[renovate]

This PR contains the following updates:

Package	Change	Age	Adoption	Passing	Confidence
@narumitw/pi-starship (source)	0.55.0 → 0.67.1
@tintinweb/pi-subagents	0.19.0 → 0.23.1
pi-ask-user	0.15.0 → 0.17.2
pi-web-search	1.4.0 → 1.7.0

Release notes are maintained in a PR comment by the renovate-release-notes-comment workflow.

---

Configuration

📅 Schedule: (in timezone America/Los_Angeles)

• Branch creation
  • "after 4pm on thursday"
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR is behind base branch, or you tick the rebase/retry checkbox.

👻 Immortal: This PR will be recreated if closed unmerged. Get config help if that's undesired.

---

• If you want to rebase/retry this PR, check this box

---

This PR was generated by Mend Renovate. View the repository job log.

[scottames-github-bot]

Renovate Release Notes

Generated from Renovate's update table by the renovate-release-notes-comment workflow.

Packages that cannot be summarized from GitHub releases are listed explicitly below.

tintinweb/pi-subagents (@​tintinweb/pi-subagents)

v0.23.1: v0.23.1

Compare Source

What changes for existing users

• viewerMarkdown no longer applies to tool calls (#413). Tool calls in the conversation viewer now always use their own renderers. The setting still controls the rest of the transcript.

New

• The conversation viewer draws tool calls the way pi's transcript does (#413). Each call uses the renderer registered in the agent's session, with its result folded into the same block: bash shows its command and output, read its line numbers, edit its diff, and an extension tool its own view. ctrl+o (pi's app.tools.expand) unfolds shortened output.

Fixes

• get_subagent_result finds a finished agent by handle after a reload (#429, #428). A name or a type handle such as explore came back not found once the agent's in-memory record was gone. The saved record's handle and alias now match too.

QA

• Automated: 2533 tests passed, including 95 scripted end-to-end tests (real pi sessions, scripted model).
• Live: the end-to-end suite against a real model. It covers foreground and background spawn with get_subagent_result, Explore dispatch, workflow fan-out, and structured output.

Thanks @​pf-sjtu, @​HerbertGao. Full details in CHANGELOG.md.

v0.23.0: v0.23.0

Compare Source

What changes for existing users

• Explore uses the session's model (#175, #350). On non-Anthropic sessions, its Haiku pin spent another provider's credits. To keep Haiku, eject Explore in /agents and add model: anthropic/claude-haiku-4-5.
• Agent files written for Claude Code behave as written (#396, #370). Tool names like Read and Glob now work. disallowedTools is enforced, and mcp__… entries narrow MCP tools. An agent whose tools: match nothing is refused. If you copied agents from Claude Code, review their toolsets.
• @​handle mentions honour inherit_context: true in direct mode and in the model-mode fallback (#300). Those prompts now include the conversation.
• Reopening an isolation: worktree agent no longer creates a fresh worktree (#420). If its worktree held changes, the reopen is refused and the work is on its branch. If the worktree was clean, the agent continues in the main checkout.

New

• Agent({ resume }) and workflow resume reopen agents that were cleaned up, after the ten-minute eviction or a reload (#420).
• get_subagent_result returns the saved result after eviction instead of "Agent not found" (#397).

Fixes

• The Agent tool no longer suggests Explore when Explore is disabled (#330).
• ext:<pkg> and exclude_extensions: <pkg> now match packages that declare a directory in pi.extensions.
• Workflow completion notifications count agents, not progress events. A 28-agent run used to report "27/138 agents".

QA

• Automated: 2519 unit tests and 95 scripted end-to-end tests (real pi sessions, scripted model). New end-to-end checks cover isolated: true keeping MCP out of the subagent and the workflow completion count.
• Live: the end-to-end suite against a real model. It covers background spawn with get_subagent_result, Explore dispatch, workflow fan-out, and structured output.
• Manual: a pi session on Haiku exercised every Agent parameter (types, name, max_turns, thinking, model, inherit_context, isolated, run_in_background, schedule with offset/cron/interval, worktree isolation with file changes) and every workflow agent() option (effort, model, agentType, schema, gate pass/fail, invalid inputs). It found the workflow count bug above. Everything else it flagged was documented behavior.

Thanks @​DraconDev, @​renatomoselli, @​yukukotani, @​vlamai, @​shkrabov, @​lluppi, @​zenworr. Full details in CHANGELOG.md.

v0.22.1: v0.22.1

Compare Source

No release notes body was provided by GitHub.

v0.22.0: v0.22.0

Compare Source

No release notes body was provided by GitHub.

v0.21.0: v0.21.0

Compare Source

No release notes body was provided by GitHub.

v0.20.0: v0.20.0

Compare Source

No release notes body was provided by GitHub.

edlsh/pi-ask-user (pi-ask-user)

v0.17.2: v0.17.2

Compare Source

0.17.2 (2026-11-01)

Bug Fixes

• ignore top-level single-question fields in questions batches (#88) (7e2cf4d)
• wrap single-select text by display width (#95) (d0a16b3)

v0.17.1: v0.17.1

Compare Source

Fixed

• In Tern, the native prompt is drawn in the composer's place (omp's current omp.editor.ask shape) instead of a modal sheet over the transcript. Option marks no longer overlap the option labels, the prompt stays within 60% of the pane with the context and options scrolling inside it, the transcript stays visible, batch tabs no longer show raw ANSI resets, and the review page keeps its answers in view beside short question keys. Needs a pi-tern that places composer-shaped presentations in the dock. (#89)

v0.17.0: v0.17.0

Compare Source

Changed

• ask_user declares exposure: "model-only", so codemode and other scripts cannot open user prompts. Hosts older than Pi 1.0.2 ignore the field at runtime. (#78)

Added

• Optional ask-user.json settings file in the agent directory (~/.pi/agent, or PI_CODING_AGENT_DIR when set) for personal defaults: displayMode, singleSelectLayout, allowComment, contextExpanded, overlayToggleKey, commentToggleKey, and emitFullEvents. Precedence is per-call parameter, then environment variable, then the file, then the built-in default. The file is re-read on every call; unknown keys and invalid values are ignored, and malformed JSON fails the call with the file path. (#71)
• Native presentation in Tern through pi-tern's optional hooks: the prompt is a native sheet in omp's ask style (question tabs, markdown question and context, clickable options, a native editor bound to the custom-response and comment editor, the review page, an action bar), and the tool card shows every option with the chosen ones checked. Pointer events run the same transitions as the keys. Without pi-tern nothing changes; the hooks are global symbols and nothing imports pi-tern.

Fixed

• A call that sends both question and questions now treats the top-level question as the batch heading instead of failing. The heading appears in the frame title, review page, native Tern sheets, fallback dialogs, and result text, and is returned as details.heading. (#83)

v0.16.0: v0.16.0

Compare Source

Added

• questions parameter for asking 2-4 independent questions in one call. Each entry has its own question, context, options, allowMultiple, and allowFreeform; the remaining parameters apply to the whole batch. The prompt shows a page per question plus a review page: tab / shift+tab switch pages without losing filters or drafts, and only the review page submits. Unanswered questions can be submitted as skipped after a second confirmation. Batches return { kind: "batch", questions, answers, cancelled } details and publish their events only on submit, with a batch: { index, total } position. In RPC mode the questions are asked one after another with the fallback dialogs. One timeout covers the whole batch. Closes chore(deps): update aqua #62. The page and review flow follows the design first proposed in chore(deps): update ⇡ golang module github.com/magefile/mage to v1.15.0 #39.

Changed

• question is no longer required by the schema; each call must pass exactly one of question or questions, which is checked before any UI opens.
• The tool description and prompt guidelines now describe batches. They still default to one focused question per call and keep dependent questions for later calls, but every host's system prompt changes once on upgrade.
• In a batch, tab / shift+tab switch questions instead of moving the option selection. Single questions are unchanged.

v0.15.1: v0.15.1

Compare Source

Fixes

• Abort freeform and RPC prompts consistently, reject late answers, and stop follow-up dialogs after cancellation.
• Treat cancellation or timeout of an optional RPC comment as cancellation, while preserving Enter-to-skip.
• Release prompt timers and abort listeners and settle each custom prompt once.
• Emit redacted cancellation events consistently for displayed freeform prompts.
• Report tool failures through Pi's error contract instead of treating them as successful calls.
• Correct bundled skill examples, parameter/privacy guidance, and installed-package documentation links.

Verification

133 automated tests pass. CI covers Node.js 22/24 and Pi 0.74.0/0.87.0, with strict types, real-host/native-TUI smoke, schema-validated examples, and package-content checks.

Thanks to Riley S.G. Yoo for the contributions in #49 and #53, incorporated with attribution into #66 and #67.

Runtime peer ranges are unchanged. Whole-call fallback timeout policy, RPC input hardening, Unicode wrapping, and larger feature proposals remain separate work.

Skipped Packages

GitHub Release Notes Unavailable

• @​narumitw/pi-starship: No GitHub release was found for 0.67.1

• pi-web-search: No GitHub release was found for 1.7.0

[Musthaq22]

066e3f1