@fyeeme/pi-goal 1.0.0

package/CHANGELOG.md

@@ -0,0 +1,43 @@
+# Changelog
+
+## [Unreleased]
+
+## [1.0.0] - 2026-09-09
+
+### Added
+
+- Initial release: oh-my-pi goal mode ported to a pi extension.
+- `goal` tool (`create`/`get`/`complete`/`resume`/`drop`/`impossible`) with omp's exact operation semantics, registered at load and kept out of the active toolset until goal mode or a guided interview needs it.
+- **Independent completion/impossibility evaluator** (`src/evaluator.ts` + `evaluator-complete.md` / `evaluator-impossible.md`), absorbed from Claude Code 2.1.261's goal design (bin/claude.exe, 2026-09-07) and going one step further than CC:
+  - `goal({op:"complete"})` requires `evidence` — the agent's per-deliverable audit of the current repo state — and is gated by an independent evaluator: a fresh `pi -p --no-session` subprocess (spawn resolution ported from pi-subagents' `getPiInvocation`) that re-verifies the repository itself (runs the checks; grounded, where CC's evaluator is transcript-only). Its JSON contract follows CC's: `{ok, reason}` with evidence quoting and the default direction "insufficient evidence = NOT met". A refuted claim keeps the goal active, returns the evaluator's findings in the tool result, and forbids re-claiming with the same evidence; omp's complete error semantics (no goal / already complete / dropped) are preserved before any evaluator spend. Evaluator unavailability (spawn failure, ~5 min timeout, unparseable/out-of-contract output) falls back to the omp self-audit completion, honestly labeled (`unavailable-fallback`); caller aborts rethrow.
+  - New `goal({op:"impossible", reason})` op — CC's third verdict channel with its "the claim is evidence, not proof — independently confirm" rule: the evaluator adjudicates the impossibility claim; confirmation pauses the goal (`state.reason: "impossible-confirmed"`) with the tool result instructing an honest report to the user; refutation keeps it working; after 2 unconfirmed disputes (`Goal.impossibleReports`, persisted) the goal pauses for a human decision (`"impossible-disputed"`). An unadjudicated claim (evaluator unavailable) changes no state and routes to the user.
+  - Tool details gain `evaluator: {verdict, reason}` (`confirmed` / `rejected` / `unavailable-fallback` / `impossible-confirmed` / `impossible-refuted` / `impossible-disputed` / `unavailable`), rendered as an evaluator line in the tool result.
+  - `pauseGoal` accepts an explicit `reason`; `recordImpossibleReport` bumps + persists the dispute counter; `isGoal` accepts the optional `impossibleReports` field (old snapshots restore unchanged).
+- `GoalRuntime` with omp's budget accounting verbatim: token deltas include cache writes and exclude cache reads, wall-clock seconds advance in whole persisted steps, budget-limit steering fires once per goal id, interrupts pause, cold resume auto-pauses, tree navigation preserves.
+- Autonomous continuation loop with no-tool-call suppression and user-message re-arm.
+- `/goal` (`set`/`show`/`pause`/`resume`/`drop`/`budget`) with interactive menu, and `/guided-goal` interview kickoff.
+- Session persistence via `goal-state` / `goal-cleared` / `goal-completed` custom entries with branch-aware restore.
+- pi-todo integration: live `<todo_context>` block inside the goal context message.
+- `goal_updated` event bus broadcast for other extensions; footer status segment.
+- `ui_prompt_start`/`ui_prompt_end` integration: continuation turns are withheld while a blocking dialog is open and resume when it closes while idle.
+- `agent_settled` status refresh, `/goal` argument autocomplete (`getArgumentCompletions`), multi-line objective editor, and a persistent `goal-completed` entry renderer (all per docs/extensions.md v0.84.4).
+- `test/omp-alignment.test.ts`: 47-test alignment suite pinning pi-goal's commands, status segment, renderer strings, persistence, and escaping to omp sources verbatim.
+
+### Changed
+
+- `goal` tool `op` parameter uses `StringEnum` instead of `Type.Union`/`Type.Literal` for Google API compatibility (docs/extensions.md, Custom Tools).
+- Peer dependencies now require pi 0.84.4+ (`ui_prompt_start`/`ui_prompt_end` events).
+
+### Fixed
+
+- omp alignment pass (test + fix until behavior is identical to oh-my-pi goal mode):
+  - `escapeXmlText` escapes only the `& < >` trio; quotes now stay verbatim in goal prompts, matching omp `sanitize-text.ts` (omp's goal-runtime tests pin this).
+  - `/goal <objective>` while a goal is active now reports omp's status message instead of opening the menu; while paused it warns "Resume the current goal first...".
+  - Dropping a goal now clears session state entirely (omp `commitState(undefined)`): `/goal show`, `/goal resume`, the menu, and `/guided-goal` treat a dropped goal like no goal, and a new `/goal <objective>` starts fresh.
+  - Menu item label is omp's `Adjust budget…` (ellipsis); the budget prompt is an editor prefilled with the current budget (omp `#promptGoalBudgetEdit`).
+  - Notification severities mirror omp: `showStatus` → info, `showWarning` → warning, `showError` → error (e.g. invalid budget input renders as an error).
+  - Footer status segment reworked to omp `renderGoalMode`: `<icon> Goal <used>[/<budget>]` with the omp unicode icon set (🎯/⏸/⚠), visible only while enabled or paused; the pi-specific `waiting for user` suffix was removed.
+  - Goal tool renderer strings match omp: call `⏳ Goal: set "…" · budget 5K`, result `◎ Goal: set ⟦active⟧` (⟦⟧ badge brackets, `goalBadgeColor`), error `✘ Goal: <op>` + two-space detail, no-goal `⚠ Goal: <op> · no active goal`; objective caps at omp TRUNCATE_LENGTHS (60 call / 100 result).
+  - `/goal` description is omp's "Toggle goal mode (persistent autonomous objective for this session)"; argument completions mirror omp `buildArgumentCompletions` (trailing-space values, null past the subcommand word).
+  - `/guided-goal` no longer emits a kickoff status message (omp shows none); `/goal resume` guard requires an actually-paused goal.
+  - `/guided-goal` kickoff queues as a follow-up behind an in-flight run (omp `session.followUp`), never a steer.

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 fyeeme
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,59 @@
+# pi-goal
+
+oh-my-pi's **goal mode** migrated to a [pi](https://github.com/earendil-works/pi-coding-agent) extension: one persistent autonomous objective looped until the agent can *prove* the success criteria — not until it runs out of things to say.
+
+Source: [oh-my-pi](https://github.com/can1357/oh-my-pi) (a fork of badlogic/pi-mono), `packages/coding-agent/src/goals` + the goal wiring in its agent session and interactive mode. The runtime semantics are ported behavior-for-behavior; host-internal surfaces are mapped to pi's extension API (see the adaptation table in `index.ts`). One deliberate deviation, absorbed from Claude Code 2.1.261's goal design (binary-verified 2026-09-07): **completion is gated by an independent evaluator** instead of self-graded — see below.
+
+## What it does
+
+- **`goal` tool** — `create` / `get` / `complete` / `resume` / `impossible` / `drop`. Creating a goal enables goal mode; completing it requires verified evidence of every deliverable (see the evaluator gate below); the runtime refuses double-completes.
+- **Independent evaluator gate (CC 2.1.261 absorption)** — `goal({op:"complete", evidence})` passes the agent's per-deliverable audit to a fresh `pi -p` subprocess that re-verifies the repository itself (it can run the tests — grounded, unlike CC's transcript-only evaluator). Its JSON contract defaults to *insufficient evidence = not met*. A refuted claim keeps the goal active and returns the evaluator's findings; the tool result tells the model to fix the gaps and re-claim with stronger evidence. If the evaluator subprocess cannot run (spawn failure, timeout ≈ 5 min, unparseable output), completion falls back to the omp self-audit behavior, honestly labeled in the result.
+- **`impossible` channel (CC 2.1.261 absorption)** — when the agent believes the goal genuinely cannot be achieved this session, `goal({op:"impossible", reason})` sends the claim to the evaluator, which independently confirms or refutes it ("the claim is evidence, not proof"). Confirmed → the goal pauses (`reason: impossible-confirmed`) and the model must report honestly to the user. Refuted → keep working; after 2 unconfirmed disputes the goal pauses for a human decision (`reason: impossible-disputed`). An unadjudicated claim (evaluator unavailable) changes nothing.
+- **Autonomous continuation** — when the agent yields while the goal is still active, pi-goal re-submits a hidden continuation prompt that restates the objective and the verification checklist, so work continues across turns without user nudges. A continuation turn that produced no tool calls suppresses the next one (no infinite idle loops); a real user message re-arms the loop. While a blocking dialog is open (`/goal` menus, confirmations, or any extension's `ask_user`), continuations are withheld via `ui_prompt_start`/`ui_prompt_end` and resume when the dialog closes.
+- **Budget accounting** — optional `token_budget` per goal. pi-goal counts input + output + cache writes (cache reads are reused prefix, not new work) plus wall-clock seconds. Crossing the budget flips the goal to `budget-limited` and steers the agent once with a wrap-up instruction. Budget exhaustion is explicitly *not* completion.
+- **Interrupt safety** — Esc interrupts pause the goal (never complete it); usage accumulated so far is kept. Resuming a session auto-pauses a still-active goal, since the run that owned it is gone.
+- **`/goal` command** — `set <objective>`, `show`, `pause`, `resume`, `drop`, `budget <N|off>`, or an interactive menu when called without arguments. Typing an objective while a goal is active (or paused) is rejected with omp's status/warning; `/goal drop` and the menu's Drop confirm before discarding.
+- **`/guided-goal [rough objective]`** — the agent interviews you in normal chat (success criteria, verification method, attempt cap, scope boundaries, stop conditions) and then creates the goal itself via `goal create`.
+- **Footer status** — a `<icon> Goal <used/budget>` segment (omp `renderGoalMode` unicode icon set: 🎯 active, ⏸ paused, ⚠ budget-limited) while the goal is enabled or paused. Completed goals stay visible in the transcript after reloads via a persistent `goal-completed` entry renderer.
+- **pi-todo integration** — when [pi-todo](../pi-todo) is installed, the goal context message includes a live `<todo_context>` block so continuations treat todos as current state, not stale transcript decoration.
+
+## Install
+
+```bash
+pi install npm:@fyeeme/pi-goal
+```
+
+or copy/link this directory into `~/.pi/agent/extensions/` (global) or `.pi/extensions/` (project).
+
+## Usage
+
+```
+/goal Ship the parser rewrite with all tests passing    # start working
+/goal show                                              # objective, status, tokens, time
+/goal budget 200000                                     # cap the spend
+/goal pause            # ... later ...  /goal resume
+/goal drop
+```
+
+Status line: a `<icon> Goal <used/budget>` segment appears in the footer while a goal is enabled or paused (omp footer segment semantics).
+
+## Session persistence
+
+Every transition writes a full snapshot entry (`goal-state`, `goal-cleared`, `goal-completed`) to the session. Restore scans the current branch backward, so forks and tree navigation restore branch-local state. Usage totals are derived from the session's own message entries — they survive compaction.
+
+## Integration contract for other extensions
+
+- Emits `goal_updated` on the shared extension event bus (`pi.events`) after every runtime transition: `{ goal: Goal | null, state: GoalModeState | undefined }`.
+- Reads `todo_updated` events and `todo-phases` entries from pi-todo (works without pi-todo installed).
+
+## Development
+
+```bash
+npm install --ignore-scripts
+npm test        # vitest, 118 tests
+npm run typecheck
+```
+
+## License
+
+MIT