Conversation
Build a self-contained pi extension (pi-extension/) that ports the Claude Code autoresearch plugin to pi's extension event system, keeping the same spirit as the claude and opencode integrations. Guardrails (hooks -> pi events), all fail-open: PreToolUse (Bash) -> tool_call (bash) scout/privacy/dangerous-cmd block PreToolUse (Read/Write) -> tool_call (read/...) scout/privacy block UserPromptSubmit -> before_agent_start iteration-context + dev-rules inject UserPromptSubmit -> input simplify-gate (block/warn shipping verbs) SessionStart -> session_start session-init SessionEnd -> session_shutdown stop-notify Commands (14) are exposed as pi prompt templates (/autoresearch, /autoresearch_debug, ...) and the dispatcher as a pi skill (/skill:autoresearch), contributed via resources_discover. The AR_DISABLE_* per-hook env flags are preserved. Verified end-to-end: tsc --strict passes (0 errors), 22/22 logic tests pass, extension loads under pi's jiti loader, and session-init/iteration-context/ scout-block/dangerous-cmd-block/simplify-gate/stop-notify all fire correctly (confirmed in ~/.pi/agent/hooks/.logs/). AGENTS.md documents the pi install path, agent-specific notes, and repo layout.
Declaring skills and prompts in both the package.json `pi` manifest AND the extension's `resources_discover` handler caused pi to register each resource twice, emitting `collision` diagnostics for every prompt (each colliding with itself) and the skill. Drop `pi.skills`/`pi.prompts` from the manifest; keep only `pi.extensions` (the entry point). The skill + 14 prompt templates are contributed solely via resources_discover at runtime, which works for `pi install`, `pi -e ./pi-extension` (dir), and `pi -e ./src/index.ts` (bare file). Verified /autoresearch_plan and /skill:autoresearch still load correctly with no manifest declaration. README documents the single-contribution mechanism and the skill-name dedup behavior (extension copy wins; stale global/project copies can be removed).
The extension wrote its bounded runtime log to ~/.pi/agent/hooks/.logs/, which
created a legacy hooks/ directory and triggered pi's migration warning
("Global hooks/ directory found. Hooks have been renamed to extensions.").
Move the log to ~/.pi/agent/autoresearch/.logs/<projectHash>/hook-log.jsonl —
a namespaced dir that does not trip the hooks→extensions deprecation. Verified
the hooks/ directory is no longer created and guardrails still fire + log.
Wire the pi extension into the same distribution, test, and release surface as Claude/OpenCode/Codex: - scripts/install.sh: add `--pi` mode (global/local). Copies the pi-extension/ bundle to <config>/extensions/autoresearch-pi so pi auto-discovers it. Honors PI_AGENT_DIR for the config dir; pi-aware overwrite confirmation. - tests/test-pi.sh: 53 cases covering the ported guardrails (shell tokenizer, ignore matcher, dangerous-cmd-block, scout-block, privacy-block, simplify-gate, iteration-context, env-flag disables) + a tsc --strict type-check against pi-coding-agent types. Gracefully skips the type-check when pi types aren't installed. All 53 pass; cleans up its own artifacts. - .github/workflows/release-readiness.yml: install pi-coding-agent and run tests/test-pi.sh on every OS in the conformance matrix. - scripts/release.sh + scripts/release.md: add test-pi.sh to the release gate. - README.md: pi badge, intro, hook-parity note, and a pi Quick Start section (guided install / pi install / try-without-installing). All five test suites green: test-pi 53, test-hooks 228, test-orchestrator 195, test-regression 65, test-maintenance 50.
Bring the pi extension fully into the canonical-source model documented in CONTRIBUTING.md: .claude/ is the source of truth, transform.sh regenerates every platform port. Previously the pi-extension skill + prompts were hand-copied from the Codex .agents/ bundle, which used $autoresearch syntax and request_user_input — wrong conventions for pi (pi uses /autoresearch slash commands and ctx.ui). - scripts/transform.sh: add transform_pi() + --pi flag. Regenerates pi-extension/skills/autoresearch/ (SKILL.md + references) and pi-extension/prompts/ (14 flat underscore templates) from .claude/ with pi adaptations (colon→underscore, /autoresearch:X→/autoresearch_X, AskUserQuestion→ctx.ui). Adds the pi skills dir to sync_runtime_helpers so orchestrator/regression scripts stay in sync. Idempotent. - pi-extension/: regenerated bundle now uses /autoresearch slash commands and ctx.ui (no more $autoresearch drift from Codex). - CONTRIBUTING.md: document pi in the repo structure, file table, quick start, multi-platform sync, testing checks, and add a "pi Hook Development" section mirroring the Claude hook guide. All five suites green (591 tests). transform.sh idempotent. /autoresearch_fix prompt loads and expands correctly with the regenerated bundle.
The subagent-context hook was the one Claude hook not yet ported. pi has no native subagent-spawn event in the extension API, but the widely-used pi-subagents extension provides two clean connection points: - Child side: pi subagents are full pi sessions that load this same extension and fire before_agent_start. When PI_SUBAGENT_CHILD=1, the hook injects the project/branch/active-TSV/iteration block — the direct port of Claude's SubagentStart injection. iteration-context already flows into children via the same handler; subagent-context adds the structured header. - Parent side: listen for pi-subagents' `prompt-template:subagent:request` event on the shared pi.events bus to log delegation launches (bounded metadata only). Notification only — the request isn't mutated. Both paths fail open if pi-subagents isn't installed (event never fires / env not set). AR_DISABLE_SUBAGENT_CONTEXT gates both. - pi-extension/src/hooks/subagent-context.ts: buildSubagentContext + isSubagentChild - pi-extension/src/guardrails.ts: wire child-side injection at before_agent_start + parent-side pi.events listener - tests/test-pi.sh: 5 new cases (58 total, all green) - README/index.ts: document the pi-subagents integration All five suites green (596 tests). transform.sh idempotent.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Hi 👋
This PR adds a pi (coding agent) extension that ports the Claude Code autoresearch plugin to pi's extension event system, alongside the existing Claude/OpenCode/Codex integrations.
Guardrails
All 9 Claude hooks ported to pi events. All fail-open, all gated by the same
AR_DISABLE_*env flags:subagent-contextintegrates with the widely-used pi-subagents extension: on the child side (PI_SUBAGENT_CHILD=1) it injects the project/branch/active-TSV/iteration block atbefore_agent_start; on the parent side it logs delegation launches via theprompt-template:subagent:requestevent. Both fail open if pi-subagents isn't installed.Commands
The 14 autoresearch commands as pi prompt templates (
/autoresearch,/autoresearch_debug, …) + the dispatcher as a pi skill (/skill:autoresearch), contributed viaresources_discover.Canonical-source model
The pi bundle is generated from
.claude/bytransform.sh(newtransform_pi()+--piflag), same as the OpenCode/Codex ports, not hand-copied. Keeps the slash-command +ctx.uiconventions correct and prevents drift. Idempotent; covered bytest-maintenance.sh.Install
Tests
New
tests/test-pi.sh(58 cases): guardrail logic for every ported hook, env-flag disables, subagent-context, and atsc --strict type-checkagainstpi-coding-agenttypes.Wired into the release-readiness CI (installs pi, runs on all 3 OS) and the release gate.
All five suites green: test-pi 58, test-hooks 228, test-orchestrator 195, test-regression 65, test-maintenance 50.
Files