Repository navigation
0.2.0 - the loader an account can actually read
What changed
The loader line an account is given now works on the surface that reads it. The installed
skill and the managed AGENTS block used to say fs.readFile('../../codemode/cm.js', ...).
That resolves from the session directory under aside repl and does work there, but the
in-app agent REPL resolves a relative read from the account root, where the same line leaves
the account root and the fs guard refuses it with Path escapes Project and session roots.
The installer now fills each account's own absolute path, quoted, so a home directory with an
apostrophe or a Windows path with backslashes still produces a line that parses. Helper 1.1.0.
How far that is verified, precisely: on three machines the line was pulled out of the
installed skill and run in a live CLI aside repl session, loading cm 1.1.0 each time
(MacBook u/0, Windows MINI u/0, macmini-cf u/1). The in-app agent REPL is the surface that
was failing, and the evidence there is the user's confirmation on 2026-09-15 that the absolute
path loads and cm.run attaches. We have no way to drive that surface automatically, and this
release does not claim we did. MacBook u/1 and u/2 report skipped for the same reason a
foreign account's file is refused: the CLI session is guarded to the account it is signed in as.
The installer can be rolled back after a release that changes bytes. snapshotForRollback
ran after the write, so previous held the bytes that had just been written. It now snapshots
before writing, hashes what it actually holds, and records disagreedWithManifest when the
manifest claimed something else. A rollback hashes what it restored rather than repeating the
snapshot's claim. doctor reports stale for a machine a release left behind instead of
calling it healthy, and repair no longer records hashes for files it did not write - which
used to make the next upgrade treat a stale file as a user edit and preserve it forever.
The G3 fixtures this surface can run were judged against live DOM. scripts/probe-g3.mjs
runs four of them on a real browser: a ref inside an iframe (the click lands in the child and
the identically labelled parent stays untouched, with a positive control proving the parent
button does respond), a native action with a batch in the middle (the tab keeps its result and
is still drivable afterwards, each item under its own jobId), an image handed to display()
(read back as a real PNG - IHDR 2880x1800 on macOS, 1440x900 on Windows), and two calls racing
on one page (the overlap is verified by timestamps, and both fields end up carrying the same
writer). 14 checks, passing on macOS and Windows.
Deployed before this tag
Three machines, five account roots, all on helper f5584a8081bdd41f…: MacBook u/0-u/2,
Windows MINI u/0, macmini-cf u/1. The macmini has seven accounts and its current account is
u/1, so the installer following currentAccountId rather than assuming u/0 was tested for
real.
What this release does not claim
G5, the usability and path-selection gate, was not run and is not counted as passed. There
is no session where a model chose between native and batch across a mixed set of requests and
was scored on keeping the answer right without an unnecessary detour. The 480-run cold/warm
campaign published earlier is performance and correctness data on paths a script fixed in
advance, and the native-first wording in the installed skill is documentation, not evidence of
behaviour. On 2026-09-15 the user directed the deployment and this release, which takes G5 off
the bar for promotion - a change of release criteria, not a passed gate. R6 is therefore not
claimed as closed, since its completion evidence is a G0-G5 receipt.
Also not measured: a downscaled image mapping coordinates back, DPI 100/125/150/200%, zoom and
clip (no API on this surface produces the stimulus), whether a model actually received an image
handed to display() (it returns undefined), native mouse/keyboard equivalence on an
explicitly chosen page, captcha solving and vault autofill. Windows viewport screenshots need
the tab brought to front and still time out occasionally at the CDP layer.
Records
evidence/release-260915/ holds the release manifest, the capability receipt with what each
surface was seen to do and what was not checked, operating notes, and a recovery document -
including the two accounts (MacBook u/0, MINI u/0) whose rollback target is not what its entry
declares, because a probe had replaced the helper on disk before the snapshot was taken.
The copies inside this tag were generated just before the version commit, so they carry that
commit rather than the tag. The corrected records, stamped with v0.2.0 and 0cfbc86, are on
dev.