claude-code-sessions 0.16.5__tar.gz → 0.18.0__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.
Files changed (42) hide show
  1. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/LICENSE +21 -21
  2. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/PKG-INFO +294 -12
  3. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/README.md +293 -11
  4. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/claude_code_sessions.egg-info/PKG-INFO +294 -12
  5. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/claude_code_sessions.egg-info/SOURCES.txt +4 -0
  6. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/claude_code_sessions.py +5973 -265
  7. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/claude_code_sessions_gui.py +337 -109
  8. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/pyproject.toml +50 -50
  9. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_alignment.py +144 -144
  10. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_anonymize.py +140 -0
  11. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_cli.py +172 -172
  12. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_containment.py +71 -71
  13. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_converge.py +2183 -1
  14. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_discovery.py +48 -48
  15. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_encoding.py +41 -41
  16. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_engine.py +274 -274
  17. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_golden.py +18 -18
  18. claude_code_sessions-0.18.0/tests/test_gui_dpi.py +171 -0
  19. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_gui_models.py +292 -0
  20. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_helpers.py +39 -39
  21. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_journal.py +87 -87
  22. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_journal_reroot.py +42 -0
  23. claude_code_sessions-0.18.0/tests/test_lineage.py +1092 -0
  24. claude_code_sessions-0.18.0/tests/test_realign.py +1830 -0
  25. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_rows.py +73 -73
  26. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_scaffold.py +18 -18
  27. claude_code_sessions-0.18.0/tests/test_split_rows.py +654 -0
  28. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_sync.py +2140 -0
  29. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_transcripts.py +24 -24
  30. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_undo.py +55 -0
  31. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/claude_code_sessions.egg-info/dependency_links.txt +0 -0
  32. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/claude_code_sessions.egg-info/entry_points.txt +0 -0
  33. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/claude_code_sessions.egg-info/top_level.txt +0 -0
  34. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/setup.cfg +0 -0
  35. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_commands.py +0 -0
  36. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_data_root.py +0 -0
  37. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_duplicate_transcripts.py +0 -0
  38. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_gui_destination.py +0 -0
  39. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_gui_env_rebuild.py +0 -0
  40. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_plan_move.py +0 -0
  41. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_recover.py +0 -0
  42. {claude_code_sessions-0.16.5 → claude_code_sessions-0.18.0}/tests/test_retitle.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: 0.16.5
3
+ Version: 0.18.0
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
- Twelve commands. All mutating commands default to a dry run; add `--apply` to execute.
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, and **conversations no
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 stays fast
241
- on a store with gigabytes behind it. **It always exits 0** when it could produce the report. 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
 
@@ -423,6 +450,22 @@ is always a preview will misread `sync --apply --json`.
423
450
  grants** across the account boundary unchanged. The permission half is the more
424
451
  security-relevant of the two: use `--verbatim` only when you actually want the second account
425
452
  to inherit what the first one had allowed.
453
+ - **A row's link to its parent is translated, not copied.** A row `sync` adds carries the same
454
+ translated link and chain that `converge` writes (see `converge`, below), or none, and the plan
455
+ prints the same `links :` and `chains :` lines. A refresh replaces the whole row, link
456
+ included, so the destination's own link is dropped or translated, never merged. A refresh that
457
+ drops it or points it at another row says so, one line per row (`link : local_c1 loses its
458
+ link`, `link : local_c1 now links to another row`, and `chain :` for the chain), and the
459
+ fields line leaves the link out so it is reported once. One exception: a destination link that
460
+ already names the same row and `taskId` is kept as it is. A refresh whose bytes already match
461
+ the destination's row is reported unchanged, unless the source row carries a link or a chain:
462
+ then the link is translated first, and the row is written when the translated link differs, so
463
+ a copy that names a row in the source account is corrected. (Under `--newer-only` such a row
464
+ meets the age holds first, like any refresh.) In the window the sync pane shows the
465
+ same lines. An apply that ends with nothing to write says "Nothing to copy: the destination
466
+ already holds these rows" and leaves Undo as it was.
467
+ **`--verbatim` copies links as they are**, so a link may name a row that does not exist in the
468
+ destination. Such an add is counted: `links : 2 copied verbatim (not translated)`.
426
469
  - **It needs the app closed too, the same as `move`, `undo`, and `recover` - regardless of which
427
470
  file resolved the signed-in account.** `sync` only ever writes the store you are *not* signed
428
471
  into, and re-checks that at the moment it writes, not just when it planned; that re-check
@@ -579,6 +622,154 @@ stopped being redundant is presence it leaves alone. `doctor`'s vanished-row che
579
622
  converge-created rows the same way it watches `new-row`'s, with the same journal-retention
580
623
  window.
581
624
 
625
+ **Side sessions keep their parent.** A side session's row names its parent's row (`spawnedFrom`),
626
+ and a row deep in a chain of them names the chain's first row (`startChain`). Both name a row by
627
+ its id in the same account, so a copy would point at a row the other account does not have.
628
+ `converge` and `sync` therefore write them translated. The link becomes
629
+ `{"sessionId": <that account's own row for the parent>, "taskId": <copied unchanged>}`, and the
630
+ chain becomes `{"firstSessionId": <that account's own row for the root>, "depth": <copied>}`.
631
+ Everything else in the source link (`title`, `sideSession`, `name`, `sharedFolder`, `wakeParent`,
632
+ `pendingPop` and any key this tool has not seen) is left behind and recorded. The `taskId` is not
633
+ optional: on the one side session measured per link shape, a link without it hid the child from
634
+ the sidebar, and a link with it nested the child under its parent. Only the link shapes that were
635
+ measured are carried: an app-written side-session link, the older `{sessionId, taskId, title}`,
636
+ and this tool's own output, so a carried link can be carried again.
637
+
638
+ The row is still created when its link is left out, and the plan says why. A link is left out
639
+ when the holders disagree on the parent or on `taskId`; when it names the conversation itself;
640
+ when the parent's transcript is gone; when that account has no row for the parent, or the
641
+ parent's own row is held in this plan, or its row there opens another conversation; when the
642
+ link's shape is not one that was measured; or when it would close a cycle. A link is written only
643
+ to a row that will exist, including a parent created in the same run, which is written first.
644
+ The plan reports it:
645
+
646
+ ```
647
+ links : 4 to carry, 1 left out (holders disagree on the parent 1)
648
+ chains : 2 to carry
649
+ cycles : 1 already present, left as they are
650
+ ```
651
+
652
+ Each part prints only when its count is not zero, the most frequent reason first. A row whose
653
+ link was left out carries `[link: <reason>]` on its own line. The apply resolves every link
654
+ again under the lock, against the store as it then stands, and its last lines say what changed
655
+ since the plan: `links : 4 carried, 1 left out (...); since the plan: 1 changed, 1 gained`.
656
+ `changed` and `dropped` count links the plan had, `gained` one it did not. With `--verbose` the
657
+ apply also lists them, one line per written row whose link or chain was left out or changed
658
+ since the plan, with its account:
659
+
660
+ ```
661
+ local_c3 -> bob@example.com (cccc1111/dddd2222) [link: the parent's own row is held in this plan; dropped since the plan]
662
+ ```
663
+
664
+ The journal records the final link per row, and recovery never resolves again. It writes the
665
+ journaled bytes, skips a row (`parent_not_written`) whose planned parent it did not write, and
666
+ skips one (`parent_changed`) whose parent's row is gone or now opens another conversation.
667
+ `recover --forward` names each row it skipped that way, with its reason, in the `skipped :` line
668
+ `converge` prints for a skipped row.
669
+
670
+ `undo` can leave a link dangling. When a row it created has become the last pointer to its
671
+ conversation, `undo` keeps it, while it still removes the parent row the same operation created.
672
+ It says so, for `undo` of a `converge`, a `sync` or a `new-row`, and for `recover --back` and a
673
+ finished `recover --forward`:
674
+
675
+ ```
676
+ link left dangling: local_c1 (its parent's row was removed)
677
+ ```
678
+
679
+ A child whose parent's row is missing is listed on its own and opens, which was measured. A
680
+ chain root's row that is missing is reported the same way ("its chain root's row"), but how the
681
+ app shows that was not measured. `link redirected: ... (its parent's row now opens another
682
+ conversation)` names a link whose target row survives and now opens a different conversation,
683
+ and `links not checked: N could not be read` counts what the report could not read. The report
684
+ covers rows that link to a row the command removed, and for `recover --forward` the operation's
685
+ own targets. It is not a check of every link in the store. A reversal cut short and run again
686
+ counts a row the first run removed as removed, so a link to it is still named. The window's Undo
687
+ shows the same lines, for a copy and a converge alike.
688
+
689
+ **`realign`** - point every account's row of a split conversation at the folder holding its
690
+ live transcript copy:
691
+
692
+ ```
693
+ claude-code-sessions realign
694
+ claude-code-sessions realign --apply
695
+ ```
696
+
697
+ These are the split rows `doctor` reports. A probe on Desktop 5.11492.5, run on one plain and
698
+ one worktree conversation, found that a row's `cwd` alone decides which copy it opens; that the
699
+ app rewrites the row's `folderPlace` to match `cwd`, and never writes `cwd` back; and that a
700
+ turn typed after `cwd` moved runs in the new folder, in the same transcript. So `realign`
701
+ rewrites `cwd`, and `folderPlace.cwd` and `folderPlace.realPath` where the row has a
702
+ `folderPlace` (it never adds one), and leaves every other key alone, including `originCwd` and
703
+ the title. The rewritten rows then open the live copy. No row opens the stale copy any more, so
704
+ it moves into the operation's journal, with its sidecar folder, where `undo` can put it back
705
+ and `doctor` lists it. Without that step, `move`, `repoint`, `new-row` and `converge` would
706
+ keep refusing the conversation for its duplicated transcript. A copy that no row opened even
707
+ before `realign` stays where it is: that is the duplicated-transcript case, which `realign` did
708
+ not cause.
709
+
710
+ By default it plans conversations with a copy behind the live one, and rows whose folder holds
711
+ no copy. Copies that hold the same conversation (byte-identical, or different only in
712
+ bookkeeping records) are realigned only with `--include-identical`, onto the copy the most
713
+ recently active row opens. `--only` narrows to one conversation, resolved the way `converge
714
+ --only` is. A conversation is **held**, with the reason printed, when:
715
+
716
+ - its copies have diverged (each holds turns the other lacks), so which history to keep is
717
+ your decision;
718
+ - a copy no row opens holds turns the live copy lacks;
719
+ - a row to change has a worktree link (`worktreePath` or `worktreeName`). In the probe, moving
720
+ such a row's path keys either detached it from its worktree or left it unable to start;
721
+ - the live copy is reached only through the worktree placement, sits in a legacy-encoded
722
+ folder, or has no path on record that encodes to its folder;
723
+ - the paths that reach the live copy's folder disagree, or its directory no longer exists;
724
+ - an unfinished operation involves it, or anything about it changed between plan and apply.
725
+
726
+ `--apply` needs the desktop app fully closed, re-plans under its lock and writes only the
727
+ conversations whose rows, copies and target are unchanged, prints any it held at that point,
728
+ and exits 3 when any conversation was held. An interrupted run recovers in both directions:
729
+ `recover --forward` finishes from the journal, and is offered only when every row and copy it
730
+ would touch is as the journal expects; `recover --back` puts each conversation's copies and
731
+ rows back, and keeps realign's placement for a conversation whose copy cannot go back (its
732
+ place is taken, for instance), saying why. `undo` works per conversation. It puts back only
733
+ the keys realign changed, keeps whatever the app wrote to other keys since, and refuses a
734
+ conversation, naming the reason, when its rows no longer hold what realign wrote, now open
735
+ another conversation, were written by a later operation, or when the conversation has gained
736
+ turns since (bookkeeping records do not count). The other conversations go ahead, and a later
737
+ `undo --apply` retries the refused ones.
738
+
739
+ A copy realign sets aside stays in its journal until `undo` puts it back or `realign --release`
740
+ sends it to the Recycle Bin; `doctor` lists every copy held. Once a conversation has gained
741
+ turns, undo refuses it for good, which is what release is for:
742
+
743
+ ```
744
+ claude-code-sessions realign --release
745
+ claude-code-sessions realign --release --apply
746
+ ```
747
+
748
+ A copy is released only when every row of its conversation now opens one copy, that copy holds
749
+ the set-aside copy's whole conversation (the bookkeeping records `tools-used` and `cost-state`
750
+ aside, which carry no conversation), and its sidecar folder holds every file the set-aside
751
+ copy's does, with the same content: a copy that holds anything else the live one lacks stays in
752
+ the journal, with the reason printed. It goes to the Windows Recycle Bin, and the step is built
753
+ so that a permanent delete cannot happen silently. A shell that cannot recycle a file destroys
754
+ it instead. So release refuses a bin turned off (by you or by policy), one whose capacity the
755
+ registry does not record, one without room for the whole run without pushing out your older
756
+ deleted files (asked again just before each move), and a path too long for the shell, and it
757
+ asks Windows to warn rather than delete in any case it cannot see. After each move it checks
758
+ that a new bin record names the file, with its size and its contents present. It needs the
759
+ desktop app closed, since its checks read the rows and transcripts the app rewrites. A copy
760
+ whose release is interrupted is finished by the next run, and only after the same checks; once
761
+ its transcript is in the bin, only while the live copy still starts with everything it held
762
+ when the release began. One gap is stated rather than closed: a case the checks cannot see
763
+ rests on Windows' warning, whose behaviour beside the silent flags release uses is untested. Once
764
+ every copy of an operation is released, the journal ages the operation out like any other;
765
+ `undo` refuses a conversation whose copy was released or is part-way through release, and
766
+ `--json` gives each copy's outcome, a failed run included.
767
+
768
+ A released copy can be restored from the Recycle Bin by hand until the bin is emptied or
769
+ Windows clears it to make room. It comes back to its journal folder, not to its project folder;
770
+ the operation's record names where each part went. Windows only, like every other store mutation
771
+ here.
772
+
582
773
  ## Safety design
583
774
 
584
775
  - **Dry-run by default.** Every mutating command prints its plan and does nothing until you
@@ -620,13 +811,20 @@ window.
620
811
  | Platform | Status | Mutations |
621
812
  |---|---|---|
622
813
  | 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` not yet verified on this build |
814
+ | 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) |
815
+ | 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 |
816
+ | 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 |
817
+ | 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 |
818
+ | 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. From 0.18.0: `converge --apply` with carried links verified 2026-11-19, the apply with the app closed; the app nested both children opened, and after a reopen every carried link and chain read back as written. Needs 0.16.3 or later to find the store | read-only + mutations |
624
819
  | macOS / Linux desktop | unverified | **read-only only** - desktop-store mutations refuse, with no override |
625
820
  | CLI-only sessions (any OS) | transcript layout verified | mutations allowed via `--transcript-only` |
626
821
 
627
- The dates are measurements against specific builds - most recently Claude Desktop
628
- 2.42304.2.0 (Microsoft Store MSIX install) for the read-only paths, and 2.42304.0.0 with Claude
629
- Code CLI 2.1.302 for the last `--apply` runs. On 2026-10-14 the
822
+ The dates are measurements against specific builds of the MSIX that Anthropic signs and
823
+ distributes itself from claude.ai - most recently Claude Desktop 6.6395.0.0 with Claude Code
824
+ CLI 2.1.326 for `doctor` and the last two `--apply` runs, a `realign` on 2026-11-17 and a
825
+ `converge` on 2026-11-19. The Microsoft Store build is a separate package, `AnthropicPBC.Claude_<id>`.
826
+ This tool does not look in that package's folder and has never been run against it.
827
+ On 2026-10-14 the
630
828
  read-only paths were verified against a real Desktop 2.x store, build 2.37337.1.0 (`list`,
631
829
  `doctor`, `alignment`, `recover` without `--apply`, and a `converge` dry run): the row format
632
830
  had only gained keys, and every row this tool had written under 1.x was still present, some
@@ -647,6 +845,30 @@ the move. 0.16.4 reads those paths under the moved store: `doctor` checks
647
845
  the rows where they now are, and `undo` and `recover` treat such an operation as they did
648
846
  before the move. A row in a store the tool cannot read at all is listed by `doctor` apart, not
649
847
  counted in its exit code, and `undo` refuses it.
848
+ On 2026-10-23, under 3.6556.2.0, the data folder had not moved again, and rows written by the
849
+ 3.x builds had only gained two keys. A `converge --apply` on 0.16.5 ran with the app closed while the browser
850
+ helper kept running, and the running-app guard excused the helper under the
851
+ `trust-signed-helper` opt-in. After the app restarted, every row the apply had written was
852
+ present and byte-identical to what it wrote, and `doctor` reported none vanished.
853
+ On 2026-11-01, under 4.4212.0.0, the data folder had not moved again. A `converge --apply` on
854
+ 0.16.5 ran with the app closed and the browser helper running, and the guard excused the
855
+ helper as it had under 3.x. After the app restarted, every row the apply had written was
856
+ present and `doctor` reported none vanished. Within a few minutes the app had re-saved six of
857
+ those rows, and each had only gained keys. By 2026-11-02 the 4.x app had created 36 rows for
858
+ itself. None lacked a key that every older row carries, and four had gained one new key,
859
+ `isStarred`.
860
+ On 2026-11-10, under 5.1231.0.0, the data folder had not moved again, and `doctor` and
861
+ `alignment` on 0.16.5 found the store where 4.x left it. That build's release notes say
862
+ Windows 10 21H2 or later and Windows Server 2022 now keep the app's data in
863
+ `%LOCALAPPDATA%\Claude-Data` too, as the Windows 11 machine these measurements come from has
864
+ since 2.42304.2, so on those systems the move happens at 5.1231.0. If the folder layout there
865
+ matches that machine's, the same versions apply: 0.16.3 or later to find the store, 0.16.4 or
866
+ later for operations from before the move. The notes also say the Microsoft Store build keeps
867
+ its data where it was, and they mention moves that copied the folder instead of moving it.
868
+ Where such a copy leaves the old store in place, run this tool from an ordinary terminal: from
869
+ a shell the desktop app started, `%APPDATA%\Claude` is the app's view of its old folder, and
870
+ 0.16.5 can read the old copy through it. Not measured here: a Windows 10 machine, a Microsoft
871
+ Store install, and an `--apply` under 5.1231.0.
650
872
  The on-disk format has changed once already during this tool's own development; treat a
651
873
  much newer build as unverified territory.
652
874
 
@@ -1007,7 +1229,9 @@ window into `journaled`/`completed`/`rolled_back`.
1007
1229
  README, the window's checkbox and the command's help, says Chrome or Edge, since either can
1008
1230
  launch the helper. Three constants left unused since the August anti-fabrication fix are gone.
1009
1231
  - **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, from
1232
+ app's whole user-data folder out of the Microsoft Store package (corrected 2026-11-10: out of
1233
+ the MSIX package Anthropic distributes itself, not the Microsoft Store build, whose data the
1234
+ 5.1231.0 release notes say does not move), from
1011
1235
  `%LOCALAPPDATA%\Packages\Claude_<pkg>\LocalCache\Roaming\Claude` to
1012
1236
  `%LOCALAPPDATA%\Claude-Data\claude_<pkg>`, and left a 0-byte file in the old folder naming
1013
1237
  the new one. 0.16.2 looked only in the old places, so `doctor` and `alignment` reported no
@@ -1058,6 +1282,64 @@ window into `journaled`/`completed`/`rolled_back`.
1058
1282
  left every route refusing until a restart, or, when the app copied its data instead of moving
1059
1283
  it, left the window writing into the old copy. A press whose store layout changed since its
1060
1284
  plan now stops and replans before its next write.
1285
+ - **0.17.0 shows which copy of a transcript each account's row opens, and repairs it.** A
1286
+ conversation's transcript can end up in two project folders, and each account's row opens
1287
+ the copy in the folder its `cwd` names, so one sidebar can show a conversation as it was
1288
+ weeks ago while another shows it today. `doctor` now lists each such conversation, which copy
1289
+ each row opens, and how the copies relate, reading past the bookkeeping records the app
1290
+ writes to whichever copy it has open; it exits 1 when an unarchived row opens an older or a
1291
+ diverged copy. It also lists, without counting them, conversations with one copy and a row
1292
+ whose folder holds none, because opening such a row made the app copy the transcript into
1293
+ that folder (measured on Desktop 5.11492.5). The new `realign` command points those rows at
1294
+ the folder holding the live copy, by rewriting `cwd` and the `folderPlace` keys the app keeps
1295
+ in step with it, and moves the stale copy into its journal, where `undo` can put it back. It
1296
+ holds anything it cannot decide safely, with the reason: diverged copies, a row with a
1297
+ worktree link, a live copy whose folder is gone, and identical copies unless
1298
+ `--include-identical` is given. On the machine it was built on, the first apply realigned 43
1299
+ conversations (84 rows) and held 52; `doctor` then listed 26 split conversations where it
1300
+ had listed 30, and 26 waiting ones where it had listed 65. `realign --release` sends a
1301
+ set-aside copy to the Recycle Bin once every row of its conversation opens one copy and that
1302
+ copy holds everything the set-aside one did, sidecar files included. It refuses each case it
1303
+ can see in which Windows would delete the copy outright, and asks Windows to warn in any
1304
+ other; how that warning behaves beside the silent flags release passes is untested, and no
1305
+ `--release --apply` had run when this shipped. The window's health report counts the split
1306
+ conversations, the waiting ones and the copies realign holds. `realign` itself is CLI-only,
1307
+ and while one is the latest operation the window offers no undo.
1308
+ The window also declares system DPI awareness before it opens, on Windows with Tk 9.0, so
1309
+ at 250 % it draws its text and widgets at the display's resolution and at the size it had
1310
+ before, where it used to be a stretched bitmap. Every pixel constant goes through one factor,
1311
+ the process's DPI over 96. Verified at 250 % and in a faithful 100 % run, and every other
1312
+ GUI harness passes in an aware process. 150 %, a common laptop setting, is untested; Tk 8.6
1313
+ is unchanged.
1314
+ - **0.18.0 keeps a side session under its parent in every account.** A side session's row names
1315
+ its parent's row (`spawnedFrom`) and its chain's first row (`startChain`), each by an id that
1316
+ exists only in that account. `converge` used to write its rows without either, and `sync`
1317
+ copied both as they stood, so a copied link named a row the destination does not have. Both
1318
+ commands now write them translated to the destination's own rows, the link as `{sessionId,
1319
+ taskId}`. On the machine it was built on, probes found that the app hides a child whose link
1320
+ lacks `taskId` and nests it when the link has both keys, for the app's side-session links and
1321
+ for its older `{sessionId, taskId, title}` links, one candidate each. Only those measured
1322
+ shapes and this tool's own output are carried. Any other link is left out with its reason, as
1323
+ are the cases "Side sessions keep their parent" above lists: holders that disagree, a parent
1324
+ the account lacks, a link that would close a cycle. A parent created in the same run is
1325
+ written first. The apply resolves every link again under the lock, and the journal keeps the
1326
+ final link per row. Recovery never resolves again. When a pending row's parent was not
1327
+ written, or its row is gone or now opens another conversation, `converge` skips that row and
1328
+ `sync` stops offering to go forward. `undo` and
1329
+ `recover --back` remove children before their parents and report any link they leave
1330
+ dangling, or naming a row that now opens another conversation; the window's Undo shows the
1331
+ same lines. A
1332
+ `spawnedFrom` or `startChain` that is not an object, a null say, counts as no link. Before
1333
+ the release, a `converge --apply` on that machine created 20 rows and carried 20 links and 18
1334
+ chains. The app nested both children opened, one whose source link also carried
1335
+ `sharedFolder` and one with `userFork`, and after a reopen every carried link and chain read
1336
+ back as written. A child with the older link shape written by a release run, and a child
1337
+ `sync` added, have not been looked at in the app. `sync` now decodes its stores a second time
1338
+ for the links. On a store of about 3,000 rows (400 MB) its dry run took about 30% longer and
1339
+ peaked at 879 MB against 229 MB, and with `--update` at 1.13 GB against 452 MB. A fix is
1340
+ planned. `converge` recovery also keeps a row that landed before a hard kill: it records the
1341
+ row as written before re-checking anything, so `undo` still removes it, even after the app
1342
+ has rewritten it.
1061
1343
  - Platform rows above move from "unverified" to "verified" as contributors confirm the store
1062
1344
  paths and behavior on their own machines.
1063
1345