Skip to content

fix(progress): tell agents how to load the answer tool and which fields it takes; bump to 1.14.5 - #135

Merged
stevecastaneda merged 1 commit into
mainfrom
diagnose-progress-plugin-failures
Nov 8, 2026
Merged

stevecastaneda merged 1 commit into
mainfrom
diagnose-progress-plugin-failures

Conversation

@stevecastaneda

Copy link
Copy Markdown
Owner

Why

In Claude Code, recording a question's answer often failed on the first call. Claude Code hides a tool's inputs until the agent loads it. Agents loaded the setup tools at the start and left out progress_question_answer, or lost it when a long chat was compacted. They then guessed the field names (ref, answer, option) and got a validation error. The retry about 3 seconds later worked, so no answers were lost. Four recent sessions show this pattern.

What changed

  • The skill gives one ToolSearch query that loads all 11 Progress tools. It tells agents to run it again after each compaction. It also maps question answer Q7 A to the tool's id and choice.
  • After progress_question_ask succeeds, its reply names progress_question_answer and its fields.
  • Two tests. One fails when the skill's query and the tool list drift apart. The other fails when the note names a field the answer tool doesn't have.

Scope

  • Claude Code agents. Codex agents see the tools without loading them.
  • The tool still rejects guessed field names. Accepting aliases would hide a wrong call, not fix it.

Verification

  • npm test in progress: 232 pass. npm run typecheck, oxlint and oxfmt --check are clean.
  • npm run check:install installs with npm 12.2.2.

Security review

No release blockers found.

Full report

Security review: @stevecastaneda/paseo-progress@1.14.5

  • Reviewer: claude --model opus
  • Commit: 409ef96
  • Tree: 1f18603e590cea7b49b943a10f605e56e69a492d
  • Diff: f54bcdd..409ef96 -- progress/
  • Date: 2026-11-08T02:37:04.296Z
  • Minutes: 1.3
  • Key: 5019d0baa02881f7cb6737d0513f535ae5602bd0546a87743b06a9e6c990cb46

I found no release blockers in 1.14.5. I reasoned everything below from the source at 409ef96. I had no shell in this session, so I ran no tests and nothing was executed.

What changed, and what it can reach

  • server/tool.ts:61-72, withNote and ANSWER_NOTE. This wraps progress_question_ask and adds a fixed sentence to the end of a successful reply. Failed replies come back unchanged. The added text is a constant the plugin writes; nothing from the repo or the agent goes into it. The wrapper keeps the tool's name, input check and description, and calls the same run() → runCli() path with the same fixed "question ask" command. It adds no file, program, HTML or network access. The chat row (client/chat-row.ts:81-93) shows the reply as plain text, so the note appears as the plugin's own wording.
  • skills/paseo-progress/SKILL.md:10-18. These are new instructions telling the agent to load the 11 Progress tools with one ToolSearch call. That only loads the plugin's own tools. It doesn't tell the agent to run commands, open files or install anything. The skill is shown in the panel (server/skill.ts:131-133) from the plugin's own folder, never from the repo.
  • package.json and package-lock.json. Version 1.14.5, and zod 4.6.7 as a dev dependency, pinned by an integrity hash. The repo being reviewed has no say in it.
  • server/tool.test.ts. Left out of the package by the !**/*.test.ts rule in files.

Every place that touches files, programs, HTML or the network

  • server/system.ts:24-31, runProgram and runCommandLine: programs start in Node's own folder, never the worktree, so a program committed to the repo can't be picked up instead. Safe.
  • server/system.ts:37-113, the file-writing helpers: each write names a trusted folder plus a relative name, rejects absolute paths and .., and refuses to write through any link along the way. Safe.
  • server/store.ts:20-28, git: runs git -C <root> from Node's folder, so a git.exe in the repo can't run. Safe.
  • server/store.ts:37-51, the working folder: always under ~/.local/state/paseo-progress/<name>-<hash>, never inside the repo. Safe.
  • server/store.ts:77-98, the history folder: read from the repo's local git config, which the attacker doesn't control. It must be inside the worktree and not under .git (checked in any letter case). Safe.
  • server/store.ts:134-151, saveRunCopy: folders are made one step at a time, and it stops if any step is a link or a file. The write goes to a temporary file that is renamed into place, so a link sitting at the target gets replaced, not written through. Safe.
  • server/store.ts:101-113, setHistoryFolder: runs only when the user clicks in the panel. It writes to the local git config through git config --local, not by editing files directly. Safe.
  • server/store.ts:155-164, listFolders: read-only. It checks the folder and leaves out .git. Safe.
  • server/progress-file.ts:59-87 and server/lock.ts: append to the progress file, write the panel-opened marker, and create and remove the lock. All of it happens in the plugin's own state folder through the system.ts helpers, and every event is checked against a schema before it's written. Safe.
  • server/cli.ts:510-523, rootRelative and guessKind: they only check whether paths exist (stat); nothing is opened or sent. Paths outside the worktree are stored as absolute paths, and the open and preview code below refuses them. Safe.
  • server/cli.ts:574-575, install-launcher from the command line: it writes wherever it's told. Agents can't reach it, because server/tool.ts:207-213 refuses it as the first argument. The other tools always pass fixed command words, and runCli only matches argv[0]. Safe.
  • server/launcher.ts:128-173, launcher install: runs only from the panel's Install button. It writes a fixed script to ~/.local/bin and won't overwrite a file the plugin didn't write. Safe.
  • server/skill.ts:24-143, skill links: they point into the plugin's own folder under the home folder. They're re-pointed automatically only if the plugin made them; new links wait for the user's click. Paseo's command comes from the user's environment or PATH and runs from Node's folder. Safe.
  • server/open.ts:67-81, resolveAttachment: looks up an attachment by its reference in the dashboard, then resolves real paths and refuses anything outside the real worktree. Safe.
  • server/open.ts:85-91, openDeliverable: runs on the user's click. It checks the real file is a regular file with an extension that only displays, never runs. HTML, SVG and scripts are refused, then it's opened through explorer, xdg-open or open started from Node's folder. Safe.
  • server/open.ts:117-178, preview and read-in-pieces: read-only, inside the worktree, with size limits. A preview never serves HTML. Images are sent as data URIs. Safe.
  • server/open.ts:38-47, defaultBrowser: reads the user's own Launch Services file with plutil. Safe.
  • server/whats-new.ts and server/question-mode.ts: read the plugin's own README and package.json, and write fixed files in the state folder through writeFileIn. Safe.
  • client/attachments.tsx:66-75: http(s) links go to Paseo's browser or openExternalUrl on the user's click. The command only accepts http(s) links. Safe.
  • client/attachments.tsx:96-110, PageLink: an HTML deliverable opens in Paseo's sandboxed preview, which blocks network access. The page is never handed to the host's default app (opensOnHost and canPreview both refuse HTML). Safe.
  • client/chat-row.ts: shows tool replies as plain text, with no HTML. Safe.

No release blockers found.

…ds it takes; bump to 1.14.5

In Claude Code an agent that hadn't loaded progress_question_answer, or lost it to compaction, guessed its fields and failed once before retrying. The skill now loads all Progress tools in one search and again after compaction, and the question_ask tool's reply names the answer tool's fields.
@stevecastaneda
stevecastaneda merged commit fe1ffac into main Nov 8, 2026
1 check passed
@stevecastaneda
stevecastaneda deleted the diagnose-progress-plugin-failures branch November 8, 2026 02:38
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant