` no longer corrupts the in-memory config when the model is not found — validation now happens before the provider/model are assigned, so a later `/advisor config key=value` can no longer persist an invalid model.
- Failed advisor calls (model not found, missing API key, empty context) no longer consume the per-run usage quota; only calls that actually reach the model count.

## [0.2.1] - 2026-05-09

### Added
- Optional `stage` parameter for the `advisor` tool (`"initial"`, `"recovery"`, `"final-check"`) so executors can explicitly signal the current phase while preserving auto-detection as a fallback.

### Changed
- Migrated package imports and `peerDependencies` from `@mariozechner/*` to `@earendil-works/*` for the Pi namespace migration (v0.74+).
- Migrated `@sinclair/typebox` to `typebox` (TypeBox v1.x) for the Pi dependency migration.
- Rewrote `promptGuidelines` with actual function-call syntax to remove executor guesswork around advisor parameters.

## [0.2.0] - 2026-04-13

### Added
- Reasoning effort control for advisor model calls (`minimal`–`xhigh`, default `high`).
- `maxContextMessages` config to tune transcript size sent to the advisor (default `18`).
- Tab completion for `/advisor` subcommands (`on`, `off`, `config`, `ask`) and config keys (`provider=`, `model=`, `reasoning=`, etc.).

### Fixed
- Advisor tool crash on OpenAI-compatible providers caused by dangling `toolCall` blocks in the transcript — historical tool calls are now stripped from assistant messages, matching the already-skipped `toolResult` blocks.
- TUI rendering crash caused by `Box+Markdown` depth bug — replaced with `Container+Text` and corrected `renderResult` signature.
- Silent empty responses when advisor model returns only `thinking` blocks — thinking content is now used as a fallback.
- Missing error propagation when `completeSimple` returns `response.errorMessage`.

### Changed
- Moved `advisor-messages.ts` into `src/` directory for cleaner package layout.
- Rewrote README with architecture diagram, configuration reference table, and stage detection docs.
- Added LICENSE copyright holder and updated package manifest `files` field.
- Advisor prompts restructured to a consistent verdict + action items format (`On track` / `Course-correct` / `Not done yet`), with stage-specific directives for exploration, recovery, and final-check phases.
- Switched from `complete()` to `completeSimple()` to support `reasoning` parameter.

## [0.1.0] - 2026-04-10

### Added
- Initial `pi-advisor` package with a Pi extension that adds a Claude-style advisor tool for strategic guidance during complex coding tasks.
- `/advisor on`, `/advisor off`, `/advisor config`, and `/advisor ask` commands.
- Curated advisor context: bounded system prompt, active tools summary, and recent tool activity.
- Stage-aware guidance that adapts the advisor prompt based on executor activity.
- Compact advisor result rendering in the Pi TUI with token usage and expand-to-read hint.
- Per-run advisor usage limit (default 3 calls).
- Package metadata, smoke tests, MIT license, and install/use documentation for npm and GitHub distribution.