# pi-tasks Dogfood Checklist

This document tracks real Pi dogfood evidence. Skipped items are not counted as passed.

## Current Status

Date: 2026-07-22

Result: passed for the scoped MVP dogfood gate, release-hardening dogfood gate, weak-model release gate, installed-package smoke, 0.1.5 receiver-bound append compatibility release gate, and 0.2.0 state-event source/resume/fork/live-TTY/manual-compaction/installed-package smoke.

Environment:

- Extension path: `./index.ts`
- Session directory: `/private/tmp/pi-tasks-dogfood/sessions`
- Primary session ID: `pi-tasks-dogfood-mvp-2`
- Primary transcript: `/private/tmp/pi-tasks-dogfood/sessions/2026-06-18T09-50-13-716Z_pi-tasks-dogfood-mvp-2.jsonl`
- Blocked-task transcript: `/private/tmp/pi-tasks-dogfood/sessions/2026-06-18T09-51-59-892Z_pi-tasks-dogfood-blocked-2.jsonl`
- Forked replay transcript: `/private/tmp/pi-tasks-dogfood/sessions/2026-06-18T10-00-01-165Z_019eda2c-5d8d-7544-a5ea-630b98aae3d5.jsonl`
- Release-hardening session directory: `/private/tmp/pi-tasks-release-dogfood/sessions`
- Normal ordered-step session ID: `pi-tasks-release-normal`
- Adversarial dedupe session ID: `pi-tasks-release-adversarial-dedupe`
- Blocked/decision session ID: `pi-tasks-release-blocked-decision`
- Installed-package session ID: `pi-tasks-release-installed-package`
- Commercial contract session directory: `/private/tmp/pi-tasks-commercial-dogfood/sessions`
- Commercial contract session ID: `pi-tasks-commercial-contract`
- Recursive decomposition session directory: `/private/tmp/pi-tasks-decomposition-dogfood/sessions`
- Recursive decomposition session ID: `pi-tasks-recursive-decomposition-step-evidence`
- Installed decomposition tools session ID: `pi-tasks-installed-decomposition-tools`
- Compaction resume session directory: `/private/tmp/pi-tasks-compaction-dogfood/sessions`
- Compaction resume session ID: `pi-tasks-compaction-resume`
- Weak-model quality session directory: `/private/tmp/pi-tasks-weakmodel-dogfood/sessions`
- English plan gate session ID: `weakmodel-plan-gate-en`
- English evidence gate session ID: `weakmodel-evidence-gate-en2`
- Chinese plan gate session ID: `weakmodel-plan-gate-zh`
- Installed package quality gate session ID: `weakmodel-installed-plan-gate`
- Readiness regression session directory: `/private/tmp/pi-tasks-id-regression/sessions`
- Readiness regression session ID: `id-regression-1`
- Branch navigation session ID: `branch-live-fork-1`
- Installed ID regression session directory: `/private/tmp/pi-tasks-installed-smoke-3/sessions`
- Installed ID regression session ID: `installed-id-regression-1`
- Final release session directory: `/private/tmp/pi-tasks-final-release/sessions`
- Final installed package smoke session ID: `final-installed-smoke-1`
- Token output session directory: `/private/tmp/pi-tasks-token-smoke/sessions`
- Token output session ID: `token-output-smoke-1`
- Installed token output session directory: `/private/tmp/pi-tasks-token-release-check/sessions`
- Installed token output session ID: `token-installed-smoke-1`
- Weak-model hardening session directory: `/private/tmp/pi-tasks-weakmodel-hardening/sessions`
- Weak-model hardening session ID: `weakmodel-hardening-1`
- Installed weak-model hardening session directory: `/private/tmp/pi-tasks-weakmodel-release-check/sessions`
- Installed weak-model hardening session ID: `weakmodel-installed-hardening-1`
- 0.1.3 release dogfood session directory: `/private/tmp/pi-tasks-release-013-dogfood/sessions`
- 0.1.3 English weak-model session ID: `release-013-weak-en`
- 0.1.3 Chinese weak-model session ID: `release-013-weak-zh`
- 0.1.3 final installed package session directory: `/private/tmp/pi-tasks-release-013-installed-final/sessions`
- 0.1.3 final installed weak-model session ID: `release-013-installed-final-weak`
- 0.1.5 release dogfood session directory: `/private/tmp/pi-tasks-release-015-dogfood/sessions`
- 0.1.5 source lifecycle session ID: `release-015-source-lifecycle`
- 0.1.5 fork replay session name: `release-015-fork-replay`
- 0.1.5 final installed package smoke directory: `/private/tmp/pi-tasks-release-015-installed-final2`
- 0.1.5 final installed package smoke session ID: `release-015-installed-final2-smoke`
- 0.2.0 release dogfood session directory: `/private/tmp/pi-tasks-release-020-dogfood/sessions`
- 0.2.0 source lifecycle session ID: `release-020-source-lifecycle-2`
- 0.2.0 fork replay session ID: `release-020-fork-replay`
- 0.2.0 observer extension path: `/private/tmp/pi-tasks-state-observer.ts`
- 0.2.0 observer event log: `/private/tmp/pi-tasks-state-events-020.jsonl`
- 0.2.0 installed package smoke directory: `/private/tmp/pi-tasks-release-020-installed`
- 0.2.0 installed package session ID: `release-020-installed-smoke`
- 0.2.0 large compaction session directory: `/private/tmp/pi-tasks-release-020-compact/sessions`
- 0.2.0 large compaction session ID: `release-020-large-compact`
- 0.2.0 large compaction seed prompt: `/private/tmp/pi-tasks-large-compact-prompt.md`

## Passed Scenarios

- Created an active task with two acceptance criteria through `task_plan`.
- Updated progress and next action through `task_update`.
- Attempted `task_complete` without evidence and received the expected rejection: `Task T1 has no evidence`.
- Listed the task with `include_evidence=true` and confirmed criterion IDs were visible for evidence attachment.
- Attached passing evidence to both criteria with `task_evidence`.
- Completed the task with evidence through `task_complete`.
- Listed completed task state with `include_done=true` and `include_evidence=true`.
- Resumed the same session ID and confirmed custom-entry replay restored the completed task and evidence records.
- Created a blocked task and confirmed `/task_list` output includes blocked status, blocker source, blocker reason, and unblock condition.
- Opened `/tasks` in a live TTY session and confirmed it rendered the replayed completed task and evidence details.
- Exited the live TTY session with `/quit` and received a clean exit.
- Forked the primary session and confirmed the fork transcript replayed copied `pi-tasks:event` entries into completed task state.
- Confirmed ordered plan-step flow: T1-S1 completed before T1-S2, then evidence and completion.
- Confirmed out-of-order step update is rejected: `Plan step T1-S3 cannot be updated before T1-S1`.
- Confirmed premature completion is rejected while a plan step remains open.
- Confirmed duplicate `task_evidence` call returns `Evidence already recorded as E1` and does not create E2.
- Confirmed blocked task flow records a resolved blocker plus explicit user decision in detailed task output.
- Confirmed skipping a plan step without a reason is rejected, while skipping with a note is accepted.
- Confirmed `pi install ./` records the local package in an isolated Pi config.
- Confirmed npm tarball install contains `dist/` runtime and supports `import("pi-tasks")`.
- Confirmed installed package Pi smoke passes with `--extension ./node_modules/pi-tasks/dist/index.js`.
- Confirmed commercial `plan_steps` flow uses expected output, linked criteria, evidence requirement, and allowed actions.
- Confirmed `task_focus` returns the current active step contract before work proceeds.
- Confirmed evidence-required steps reject `done` before linked evidence exists.
- Confirmed `off_plan` activity without `scope_reason` is rejected, and `off_plan` with `scope_reason` is recorded as a warning.
- Confirmed non-atomic steps reject `done` until recursively decomposed with `task_decompose`.
- Confirmed `task_granularity_check` reports atomicity booleans and directs the agent to decomposition for coarse steps.
- Confirmed decomposition replaces `T1-S1` with atomic child steps `T1-S1.1` and `T1-S1.2`.
- Confirmed step-scoped evidence with `step_ids` prevents evidence for `T1-S1.1` from satisfying `T1-S1.2`, even when both share the same acceptance criterion.
- Confirmed installed package runtime exposes `task_granularity_check` and `task_decompose`, and can replace an installed-package non-atomic step with child steps.
- Confirmed `task_resume` reports current decomposed child step, lineage, next allowed actions, and verification gaps.
- Confirmed `task_checkpoint` persists a snapshot with resume fields and `task_resume` remains stable after checkpoint.
- Confirmed same-session replay after checkpoint restores the current child step `T1-S1.2` and next allowed action `task_evidence`.
- Confirmed English bad atomic plan is rejected by the plan quality gate and returns `pi-tasks resume` recovery guidance.
- Confirmed Chinese bad atomic plan is rejected by the plan quality gate and returns `pi-tasks resume` recovery guidance.
- Confirmed low-quality test evidence without `quality.observedOutput` is rejected and returns `pi-tasks resume` recovery guidance.
- Confirmed valid evidence with `quality.source`, `quality.reproducible`, `quality.verifier`, `quality.artifactRefs`, and `quality.observedOutput` can be recorded and linked to a step.
- Confirmed installed package runtime rejects a bad atomic plan through the same quality gate and returns `pi-tasks resume` recovery guidance.
- Confirmed rejected `task_plan` calls do not consume task IDs; after a rejected invalid plan, the first successful task was still created as `T1`.
- Confirmed weak-model guidance works better when `plan_steps.criterionIds` is omitted during new task creation; generated IDs such as `T1-AC1` are auto-linked after creation.
- Confirmed live PTY `/tasks` renders the active task, step contract, criterion, and verification gaps.
- Confirmed live PTY `/quit` exits cleanly with Pi's resume instruction.
- Confirmed installed package runtime preserves rejected-plan ID behavior through `./node_modules/pi-tasks/dist/index.js`.
- Confirmed final clean tarball install supports `import("pi-tasks")`.
- Confirmed final installed package runtime rejects invalid `task_plan`, then creates the first valid installed-package task as `T1`.
- Confirmed real Pi manual `/compact` fired a native compaction event and a `pi-tasks` snapshot with `reason: compaction`.
- Confirmed post-compaction replay with `task_resume` restores active task `T1`, current step `T1-S1`, allowed action `task_evidence`, and completion gaps.
- Confirmed interactive `/tree` navigation in a forked live PTY can move to the branch point before valid task creation, where `/tasks` correctly reports `No tasks on this branch.`
- Confirmed 50-column live Pi TUI rendering keeps status/widget lines readable with truncation and does not cover the input line.
- Confirmed token-efficient tool output returns compact `task_resume` guidance rather than full evidence dumps.
- Confirmed `/tasks` defaults to compact one-line task summary and `/tasks detail` explicitly expands step and criterion details.
- Confirmed installed 0.1.2 tarball runtime preserves compact `task_list` output through `./node_modules/pi-tasks/dist/index.js`.
- Confirmed `task_next` returns one-step weak-model guidance with only next tool, current-step lock, blocked tools, and minimum params.
- Confirmed compound atomic step wording is rejected by the plan quality gate.
- Confirmed future-step evidence is rejected unless current-step lock is overridden with an explicit reason.
- Confirmed oversized evidence summary is rejected and long output is directed to artifact references.
- Confirmed installed 0.1.2 tarball runtime exposes `task_next` and structured recovery for compound atomic rejection.
- Confirmed `npm run release:check` passes for 0.1.3, including typecheck, check, unit tests, build, source/dist import smokes, pack dry-run, clean tarball install, and audit.
- Confirmed fixed weak-model dogfood prompts are stored under `docs/dogfood-prompts/`.
- Confirmed 0.1.3 source runtime rejects English compound atomic wording, oversized evidence summaries, and future-step evidence without override while `task_next` returns the current-step lock and only next tool.
- Confirmed 0.1.3 source runtime rejects Chinese compound atomic wording and returns structured recovery with `retry_with` and `do_not_retry_same_call`.
- Confirmed 0.1.3 installed tarball runtime exposes `task_next` and structured recovery through `./node_modules/pi-tasks/dist/index.js`.
- Confirmed 0.1.5 source runtime creates and completes task `T1` with decision `D1`, evidence `E1`, status `done`, 100% progress, and criterion `1/1` for receiver-bound append compatibility dogfood.
- Confirmed 0.1.5 same-session resume restores completed task `T1`, evidence `E1`, and decision `D1`.
- Confirmed 0.1.5 fork replay restores completed task `T1`, evidence `E1`, decision `D1`, and satisfied criterion state.
- Confirmed live TTY `/tasks` default excludes done tasks, while `/tasks detail` renders `T1 [done]`, decision `D1`, evidence `E1`, and criterion `T1-AC1`.
- Confirmed live TTY `/quit` exits cleanly and prints the Pi resume command.
- Confirmed 0.1.5 clean tarball install supports `import("pi-tasks")` and installed-package Pi runtime creates task `T1` through `./node_modules/pi-tasks/dist/index.js`.
- Confirmed 0.2.0 source runtime publishes `pi-tasks:state` on `session_start` and successful task mutations with stable widget id `pi-tasks`, active task `T1`, and no raw `events` field.
- Confirmed 0.2.0 same-session resume restores active task `T1` and publishes a `session_start` state event.
- Confirmed 0.2.0 fork replay restores active task `T1` and publishes a `session_start` state event.
- Confirmed 0.2.0 live TTY `/tasks` renders active task `T1`; `/quit` exits cleanly.
- Confirmed 0.2.0 installed tarball supports `import("pi-tasks")` and installed runtime publishes `pi-tasks:state` for `session_start` and `task_mutation`.
- Confirmed 0.2.0 large-session manual `/compact` compacted from 56,203 tokens and emitted a `task_mutation` state event for active task `T1` without raw `events`.

## Commands

Primary task lifecycle:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-dogfood/sessions \
  pi --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_list,task_update,task_evidence,task_decision,task_complete \
  --session-id pi-tasks-dogfood-mvp-2 \
  --name pi-tasks-dogfood-2 \
  -p "Create one pi-tasks dogfood task with two acceptance criteria, update it to 50%, try to complete it before adding evidence, list the criteria ids, add evidence for both criteria, complete it, and list it with evidence."
```

Resume replay:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-dogfood/sessions \
  pi --extension ./index.ts --no-builtin-tools \
  --tools task_list \
  --session-id pi-tasks-dogfood-mvp-2 \
  -p "Resume verification: list current pi-tasks state with done tasks and evidence."
```

Blocked-task display:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-dogfood/sessions \
  pi --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_update,task_list \
  --session-id pi-tasks-dogfood-blocked-2 \
  --name pi-tasks-dogfood-blocked-2 \
  -p "Create one task, mark it blocked with reason, blockedBy external, and neededToUnblock, then list with evidence."
```

Live TTY command and clean exit:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-dogfood/sessions \
  pi --extension ./index.ts --session-id pi-tasks-dogfood-mvp-2
```

Then run `/tasks` and `/quit` in the TTY.

Forked replay:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-dogfood/sessions \
  pi --extension ./index.ts --no-builtin-tools \
  --tools task_list \
  --fork pi-tasks-dogfood-mvp-2 \
  -p "List current pi-tasks state with done tasks and evidence."
```

Release-hardening adversarial scenario:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-dogfood/sessions \
  pi --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_list,task_update,task_evidence,task_decision,task_complete \
  --session-id pi-tasks-release-adversarial-dedupe \
  --name pi-tasks-release-adversarial-dedupe \
  -p "Create a task with three ordered initial steps, then intentionally try out-of-order step completion, premature completion, and duplicate evidence."
```

Installed package smoke:

```sh
npm pack --pack-destination /private/tmp/pi-tasks-release-tarball
cd /private/tmp/pi-tasks-release-consumer
npm install /private/tmp/pi-tasks-release-tarball/pi-tasks-0.1.0.tgz
node -e "import('pi-tasks')"
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-dogfood/sessions \
  pi --no-extensions --extension ./node_modules/pi-tasks/dist/index.js \
  --no-builtin-tools \
  --tools task_plan,task_update,task_evidence,task_complete,task_list \
  --session-id pi-tasks-release-installed-package \
  -p "Installed package smoke: create one task with one initial step and one acceptance criterion, mark the step done, add evidence, complete it, and list the final state with evidence."
```

Commercial contract scenario:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-commercial-dogfood/sessions \
  pi --no-extensions --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_focus,task_update,task_evidence,task_complete,task_list \
  --session-id pi-tasks-commercial-contract \
  --name pi-tasks-commercial-contract \
  -p "Create one task using structured plan_steps, call task_focus, reject step done before evidence, record evidence, close steps, reject off_plan without scope_reason, record off_plan with scope_reason, complete, and list final evidence."
```

Recursive decomposition and step-scoped evidence scenario:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-decomposition-dogfood/sessions \
  pi --no-extensions --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_focus,task_granularity_check,task_decompose,task_update,task_evidence,task_complete,task_list \
  --session-id pi-tasks-recursive-decomposition-step-evidence \
  --name pi-tasks-recursive-decomposition-step-evidence \
  -p "Create one non-atomic plan step, confirm task_focus and task_granularity_check show needs_breakdown, reject marking it done, decompose it into two atomic child steps that share one criterion, record evidence for only T1-S1.1 with step_ids, reject T1-S1.2 done before its own evidence, then record T1-S1.2 evidence, complete, and list final evidence."
```

Compaction-resilient resume scenario:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-compaction-dogfood/sessions \
  pi --no-extensions --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_decompose,task_resume,task_checkpoint,task_evidence,task_update,task_list \
  --session-id pi-tasks-compaction-resume \
  --name pi-tasks-compaction-resume \
  -p "Create one non-atomic task, decompose it into atomic child steps, call task_resume, create a task_checkpoint, then report the current child step, lineage, next allowed actions, and verification gaps."
```

Weak-model quality gate scenarios:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-weakmodel-dogfood/sessions \
  pi --no-extensions --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_resume \
  --session-id weakmodel-plan-gate-en \
  --name weakmodel-plan-gate-en \
  -p "Plan quality gate dogfood. Call task_plan once with a deliberately bad atomic step and report the exact error and recovery guidance."
```

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-weakmodel-dogfood/sessions \
  pi --no-extensions --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_evidence,task_resume \
  --session-id weakmodel-evidence-gate-en2 \
  --name weakmodel-evidence-gate-en2 \
  -p "Create a valid task, then record test evidence without quality.observedOutput and report the exact rejection."
```

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-weakmodel-dogfood/sessions \
  pi --no-extensions --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_resume \
  --session-id weakmodel-plan-gate-zh \
  --name weakmodel-plan-gate-zh \
  -p "中文品質閘門測試。請呼叫 task_plan 建立一個故意很差的 atomic step，回報 exact error，並確認 recovery guidance 是否包含 pi-tasks resume。"
```

Rejected-plan ID regression and live PTY command:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-id-regression/sessions \
  pi --no-extensions --extension ./index.ts --no-builtin-tools \
  --tools task_plan,task_list \
  --session-id id-regression-1 \
  --name id-regression-1 \
  -p "First call task_plan with a deliberately invalid vague atomic step missing allowedActions so it is rejected. Then call task_plan with a valid task named ID Regression, one acceptance criterion, and one atomic plan step with allowedActions [\"task_evidence\"]. Do not set criterionIds manually. Finally call task_list and report the created task id."
```

Then open the same session in a PTY and run `/tasks` and `/quit`:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-id-regression/sessions \
  pi --no-extensions --extension ./index.ts --session-id id-regression-1
```

Manual compaction and post-compaction resume:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-id-regression/sessions \
  pi --no-extensions --extension ./index.ts --session-id id-regression-1
```

Then run `/compact pi-tasks commercial release dogfood: preserve task state and resume contract` and `/quit`.

Observed transcript evidence:

- transcript: `/private/tmp/pi-tasks-id-regression/sessions/2026-06-19T10-15-30-295Z_id-regression-1.jsonl`
- native compaction: `2cfcc4a9`, `tokensBefore: 3663`
- pi-tasks snapshot: `snapshot-2026-06-19T13:31:35.340Z`
- snapshot reason: `compaction`
- snapshot resume task: `T1`
- snapshot gaps: `T1-S1 step pending`, `T1-AC1 pending`, `no evidence`

Then verify replay after compaction:

```sh
gtimeout 60s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-id-regression/sessions \
  pi --no-extensions --extension ./index.ts --no-builtin-tools \
  --tools task_resume,task_list \
  --session-id id-regression-1 \
  -p "After manual compaction, call task_resume and task_list. Report the active task id, current step, and verification gaps."
```

Observed result: active task `T1`, current step `T1-S1`, allowed action `task_evidence`, gaps `no evidence` and `T1-AC1 pending`.

Interactive branch divergence navigation:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-id-regression/sessions \
  pi --no-extensions --extension ./index.ts --fork id-regression-1 --name branch-live-fork-1
```

Then run `/tasks`, run `/tree`, select the point after the rejected invalid `task_plan` but before valid `T1` creation, choose `No summary`, and run `/tasks`.

Observed result: Pi reports `Navigated to selected point`, and `/tasks` reports `No tasks on this branch.`

Narrow terminal live TUI:

```sh
stty cols 50 rows 24
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-id-regression/sessions \
  pi --no-extensions --extension ./index.ts --session-id id-regression-1
```

Then run `/tasks` and `/quit`.

Observed result: at 50 columns, the active-task widget wraps/truncates long `next` and `gaps` text, `/tasks` output wraps inside the terminal width, and the input line remains usable.

Installed package ID regression:

```sh
rm -rf /private/tmp/pi-tasks-installed-smoke-3
mkdir -p /private/tmp/pi-tasks-installed-smoke-3/consumer /private/tmp/pi-tasks-installed-smoke-3/sessions
cd /private/tmp/pi-tasks-installed-smoke-3/consumer
npm init -y
npm install /private/tmp/pi-tasks-release-check-3/tarball/pi-tasks-0.1.0.tgz
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-installed-smoke-3/sessions \
  pi --no-extensions --extension ./node_modules/pi-tasks/dist/index.js \
  --no-builtin-tools \
  --tools task_plan,task_list \
  --session-id installed-id-regression-1 \
  --name installed-id-regression-1 \
  -p "First call task_plan with a deliberately invalid vague atomic step missing allowedActions so it is rejected. Then create a valid task named Installed ID Regression with one criterion and one atomic plan step with allowedActions [\"task_evidence\"]. Do not set criterionIds manually. Call task_list and report the created task id."
```

Final clean tarball install and installed-package runtime:

```sh
rm -rf /private/tmp/pi-tasks-final-release
mkdir -p /private/tmp/pi-tasks-final-release/tarball /private/tmp/pi-tasks-final-release/consumer /private/tmp/pi-tasks-final-release/sessions
npm pack --pack-destination /private/tmp/pi-tasks-final-release/tarball
cd /private/tmp/pi-tasks-final-release/consumer
npm init -y
npm install /private/tmp/pi-tasks-final-release/tarball/pi-tasks-0.1.0.tgz
node -e "import('pi-tasks')"
gtimeout 120s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-final-release/sessions \
  pi --no-extensions --extension ./node_modules/pi-tasks/dist/index.js \
  --no-builtin-tools \
  --tools task_plan,task_list \
  --session-id final-installed-smoke-1 \
  --name final-installed-smoke-1 \
  -p "Installed package final smoke. First call task_plan with a deliberately invalid vague atomic step missing allowedActions so it is rejected. Then create a valid task named Final Installed Smoke with one acceptance criterion and one atomic plan step with allowedActions [\"task_evidence\"]. Do not set criterionIds manually. Call task_list and report the created task id."
```

Observed result: invalid `task_plan` rejected with `expected output is too short; step text uses vague or broad wording; allowedActions are required`; valid installed-package task created as `T1: Final Installed Smoke`; `task_list` confirmed `T1` active.

Token-efficient output smoke:

```sh
rm -rf /private/tmp/pi-tasks-token-smoke
mkdir -p /private/tmp/pi-tasks-token-smoke/sessions
gtimeout 120s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-token-smoke/sessions \
  pi --no-extensions --extension ./index.ts \
  --no-builtin-tools \
  --tools task_plan,task_list,task_resume \
  --session-id token-output-smoke-1 \
  --name token-output-smoke-1 \
  -p "Token output smoke. Create a valid task named Token Output Smoke with one acceptance criterion and one atomic plan step with allowedActions [\"task_evidence\"]. Do not set criterionIds manually. After task_plan, call task_list without include_evidence and task_resume. Report whether the task_plan result was compact resume guidance rather than a full evidence dump."
```

Observed result: created `T1: Token Output Smoke`; `task_plan` returned compact `pi-tasks resume` guidance; `task_list` without `include_evidence` returned only `T1 [active] 1% criteria:0/1 - Token Output Smoke`.

Then open the same session in a PTY:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-token-smoke/sessions \
  pi --no-extensions --extension ./index.ts --session-id token-output-smoke-1
```

Observed result:

- `/tasks` returned compact summary only: `T1 [active] 1% criteria:0/1 - Token Output Smoke`.
- `/tasks detail` expanded step and criterion details, including `criteria:T1-AC1`.
- `/quit` exited cleanly.

Installed 0.1.2 token output smoke:

```sh
rm -rf /private/tmp/pi-tasks-token-release-check
mkdir -p /private/tmp/pi-tasks-token-release-check/tarball /private/tmp/pi-tasks-token-release-check/consumer /private/tmp/pi-tasks-token-release-check/sessions
npm pack --pack-destination /private/tmp/pi-tasks-token-release-check/tarball
cd /private/tmp/pi-tasks-token-release-check/consumer
npm init -y
npm install /private/tmp/pi-tasks-token-release-check/tarball/pi-tasks-0.1.2.tgz
node -e "import('pi-tasks')"
gtimeout 120s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-token-release-check/sessions \
  pi --no-extensions --extension ./node_modules/pi-tasks/dist/index.js \
  --no-builtin-tools \
  --tools task_plan,task_list \
  --session-id token-installed-smoke-1 \
  --name token-installed-smoke-1 \
  -p "Installed token output smoke. Create a valid task named Installed Token Smoke with one acceptance criterion and one atomic plan step with allowedActions [\"task_evidence\"]. Do not set criterionIds manually. Then call task_list without include_evidence and report whether the result is compact."
```

Observed result: installed runtime created `T1: Installed Token Smoke`; `task_list` without `include_evidence` returned compact single-line summary without evidence details.

Weak-model hardening dogfood:

```sh
rm -rf /private/tmp/pi-tasks-weakmodel-hardening
mkdir -p /private/tmp/pi-tasks-weakmodel-hardening/sessions
gtimeout 180s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-weakmodel-hardening/sessions \
  pi --no-extensions --extension ./index.ts \
  --no-builtin-tools \
  --tools task_next,task_plan,task_evidence,task_list \
  --session-id weakmodel-hardening-1 \
  --name weakmodel-hardening-1 \
  -p "Weak-model hardening dogfood. First call task_next with no active task. Then call task_plan once with a deliberately invalid atomic step text 'Run tests and update docs' so compound wording is rejected. Then create a valid task named Weak Model Hardening with one acceptance criterion and two atomic plan steps; omit criterionIds manually; each step must have one action, one output, one verification method, allowedActions [\"task_evidence\"], and no compound wording. Then call task_evidence once targeting future step T1-S2 while current step is T1-S1 so current-step lock rejects it. Then call task_evidence once with a summary longer than 501 characters so evidence budget rejects it. Finally call task_next and task_list without include_evidence, and report exact rejections plus the only next tool."
```

Observed result:

- compound atomic rejection: `Plan step 1 failed quality gate: step text uses vague or broad wording; step text appears to contain multiple actions`
- future-step evidence rejection: `Evidence step_ids must target current step T1-S1; overrideReason is required for T1-S2`
- oversized evidence rejection: `Evidence summary exceeds 500 characters; put long output in artifactRefs`
- final `task_next`: only next tool `task_evidence`, current step lock `T1-S1`, minimum params include `task_id`, `step_ids`, `criterion_ids`, and `references`
- compact `task_list`: `T1 [active] 1% criteria:0/1 - Weak Model Hardening`

Installed 0.1.2 weak-model hardening smoke:

```sh
rm -rf /private/tmp/pi-tasks-weakmodel-release-check
mkdir -p /private/tmp/pi-tasks-weakmodel-release-check/tarball /private/tmp/pi-tasks-weakmodel-release-check/consumer /private/tmp/pi-tasks-weakmodel-release-check/sessions
npm pack --pack-destination /private/tmp/pi-tasks-weakmodel-release-check/tarball
cd /private/tmp/pi-tasks-weakmodel-release-check/consumer
npm init -y
npm install /private/tmp/pi-tasks-weakmodel-release-check/tarball/pi-tasks-0.1.2.tgz
node -e "import('pi-tasks')"
gtimeout 120s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-weakmodel-release-check/sessions \
  pi --no-extensions --extension ./node_modules/pi-tasks/dist/index.js \
  --no-builtin-tools \
  --tools task_next,task_plan \
  --session-id weakmodel-installed-hardening-1 \
  --name weakmodel-installed-hardening-1 \
  -p "Installed weak-model smoke. Call task_next with no active task. Then call task_plan with one deliberately invalid atomic step text 'Run tests and update docs' and report whether compound wording is rejected with structured recovery."
```

Observed result: installed runtime exposed `task_next`; compound atomic step was rejected; structured recovery included `retry_with: task_plan`, `do_not_retry_same_call: true`, a reason, and resume guidance.

0.1.3 release check:

```sh
npm run release:check
```

Observed result: passed. The gate ran typecheck, Biome check, 48 unit tests, build, source import smoke, dist import smoke, npm pack dry-run, clean tarball install plus `node -e "import('pi-tasks')"`, and `npm audit --audit-level=low`.

0.1.3 source weak-model hardening dogfood:

```sh
rm -rf /private/tmp/pi-tasks-release-013-dogfood
mkdir -p /private/tmp/pi-tasks-release-013-dogfood/sessions
gtimeout 180s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-013-dogfood/sessions \
  pi --no-extensions --extension ./index.ts \
  --no-builtin-tools \
  --tools task_next,task_plan,task_evidence,task_list \
  --session-id release-013-weak-en \
  --name release-013-weak-en \
  -p "<docs/dogfood-prompts/weak-model-en.md source extension prompt>"
```

Observed result:

- compound atomic rejection: `Plan step 1 failed quality gate: step text uses vague or broad wording; step text appears to contain multiple actions`
- oversized evidence rejection: `Evidence summary exceeds 500 characters; put long output in artifactRefs`
- final `task_next`: only next tool `task_evidence`, current step lock `T1-S1`
- compact `task_list`: `T1 [active] 1% criteria:0/1 - Weak Model Hardening`

0.1.3 current-step lock probe:

```sh
gtimeout 90s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-013-dogfood/sessions \
  pi --no-extensions --extension ./index.ts \
  --no-builtin-tools \
  --tools task_evidence,task_next \
  --session-id release-013-weak-en \
  --name release-013-weak-en \
  -p "Current-step lock probe. The active task is T1 and the current step should be T1-S1. Call task_evidence once for task_id T1 targeting future step_ids [\"T1-S2\"] and criterion_ids [\"T1-AC1\"]. Use complete quality fields and do not provide override_reason. Report the exact rejection and then call task_next."
```

Observed result: `Evidence step_ids must target current step T1-S1; overrideReason is required for T1-S2`; `task_next` still reported only next tool `task_evidence` and current step lock `T1-S1`.

0.1.3 Chinese weak-model smoke:

```sh
gtimeout 120s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-013-dogfood/sessions \
  pi --no-extensions --extension ./index.ts \
  --no-builtin-tools \
  --tools task_next,task_plan,task_list \
  --session-id release-013-weak-zh \
  --name release-013-weak-zh \
  -p "<docs/dogfood-prompts/weak-model-zh-TW.md installed package smoke prompt>"
```

Observed result: no active task caused `task_next` to recommend `task_plan`; Chinese compound wording was rejected; structured recovery included `retry_with: task_plan` and `do_not_retry_same_call: true`.

0.1.3 final installed package weak-model smoke:

```sh
rm -rf /private/tmp/pi-tasks-release-013-installed-final
mkdir -p /private/tmp/pi-tasks-release-013-installed-final/tarball /private/tmp/pi-tasks-release-013-installed-final/consumer /private/tmp/pi-tasks-release-013-installed-final/sessions
npm pack --pack-destination /private/tmp/pi-tasks-release-013-installed-final/tarball
cd /private/tmp/pi-tasks-release-013-installed-final/consumer
npm init -y
npm install /private/tmp/pi-tasks-release-013-installed-final/tarball/pi-tasks-0.1.3.tgz
node -e "import('pi-tasks')"
gtimeout 120s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-013-installed-final/sessions \
  pi --no-extensions --extension ./node_modules/pi-tasks/dist/index.js \
  --no-builtin-tools \
  --tools task_next,task_plan \
  --session-id release-013-installed-final-weak \
  --name release-013-installed-final-weak \
  -p "<docs/dogfood-prompts/weak-model-en.md installed package smoke prompt>"
```

Observed result: installed runtime exposed `task_next`; compound atomic step was rejected; structured recovery included `retry_with: task_plan`, `do_not_retry_same_call: true`, and no-active-task resume guidance. Final tarball shasum: `9a3d50a4b97088fb551e36e024b29f5547091e59`.

0.1.5 receiver-bound append compatibility source dogfood:

```sh
gtimeout 180s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-015-dogfood/sessions \
  pi --provider openai-codex --model gpt-5.5 \
  --no-extensions --extension ./index.ts \
  --no-builtin-tools \
  --tools task_plan,task_update,task_evidence,task_decision,task_complete,task_list \
  --session-id release-015-source-lifecycle \
  --name release-015-source-lifecycle \
  -p "Release 0.1.5 source dogfood. Create a valid task named Receiver Append Compatibility with one acceptance criterion and one atomic plan step. Do not set criterionIds manually. Update progress to 50 and next_action to record evidence. Record one agent decision about supporting receiver-bound appendEntry methods. Record passing test evidence for T1-AC1 and T1-S1 with complete quality fields. Mark T1-S1 done with evidence E1. Complete T1 with evidence E1. Call task_list with include_done=true and include_evidence=true. Report created task id, decision id, evidence id, and final status."
```

Observed result: created task `T1`, decision `D1`, evidence `E1`, final status `done`, progress `100%`, criteria `1/1`, evidence `1`.

0.1.5 same-session resume:

```sh
gtimeout 90s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-015-dogfood/sessions \
  pi --provider openai-codex --model gpt-5.5 \
  --no-extensions --extension ./index.ts \
  --no-builtin-tools \
  --tools task_list,task_resume \
  --session-id release-015-source-lifecycle \
  -p "Release 0.1.5 resume dogfood. Call task_resume and task_list with include_done=true and include_evidence=true. Report whether task T1 is restored as done with evidence E1 and decision D1."
```

Observed result: `T1` restored as done, evidence `E1` present, decision `D1` present, and no active task after completion.

0.1.5 fork replay:

```sh
gtimeout 90s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-015-dogfood/sessions \
  pi --provider openai-codex --model gpt-5.5 \
  --no-extensions --extension ./index.ts \
  --no-builtin-tools \
  --tools task_list \
  --fork release-015-source-lifecycle \
  --name release-015-fork-replay \
  -p "Release 0.1.5 fork replay dogfood. Call task_list with include_done=true and include_evidence=true. Report whether fork replay restored task T1 as done with evidence E1."
```

Observed result: fork replay restored `T1` as done with evidence `E1`, decision `D1`, and acceptance criteria `1/1` satisfied.

0.1.5 live TTY `/tasks` and clean exit:

```sh
env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-015-dogfood/sessions \
  pi --provider openai-codex --model gpt-5.5 \
  --no-extensions --extension ./index.ts \
  --session-id release-015-source-lifecycle
```

Observed result: `/tasks` returned `No tasks on this branch.` because compact default excludes completed tasks; `/tasks detail` rendered `T1 [done]`, decision `D1`, evidence `E1`, and `T1-AC1`; `/quit` exited cleanly and printed the Pi resume command.

0.1.5 final installed package smoke:

```sh
mkdir -p /private/tmp/pi-tasks-release-015-installed-final2/tarball /private/tmp/pi-tasks-release-015-installed-final2/consumer /private/tmp/pi-tasks-release-015-installed-final2/sessions
env npm_config_cache=/private/tmp/pi-tasks-release-015-installed-final2/npm-cache \
  npm pack --pack-destination /private/tmp/pi-tasks-release-015-installed-final2/tarball
cd /private/tmp/pi-tasks-release-015-installed-final2/consumer
npm init -y
env npm_config_cache=/private/tmp/pi-tasks-release-015-installed-final2/npm-cache \
  npm install /private/tmp/pi-tasks-release-015-installed-final2/tarball/pi-tasks-0.1.5.tgz
node -e "import('pi-tasks')"
gtimeout 120s env PI_CODING_AGENT_SESSION_DIR=/private/tmp/pi-tasks-release-015-installed-final2/sessions \
  pi --provider openai-codex --model gpt-5.5 \
  --no-extensions --extension ./node_modules/pi-tasks/dist/index.js \
  --no-builtin-tools \
  --tools task_plan,task_list \
  --session-id release-015-installed-final2-smoke \
  --name release-015-installed-final2-smoke \
  -p "Release 0.1.5 final installed package smoke. Create a valid task named Final Installed Receiver Smoke with one acceptance criterion and one atomic plan step. Do not set criterionIds manually. Call task_list and report the created task id and current status."
```

Observed result: tarball `pi-tasks-0.1.5.tgz` was created; clean consumer import passed; installed runtime created task `T1` with status `active`.

0.2.0 state-event source dogfood:

```sh
pi --no-extensions \
  --extension ./index.ts \
  --extension /private/tmp/pi-tasks-state-observer.ts \
  --no-builtin-tools \
  --tools task_plan,task_update,task_list,state_event_report \
  --session-dir /private/tmp/pi-tasks-release-020-dogfood/sessions \
  --session-id release-020-source-lifecycle-2 \
  --name release-020-source-lifecycle-2 \
  -p "Release dogfood for pi-tasks 0.2.0. Use only the enabled tools. Step 1 call task_plan to create an active task titled \"0.2.0 state hook dogfood\" with objective \"Verify state hook publication\" and one acceptance criterion \"State hook publishes after mutation\". Include one atomic plan step with expected output \"Observed state_event_report output\" and evidenceRequired false. Step 2 call task_update to set progress to 50 and next_action to \"Inspect state hook report\". Step 3 call state_event_report. Step 4 call task_list. In your final answer, include the state_event_report JSON exactly."
```

Observed result: `state_event_report` returned `count: 3`, reasons `session_start`, `task_mutation`, `task_mutation`, stable widget id `pi-tasks`, active task `T1`, and `lastHasRawEvents: false`.

0.2.0 same-session resume:

```sh
pi --no-extensions \
  --extension ./index.ts \
  --extension /private/tmp/pi-tasks-state-observer.ts \
  --no-builtin-tools \
  --tools task_list,state_event_report \
  --session-dir /private/tmp/pi-tasks-release-020-dogfood/sessions \
  --session-id release-020-source-lifecycle-2 \
  -p "Resume dogfood for pi-tasks 0.2.0. Use task_list to list the current active task, then call state_event_report. Final answer must include the state_event_report JSON exactly and mention whether T1 replayed."
```

Observed result: `T1` replayed; `state_event_report` returned `count: 1`, reason `session_start`, stable widget id `pi-tasks`, active task `T1`, and `lastHasRawEvents: false`.

0.2.0 fork replay:

```sh
pi --no-extensions \
  --extension ./index.ts \
  --extension /private/tmp/pi-tasks-state-observer.ts \
  --no-builtin-tools \
  --tools task_list,state_event_report \
  --session-dir /private/tmp/pi-tasks-release-020-dogfood/sessions \
  --fork release-020-source-lifecycle-2 \
  --session-id release-020-fork-replay \
  -p "Fork replay dogfood for pi-tasks 0.2.0. Use task_list to list the current active task, then call state_event_report. Final answer must include the state_event_report JSON exactly and mention whether T1 replayed in the fork."
```

Observed result: `T1` replayed in the fork; `state_event_report` returned `count: 1`, reason `session_start`, stable widget id `pi-tasks`, active task `T1`, and `lastHasRawEvents: false`.

0.2.0 live TTY `/tasks`, `/compact`, and clean exit:

```sh
pi --no-extensions \
  --extension ./index.ts \
  --extension /private/tmp/pi-tasks-state-observer.ts \
  --session-dir /private/tmp/pi-tasks-release-020-dogfood/sessions \
  --session-id release-020-source-lifecycle-2
```

Observed result: `/tasks` rendered active task `T1`; `/compact` on the small source lifecycle session reported `Compaction failed: Nothing to compact (session too small)` and was not counted as manual compaction coverage; `/quit` exited cleanly with exit code 0.

0.2.0 large-session manual `/compact`:

```sh
pi --no-extensions \
  --extension ./index.ts \
  --extension /private/tmp/pi-tasks-state-observer.ts \
  --no-builtin-tools \
  --tools task_plan,state_event_report \
  --session-dir /private/tmp/pi-tasks-release-020-compact/sessions \
  --session-id release-020-large-compact \
  --name release-020-large-compact \
  -p "Create a pi-tasks 0.2.0 large compaction dogfood task. Use task_plan to create an active task titled \"0.2.0 large compaction dogfood\" with objective \"Verify manual compact emits state\" and one acceptance criterion \"Manual compact emits pi-tasks state\". Include one atomic plan step with expected output \"Observer log contains task_mutation after compact\" and evidenceRequired false. Then call state_event_report and include its JSON exactly."
pi --no-extensions \
  --extension ./index.ts \
  --extension /private/tmp/pi-tasks-state-observer.ts \
  --no-tools \
  --session-dir /private/tmp/pi-tasks-release-020-compact/sessions \
  --session-id release-020-large-compact \
  -p @/private/tmp/pi-tasks-large-compact-prompt.md
pi --no-extensions \
  --extension ./index.ts \
  --extension /private/tmp/pi-tasks-state-observer.ts \
  --session-dir /private/tmp/pi-tasks-release-020-compact/sessions \
  --session-id release-020-large-compact
```

Observed result: live `/compact` compacted from 56,203 tokens; observer log recorded `session_start` and `task_mutation` events with stable widget id `pi-tasks`, active task `T1`, and `hasRawEvents: false`; `/quit` exited cleanly with exit code 0.

0.2.0 final installed package smoke:

```sh
mkdir -p /private/tmp/pi-tasks-release-020-installed/tarball /private/tmp/pi-tasks-release-020-installed/consumer /private/tmp/pi-tasks-release-020-installed/sessions
npm pack --pack-destination /private/tmp/pi-tasks-release-020-installed/tarball
cd /private/tmp/pi-tasks-release-020-installed/consumer
npm init -y
npm install /private/tmp/pi-tasks-release-020-installed/tarball/pi-tasks-0.2.0.tgz
node -e "import('pi-tasks')"
pi --no-extensions \
  --extension ./node_modules/pi-tasks/dist/index.js \
  --extension /private/tmp/pi-tasks-state-observer.ts \
  --no-builtin-tools \
  --tools task_plan,task_list,state_event_report \
  --session-dir /private/tmp/pi-tasks-release-020-installed/sessions \
  --session-id release-020-installed-smoke \
  --name release-020-installed-smoke \
  -p "Installed package dogfood for pi-tasks 0.2.0. Use task_plan to create an active task titled \"Installed 0.2.0 state hook smoke\" with one acceptance criterion \"Installed state hook publishes\" and one atomic plan step with expected output \"state_event_report output observed\" and evidenceRequired false. Then call state_event_report and task_list. Final answer must include the state_event_report JSON exactly."
```

Observed result: tarball `pi-tasks-0.2.0.tgz` was created; clean consumer import passed; installed runtime created task `T1`; `state_event_report` returned `count: 2`, reasons `session_start`, `task_mutation`, stable widget id `pi-tasks`, active task `T1`, and `lastHasRawEvents: false`.

## Package Gates

Also passed on 2026-06-19:

- `npm run typecheck`
- `npm run check`
- `npm test` (42 unit tests)
- `npm run build`
- `node --experimental-strip-types -e "import('./index.ts')"`
- `node -e "import('./dist/index.js')"`
- `npm pack --dry-run` with `npm_config_cache=/private/tmp/pi-tasks-npm-cache`
- `npm pack --pack-destination /private/tmp/pi-tasks-final-release/tarball`
- clean consumer `npm install /private/tmp/pi-tasks-final-release/tarball/pi-tasks-0.1.0.tgz`
- clean consumer `node -e "import('pi-tasks')"`
- `npm audit --audit-level=low` with `npm_config_cache=/private/tmp/pi-tasks-npm-cache`
- installed-package Pi smoke through `./node_modules/pi-tasks/dist/index.js`
- 0.1.2 clean tarball install plus `node -e "import('pi-tasks')"`
- 0.1.2 installed-package token-output Pi smoke through `./node_modules/pi-tasks/dist/index.js`
- weak-model hardening Pi dogfood for `task_next`, compound plan rejection, current-step evidence lock, and evidence budget rejection
- 0.1.2 installed-package weak-model smoke for `task_next` and structured recovery
- 0.1.3 `npm run release:check`
- 0.1.3 source weak-model English and Chinese dogfood
- 0.1.3 installed-package weak-model smoke through `./node_modules/pi-tasks/dist/index.js`
- 0.1.5 source lifecycle, same-session resume, fork replay, live `/tasks detail`, and clean `/quit`
- 0.1.5 clean tarball install plus installed-package Pi smoke through `./node_modules/pi-tasks/dist/index.js`
- 0.2.0 `npm run release:check`
- 0.2.0 source state-event lifecycle, same-session resume, fork replay, live `/tasks`, large-session manual `/compact`, and clean `/quit`
- 0.2.0 clean tarball install plus installed-package Pi state-event smoke through `./node_modules/pi-tasks/dist/index.js`

## Remaining Runtime Coverage

No remaining runtime coverage gaps are known for the 0.1.0 or 0.2.0 release scope.

Known runtime note:

- Long multi-step Pi self-correction prompts can still stall at the external agent layer; short staged prompts are verified and should remain the recommended release dogfood style.