Skip to content

0.3.0 - the calls an agent gets wrong on the first try

Choose a tag to compare

@lidge-jun lidge-jun released this 15 Sep 05:15
· 117 commits to main since this release

Six friction points an agent hit while actually using 0.2.0, plus issue #24, and a
package that can now be published without redistributing somebody else's binary in
silence. Every change below was measured against the installed build, not inferred.

Breaking

  • browse.readText no longer returns markdown. The body is text, and format
    says what it is: markdown when the fetch path converted the html, text when the
    browser returned its rendered body. Two body fields doubled the JSON for a 10k page
    and let the output shrinker drop the copy you were reading. A 0.2.x cache entry is
    renamed on read, so nothing has to be refetched.
  • A search that steps over a symlinked directory now reports complete: false and
    scope.skippedSymlinks (issue #24). A directory of 37 entries where 35 were links
    used to answer with 2 rows and complete: true, and no option revealed the
    difference. A skipped file link is counted without lowering completeness: it
    cannot hide a subtree, and three symlinked bin stubs used to make the signal useless.
  • The npm tarball no longer ships the maintainer scripts. probe-g3 and
    probe-native-helper default to a live ~/.aside/u/0 and write into it;
    release-records reads evidence/ from a git checkout. Nothing under src/ or
    bin/ imported any of them. The installer, the three register entry points and
    run-tests remain.

Guest sandbox

  • A dynamic import inside --code is refused with EGUESTIMPORT and the list of
    injected globals, instead of Node's ERR_VM_DYNAMIC_IMPORT_CALLBACK_MISSING. The
    list is built from the injection table, not hardcoded. No vm module loader was
    opened: importModuleDynamically is ignored without an experimental flag, and the
    main-context loader is a sandbox escape.
  • The refusal names the codes it translates by Node error code and name, never by
    prose, so a user error carrying the same words is not intercepted. A guest that
    catches its own rejection still sees Node's text; that limit is written down.

Call shapes

  • search.content and search.count refuse pattern with the name that belongs to
    the method you called, and search.files refuses query the same way. When the
    value itself looks like a glob the refusal offers both, because either could be what
    you meant. The existing valid: list is unchanged.
  • browse.readText accepts a URL string or { url }. Both shapes work.
  • browse.exec and browse.attach accept treeNodes: true, which returns the
    accessibility tree parsed into snapshot.nodes as { depth, role, name, ref, attrs, line } rows. Hierarchical data groups by depth instead of by a regex state machine
    over the string tree. Off by default: shipping it on every snapshot dropped a
    280-row page out of the output envelope, and the parser itself is only injected when
    asked for, because one Windows job had 571 characters of command-line headroom.

Browsing is still opt-in, and says so once

  • codemode --enable-browse turns it on. It writes browseCaps.enabled into the user
    config and touches nothing else, no-ops when already on, refuses invalid JSON with
    EBADCONFIG, and runs before the config is loaded so a broken file can still be
    repaired. All four refusal sites quote the same command, and a test keeps them equal.

Guidance an installed account receives

  • New references/call-shapes.md collects the calls that cost a first attempt, and the
    installer owns it, so upgrade, doctor, uninstall and rollback all know about it.
  • The AGENTS block now names the guest sandbox inside its 35-line budget: no import,
    require, process, fetch, setTimeout or Buffer; the twelve injected globals
    including console; and that browse is injected but refuses until enabled.

Installer

  • uninstall no longer leaves its own empty references/ behind. rmSync without
    recursive throws EISDIR on a directory and the catch swallowed it.
  • rollback now removes files the generation it restores never had, and reports them
    as dropped. A file the user edited is kept. Restoring only the snapshot left newer
    files on disk under a manifest that did not know them, and doctor called our own file
    "new".

Packaging

  • bin/THIRD-PARTY-NOTICES.md carries ripgrep's MIT text and, because the Windows
    build links PCRE2, its BSD notice, and both ship inside the package. A link is not a
    notice, and the repository LICENSE covers this project's code only.
  • On Windows the packaged bin/rg.exe is now the resolver's last candidate. rgPath
    is written by register, and an npm install has no register step, so the one copy of
    ripgrep we shipped was the one PATH could not find.
  • package.json gained repository, homepage, bugs, keywords and author, and
    npm test from an installed package explains itself instead of failing on a test/
    directory the consumer was never given.
  • MCP serverInfo.version reads the package instead of the frozen string 0.1.0.

npm

Publishable now. The name is free, a packed tarball installs globally and answers
--doctor, and plannedFiles() / install --dry-run resolve templates from outside
the repository. 77 entries, 1.95MB. This release does not publish — the judgement
and the preparation are what shipped. Reasoning is in
devlog/_plan/260915_agent-friction/070_w8_npm.md.

Not run

The G5 usability gate is still unrun and is not claimed anywhere. Captcha/vault,
downscaled-image coordinate mapping, DPI/zoom/clip and native mouse/keyboard
equivalence remain out of scope.

Full suite: 722 tests, 721 pass, 1 skip, on Ubuntu 18/20/22, macOS 22 and Windows 22.