Fix/0.6.1 os aware scripts - #14
Merged
Merged
Conversation
…p did A macOS run of the React TODO spec on 0.6.0, ornith:9b, step sessions on. Eight steps, four refused, stopped at item one of six reporting "nothing was written" — for a step whose two commands had both exited 0 with the project on disk. Four faults, none of them the model being small. The agent is told what machine it is on. Three of the four refusals were the model reaching for shell syntax it had no way to know was unavailable: `mkdir -p`, then `cd x && npm install`. It was guessing at the platform and at the execution model, and the extension knew the first before the first token was generated. core/environmentProfile detects the OS, release, arch, and Node version and states them in the system prompt, along with the sentence that matters more — there is no shell on any platform, cwd is how you run elsewhere, and create_folder/list_files/ read_file/write_file replace the shell utilities. Tool-using routes only; a Plan prompt gets the facts without the tool names it structurally does not have. .hirayacoder/ is added to the workspace .gitignore at startup. It holds transcripts, memory, and the audit log, and has never been ignored anywhere. An existing file is appended to, never rewritten, and left alone if any line already covers it — including a !.hirayacoder negation. npm create/init/exec is no longer refused for the absence of the project it creates. NO_PACKAGE_JSON is right about `npm install`, which climbs out of the workspace, and exactly wrong about the scaffolders, where the missing manifest is the point. This was the first failure of the run and everything after it was the model trying to satisfy a precondition that was never real. `cd folder && command` is rewritten into the two arguments run_script already takes. 0.6.0 added cwd, documented it, and made the refusal name it; the model sent the chained form anyway, was told, and sent it again. Exactly one shape is accepted and nothing is relaxed — the rewrite goes through pre-flight, the gate, the allow-list, and pathGuard, and the model is told what actually ran. A step whose work was a command is no longer scored as having done nothing. ChangeSet.recordCommand appends to commands, ChangeSet.since reads files, and the two never met, so a scaffold step could not pass. Commands are now revision-stamped, commandsSince is the counterpart to since, and stepGuard accepts a command that succeeded as evidence — still judged from what the extension recorded, never from the model's account.
Rendered from self-contained HTML in docs/images/src — no fonts, scripts, or images are fetched and the app icon is inlined, so any Chromium produces the same output. The HTML is the source of truth; regenerating on a version bump is a one-line edit rather than a redesign, and the how and the what-to-update are in the sources' own README. Excluded from the .vsix. The listing pulls them by absolute raw.githubusercontent URL exactly as it already does the icon, so there is no reason for a megabyte of PNG to ride along in every install.
npm audit --omit=dev is clean — the extension ships no runtime dependencies. The three findings are an allow-listed npx that reaches the registry without ever asking (pre-existing, and the one that matters), two new filesystem writes that skip pathGuard's symlink check, and dev-only advisories in the mocha tree. Nothing blocks 0.6.1. The eslint security warnings were each checked rather than waved through; the new unchainCd regex was measured at 1ms on 50,000 characters of adversarial input.
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.
This pull request introduces several important improvements and fixes to HirayaCoder, focusing on better platform awareness, improved handling of project scaffolding commands, and more accurate step completion tracking. The changes enhance agent reliability, especially when running commands that scaffold new projects, and ensure that the extension correctly records and reports work done by commands, not just file writes.
Platform awareness and prompt improvements:
core/environmentProfile, and includes this information in the system prompt and session records. This helps the model avoid guessing the platform or shell environment and provides accurate context for tool-using routes. (app/agent/agentSession.js,CHANGELOG.md,README.md) (F52e7a49R476, F52e7a49R548, F52e7a49R1002, F5e9ba66R5, F6f9cc61R4, F6f9cc61R133)Project scaffolding command handling:
npm create,yarn create,pnpm create, and their equivalents) are now exempt from the package manifest existence check inscriptPreflight. This prevents false refusals when scaffolding a new project, fixing a major source of failure in previous runs. (app/agent/scriptPreflight.js) (Fa33e1deR35, Fa33e1deR186)Step completion and audit improvements:
ChangeSet, and the step guard (stepGuard.verify) recognizes successful commands as evidence that work was done, even if no files were written directly. This ensures that steps like project scaffolding are properly credited and not incorrectly marked as "nothing was written." (app/agent/agentSession.js,app/agent/stepGuard.js) (F52e7a49R71, F52e7a49R119, F52e7a49R887, F761fe82R46, F761fe82R148)User experience and documentation:
.hirayacoder/is now automatically added to.gitignoreon session start, preventing users from accidentally committing session logs and transcripts. (CHANGELOG.md) (F5e9ba66R5)README.md,.vscodeignore) (F6f9cc61R4, F6f9cc61R133, F7bc3fc3R67)These changes collectively make HirayaCoder more robust, transparent, and user-friendly, especially when handling cross-platform scenarios and new project scaffolding workflows.