macOS ships bash 3.2, and the port used three things it does not have:
mapfile, fractional `read -t`, and GNU `stat -c`. The first two took the
interactive menu down - `mapfile: command not found` in the preset
picker, and an arrow key that parsed as a bare Escape, leaving `[C` in
the tty for zsh to report as a bad pattern.
read_key now reads the escape tail through the terminal itself: icanon
off with min 0 / time 1, so a plain read returns the moment a byte
arrives and gives up after ~0.1s. bash 3.2's `read -t 0` cannot be used
to poll for this - it reports nothing even with bytes buffered (checked
against 3.2.57). CSI sequences are consumed whole, so modified keys like
Ctrl+Right no longer leak their tail either. Verified interactively
under bash 3.2.57 in a pty: arrows, pgup/pgdn, home/end, enter,
backspace, and a bare Escape all parse, with no stray bytes in the tty.
Session listing reads /proc, which macOS does not have; it now says so
rather than silently reporting no running sessions - the answer that
gets people to switch out from under a live session. install.sh swaps
GNU `find -print -quit` for a glob, and the README notes both.
Co-Authored-By: Claude Code <noreply@anthropic.com>
apiKeyHelper is a shell command line, not a path, so the raw value written
into settings.json was split at the first space. A Windows profile named
"Mohammed Ahmed" produced an attempt to run C:\Users\Mohammed, surfacing as
"your apiKeyHelper script is failing" with nothing to go on. Quote the value
when it contains anything a shell cares about, and leave it bare otherwise so
no existing settings.json churns on the next switch. The POSIX port had the
same bug against a /Users/First Last home; shlex.quote has exactly the wanted
"leave ordinary paths alone" behaviour.
doctor could not see any of this. It quoted the path itself before running it,
so it exercised a command line Claude Code never uses and passed while the
real one failed. It now reads the string out of settings.json, reports it when
it is not what a switch would write, and runs that string through a shell.
The helper itself exited 1 in silence on four distinct faults - no state, no
preset, no key, undecryptable key - collapsing them into one indistinguishable
message. Each now names itself on stderr, which is what /status displays. The
DPAPI case says what it actually means: a key stored by a different Windows
account than the one Claude Code runs as. Success paths stay silent, so stdout
still carries the key and nothing else.
Also make install.ps1 survive a Restricted execution policy: piped through
iex it is not subject to the policy, but invoking the installed script for the
key prompt is, which is where a fresh install died. Set Process scope for the
install, offer to set CurrentUser to RemoteSigned, and clear the
mark-of-the-web that Expand-Archive can leave on the extracted scripts.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Nothing tells you a session is unresumable until you try to resume it, and by
then you have usually forgotten which one it was. The widget now scans every
project on a timer and whenever the panel opens, marks its icon when something
needs fixing, and lists the affected sessions with a repair button that says
what it will drop and what it will keep before doing anything.
The scan had to get roughly eighty times cheaper first. Classifying a transcript
needs two facts - what its last message id is, and whether any Anthropic id
exists at all - and the first settles the common case alone. Reading the tail of
each file and only opening the whole thing when the tail already looks wrong
takes the sweep from ~5s to ~60ms across 59 transcripts, which is the difference
between something that can sit on a timer and something that cannot. `--all`
uses the same path, and `--json` exposes it.
The badge is a dot beside the mark rather than a recolouring of it: this widget's
job is to report which provider is active, and tinting it red to mean something
else entirely would be a lie about that.
Verified by planting a genuinely corrupted transcript, watching the scan find it
and the dot appear, then removing it and watching both clear.
repair-session only ever listed the project you were standing in, which is a
poor place to start from when the thing you cannot resume is a session whose
project you no longer remember. --all drops the working-directory scoping and
checks all of them. (A named session id was already looked up across every
project; only the listing was scoped.)
The first version of that scan reported 16 of 59 transcripts as unresumable,
which was true in the narrowest sense and useless in every other. Checking them:
5 had never received an assistant reply at all, and 10 had run start to finish
on a gateway, so every id in them is that provider's by design. Those resume
perfectly well under the provider they were born on, have nothing to truncate
back to, and are only a problem if you try to resume them as Anthropic. Calling
either of them damage buries the one case that is.
So the verdict now carries a kind - healthy, repairable, gateway-native,
synthetic-only, no-messages - and only `repairable` is reported or acted on: a
transcript with a genuine msg_ message and junk after it. Across the same 59 it
now correctly reports nothing to repair, while still classifying the real
corrupted transcript from the incident as repairable with 108 lines to drop.
A gateway-native session asked about directly now says what it is and that there
is nothing to fix, rather than failing with "no Anthropic-issued message".
Truncating the transcript is the mechanical fix, but the turns being cut are the
work itself. Losing the conversation that produced a morning's changes is most
of the damage, and a session that resumes with a hole in its memory is barely
resumed - it has no idea what it just did or what it was asked.
So --apply now writes the dropped turns out as <session>.recovered-<stamp>.md,
a readable record of what was asked, what was answered and what was run, and
appends the same text back to the truncated transcript as a single note. Tool
results are deliberately excluded: they are most of a transcript by volume and
the least useful part of a summary.
The note is a user entry marked isMeta - the marker Claude Code already uses for
its own local-command caveats, meaning context rather than something to answer -
and carries no message.id, so it cannot recreate the previous_message_id
condition being repaired. Verified on the real corrupted transcript: 1921 lines
in, 1813 kept plus the note, chained to the last good assistant uuid, no id on
the message, and the file reads back as healthy. The generated digest recovers
the actual instruction that was lost in the incident, which is the thing that
made this worth doing.
--no-reinject writes the markdown but leaves the session alone.
Acts on docs/incident-mode-switch-corrupts-live-sessions.md, which is added here
as the record of why.
The report identifies a consequence that was not modelled. A failed call is
recoverable; a *successful* one may not be. If a running session takes even one
completion from the provider being switched to - which happens when that mode
matches the base URL it already had cached - that provider's message-id format
lands in its transcript. OpenRouter issues `gen-<epoch>-<rand>` where Anthropic
issues `msg_...`, and native Anthropic then refuses to resume the session at
all, with a 400 naming previous_message_id. The only way back is to truncate the
transcript, losing every turn after the cut. That happened here, and was fixed
by hand.
Two changes follow.
Sessions are now settled before the write, not reported after it. A switch with
anything running stops, names the sessions, explains what is about to happen to
them, and offers restart (the only answer that ends with everything on the mode
the bar now claims), close, proceed anyway, or abort - defaulting to abort.
Non-interactively it refuses outright unless given --yes. The old after-the-fact
reporter is deleted rather than left as a second, contradictory account.
`claude-mode repair-session` replaces the hand surgery: it finds a project's
transcripts, reports which are resumable, and on --apply backs the file up and
truncates to the last Anthropic-issued message. Verified against the real
corrupted transcript from the incident - it reproduces the manual cut exactly,
1921 lines to 1813, dropping the two `gen-` completions and the error
placeholders after them, leaving a transcript that ends on a genuine msg_ id. It
refuses a transcript written to in the last 90 seconds, since that one belongs
to a session still running.
The panel passes --yes, having already asked in its own card, and that card now
names the transcript risk rather than only the inconvenient one.
Requirement 4 of the report - documenting the mechanism - landed in 9c301e1;
the README now carries the unrecoverable half as well.
The openrouter default moves its hot tiers: opus to z-ai/glm-5.3-flash and
sonnet to deepseek/deepseek-v4-flash-0731. haiku and fable are unchanged.
`cheap` and `lmstudio-qwen` are gone, leaving exactly one preset per mode so
`claude-mode <mode>` is never ambiguous and there is no menu to read before the
thing you asked for happens. The surviving lmstudio preset keeps the Qwen3.6
model rather than KAT-Coder: the two differed mainly in that KAT's chat template
carries the message-order assertion this README already warns about, so between
two presets that had to become one, the one that is known to work won.
More presets are still a `preset new` away; the shipped set is a starting point,
not a ceiling.
Which is the other half of this. A shipped preset was never a working
configuration - OpenRouter and Z.AI have no key stored, and lmstudio's model ids
were whatever happened to be installed on the machine this was packaged on. That
was left for the user to discover through a failure. Now the shipped presets
carry `configured: false`, preflight blocks on it, and `claude-mode setup <mode>`
walks through what is actually needed: key, server URL and auth for LM Studio,
then models chosen from the provider's own catalogue rather than typed from
memory. A switch that trips this in a terminal offers to run setup there and
then instead of printing a command to type next.
Absent means configured, deliberately: presets that predate this and any built
by hand with `preset new` do not suddenly start demanding a wizard.
The panel gets a "Set up <mode>…" button that hands the whole flow to a terminal,
since a bar popup can host neither a hidden key prompt nor a filter-select list.
Two bugs found while testing it, both real:
ask_value printed its prompt to stdout while being called inside $( ), so the
prompt text came back glued to the front of the answer and set-url rejected the
result. Moved to stderr, which is why warn and err already go there.
lms_catalogue never sent the API key. On a server with authentication switched
on - the case just added support for - /api/v0/models answers 401 like anything
else, so the catalogue came back empty and every caller silently concluded the
server had no models installed. It now sends the preset's credential, as do the
three other call sites that read it.
Every one of these said a switch leaves running sessions alone - that they keep
talking to the old provider until restarted. That is wrong, and reassuring in
exactly the wrong direction: a switch breaks them.
The static half of the config does behave as described. Base URL, model ids and
the env block are read once at startup and a running session keeps what it
started with.
The credential is not. It comes from running apiKeyHelper, which Claude Code
re-invokes on a timer - CLAUDE_CODE_API_KEY_HELPER_TTL_MS is present in the
2.1.251 binary - and claude-key-helper.sh answers for whatever state.json says at
that moment. So the switch reaches into a live session through the one part that
was never cached. Into anthropic, the helper returns nothing by design and the
next refresh comes back with no credential; into another provider, it hands over
the new key while the session is still pointed at the old base URL, which
rejects it. Either way the session starts failing calls whenever the TTL happens
to expire, mid-turn as easily as between turns.
One case does survive, and is now stated rather than glossed: switching between
two presets of the same provider that share a keyRef keeps the same key and the
same endpoint, so the session carries on with the model ids it started with.
The claim in "Restarting sessions" predates this work - it came in with the
Windows build - but it is wrong for the same reason and is corrected too.
GLM 5.3 supersedes 5.2, so every reference to the older id follows it: the zai
preset's opus, sonnet and fable tiers, and the fable tier of `default`, which
routes to the same model through OpenRouter.
Also the hardcoded Z.AI entries in both builds' model pickers, which are the one
place a model id is not read from a preset - leaving those would have gone on
offering a superseded model from the menu while the presets used the new one.
haiku stays on glm-4.7: that is the fast/cheap tier, not a 5.x.
Four changes to the POSIX build, found while getting it working on Omarchy.
Colour follows the desktop theme. The sixteen ANSI slots carry no guarantee
about relative brightness and monochrome themes exploit that: under Omarchy's
Solitude, slot 36 (headings) resolves to #707070 and slot 31 (FAIL) to #565d60,
which against #cacccc body text on a #101315 ground is 3.8:1 and 2.8:1 where the
body text is 11.6:1. Headings rendered as fine print and errors as the quietest
thing on screen. The palette is now derived from the theme's own colors.toml,
with each role measured against the background it will actually be drawn on and
lifted toward the foreground when it falls short - hue kept where the theme has
any, weight substituted where it does not. Headings go 3.8:1 -> 9.4:1 and FAIL
2.8:1 -> 5.2:1. Falls back to the ANSI slots off Omarchy, with the two roles the
slots get wrong corrected.
health.json was only ever written by a switch, so a fresh install had none at
all and any reader had to guess. It is now refreshed by `status` and seeded at
install, and `claude-mode health` forces it. The installer also copies VERSION,
which cm_version() has always read and nothing ever wrote - every health.json
until now reported 0.0.0.
Preflight, because a switch that cannot work does not fail loudly: it succeeds,
and every session started afterwards breaks in a way that points at Claude Code
rather than at here. Keys, the helper, the preset's provider and the server are
all checked before the write. LM Studio is the sharp case - its token is an
inline placeholder, so nothing about the switch needs the server to exist.
Session control, because Claude Code reads settings.json once at startup: a
switch leaves running sessions on the old provider until they are restarted, and
one mid-request can lose that turn outright. `sessions` lists them, `--stop` and
`--restart` act on them behind a confirmation, `--dry-run` shows the plan.
Sessions are found through /proc/<pid>/exe rather than by process name, which
would sweep up every shell that merely mentions claude - including the one this
runs from. Two exclusions: the calling session, and forks of a session. A busy
session spawns children off its own binary that inherit the same exe, and
without filtering those the count climbed and fell with load - it read 2, 5, 11
and 40 for the same two sessions before the parent check went in.
LM Studio is no longer assumed to be on this machine. `preset url` and
`preset auth` move it to a LAN box, a tunnel or a proxy and turn authentication
on, and the probe distinguishes ok / auth / notfound / refused, because "start
the server" and "your key is wrong" are opposite remedies. It is probed wherever
it lives - a sleeping LAN box is exactly as absent as an empty loopback port -
while remote gateways are not, since those being briefly unreachable is the
network's problem and a missing key never fixes itself.
Source of truth so far has been c:\Users\smoido\projects\cli on the Windows
box, which has no git history of its own. This is that tree copied verbatim over
SSH, minus dist/ - the PowerShell build, the POSIX port under linux/, and the
presets both share.
Recorded as its own commit so that everything after it is a reviewable diff
rather than an undifferentiated first drop.
