claude-code-sessions 0.16.5__tar.gz → 0.17.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/LICENSE +21 -21
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/PKG-INFO +185 -12
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/README.md +184 -11
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/claude_code_sessions.egg-info/PKG-INFO +185 -12
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/claude_code_sessions.egg-info/SOURCES.txt +3 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/claude_code_sessions.py +3125 -22
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/claude_code_sessions_gui.py +257 -85
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/pyproject.toml +50 -50
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_alignment.py +144 -144
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_cli.py +172 -172
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_containment.py +71 -71
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_discovery.py +48 -48
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_encoding.py +41 -41
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_engine.py +274 -274
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_golden.py +18 -18
- claude_code_sessions-0.17.0/tests/test_gui_dpi.py +171 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_gui_models.py +153 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_helpers.py +39 -39
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_journal.py +87 -87
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_journal_reroot.py +8 -0
- claude_code_sessions-0.17.0/tests/test_realign.py +1826 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_rows.py +73 -73
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_scaffold.py +18 -18
- claude_code_sessions-0.17.0/tests/test_split_rows.py +654 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_transcripts.py +24 -24
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/claude_code_sessions.egg-info/dependency_links.txt +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/claude_code_sessions.egg-info/entry_points.txt +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/claude_code_sessions.egg-info/top_level.txt +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/setup.cfg +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_anonymize.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_commands.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_converge.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_data_root.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_duplicate_transcripts.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_gui_destination.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_gui_env_rebuild.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_plan_move.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_recover.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_retitle.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_sync.py +0 -0
- {claude_code_sessions-0.16.5 → claude_code_sessions-0.17.0}/tests/test_undo.py +0 -0
|
@@ -1,21 +1,21 @@
|
|
|
1
|
-
MIT License
|
|
2
|
-
|
|
3
|
-
Copyright (c) 2026 Craig Stoller
|
|
4
|
-
|
|
5
|
-
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
-
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
-
in the Software without restriction, including without limitation the rights
|
|
8
|
-
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
-
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
-
furnished to do so, subject to the following conditions:
|
|
11
|
-
|
|
12
|
-
The above copyright notice and this permission notice shall be included in all
|
|
13
|
-
copies or substantial portions of the Software.
|
|
14
|
-
|
|
15
|
-
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
-
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
-
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
-
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
-
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
-
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
-
SOFTWARE.
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Craig Stoller
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: claude-code-sessions
|
|
3
|
-
Version: 0.
|
|
3
|
+
Version: 0.17.0
|
|
4
4
|
Summary: Unofficial tool for Claude Code sessions on disk: make a session visible under your other Claude account on the same machine, or move one between project folders. Command line or GUI. Journaled, reversible, fails closed.
|
|
5
5
|
Author: Craig Stoller
|
|
6
6
|
License: MIT
|
|
@@ -157,7 +157,7 @@ read. See `docs/internals.md`, RULING 7.
|
|
|
157
157
|
|
|
158
158
|
## Before any move: close the Claude app.
|
|
159
159
|
|
|
160
|
-
`move`, `undo`, `recover`, `repoint --apply`, `new-row --apply`, and every mutating `sync`
|
|
160
|
+
`move`, `undo`, `recover`, `repoint --apply`, `new-row --apply`, `realign --apply`, and every mutating `sync`
|
|
161
161
|
route (`--apply`, `undo` of a completed sync, `recover --back` on a stuck one) all refuse
|
|
162
162
|
while they can see the Claude desktop app running - whoever the identity files currently say
|
|
163
163
|
is signed in. That includes `undo` of a completed `repoint` or `new-row`, and `recover` on an
|
|
@@ -183,7 +183,7 @@ itself, and the guard counts it - if a refusal names that process, fully exit th
|
|
|
183
183
|
|
|
184
184
|
## Usage
|
|
185
185
|
|
|
186
|
-
|
|
186
|
+
Thirteen commands. All mutating commands default to a dry run; add `--apply` to execute.
|
|
187
187
|
(`trust-signed-helper` is the opt-in described above; it changes a setting, not your
|
|
188
188
|
sessions.)
|
|
189
189
|
|
|
@@ -204,8 +204,9 @@ claude-code-sessions list gate
|
|
|
204
204
|
```
|
|
205
205
|
|
|
206
206
|
**`doctor`** - read-only health report (stale locks, unresolved operations, orphaned rows,
|
|
207
|
-
encoding-scheme ambiguity, a synthesized row that has since vanished,
|
|
208
|
-
account's sidebar points at**
|
|
207
|
+
encoding-scheme ambiguity, a synthesized row that has since vanished, **conversations no
|
|
208
|
+
account's sidebar points at**, and **conversations whose accounts open different copies of one
|
|
209
|
+
transcript**):
|
|
209
210
|
|
|
210
211
|
```
|
|
211
212
|
claude-code-sessions doctor
|
|
@@ -219,6 +220,30 @@ Resuming a session while signed into a different account repoints its row, which
|
|
|
219
220
|
conversation you were in an hour ago reachable from no sidebar at all. That is what this
|
|
220
221
|
ranking is for, and `repoint` is how you put it back.
|
|
221
222
|
|
|
223
|
+
The copies check covers a case every other number misses. A conversation's transcript can end
|
|
224
|
+
up in two project folders, usually after a session moves into a second worktree. Each account's
|
|
225
|
+
row names a folder through its `cwd`, and the app opens the copy in that folder, even a stale
|
|
226
|
+
one. So one sidebar can show a conversation as it was weeks ago while another shows it today,
|
|
227
|
+
and `alignment` still counts it complete. `doctor` lists each such conversation, which
|
|
228
|
+
copy each account's row opens, and how each pair of copies relates: one behind the other (one
|
|
229
|
+
history, one copy stopped growing), the same conversation, byte-identical, or diverged (each
|
|
230
|
+
holds turns the other lacks). It compares the bytes first and then reads what lies past the
|
|
231
|
+
bytes two copies share, because the app writes bookkeeping records (`tools-used` when it opens
|
|
232
|
+
a copy, `cost-state` when it quits) to whichever copy it has open, the stale one included. A
|
|
233
|
+
copy ahead only by those records holds the same conversation; a stale copy the app has opened
|
|
234
|
+
and quit is still behind, although its bytes no longer match. Any other record type counts as
|
|
235
|
+
conversation. `doctor` exits 1 when an unarchived row opens an older or a diverged copy.
|
|
236
|
+
Identical copies and copies that differ only in bookkeeping are listed and do not change the
|
|
237
|
+
exit code. Neither does a row whose folder holds no copy at all. A row whose `cwd` names a main repository while the transcript sits in
|
|
238
|
+
one of its worktree folders is the app's normal placement, not a fault.
|
|
239
|
+
|
|
240
|
+
Deleting the stale copy is a catch-up, not a fix. On one conversation, measured on Desktop
|
|
241
|
+
5.11492.5, opening a row whose folder held no copy made the app copy the transcript back into
|
|
242
|
+
that folder, so the split returns once the conversation is worked from one account and opened
|
|
243
|
+
from another. For the same reason `doctor` also lists, without counting them, conversations with
|
|
244
|
+
a single copy and a row whose folder holds none: opening that row starts a split. `realign`
|
|
245
|
+
(below) is the repair for both.
|
|
246
|
+
|
|
222
247
|
**`alignment`** - read-only scoreboard: how close your accounts are to holding one shared
|
|
223
248
|
history. `doctor` answers "is anything broken"; this answers "how far apart are my sidebars,
|
|
224
249
|
and which way is it moving":
|
|
@@ -237,8 +262,10 @@ Two of those actively trade against each other: making a conversation reachable
|
|
|
237
262
|
both halves of a disagreeing pair in every sidebar under one name, which is a direct hit to
|
|
238
263
|
*distinguishable*. A single "aligned: yes/no" would hide exactly the decision you have to make.
|
|
239
264
|
|
|
240
|
-
It reads row files and transcript **filenames** only - no transcript content - so it
|
|
241
|
-
on a store with gigabytes behind it.
|
|
265
|
+
It reads row files and transcript **filenames and sizes** only - no transcript content - so it
|
|
266
|
+
stays fast on a store with gigabytes behind it. That is why its line on conversations whose
|
|
267
|
+
accounts open different copies only counts them, and how many differ in length; telling a stale
|
|
268
|
+
copy from a diverged one means reading both, which `doctor` does. **It always exits 0** when it could produce the report. It
|
|
242
269
|
is a scoreboard, not a check: these numbers stay non-zero for as long as the work takes, and a
|
|
243
270
|
command that exits 1 for months trains you to stop reading it.
|
|
244
271
|
|
|
@@ -579,6 +606,90 @@ stopped being redundant is presence it leaves alone. `doctor`'s vanished-row che
|
|
|
579
606
|
converge-created rows the same way it watches `new-row`'s, with the same journal-retention
|
|
580
607
|
window.
|
|
581
608
|
|
|
609
|
+
**`realign`** - point every account's row of a split conversation at the folder holding its
|
|
610
|
+
live transcript copy:
|
|
611
|
+
|
|
612
|
+
```
|
|
613
|
+
claude-code-sessions realign
|
|
614
|
+
claude-code-sessions realign --apply
|
|
615
|
+
```
|
|
616
|
+
|
|
617
|
+
These are the split rows `doctor` reports. A probe on Desktop 5.11492.5, run on one plain and
|
|
618
|
+
one worktree conversation, found that a row's `cwd` alone decides which copy it opens; that the
|
|
619
|
+
app rewrites the row's `folderPlace` to match `cwd`, and never writes `cwd` back; and that a
|
|
620
|
+
turn typed after `cwd` moved runs in the new folder, in the same transcript. So `realign`
|
|
621
|
+
rewrites `cwd`, and `folderPlace.cwd` and `folderPlace.realPath` where the row has a
|
|
622
|
+
`folderPlace` (it never adds one), and leaves every other key alone, including `originCwd` and
|
|
623
|
+
the title. The rewritten rows then open the live copy. No row opens the stale copy any more, so
|
|
624
|
+
it moves into the operation's journal, with its sidecar folder, where `undo` can put it back
|
|
625
|
+
and `doctor` lists it. Without that step, `move`, `repoint`, `new-row` and `converge` would
|
|
626
|
+
keep refusing the conversation for its duplicated transcript. A copy that no row opened even
|
|
627
|
+
before `realign` stays where it is: that is the duplicated-transcript case, which `realign` did
|
|
628
|
+
not cause.
|
|
629
|
+
|
|
630
|
+
By default it plans conversations with a copy behind the live one, and rows whose folder holds
|
|
631
|
+
no copy. Copies that hold the same conversation (byte-identical, or different only in
|
|
632
|
+
bookkeeping records) are realigned only with `--include-identical`, onto the copy the most
|
|
633
|
+
recently active row opens. `--only` narrows to one conversation, resolved the way `converge
|
|
634
|
+
--only` is. A conversation is **held**, with the reason printed, when:
|
|
635
|
+
|
|
636
|
+
- its copies have diverged (each holds turns the other lacks), so which history to keep is
|
|
637
|
+
your decision;
|
|
638
|
+
- a copy no row opens holds turns the live copy lacks;
|
|
639
|
+
- a row to change has a worktree link (`worktreePath` or `worktreeName`). In the probe, moving
|
|
640
|
+
such a row's path keys either detached it from its worktree or left it unable to start;
|
|
641
|
+
- the live copy is reached only through the worktree placement, sits in a legacy-encoded
|
|
642
|
+
folder, or has no path on record that encodes to its folder;
|
|
643
|
+
- the paths that reach the live copy's folder disagree, or its directory no longer exists;
|
|
644
|
+
- an unfinished operation involves it, or anything about it changed between plan and apply.
|
|
645
|
+
|
|
646
|
+
`--apply` needs the desktop app fully closed, re-plans under its lock and writes only the
|
|
647
|
+
conversations whose rows, copies and target are unchanged, prints any it held at that point,
|
|
648
|
+
and exits 3 when any conversation was held. An interrupted run recovers in both directions:
|
|
649
|
+
`recover --forward` finishes from the journal, and is offered only when every row and copy it
|
|
650
|
+
would touch is as the journal expects; `recover --back` puts each conversation's copies and
|
|
651
|
+
rows back, and keeps realign's placement for a conversation whose copy cannot go back (its
|
|
652
|
+
place is taken, for instance), saying why. `undo` works per conversation. It puts back only
|
|
653
|
+
the keys realign changed, keeps whatever the app wrote to other keys since, and refuses a
|
|
654
|
+
conversation, naming the reason, when its rows no longer hold what realign wrote, now open
|
|
655
|
+
another conversation, were written by a later operation, or when the conversation has gained
|
|
656
|
+
turns since (bookkeeping records do not count). The other conversations go ahead, and a later
|
|
657
|
+
`undo --apply` retries the refused ones.
|
|
658
|
+
|
|
659
|
+
A copy realign sets aside stays in its journal until `undo` puts it back or `realign --release`
|
|
660
|
+
sends it to the Recycle Bin; `doctor` lists every copy held. Once a conversation has gained
|
|
661
|
+
turns, undo refuses it for good, which is what release is for:
|
|
662
|
+
|
|
663
|
+
```
|
|
664
|
+
claude-code-sessions realign --release
|
|
665
|
+
claude-code-sessions realign --release --apply
|
|
666
|
+
```
|
|
667
|
+
|
|
668
|
+
A copy is released only when every row of its conversation now opens one copy, that copy holds
|
|
669
|
+
the set-aside copy's whole conversation (the bookkeeping records `tools-used` and `cost-state`
|
|
670
|
+
aside, which carry no conversation), and its sidecar folder holds every file the set-aside
|
|
671
|
+
copy's does, with the same content: a copy that holds anything else the live one lacks stays in
|
|
672
|
+
the journal, with the reason printed. It goes to the Windows Recycle Bin, and the step is built
|
|
673
|
+
so that a permanent delete cannot happen silently. A shell that cannot recycle a file destroys
|
|
674
|
+
it instead. So release refuses a bin turned off (by you or by policy), one whose capacity the
|
|
675
|
+
registry does not record, one without room for the whole run without pushing out your older
|
|
676
|
+
deleted files (asked again just before each move), and a path too long for the shell, and it
|
|
677
|
+
asks Windows to warn rather than delete in any case it cannot see. After each move it checks
|
|
678
|
+
that a new bin record names the file, with its size and its contents present. It needs the
|
|
679
|
+
desktop app closed, since its checks read the rows and transcripts the app rewrites. A copy
|
|
680
|
+
whose release is interrupted is finished by the next run, and only after the same checks; once
|
|
681
|
+
its transcript is in the bin, only while the live copy still starts with everything it held
|
|
682
|
+
when the release began. One gap is stated rather than closed: a case the checks cannot see
|
|
683
|
+
rests on Windows' warning, whose behaviour beside the silent flags release uses is untested. Once
|
|
684
|
+
every copy of an operation is released, the journal ages the operation out like any other;
|
|
685
|
+
`undo` refuses a conversation whose copy was released or is part-way through release, and
|
|
686
|
+
`--json` gives each copy's outcome, a failed run included.
|
|
687
|
+
|
|
688
|
+
A released copy can be restored from the Recycle Bin by hand until the bin is emptied or
|
|
689
|
+
Windows clears it to make room. It comes back to its journal folder, not to its project folder;
|
|
690
|
+
the operation's record names where each part went. Windows only, like every other store mutation
|
|
691
|
+
here.
|
|
692
|
+
|
|
582
693
|
## Safety design
|
|
583
694
|
|
|
584
695
|
- **Dry-run by default.** Every mutating command prints its plan and does nothing until you
|
|
@@ -620,13 +731,20 @@ window.
|
|
|
620
731
|
| Platform | Status | Mutations |
|
|
621
732
|
|---|---|---|
|
|
622
733
|
| Windows 11 + Claude Desktop | verified 2026-07-31; sync end-to-end incl. live continuation 2026-08-03; sync-undo drift refusal live 2026-08-04 | read-only + mutations |
|
|
623
|
-
| Windows 11 + Claude Desktop 2.42304.2 (user data moved to `%LOCALAPPDATA%\Claude-Data`) | from 0.16.3: read-only paths probed 2026-10-19 against the moved store. 0.16.2 and earlier find no store | read-only + mutations, `--apply`
|
|
734
|
+
| Windows 11 + Claude Desktop 2.42304.2 (user data moved to `%LOCALAPPDATA%\Claude-Data`) | from 0.16.3: read-only paths probed 2026-10-19 against the moved store. 0.16.2 and earlier find no store where the app moved the folder, and read the old copy where it copied the folder and left the original in place | read-only + mutations, `--apply` never run on this build (the app updated to 3.x first) |
|
|
735
|
+
| Windows 11 + Claude Desktop 3.6556.2 (user data still in `%LOCALAPPDATA%\Claude-Data`) | on 0.16.5: read-only paths and `converge --apply` verified 2026-10-23, the apply with the app closed and the browser helper running. Needs 0.16.3 or later to find the store | read-only + mutations |
|
|
736
|
+
| Windows 11 + Claude Desktop 4.4212.0 (user data still in `%LOCALAPPDATA%\Claude-Data`) | on 0.16.5: `doctor`, `alignment` and `converge --apply` verified 2026-11-01, the apply with the app closed and the browser helper running. Needs 0.16.3 or later to find the store | read-only + mutations |
|
|
737
|
+
| Windows 11 + Claude Desktop 5.1231.0 (user data still in `%LOCALAPPDATA%\Claude-Data`) | on 0.16.5: `doctor` and `alignment` verified read-only 2026-11-10. Needs 0.16.3 or later to find the store | read-only + mutations, `--apply` not yet run on this build |
|
|
738
|
+
| Windows 11 + Claude Desktop 6.6395.0 (user data still in `%LOCALAPPDATA%\Claude-Data`) | from 0.17.0: `doctor` and `realign --apply` verified 2026-11-17, the apply with the app closed; after the app reopened, `doctor` and a fresh `realign` plan found every rewritten row opening the live copy and no set-aside copy back in its old folder. Needs 0.16.3 or later to find the store | read-only + mutations |
|
|
624
739
|
| macOS / Linux desktop | unverified | **read-only only** - desktop-store mutations refuse, with no override |
|
|
625
740
|
| CLI-only sessions (any OS) | transcript layout verified | mutations allowed via `--transcript-only` |
|
|
626
741
|
|
|
627
|
-
The dates are measurements against specific builds
|
|
628
|
-
|
|
629
|
-
|
|
742
|
+
The dates are measurements against specific builds of the MSIX that Anthropic signs and
|
|
743
|
+
distributes itself from claude.ai - most recently Claude Desktop 6.6395.0.0 with Claude Code
|
|
744
|
+
CLI 2.1.326 for `doctor` and the last `--apply` run, a `realign`; `converge --apply` last ran
|
|
745
|
+
on 4.4212.0.0 with CLI 2.1.309. The Microsoft Store build is a separate package, `AnthropicPBC.Claude_<id>`.
|
|
746
|
+
This tool does not look in that package's folder and has never been run against it.
|
|
747
|
+
On 2026-10-14 the
|
|
630
748
|
read-only paths were verified against a real Desktop 2.x store, build 2.37337.1.0 (`list`,
|
|
631
749
|
`doctor`, `alignment`, `recover` without `--apply`, and a `converge` dry run): the row format
|
|
632
750
|
had only gained keys, and every row this tool had written under 1.x was still present, some
|
|
@@ -647,6 +765,30 @@ the move. 0.16.4 reads those paths under the moved store: `doctor` checks
|
|
|
647
765
|
the rows where they now are, and `undo` and `recover` treat such an operation as they did
|
|
648
766
|
before the move. A row in a store the tool cannot read at all is listed by `doctor` apart, not
|
|
649
767
|
counted in its exit code, and `undo` refuses it.
|
|
768
|
+
On 2026-10-23, under 3.6556.2.0, the data folder had not moved again, and rows written by the
|
|
769
|
+
3.x builds had only gained two keys. A `converge --apply` on 0.16.5 ran with the app closed while the browser
|
|
770
|
+
helper kept running, and the running-app guard excused the helper under the
|
|
771
|
+
`trust-signed-helper` opt-in. After the app restarted, every row the apply had written was
|
|
772
|
+
present and byte-identical to what it wrote, and `doctor` reported none vanished.
|
|
773
|
+
On 2026-11-01, under 4.4212.0.0, the data folder had not moved again. A `converge --apply` on
|
|
774
|
+
0.16.5 ran with the app closed and the browser helper running, and the guard excused the
|
|
775
|
+
helper as it had under 3.x. After the app restarted, every row the apply had written was
|
|
776
|
+
present and `doctor` reported none vanished. Within a few minutes the app had re-saved six of
|
|
777
|
+
those rows, and each had only gained keys. By 2026-11-02 the 4.x app had created 36 rows for
|
|
778
|
+
itself. None lacked a key that every older row carries, and four had gained one new key,
|
|
779
|
+
`isStarred`.
|
|
780
|
+
On 2026-11-10, under 5.1231.0.0, the data folder had not moved again, and `doctor` and
|
|
781
|
+
`alignment` on 0.16.5 found the store where 4.x left it. That build's release notes say
|
|
782
|
+
Windows 10 21H2 or later and Windows Server 2022 now keep the app's data in
|
|
783
|
+
`%LOCALAPPDATA%\Claude-Data` too, as the Windows 11 machine these measurements come from has
|
|
784
|
+
since 2.42304.2, so on those systems the move happens at 5.1231.0. If the folder layout there
|
|
785
|
+
matches that machine's, the same versions apply: 0.16.3 or later to find the store, 0.16.4 or
|
|
786
|
+
later for operations from before the move. The notes also say the Microsoft Store build keeps
|
|
787
|
+
its data where it was, and they mention moves that copied the folder instead of moving it.
|
|
788
|
+
Where such a copy leaves the old store in place, run this tool from an ordinary terminal: from
|
|
789
|
+
a shell the desktop app started, `%APPDATA%\Claude` is the app's view of its old folder, and
|
|
790
|
+
0.16.5 can read the old copy through it. Not measured here: a Windows 10 machine, a Microsoft
|
|
791
|
+
Store install, and an `--apply` under 5.1231.0.
|
|
650
792
|
The on-disk format has changed once already during this tool's own development; treat a
|
|
651
793
|
much newer build as unverified territory.
|
|
652
794
|
|
|
@@ -1007,7 +1149,9 @@ window into `journaled`/`completed`/`rolled_back`.
|
|
|
1007
1149
|
README, the window's checkbox and the command's help, says Chrome or Edge, since either can
|
|
1008
1150
|
launch the helper. Three constants left unused since the August anti-fabrication fix are gone.
|
|
1009
1151
|
- **0.16.3 finds the session store where Desktop 2.42304.2 moved it.** That build moved the
|
|
1010
|
-
app's whole user-data folder out of the Microsoft Store package
|
|
1152
|
+
app's whole user-data folder out of the Microsoft Store package (corrected 2026-11-10: out of
|
|
1153
|
+
the MSIX package Anthropic distributes itself, not the Microsoft Store build, whose data the
|
|
1154
|
+
5.1231.0 release notes say does not move), from
|
|
1011
1155
|
`%LOCALAPPDATA%\Packages\Claude_<pkg>\LocalCache\Roaming\Claude` to
|
|
1012
1156
|
`%LOCALAPPDATA%\Claude-Data\claude_<pkg>`, and left a 0-byte file in the old folder naming
|
|
1013
1157
|
the new one. 0.16.2 looked only in the old places, so `doctor` and `alignment` reported no
|
|
@@ -1058,6 +1202,35 @@ window into `journaled`/`completed`/`rolled_back`.
|
|
|
1058
1202
|
left every route refusing until a restart, or, when the app copied its data instead of moving
|
|
1059
1203
|
it, left the window writing into the old copy. A press whose store layout changed since its
|
|
1060
1204
|
plan now stops and replans before its next write.
|
|
1205
|
+
- **0.17.0 shows which copy of a transcript each account's row opens, and repairs it.** A
|
|
1206
|
+
conversation's transcript can end up in two project folders, and each account's row opens
|
|
1207
|
+
the copy in the folder its `cwd` names, so one sidebar can show a conversation as it was
|
|
1208
|
+
weeks ago while another shows it today. `doctor` now lists each such conversation, which copy
|
|
1209
|
+
each row opens, and how the copies relate, reading past the bookkeeping records the app
|
|
1210
|
+
writes to whichever copy it has open; it exits 1 when an unarchived row opens an older or a
|
|
1211
|
+
diverged copy. It also lists, without counting them, conversations with one copy and a row
|
|
1212
|
+
whose folder holds none, because opening such a row made the app copy the transcript into
|
|
1213
|
+
that folder (measured on Desktop 5.11492.5). The new `realign` command points those rows at
|
|
1214
|
+
the folder holding the live copy, by rewriting `cwd` and the `folderPlace` keys the app keeps
|
|
1215
|
+
in step with it, and moves the stale copy into its journal, where `undo` can put it back. It
|
|
1216
|
+
holds anything it cannot decide safely, with the reason: diverged copies, a row with a
|
|
1217
|
+
worktree link, a live copy whose folder is gone, and identical copies unless
|
|
1218
|
+
`--include-identical` is given. On the machine it was built on, the first apply realigned 43
|
|
1219
|
+
conversations (84 rows) and held 52; `doctor` then listed 26 split conversations where it
|
|
1220
|
+
had listed 30, and 26 waiting ones where it had listed 65. `realign --release` sends a
|
|
1221
|
+
set-aside copy to the Recycle Bin once every row of its conversation opens one copy and that
|
|
1222
|
+
copy holds everything the set-aside one did, sidecar files included. It refuses each case it
|
|
1223
|
+
can see in which Windows would delete the copy outright, and asks Windows to warn in any
|
|
1224
|
+
other; how that warning behaves beside the silent flags release passes is untested, and no
|
|
1225
|
+
`--release --apply` had run when this shipped. The window's health report counts the split
|
|
1226
|
+
conversations, the waiting ones and the copies realign holds. `realign` itself is CLI-only,
|
|
1227
|
+
and while one is the latest operation the window offers no undo.
|
|
1228
|
+
The window also declares system DPI awareness before it opens, on Windows with Tk 9.0, so
|
|
1229
|
+
at 250 % it draws its text and widgets at the display's resolution and at the size it had
|
|
1230
|
+
before, where it used to be a stretched bitmap. Every pixel constant goes through one factor,
|
|
1231
|
+
the process's DPI over 96. Verified at 250 % and in a faithful 100 % run, and every other
|
|
1232
|
+
GUI harness passes in an aware process. 150 %, a common laptop setting, is untested; Tk 8.6
|
|
1233
|
+
is unchanged.
|
|
1061
1234
|
- Platform rows above move from "unverified" to "verified" as contributors confirm the store
|
|
1062
1235
|
paths and behavior on their own machines.
|
|
1063
1236
|
|
|
@@ -127,7 +127,7 @@ read. See `docs/internals.md`, RULING 7.
|
|
|
127
127
|
|
|
128
128
|
## Before any move: close the Claude app.
|
|
129
129
|
|
|
130
|
-
`move`, `undo`, `recover`, `repoint --apply`, `new-row --apply`, and every mutating `sync`
|
|
130
|
+
`move`, `undo`, `recover`, `repoint --apply`, `new-row --apply`, `realign --apply`, and every mutating `sync`
|
|
131
131
|
route (`--apply`, `undo` of a completed sync, `recover --back` on a stuck one) all refuse
|
|
132
132
|
while they can see the Claude desktop app running - whoever the identity files currently say
|
|
133
133
|
is signed in. That includes `undo` of a completed `repoint` or `new-row`, and `recover` on an
|
|
@@ -153,7 +153,7 @@ itself, and the guard counts it - if a refusal names that process, fully exit th
|
|
|
153
153
|
|
|
154
154
|
## Usage
|
|
155
155
|
|
|
156
|
-
|
|
156
|
+
Thirteen commands. All mutating commands default to a dry run; add `--apply` to execute.
|
|
157
157
|
(`trust-signed-helper` is the opt-in described above; it changes a setting, not your
|
|
158
158
|
sessions.)
|
|
159
159
|
|
|
@@ -174,8 +174,9 @@ claude-code-sessions list gate
|
|
|
174
174
|
```
|
|
175
175
|
|
|
176
176
|
**`doctor`** - read-only health report (stale locks, unresolved operations, orphaned rows,
|
|
177
|
-
encoding-scheme ambiguity, a synthesized row that has since vanished,
|
|
178
|
-
account's sidebar points at**
|
|
177
|
+
encoding-scheme ambiguity, a synthesized row that has since vanished, **conversations no
|
|
178
|
+
account's sidebar points at**, and **conversations whose accounts open different copies of one
|
|
179
|
+
transcript**):
|
|
179
180
|
|
|
180
181
|
```
|
|
181
182
|
claude-code-sessions doctor
|
|
@@ -189,6 +190,30 @@ Resuming a session while signed into a different account repoints its row, which
|
|
|
189
190
|
conversation you were in an hour ago reachable from no sidebar at all. That is what this
|
|
190
191
|
ranking is for, and `repoint` is how you put it back.
|
|
191
192
|
|
|
193
|
+
The copies check covers a case every other number misses. A conversation's transcript can end
|
|
194
|
+
up in two project folders, usually after a session moves into a second worktree. Each account's
|
|
195
|
+
row names a folder through its `cwd`, and the app opens the copy in that folder, even a stale
|
|
196
|
+
one. So one sidebar can show a conversation as it was weeks ago while another shows it today,
|
|
197
|
+
and `alignment` still counts it complete. `doctor` lists each such conversation, which
|
|
198
|
+
copy each account's row opens, and how each pair of copies relates: one behind the other (one
|
|
199
|
+
history, one copy stopped growing), the same conversation, byte-identical, or diverged (each
|
|
200
|
+
holds turns the other lacks). It compares the bytes first and then reads what lies past the
|
|
201
|
+
bytes two copies share, because the app writes bookkeeping records (`tools-used` when it opens
|
|
202
|
+
a copy, `cost-state` when it quits) to whichever copy it has open, the stale one included. A
|
|
203
|
+
copy ahead only by those records holds the same conversation; a stale copy the app has opened
|
|
204
|
+
and quit is still behind, although its bytes no longer match. Any other record type counts as
|
|
205
|
+
conversation. `doctor` exits 1 when an unarchived row opens an older or a diverged copy.
|
|
206
|
+
Identical copies and copies that differ only in bookkeeping are listed and do not change the
|
|
207
|
+
exit code. Neither does a row whose folder holds no copy at all. A row whose `cwd` names a main repository while the transcript sits in
|
|
208
|
+
one of its worktree folders is the app's normal placement, not a fault.
|
|
209
|
+
|
|
210
|
+
Deleting the stale copy is a catch-up, not a fix. On one conversation, measured on Desktop
|
|
211
|
+
5.11492.5, opening a row whose folder held no copy made the app copy the transcript back into
|
|
212
|
+
that folder, so the split returns once the conversation is worked from one account and opened
|
|
213
|
+
from another. For the same reason `doctor` also lists, without counting them, conversations with
|
|
214
|
+
a single copy and a row whose folder holds none: opening that row starts a split. `realign`
|
|
215
|
+
(below) is the repair for both.
|
|
216
|
+
|
|
192
217
|
**`alignment`** - read-only scoreboard: how close your accounts are to holding one shared
|
|
193
218
|
history. `doctor` answers "is anything broken"; this answers "how far apart are my sidebars,
|
|
194
219
|
and which way is it moving":
|
|
@@ -207,8 +232,10 @@ Two of those actively trade against each other: making a conversation reachable
|
|
|
207
232
|
both halves of a disagreeing pair in every sidebar under one name, which is a direct hit to
|
|
208
233
|
*distinguishable*. A single "aligned: yes/no" would hide exactly the decision you have to make.
|
|
209
234
|
|
|
210
|
-
It reads row files and transcript **filenames** only - no transcript content - so it
|
|
211
|
-
on a store with gigabytes behind it.
|
|
235
|
+
It reads row files and transcript **filenames and sizes** only - no transcript content - so it
|
|
236
|
+
stays fast on a store with gigabytes behind it. That is why its line on conversations whose
|
|
237
|
+
accounts open different copies only counts them, and how many differ in length; telling a stale
|
|
238
|
+
copy from a diverged one means reading both, which `doctor` does. **It always exits 0** when it could produce the report. It
|
|
212
239
|
is a scoreboard, not a check: these numbers stay non-zero for as long as the work takes, and a
|
|
213
240
|
command that exits 1 for months trains you to stop reading it.
|
|
214
241
|
|
|
@@ -549,6 +576,90 @@ stopped being redundant is presence it leaves alone. `doctor`'s vanished-row che
|
|
|
549
576
|
converge-created rows the same way it watches `new-row`'s, with the same journal-retention
|
|
550
577
|
window.
|
|
551
578
|
|
|
579
|
+
**`realign`** - point every account's row of a split conversation at the folder holding its
|
|
580
|
+
live transcript copy:
|
|
581
|
+
|
|
582
|
+
```
|
|
583
|
+
claude-code-sessions realign
|
|
584
|
+
claude-code-sessions realign --apply
|
|
585
|
+
```
|
|
586
|
+
|
|
587
|
+
These are the split rows `doctor` reports. A probe on Desktop 5.11492.5, run on one plain and
|
|
588
|
+
one worktree conversation, found that a row's `cwd` alone decides which copy it opens; that the
|
|
589
|
+
app rewrites the row's `folderPlace` to match `cwd`, and never writes `cwd` back; and that a
|
|
590
|
+
turn typed after `cwd` moved runs in the new folder, in the same transcript. So `realign`
|
|
591
|
+
rewrites `cwd`, and `folderPlace.cwd` and `folderPlace.realPath` where the row has a
|
|
592
|
+
`folderPlace` (it never adds one), and leaves every other key alone, including `originCwd` and
|
|
593
|
+
the title. The rewritten rows then open the live copy. No row opens the stale copy any more, so
|
|
594
|
+
it moves into the operation's journal, with its sidecar folder, where `undo` can put it back
|
|
595
|
+
and `doctor` lists it. Without that step, `move`, `repoint`, `new-row` and `converge` would
|
|
596
|
+
keep refusing the conversation for its duplicated transcript. A copy that no row opened even
|
|
597
|
+
before `realign` stays where it is: that is the duplicated-transcript case, which `realign` did
|
|
598
|
+
not cause.
|
|
599
|
+
|
|
600
|
+
By default it plans conversations with a copy behind the live one, and rows whose folder holds
|
|
601
|
+
no copy. Copies that hold the same conversation (byte-identical, or different only in
|
|
602
|
+
bookkeeping records) are realigned only with `--include-identical`, onto the copy the most
|
|
603
|
+
recently active row opens. `--only` narrows to one conversation, resolved the way `converge
|
|
604
|
+
--only` is. A conversation is **held**, with the reason printed, when:
|
|
605
|
+
|
|
606
|
+
- its copies have diverged (each holds turns the other lacks), so which history to keep is
|
|
607
|
+
your decision;
|
|
608
|
+
- a copy no row opens holds turns the live copy lacks;
|
|
609
|
+
- a row to change has a worktree link (`worktreePath` or `worktreeName`). In the probe, moving
|
|
610
|
+
such a row's path keys either detached it from its worktree or left it unable to start;
|
|
611
|
+
- the live copy is reached only through the worktree placement, sits in a legacy-encoded
|
|
612
|
+
folder, or has no path on record that encodes to its folder;
|
|
613
|
+
- the paths that reach the live copy's folder disagree, or its directory no longer exists;
|
|
614
|
+
- an unfinished operation involves it, or anything about it changed between plan and apply.
|
|
615
|
+
|
|
616
|
+
`--apply` needs the desktop app fully closed, re-plans under its lock and writes only the
|
|
617
|
+
conversations whose rows, copies and target are unchanged, prints any it held at that point,
|
|
618
|
+
and exits 3 when any conversation was held. An interrupted run recovers in both directions:
|
|
619
|
+
`recover --forward` finishes from the journal, and is offered only when every row and copy it
|
|
620
|
+
would touch is as the journal expects; `recover --back` puts each conversation's copies and
|
|
621
|
+
rows back, and keeps realign's placement for a conversation whose copy cannot go back (its
|
|
622
|
+
place is taken, for instance), saying why. `undo` works per conversation. It puts back only
|
|
623
|
+
the keys realign changed, keeps whatever the app wrote to other keys since, and refuses a
|
|
624
|
+
conversation, naming the reason, when its rows no longer hold what realign wrote, now open
|
|
625
|
+
another conversation, were written by a later operation, or when the conversation has gained
|
|
626
|
+
turns since (bookkeeping records do not count). The other conversations go ahead, and a later
|
|
627
|
+
`undo --apply` retries the refused ones.
|
|
628
|
+
|
|
629
|
+
A copy realign sets aside stays in its journal until `undo` puts it back or `realign --release`
|
|
630
|
+
sends it to the Recycle Bin; `doctor` lists every copy held. Once a conversation has gained
|
|
631
|
+
turns, undo refuses it for good, which is what release is for:
|
|
632
|
+
|
|
633
|
+
```
|
|
634
|
+
claude-code-sessions realign --release
|
|
635
|
+
claude-code-sessions realign --release --apply
|
|
636
|
+
```
|
|
637
|
+
|
|
638
|
+
A copy is released only when every row of its conversation now opens one copy, that copy holds
|
|
639
|
+
the set-aside copy's whole conversation (the bookkeeping records `tools-used` and `cost-state`
|
|
640
|
+
aside, which carry no conversation), and its sidecar folder holds every file the set-aside
|
|
641
|
+
copy's does, with the same content: a copy that holds anything else the live one lacks stays in
|
|
642
|
+
the journal, with the reason printed. It goes to the Windows Recycle Bin, and the step is built
|
|
643
|
+
so that a permanent delete cannot happen silently. A shell that cannot recycle a file destroys
|
|
644
|
+
it instead. So release refuses a bin turned off (by you or by policy), one whose capacity the
|
|
645
|
+
registry does not record, one without room for the whole run without pushing out your older
|
|
646
|
+
deleted files (asked again just before each move), and a path too long for the shell, and it
|
|
647
|
+
asks Windows to warn rather than delete in any case it cannot see. After each move it checks
|
|
648
|
+
that a new bin record names the file, with its size and its contents present. It needs the
|
|
649
|
+
desktop app closed, since its checks read the rows and transcripts the app rewrites. A copy
|
|
650
|
+
whose release is interrupted is finished by the next run, and only after the same checks; once
|
|
651
|
+
its transcript is in the bin, only while the live copy still starts with everything it held
|
|
652
|
+
when the release began. One gap is stated rather than closed: a case the checks cannot see
|
|
653
|
+
rests on Windows' warning, whose behaviour beside the silent flags release uses is untested. Once
|
|
654
|
+
every copy of an operation is released, the journal ages the operation out like any other;
|
|
655
|
+
`undo` refuses a conversation whose copy was released or is part-way through release, and
|
|
656
|
+
`--json` gives each copy's outcome, a failed run included.
|
|
657
|
+
|
|
658
|
+
A released copy can be restored from the Recycle Bin by hand until the bin is emptied or
|
|
659
|
+
Windows clears it to make room. It comes back to its journal folder, not to its project folder;
|
|
660
|
+
the operation's record names where each part went. Windows only, like every other store mutation
|
|
661
|
+
here.
|
|
662
|
+
|
|
552
663
|
## Safety design
|
|
553
664
|
|
|
554
665
|
- **Dry-run by default.** Every mutating command prints its plan and does nothing until you
|
|
@@ -590,13 +701,20 @@ window.
|
|
|
590
701
|
| Platform | Status | Mutations |
|
|
591
702
|
|---|---|---|
|
|
592
703
|
| Windows 11 + Claude Desktop | verified 2026-07-31; sync end-to-end incl. live continuation 2026-08-03; sync-undo drift refusal live 2026-08-04 | read-only + mutations |
|
|
593
|
-
| Windows 11 + Claude Desktop 2.42304.2 (user data moved to `%LOCALAPPDATA%\Claude-Data`) | from 0.16.3: read-only paths probed 2026-10-19 against the moved store. 0.16.2 and earlier find no store | read-only + mutations, `--apply`
|
|
704
|
+
| Windows 11 + Claude Desktop 2.42304.2 (user data moved to `%LOCALAPPDATA%\Claude-Data`) | from 0.16.3: read-only paths probed 2026-10-19 against the moved store. 0.16.2 and earlier find no store where the app moved the folder, and read the old copy where it copied the folder and left the original in place | read-only + mutations, `--apply` never run on this build (the app updated to 3.x first) |
|
|
705
|
+
| Windows 11 + Claude Desktop 3.6556.2 (user data still in `%LOCALAPPDATA%\Claude-Data`) | on 0.16.5: read-only paths and `converge --apply` verified 2026-10-23, the apply with the app closed and the browser helper running. Needs 0.16.3 or later to find the store | read-only + mutations |
|
|
706
|
+
| Windows 11 + Claude Desktop 4.4212.0 (user data still in `%LOCALAPPDATA%\Claude-Data`) | on 0.16.5: `doctor`, `alignment` and `converge --apply` verified 2026-11-01, the apply with the app closed and the browser helper running. Needs 0.16.3 or later to find the store | read-only + mutations |
|
|
707
|
+
| Windows 11 + Claude Desktop 5.1231.0 (user data still in `%LOCALAPPDATA%\Claude-Data`) | on 0.16.5: `doctor` and `alignment` verified read-only 2026-11-10. Needs 0.16.3 or later to find the store | read-only + mutations, `--apply` not yet run on this build |
|
|
708
|
+
| Windows 11 + Claude Desktop 6.6395.0 (user data still in `%LOCALAPPDATA%\Claude-Data`) | from 0.17.0: `doctor` and `realign --apply` verified 2026-11-17, the apply with the app closed; after the app reopened, `doctor` and a fresh `realign` plan found every rewritten row opening the live copy and no set-aside copy back in its old folder. Needs 0.16.3 or later to find the store | read-only + mutations |
|
|
594
709
|
| macOS / Linux desktop | unverified | **read-only only** - desktop-store mutations refuse, with no override |
|
|
595
710
|
| CLI-only sessions (any OS) | transcript layout verified | mutations allowed via `--transcript-only` |
|
|
596
711
|
|
|
597
|
-
The dates are measurements against specific builds
|
|
598
|
-
|
|
599
|
-
|
|
712
|
+
The dates are measurements against specific builds of the MSIX that Anthropic signs and
|
|
713
|
+
distributes itself from claude.ai - most recently Claude Desktop 6.6395.0.0 with Claude Code
|
|
714
|
+
CLI 2.1.326 for `doctor` and the last `--apply` run, a `realign`; `converge --apply` last ran
|
|
715
|
+
on 4.4212.0.0 with CLI 2.1.309. The Microsoft Store build is a separate package, `AnthropicPBC.Claude_<id>`.
|
|
716
|
+
This tool does not look in that package's folder and has never been run against it.
|
|
717
|
+
On 2026-10-14 the
|
|
600
718
|
read-only paths were verified against a real Desktop 2.x store, build 2.37337.1.0 (`list`,
|
|
601
719
|
`doctor`, `alignment`, `recover` without `--apply`, and a `converge` dry run): the row format
|
|
602
720
|
had only gained keys, and every row this tool had written under 1.x was still present, some
|
|
@@ -617,6 +735,30 @@ the move. 0.16.4 reads those paths under the moved store: `doctor` checks
|
|
|
617
735
|
the rows where they now are, and `undo` and `recover` treat such an operation as they did
|
|
618
736
|
before the move. A row in a store the tool cannot read at all is listed by `doctor` apart, not
|
|
619
737
|
counted in its exit code, and `undo` refuses it.
|
|
738
|
+
On 2026-10-23, under 3.6556.2.0, the data folder had not moved again, and rows written by the
|
|
739
|
+
3.x builds had only gained two keys. A `converge --apply` on 0.16.5 ran with the app closed while the browser
|
|
740
|
+
helper kept running, and the running-app guard excused the helper under the
|
|
741
|
+
`trust-signed-helper` opt-in. After the app restarted, every row the apply had written was
|
|
742
|
+
present and byte-identical to what it wrote, and `doctor` reported none vanished.
|
|
743
|
+
On 2026-11-01, under 4.4212.0.0, the data folder had not moved again. A `converge --apply` on
|
|
744
|
+
0.16.5 ran with the app closed and the browser helper running, and the guard excused the
|
|
745
|
+
helper as it had under 3.x. After the app restarted, every row the apply had written was
|
|
746
|
+
present and `doctor` reported none vanished. Within a few minutes the app had re-saved six of
|
|
747
|
+
those rows, and each had only gained keys. By 2026-11-02 the 4.x app had created 36 rows for
|
|
748
|
+
itself. None lacked a key that every older row carries, and four had gained one new key,
|
|
749
|
+
`isStarred`.
|
|
750
|
+
On 2026-11-10, under 5.1231.0.0, the data folder had not moved again, and `doctor` and
|
|
751
|
+
`alignment` on 0.16.5 found the store where 4.x left it. That build's release notes say
|
|
752
|
+
Windows 10 21H2 or later and Windows Server 2022 now keep the app's data in
|
|
753
|
+
`%LOCALAPPDATA%\Claude-Data` too, as the Windows 11 machine these measurements come from has
|
|
754
|
+
since 2.42304.2, so on those systems the move happens at 5.1231.0. If the folder layout there
|
|
755
|
+
matches that machine's, the same versions apply: 0.16.3 or later to find the store, 0.16.4 or
|
|
756
|
+
later for operations from before the move. The notes also say the Microsoft Store build keeps
|
|
757
|
+
its data where it was, and they mention moves that copied the folder instead of moving it.
|
|
758
|
+
Where such a copy leaves the old store in place, run this tool from an ordinary terminal: from
|
|
759
|
+
a shell the desktop app started, `%APPDATA%\Claude` is the app's view of its old folder, and
|
|
760
|
+
0.16.5 can read the old copy through it. Not measured here: a Windows 10 machine, a Microsoft
|
|
761
|
+
Store install, and an `--apply` under 5.1231.0.
|
|
620
762
|
The on-disk format has changed once already during this tool's own development; treat a
|
|
621
763
|
much newer build as unverified territory.
|
|
622
764
|
|
|
@@ -977,7 +1119,9 @@ window into `journaled`/`completed`/`rolled_back`.
|
|
|
977
1119
|
README, the window's checkbox and the command's help, says Chrome or Edge, since either can
|
|
978
1120
|
launch the helper. Three constants left unused since the August anti-fabrication fix are gone.
|
|
979
1121
|
- **0.16.3 finds the session store where Desktop 2.42304.2 moved it.** That build moved the
|
|
980
|
-
app's whole user-data folder out of the Microsoft Store package
|
|
1122
|
+
app's whole user-data folder out of the Microsoft Store package (corrected 2026-11-10: out of
|
|
1123
|
+
the MSIX package Anthropic distributes itself, not the Microsoft Store build, whose data the
|
|
1124
|
+
5.1231.0 release notes say does not move), from
|
|
981
1125
|
`%LOCALAPPDATA%\Packages\Claude_<pkg>\LocalCache\Roaming\Claude` to
|
|
982
1126
|
`%LOCALAPPDATA%\Claude-Data\claude_<pkg>`, and left a 0-byte file in the old folder naming
|
|
983
1127
|
the new one. 0.16.2 looked only in the old places, so `doctor` and `alignment` reported no
|
|
@@ -1028,6 +1172,35 @@ window into `journaled`/`completed`/`rolled_back`.
|
|
|
1028
1172
|
left every route refusing until a restart, or, when the app copied its data instead of moving
|
|
1029
1173
|
it, left the window writing into the old copy. A press whose store layout changed since its
|
|
1030
1174
|
plan now stops and replans before its next write.
|
|
1175
|
+
- **0.17.0 shows which copy of a transcript each account's row opens, and repairs it.** A
|
|
1176
|
+
conversation's transcript can end up in two project folders, and each account's row opens
|
|
1177
|
+
the copy in the folder its `cwd` names, so one sidebar can show a conversation as it was
|
|
1178
|
+
weeks ago while another shows it today. `doctor` now lists each such conversation, which copy
|
|
1179
|
+
each row opens, and how the copies relate, reading past the bookkeeping records the app
|
|
1180
|
+
writes to whichever copy it has open; it exits 1 when an unarchived row opens an older or a
|
|
1181
|
+
diverged copy. It also lists, without counting them, conversations with one copy and a row
|
|
1182
|
+
whose folder holds none, because opening such a row made the app copy the transcript into
|
|
1183
|
+
that folder (measured on Desktop 5.11492.5). The new `realign` command points those rows at
|
|
1184
|
+
the folder holding the live copy, by rewriting `cwd` and the `folderPlace` keys the app keeps
|
|
1185
|
+
in step with it, and moves the stale copy into its journal, where `undo` can put it back. It
|
|
1186
|
+
holds anything it cannot decide safely, with the reason: diverged copies, a row with a
|
|
1187
|
+
worktree link, a live copy whose folder is gone, and identical copies unless
|
|
1188
|
+
`--include-identical` is given. On the machine it was built on, the first apply realigned 43
|
|
1189
|
+
conversations (84 rows) and held 52; `doctor` then listed 26 split conversations where it
|
|
1190
|
+
had listed 30, and 26 waiting ones where it had listed 65. `realign --release` sends a
|
|
1191
|
+
set-aside copy to the Recycle Bin once every row of its conversation opens one copy and that
|
|
1192
|
+
copy holds everything the set-aside one did, sidecar files included. It refuses each case it
|
|
1193
|
+
can see in which Windows would delete the copy outright, and asks Windows to warn in any
|
|
1194
|
+
other; how that warning behaves beside the silent flags release passes is untested, and no
|
|
1195
|
+
`--release --apply` had run when this shipped. The window's health report counts the split
|
|
1196
|
+
conversations, the waiting ones and the copies realign holds. `realign` itself is CLI-only,
|
|
1197
|
+
and while one is the latest operation the window offers no undo.
|
|
1198
|
+
The window also declares system DPI awareness before it opens, on Windows with Tk 9.0, so
|
|
1199
|
+
at 250 % it draws its text and widgets at the display's resolution and at the size it had
|
|
1200
|
+
before, where it used to be a stretched bitmap. Every pixel constant goes through one factor,
|
|
1201
|
+
the process's DPI over 96. Verified at 250 % and in a faithful 100 % run, and every other
|
|
1202
|
+
GUI harness passes in an aware process. 150 %, a common laptop setting, is untested; Tk 8.6
|
|
1203
|
+
is unchanged.
|
|
1031
1204
|
- Platform rows above move from "unverified" to "verified" as contributors confirm the store
|
|
1032
1205
|
paths and behavior on their own machines.
|
|
1033
1206
|
|