FE-841: Pi agent foundations — in-process SDK, toolchain profiles by kostandinang · Pull Request #194 · hashintel/brunch

kostandinang · 2026-06-10T13:58:51Z

Stack Context

This is the foundation layer of the cook/brunch-serve stack. It establishes how the cook agent runs — embedded rather than shelled out — and how it targets a repo's test runner, so the brownfield-delivery and operational-hardening PRs above it can build on a stable agent runtime.

Consolidates two frontier items (FE-841, FE-843) into one PR.

What?

Embed pi as an in-process SDK (FE-841). Replace the external pi CLI dependency with an in-process SDK session — per-session cwd/tools, env-fed auth, no external install. Hardened with session guards, tightened timeouts and output caps, cleaned-up late setup timeouts, and pinned SDK patch levels (node floor + nested vulnerable patches).

Data-driven toolchain profiles (FE-843). A profile registry (node-vitest, node-test, node-jest, deno, …) replacing hardcoded toolchain assumptions. The architect classifies the profile from spec prose; selection precedence is flag ≫ spec ≫ bun, persisted into plan.yaml. Inherited profile names are rejected.

Why?

The cook agent previously depended on an external pi CLI and assumed a single toolchain. Neither is portable for autonomous brownfield delivery from a spec. Embedding the SDK removes an install-time dependency and gives per-session control; the data-driven profile registry lets the architect target the repo's real test runner instead of a hardcoded one.

Note: cook agent sandbox confinement (FE-853) was originally slated for this foundation but is un-integrated parallel work (it predates the data-driven profile refactor); it stays on its own branch/PR (#211) pending a dedicated integration.

cursor · 2026-06-10T13:58:59Z

PR Summary

High Risk
Changes the core cook LLM execution path (auth, timeouts, tool scoping) and introduces a large new dependency tree; unknown profile handling and architect fallbacks affect every plan/cook run.

Overview
Cook runs the coding agent in-process via @earendil-works/pi-coding-agent (0.79.8): runPi builds SDK sessions with sandbox cwd, comma-split tool allowlists (read-only for evaluate-done), ANTHROPIC_API_KEY-only auth in a per-run temp agent dir, buffered text_delta output (not stdout), and shared timeout/output-cap guards with abort on hang or overflow. The external pi CLI spawn path is removed.

Toolchain profiles are data-driven in project-profile.ts (bun, brunch, node-vitest, node-test, node-jest, deno). brunch plan accepts --profile=<id>; the emitter resolves flag ≫ spec profile ≫ architect-classified profile ≫ bun, stamps plan.profile on emitted YAML, and resolveToolchain throws on unknown ids at cook time (absent/empty still defaults to bun for fixtures). The build architect schema/prompt adds optional profile classification from spec prose only.

Docs (orchestration-guide.md) and planning notes reflect embedded cook + optional npx pi for plan translation. Node engine floor is >=22.19.0; lockfile adds the pi SDK tree and npm overrides for nested deps.

^{Reviewed by Cursor Bugbot for commit a565a73. Bugbot is set up for automated code reviews on this repo. Configure here.}

github-actions · 2026-06-10T13:59:29Z

Dependency Review

The following issues were found:

❌ 1 vulnerable package(s)
✅ 0 package(s) with incompatible licenses
✅ 0 package(s) with invalid SPDX license definitions
✅ 0 package(s) with unknown licenses.
⚠️ 11 packages with OpenSSF Scorecard issues.

View full job summary

kostandinang · 2026-06-10T16:01:28Z

FE-841: Pi agent foundations — in-process SDK, toolchain profiles #194 : 3 dependent PRs (#198 , #211 , #212 ) 👈 (View in Graphite)
main

This stack of pull requests is managed by Graphite. Learn more about stacking.

…endency Replace the spawn('pi', …) subprocess in pi-actions.ts with an in-process createAgentSession over @earendil-works/pi-coding-agent (pinned 0.79.1), so a fresh checkout no longer needs an external pi binary on $PATH — only ANTHROPIC_API_KEY. - buildSessionOptions maps the old CLI flags to SDK config: tools allowlist, model-registry lookup, system-prompt override, in-memory session + empty resource overrides (no context files). - Auth fed from brunch's own ANTHROPIC_API_KEY (no pi login / ~/.pi/auth.json); one reused per-process agent dir, no per-call temp leak. - Output captured off text_delta events (never brunch's stdout, keeps the cook SSE stream clean); cooperative timeout + 10MB output cap. - Read-only evaluator scoping (I126-K) preserved via the SDK tools allowlist; injectable SessionFactory keeps the drive loop offline-testable. - Drops the `which pi` prereq from the cook guide (npx pi for the interim helper). - Override transitive hono to 4.12.25 (patches the moderate advisories the dependency-review gate flags). - PI_REAL_LLM-gated smoke proves a real in-process run with no pi on $PATH. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

…ard queue Co-authored-by: Claude <noreply@anthropic.com>

…t/deno profiles Profiles are now data literals (path templates + argv template + conventions prose) compiled into the unchanged Toolchain interface — adding a runtime is one data entry. Registry invariants pinned enumerably across all profiles. Co-authored-by: Claude <noreply@anthropic.com>

…d into plan.yaml brunch plan gains --profile=<id> (validated via parseProfileId); the emitter resolves the chain once and stamps the id on both the authored and fallback plans, so cook always reads the profile the emitter used. resolveToolchain now throws UnknownProfileError on a typo'd id instead of silently running bun; absent stays lenient for hand-authored fixtures. Co-authored-by: Claude <noreply@anthropic.com>

architectPlan's schema gains an optional profile enum (registry ids, null when the spec is silent); the emitter chain becomes flag ≫ spec profile ≫ architect-classified ≫ bun. A hallucinated profile fails the schema parse and rides the existing deterministic fallback. D160-K intact — classification reads projected spec prose only. Queue exhausted: CARDS.md retired; PLAN frontier marked branch-complete. Co-authored-by: Claude <noreply@anthropic.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

pi-coding-agent 0.79.1 pinned undici 8.3.0, which trips three advisories flagged by the dependency-review check (GHSA-38rv-x7px-6hhq, GHSA-vmh5-mc38-953g, GHSA-pr7r-676h-xcf6). 0.79.8 is the first patch release that depends on the fixed undici 8.5.0. A root override can't reach the SDK's exact-pinned nested undici, so bump the SDK itself.

Detect SDK-level assistant failures after prompt resolution so cook does not advance past a failed in-process agent run. Co-authored-by: Cursor <cursoragent@cursor.com>

cursor

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 29dd141. Configure here.}

Co-authored-by: Cursor <cursoragent@cursor.com>

cursor Bot reviewed Jun 10, 2026

View reviewed changes

Comment thread src/orchestrator/src/pi-actions.ts Outdated

kostandinang force-pushed the ka/fe-841-pi-sdk-embed branch from 1b4e94a to 3e1665d Compare June 10, 2026 14:08

cursor Bot reviewed Jun 10, 2026

View reviewed changes

Comment thread src/orchestrator/src/pi-actions.ts Outdated

Comment thread src/orchestrator/src/pi-actions.ts Outdated

kostandinang mentioned this pull request Jun 10, 2026

FE-843: Expand cook toolchain profiles, selected at plan time #198

Closed

cursor Bot reviewed Jun 10, 2026

View reviewed changes

Comment thread src/orchestrator/src/pi-actions.ts Outdated

kostandinang requested a review from lunelson June 10, 2026 17:06

cursor Bot reviewed Jun 10, 2026

View reviewed changes

Comment thread src/orchestrator/src/pi-actions.ts Outdated

Comment thread src/orchestrator/src/pi-actions.ts

lunelson previously approved these changes Jun 11, 2026

View reviewed changes

kostandinang dismissed lunelson’s stale review via a6b27a5 June 12, 2026 08:55

kostandinang requested a review from lunelson June 12, 2026 09:24

kostandinang mentioned this pull request Jun 15, 2026

FE-853: Confine cook agent tools to the run sandbox #211

Open

kostandinang and others added 4 commits June 15, 2026 12:57

FE-841: Harden in-process pi session guards

28f4006

Co-authored-by: Cursor <cursoragent@cursor.com>

FE-841: Clean up late pi setup timeouts

a7a6635

Co-authored-by: Cursor <cursoragent@cursor.com>

FE-841: Tighten pi timeout and output caps

81e0833

Co-authored-by: Cursor <cursoragent@cursor.com>

kostandinang force-pushed the ka/fe-841-pi-sdk-embed branch from a6b27a5 to 81e0833 Compare June 15, 2026 11:59

FE-841: pin pi SDK nested vulnerable patches

7ed8f1e

Co-authored-by: Cursor <cursoragent@cursor.com>

FE-841: pin root ws to patched release

6449742

Co-authored-by: Cursor <cursoragent@cursor.com>

cursor Bot reviewed Jun 16, 2026

View reviewed changes

Comment thread package.json Outdated

kostandinang and others added 6 commits June 17, 2026 09:33

FE-841: require the pi SDK node floor

9d0df30

Co-authored-by: Cursor <cursoragent@cursor.com>

FE-843: Frontier setup — toolchain-profile-expansion plan entry + 3-c…

669b679

…ard queue Co-authored-by: Claude <noreply@anthropic.com>

FE-843: reject inherited profile names

8557ba7

Co-authored-by: Cursor <cursoragent@cursor.com>

This was referenced Jun 17, 2026

FE-881: Cook agent loads the target repo's sandbox-scoped skills #227

Closed

FE-883: Cook artifact lifecycle — git-merge slice composition #230

Closed

FE-884: Recover from cook epic-verification failure instead of halting #232

Open

kostandinang changed the title ~~FE-841: Embed pi as an in-process SDK to drop the external pi CLI dependency~~ FE-841: Pi agent foundations — in-process SDK, toolchain profiles Jun 18, 2026

lunelson previously approved these changes Jun 19, 2026

View reviewed changes

kostandinang dismissed lunelson’s stale review via b75cac0 June 19, 2026 11:38

cursor Bot reviewed Jun 19, 2026

View reviewed changes

Comment thread src/orchestrator/src/pi-actions.ts

kostandinang self-assigned this Jun 19, 2026

FE-841: reject failed embedded pi turns

29dd141

Detect SDK-level assistant failures after prompt resolution so cook does not advance past a failed in-process agent run. Co-authored-by: Cursor <cursoragent@cursor.com>

cursor Bot reviewed Jun 19, 2026

View reviewed changes

Comment thread src/orchestrator/src/project-profile.ts

FE-841: tolerate empty plan profiles

a565a73

Co-authored-by: Cursor <cursoragent@cursor.com>

kostandinang requested a review from lunelson June 19, 2026 15:09

lunelson approved these changes Jun 19, 2026

View reviewed changes

kostandinang added this pull request to the merge queue Jun 19, 2026

Merged via the queue into main with commit 977ea9c Jun 19, 2026
6 checks passed

kostandinang deleted the ka/fe-841-pi-sdk-embed branch June 19, 2026 15:27

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FE-841: Pi agent foundations — in-process SDK, toolchain profiles#194

FE-841: Pi agent foundations — in-process SDK, toolchain profiles#194
kostandinang merged 15 commits into
mainfrom
ka/fe-841-pi-sdk-embed

kostandinang commented Jun 10, 2026 •

edited

Loading

Uh oh!

cursor Bot commented Jun 10, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Jun 10, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

kostandinang commented Jun 10, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cursor Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

kostandinang commented Jun 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Stack Context

What?

Why?

Uh oh!

cursor Bot commented Jun 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Summary

Uh oh!

github-actions Bot commented Jun 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Dependency Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

kostandinang commented Jun 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cursor Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

kostandinang commented Jun 10, 2026 •

edited

Loading

cursor Bot commented Jun 10, 2026 •

edited

Loading

github-actions Bot commented Jun 10, 2026 •

edited

Loading

kostandinang commented Jun 10, 2026 •

edited

Loading