# 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). ## [Unreleased] ## [0.6.6] - 2026-05-20 ### Fixed - Delay fallback summary auto-refresh until Pi reports idle; session replacements cancel stale handoffs, while idle timeouts warn and still send the hidden handoff. - Queue chained boomerang steps as follow-ups so they advance from Pi 0.75.4 `agent_end` handlers without busy-agent errors. - Document the correct `pi install npm:pi-boomerang` install command. - Preserve much longer validation commands and failure details in boomerang summaries while still bounding huge logs. ## [0.6.5] - 2026-05-03 ### Fixed - Auto-refresh fallback boomerang summaries with `/reload`, including shortcut-first auto mode before any `/boomerang` command has initialized a command context, so stale raw branch content is not left visible after the summary appears. - Preserve draft editor text while auto-refreshing shortcut-first fallback summaries through the temporary `/reload` editor path. - Removed the duplicate visible `boomerang-summary` message; the branch summary remains the single visible summary while the hidden handoff still wakes the orchestrator. - Declared `@mariozechner/pi-coding-agent` as a peer dependency now that boomerang imports `CustomEditor` at runtime. ## [0.6.4] - 2026-05-02 ### Fixed - Show the boomerang summary in chat while keeping the hidden self-contained orchestrator handoff. ## [0.6.3] - 2026-05-02 ### Fixed - Made the hidden post-boomerang follow-up self-contained so agents do not search session logs for the just-created summary. ## [0.6.2] - 2026-05-02 ### Fixed - Prevented auto-compaction from immediately compacting tree-navigation boomerang summaries. ## [0.6.1] - 2026-05-01 ### Changed - Auto-boomerang mode is now one-shot and turns off after wrapping the next prompt. ## [0.6.0] - 2026-05-01 ### Added - Added session-only auto-boomerang mode, toggled with `Ctrl+Alt+B` or `/boomerang auto [on|off|toggle|status]`, so normal prompts can automatically run with boomerang summarization. ### Fixed - Auto-boomerang now also wraps extension-sourced prompt-template command output, such as `/tldr` from prompt-template-model. ## [0.5.0] - 2026-05-01 ### Added - Expanded boomerang summaries into operational handoffs with outcome, changed files, relevant reads, command counts, validation commands, failures, and config metadata. - Added a hidden follow-up trigger after successful boomerang returns so the orchestrator reads the new handoff summary immediately. ### Fixed - Kept boomerang-related summary and compaction rows expanded from creation time, so Ctrl+O is never needed after a boomerang returns. ## [0.4.5] - 2026-04-30 ### Fixed - Expanded completed boomerang branch summaries by default so users no longer need to press Ctrl+O to reveal details. - Avoided expanding the UI if a rethrow collapse is cancelled before boomerang still owns the collapse. ## [0.4.4] - 2026-04-22 ### Fixed - Migrated extension tool schemas from `@sinclair/typebox` to `typebox` 1.x so packaged installs follow Pi's current extension runtime contract. ### Changed - Added `typebox` as a runtime dependency for packaged installs. ## [0.4.3] - 2026-04-04 ### Changed - Added a `promptSnippet` for the `boomerang` tool so Pi 0.59+ includes it in the default tool prompt section and reliably discovers it during autonomous runs. ## [0.4.2] - 2026-04-03 ### Fixed - Boomerang branch summaries now keep the full `Outcome` text instead of truncating at 500 characters, so collapsed context preserves complete results for future turns. - Preserved assistant formatting in `Outcome` summaries by normalizing line endings without flattening multiline content. ## [0.4.1] - 2026-04-03 ### Changed - Migrated boomerang session reset handling to the `session_start` lifecycle event and removed the legacy `session_switch` hook. - Clarified `--rethrow` and `--loop` boomerang behavior in the README, including count requirements, alias precedence, and pass-by-pass execution details. ## [0.4.0] - 2026-03-18 ### Added - **`--rethrow N` syntax** — Run full multi-pass boomerang execution with context collapse between rethrows (for example `/boomerang /task --rethrow 3`) - **Synchronous rethrow execution model** — Rethrows run in-command via `waitForIdle`, with per-rethrow collapse and accumulated summaries - **Boomerang tool accepts a task parameter** — When enabled, the agent can call `boomerang({ task: "fix bugs --rethrow 3" })` to queue tasks with full rethrow/chain/template support ### Changed - **Replaced loop syntax with `--rethrow N`** — The `Nx` loop syntax, convergence detection, and `--converge` flag are replaced by `--rethrow N`. Boomerang handles context collapse; pi-prompt-template-model owns inner loops via `--loop`. - **Boomerang tool only registered when enabled** — Tool only appears in the agent's tool list after `/boomerang tool on` - **`--loop N` in boomerang now maps to `--rethrow N`** — Prevents loop flags from being injected into rendered template prompts and keeps boomerang execution deterministic. - **Loop alias mapping is now surfaced in UI** — Boomerang shows an info notification when `--loop` is auto-mapped to `--rethrow` so execution mode is explicit. - **`--rethrow` now takes precedence over `--loop` when both are present** — Boomerang strips loop tokens and reports the precedence decision so loop flags never leak into rendered prompts. ### Fixed - **Prevented premature collapse before agent output** — Command-mode boomerangs now wait for the queued task message to appear and for the first assistant message after that task before advancing chain steps or collapsing context, including cases where `getLeafId()` is transiently null when the task is queued. - **Preserved template read errors instead of downgrading to "not found"** — Template I/O failures now surface explicit `Failed to read template ...` errors in single, chain, and rethrow flows. - **Preserved skill read error details** — Skill load failures now include the underlying read error text (for example permissions or path/type errors) instead of a generic warning. - **Preserved config persistence failures** — Config save failures now warn in the UI, and config load failures are logged with full error context. - **Preserved model restore failures** — Restore now reports model restore failures as warnings instead of silently reporting successful restoration. - **Stripped invalid numeric loop counts when `--rethrow` is present** — Inputs like `--rethrow 2 --loop 0` no longer leak `0` into rendered task arguments. - **Stripped repeated `--loop` tokens when `--rethrow` is present** — Inputs like `--rethrow 2 --loop 0 --loop 3` now remove all loop metadata so no loop flags leak into rendered task arguments. - **Prevented queued-tool task overwrites** — The boomerang tool now rejects queueing a second task while one is already queued instead of silently replacing the first queued task. - **Model restoration no longer depends on command-context model snapshots** — Restore now uses runtime model snapshots, including tool-queued task starts, preventing stale command-context values from causing incorrect rollback targets. ## [0.3.0] - 2026-03-14 ### Added - **Loop execution** — Run tasks multiple times with `/boomerang /task 5x`. Each iteration collapses back to an auto-managed anchor point, with changes accumulating across iterations. Supports templates, chains, and plain tasks. - **Convergence detection** — `--converge` flag stops the loop early if an iteration makes no file changes (e.g., `/boomerang /deslop 5x --converge`) - **Loop-aware system prompt** — Agent is told which iteration it's on so it builds on previous work incrementally - **Combined status indicators** — Shows `loop 2/5` during loops and `loop 2/5 · chain 1/3` for chain+loop combinations ## [0.2.1] - 2026-03-09 ### Fixed - Added missing `pi` manifest to package.json so pi can auto-discover the extension ## [0.2.0] - 2026-03-04 ### Added - **Chain execution** - Run multiple templates in sequence with `/boomerang /scout -> /planner -> /impl -- "task"`. Each step can have its own args, model, skill, and thinking level. Context collapses after the final step. - **Prompt template execution** - `/boomerang /