Inject the pstack skills dir so a poteto-agent child can find poteto-mode - #1
Merged
Merged
Conversation
…mode The poteto-agent brief told a child to read poteto-mode's SKILL.md at the <location> of its <available_skills> entry. poteto-mode is hidden (disable-model-invocation: true), so the host never lists it there and the instruction could not resolve. Upstream Cursor has no such sentence: a poteto-agent child resolves the skill by name through the host. Prime Agent has no equivalent, so the extension now appends one line on every turn, mode on or off: pstack skills dir: <SKILLS_DIR> (every pstack skill, hidden ones included, is <name>/SKILL.md under it) The brief body is back to upstream's sentence plus one clause that points at that line. The poteto-agent-find seam emits the same clause, so the wording survives a sync if upstream ever moves the text under skills/. Tests: systemPromptInjection strings updated for the new first line, a skillsDirLine unit test, an end-to-end before_agent_start test that registers the extension against a fake host and asserts <dir>/poteto-mode/SKILL.md exists at the injected path, and a lever test pinning the seam output. npm run check passes; npm run sync:upstream writes 0 files under skills/.
![barnard-33[bot]](https://avatars.githubusercontent.com/in/4065997?s=60&v=4){kind=small}
| * skill paths from. `agents/poteto-agent.md` points children at it. | ||
| */ | ||
| export function skillsDirLine(skillsDir: string): string { | ||
| return `pstack skills dir: ${skillsDir} (every pstack skill, hidden ones included, is <name>/SKILL.md under it)`; |
There was a problem hiding this comment.
non_blocking (correctness, confidence 0.84): The injected line does not delimit or encode the skills directory unambiguously. If the package is installed below a path containing ( such as /opt/Prime Agent (Beta)/..., the same ( sequence appears inside the path and before the explanatory suffix. A child following the documented line-based lookup can truncate the path and fail to read poteto-mode/SKILL.md.
Suggested fix: Put the raw skills directory on a line by itself, or encode/quote it with an unambiguous format and update the brief and test parser to use that format.
| sendUserMessage: () => undefined, | ||
| }; | ||
| const previous = process.env.PRIME_AGENT_CODING_AGENT_DIR; | ||
| process.env.PRIME_AGENT_CODING_AGENT_DIR = mkdtempSync(join(tmpdir(), "pstack-agent-dir-")); |
There was a problem hiding this comment.
non_blocking (other, confidence 0.99): The new end-to-end test creates a unique temporary agent directory but never removes it. Each test run leaves an empty pstack-agent-dir-* directory under the system temporary directory.
Suggested fix: Store the result of mkdtempSync in an agentDir variable and call rmSync(agentDir, { recursive: true, force: true }) in the existing finally block, while restoring the environment variable.
Tau PR reviewVerdict: Approved Coordinator approved; risk tier full; specialists completed 14/14. Blocking findings: 0; non-blocking findings: 2. Approval bias applied: approved with comments unless a concrete blocking shipping risk remained after dedupe/downgrade. Findings
Review scope
Skipped checks
Tau PR review — risk tier |
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.
What changes for a user
A child spawned with the
poteto-agentbrief can now find and readpoteto-mode'sSKILL.md. Before, the brief told the child to look uppoteto-modein its<available_skills>list.poteto-modeis hidden (disable-model-invocation: true), so it is never listed there, and the instruction could not resolve.Every session now carries one extra line in its system prompt, whether or not
/poteto-modeis on:This matches upstream. In Cursor, a
poteto-agentchild resolvespoteto-modeby name through the host, from any parent, with no toggle. Prime Agent has no equivalent inside a child, so a stable, always-present path is the closest native form.How it was done
extensions/pstack/index.ts: newskillsDirLine();systemPromptInjectiontakesskillsDirand always leads with that line. Thebefore_agent_starthook passes the sameSKILLS_DIRthat/pstack offalready uses.agents/poteto-agent.md: body is back to upstream's sentence plus one clause pointing at the injected line. The Pi-erafffindsentence and the Prime-era<location>sentence are both gone.scripts/reground-from-pi.mjs: thepoteto-agent-findseam emits the same clause, so the wording survives a sync if upstream ever moves that text underskills/.README.md,docs/guide.md,CHANGELOG.md: describe the line and why it is unconditional.Tests:
extensions/pstack/index.test.ts: injection strings updated for the new first line; askillsDirLineunit test; an end-to-end test that registers the extension against a fake host, firesbefore_agent_start, parses the injected path, and asserts<dir>/poteto-mode/SKILL.mdexists on disk. Withindex.tsreverted, that test fails.scripts/reground-from-pi.test.mjs: pins the seam output.Verified live:
prime-agent -p --no-tools -e <this worktree>answered "What is the pstack skills dir?" with the worktree'sskills/path and "Is poteto-mode listed as an available skill?" withNo.Checks
npm run checkpasses locally (lint 47 skills, 0 problems; node 53 pass; bun 52 pass)scripts/reground-from-pi.mjsseams orscripts/prime-overrides/(not by hand underskills/).npm run sync:upstreamwrites 0 files underskills/.Note
Low Risk
Prompt and extension injection only; no auth, data, or spawn semantics changes beyond giving children a stable path to an existing skill file.
Overview
poteto-agentchildren can now locate hiddenpoteto-mode. The brief no longer tells them to use<available_skills>(wherepoteto-modenever appears because it isdisable-model-invocation: true). Instead it points at<dir>/poteto-mode/SKILL.md, with<dir>taken from a new always-on system-prompt line.The pstack extension adds
skillsDirLine()and makessystemPromptInjectionlead every turn withpstack skills dir: <path> …, including when/poteto-modeis off and forrlm.spawnchildren. Thebefore_agent_starthook always appends that block (not only when poteto mode or a role table is active).agents/poteto-agent.md,scripts/reground-from-pi.mjs(poteto-agent-findseam), and README / guide / CHANGELOG are aligned with that contract. Tests cover the new line, updated injection strings, and an integration check that the injected path containspoteto-mode/SKILL.mdon disk.Reviewed by Cursor Bugbot for commit 27e231b. Bugbot is set up for automated code reviews on this repo. Configure here.