Skip to content

fix(state): preserve damaged shared documents before replacement - #1564

Merged
devswha merged 2 commits into
mainfrom
fix/preserve-damaged-shared-state
Oct 21, 2026
Merged

devswha merged 2 commits into
mainfrom
fix/preserve-damaged-shared-state

Conversation

@devswha

@devswha devswha commented Oct 21, 2026

Copy link
Copy Markdown
Owner

Change

A damaged preferences.json or seen.json was read as empty and then silently replaced by the
next change, so the only copy of what was lost disappeared. Both stores now keep it aside first,
through one small helper (server/state-document.ts) that tells a missing file from an
unreadable one, from malformed JSON and from a document of the wrong shape.

  • Reading (startup, GET /api/preferences, GET /api/seen) never renames or writes anything.
    Parse and shape failures log a diagnostic naming the kind of failure, never the content.
    Preferences whose settings is not an object (or with invalid pins), and a seen document that
    is not a list or holds rows that are not finishes, count as damaged. Valid seen rows are still
    used.
  • Until a store has replaced its file, every write re-reads it in the same synchronous section,
    whether it was valid, damaged, unreadable or missing at startup, so the decision never rests on
    a stale flag from startup and a file damaged after a valid start is kept too:
    • a file repaired by hand meanwhile is reloaded and the change merged over it, with no backup;
      nothing of the repair is held in memory until the replacement succeeds;
    • damaged bytes (the ones on disk now, not the sample from startup) are moved untouched to
      <file>.bad with mode 0600, replacing the previous .bad only when this move succeeds, and
      then the new document is renamed into place. There is one recovery generation, never a
      growing series of timestamped copies.
  • Failure receipts:
    • a read failure (permissions, a directory in place of the file) refuses the write and replaces
      nothing;
    • if the damaged file cannot be kept aside (chmod or rename fails), the write is refused with an
      explicit 500 (preferences_unwritable / seen_unwritable) naming the .bad file, and the
      original, the older .bad, the held state and the listeners are all unchanged;
    • if the final replacement fails after the move, the complete original stays in .bad, nothing
      is published, the held state is unchanged, and a retry checks the file again: it writes the
      new document while keeping those bytes, or keeps newer damaged bytes found on disk.
  • A missing file behaves as before. Devices and VAPID handling are untouched. No HTTP/WS shape,
    demo response or UI string changes.

The exclusion this relies on is within one server process. Several servers sharing one state
directory, or another program editing the file during the final rename, are out of scope.

Validation

  • bun test ./server/preferences.test.ts ./server/seen.test.ts ./server/devices.test.ts ./server/push.test.ts:
    89 pass, 0 fail. This includes the existing devices and malformed-VAPID fail-closed tests.
  • New cases in both store test files: "preserves malformed bytes only on first effective write",
    "preserves wrong-shaped documents privately", "keeps a manual repair rather than quarantining
    it", "refuses overwrite when preservation fails", "retains one bad generation across
    replacement failure", "refuses to write over a file it cannot read", "preserves a file damaged
    after a valid start on the first write" (both stores) and, for seen, "holds nothing of a repair
    whose replacement failed, and checks the file again on retry". They assert exact byte
    equality of .bad, mode 0600 on POSIX, at most one .bad, and no listener call before a
    successful replacement. The replacement failure is injected with spyOn(fs, "renameSync"); the
    preservation failure uses a real non-empty directory at .bad. All 19 new cases fail against
    the stores on main (bun test exits 1).
  • The existing damaged-seen test, which expected the damaged file to be replaced, became
    "preserves malformed bytes only on first effective write".
  • bun run check fast exits 0: workflow syntax, generated types and themes, typecheck, build,
    and unit tests (5791 pass, 0 fail, 8 skip).

Ready for maintainer review; do not merge.

A damaged preferences.json or seen.json used to be read as empty and then
silently replaced by the next change, destroying the only copy of what was
lost. Both stores now go through a small shared helper
(server/state-document.ts) that tells a missing file from an unreadable one,
from malformed JSON and from a document of the wrong shape.

- Reading (construction, GET) never renames or writes anything; parse and
  shape failures log a diagnostic that names the failure, not the content.
  Preferences whose settings are not an object, and a seen document that is
  not a list or holds rows that are not finishes, count as damaged.
- The first write after such a start re-reads the file in the same
  synchronous section: a now-valid manual repair is reloaded and the change
  merged over it; damaged bytes (the current ones) are moved untouched to
  <file>.bad at 0600, replacing the previous generation only when this move
  succeeds, and then the new document is published.
- A read failure, or a failure to keep the damaged file aside, refuses the
  write and leaves the original in place; if the final replacement fails the
  original stays complete in .bad, nothing is published, and a retry keeps it.
- A missing file behaves as before. No HTTP/WS shape, demo response or UI
  string changes.

Checked with bun test for server/preferences.test.ts, server/seen.test.ts,
server/devices.test.ts and server/push.test.ts (86 pass). The new cases
cover malformed, wrong-shaped, read-failed, existing-backup,
preservation-failure, replacement-failure and manual-repair files with exact
byte equality, and all 16 of them fail against the previous stores.
bun run check fast: workflow syntax, generated files, typecheck and build
pass; the unit-test failures under heavy machine load were in unrelated
spawn-heavy files and failed the same way on the base commit.
The preferences and read-finish stores now read their file again before
the first replacement even when it was valid at start, so a document
damaged after the server started is kept as <file>.bad instead of being
overwritten. The read-finish store no longer holds a manually repaired
document, or forgets that the file needs checking, until its replacement
has succeeded: a failed rename leaves list() unchanged, and the retry
checks the file again.
@coderabbitai

coderabbitai Bot commented Oct 21, 2026 •

Copy link
Copy Markdown

Review in Change Stack →

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info
⚙️ Run configuration
  • Configuration used: Repository: devswha/herdr-web-ui/.coderabbit.yaml
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: b1c48ab5-74a7-4b34-994b-987f3b73601f
📥 Commits

Reviewing files that changed from the base of the PR and between 4cabde2 and da135cf.

📒 Files selected for processing: 6
  • changelog.d/preserve-damaged-shared-state.md
  • server/preferences.test.ts
  • server/preferences.ts
  • server/seen.test.ts
  • server/seen.ts
  • server/state-document.ts

Included review availability: This review used your included allowance. Your plan provides up to 4 included reviews per hour; 2 remain after this review.

📝 Summary

Summary by CodeRabbit

  • Bug Fixes

    • Damaged or malformed preference and read-history files are now preserved before recovery, preventing accidental data loss.
    • Valid data is retained when files contain a mix of valid and invalid entries.
    • Concurrent changes made during manual repairs are preserved.
    • Unreadable files are no longer overwritten, and failed recovery attempts can be retried safely.
    • Recovered files use more secure file permissions, with only one damaged backup retained.
  • Documentation

    • Added documentation describing damaged-file preservation and recovery behavior.

Shared state stores now use common helpers to read and replace JSON documents. The stores preserve damaged files during recovery, reread disk state before their first write, and have expanded tests for repair and write-failure scenarios.

Changes: Shared State Recovery

Layer / File(s) Summary
Read and replace state documents
server/state-document.ts, changelog.d/preserve-damaged-shared-state.md
Added helpers to read JSON state and atomically replace files. When preservation is requested, the helper moves the existing file to a .bad path before replacement. The changelog describes this behavior.
Recover preference documents
server/preferences.ts, server/preferences.test.ts
PreferenceStore now distinguishes missing, malformed, and invalid documents. It rereads disk state before the first write and preserves damaged files during repair. Tests cover recovery, manual repairs, unreadable files, and write failures.
Recover seen-state documents
server/seen.ts, server/seen.test.ts
SeenStore now validates and merges disk state before its first write, then persists through the shared helpers. Tests cover damaged rows, manual repairs, unreadable files, and write failures.

Priority: ⬇️ Low

Estimated code review effort: 3 (Moderate) | ~25 minutes

Change: Bug fix

Merge Risk: ⚪ Minimal · up to da135

This change preserves damaged preferences and seen-state files as .bad copies before replacing them, and it refuses to overwrite files it cannot read. No actionable merge-blocking issue was identified in the supplied context.

Pre-merge checks: Passed 4 · Failed 1

Failed Failed checks: 1 warning

Check name Status Explanation Resolution
Docstring Coverage Warning Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 2 functions across 5 files. (1 skipped: 1… Write docstrings for the functions missing them to satisfy the coverage threshold.
Passed Passed checks: 4 passed
Check name Status Explanation
Title check Passed The title clearly and concisely describes the main change: preserving damaged shared documents before replacement.
Description check Passed The description includes the required Change and Validation sections. It explains the problem, resulting behavior, scope, failure handling, and test results in sufficient detail.
Linked Issues check Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check Passed Check skipped because no linked issues were found for this pull request.
Full details: Docstring Coverage

Explanation

Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 2 functions across 5 files. (1 skipped: 1 unsupported.)

Resolution

Write docstrings for the functions missing them to satisfy the coverage threshold.

✨ Finishing Touches: 💡 1
📝 Generate docstrings: 💡
  • Commit to this branch
  • Create a new PR
🧪 Generate unit tests (beta)
  • Commit to this branch
  • Create a new PR
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@devswha
devswha merged commit 08e9d30 into main Oct 21, 2026
29 of 33 checks passed
@devswha
devswha deleted the fix/preserve-damaged-shared-state branch October 21, 2026 20:13
devswha added a commit that referenced this pull request Oct 22, 2026
## Change

Release 0.4.13: everything merged since 0.4.12 (136 commits, 109
changelog entries). Metadata only; no code changes in this PR.

Left out of this release on purpose (still open): #1542, #1597, #1598,
#1579, and the outside contributors' PRs not yet reviewed or held (#1427
#1442 #1516 #1520 #1521 #1527 #1528 #1530 #1540 #1560 #1629 #1632
#1634).

### Metadata
- `package.json` and `herdr-plugin.toml`: 0.4.13.
- CHANGELOG: 87 `changelog.d/` fragments folded into `[0.4.13] -
2026-10-22` by `scripts/changelog-fold.ts`, with `by @login` credits on
the 36 links of outside contributors' PRs; `changelog.d/` holds only
README.md. Docs/test-only PRs without a fragment (#1395-#1398, #1439,
#1440) are not listed, as before.
- `release-summaries.json`: 0.4.13 in en/ko/ja/zh (8 new, 6 improved, 8
fixed; equal in every language, each <= 90 chars). The first line tells
that the app is now called Ovis (#1454) and installs and updates as
before.
- `REMOTE_BUNDLE_VERSION` 31 -> 32: `server/` and `shared/` changed
since remote-v31.
- Telemetry: `telemetry/worker.ts` and `shared/telemetry.ts` are
unchanged since the recorded deploy, so no Worker redeploy.

### After merge
1. Push `remote-v32` on the merge commit and wait for **Remote PC
bundles** to publish it.
2. Merge nothing touching `server/`, `shared/`, `bun.lock` or `patches/`
until the Release run is done.
3. `gh workflow run release.yml --ref main -f version=0.4.13`.

## Validation
- `bun scripts/release-guards.ts`: "Release 0.4.13: remote bundle,
telemetry Worker and release metadata are in order."
- `bun scripts/release-notes.ts 0.4.13` renders the three English lists
and the folded changelog.

<details><summary>Changelog (0.4.13)</summary>


### Added
- Claude Code, grok and Antigravity messages in Chat offer Rewind in
terminal: it sends nothing,
preserves your draft and opens the pane's terminal with a hint to run
the agent's own `/rewind`.
Chat follows the native conversation when the agent records its new
branch.
  ([#1500](#1500))
- Chat sends wait while Claude Code's rewind dialog is open in the
terminal, so a message cannot
  land in the dialog's own text field.
  ([#1500](#1500))
- Sending a bare `/goal` to Claude Code from Chat now shows its goal
status (the condition and how
long it has run, or that no goal is set) in a card with Close, instead
of leaving the dialog in
  the terminal.
  ([#1450](#1450))
- Command Code (`cmd`) agent panes now show their conversation in the
web UI: transcript
resolution from `~/.commandcode/projects/`, `/tree` branch projection,
model and context
window from Command Code's catalog, permission prompts, and skill
discovery.
([#1399](#1399) by
@diegolhambi)
- An ordinary command sent from Chat says what became of it under its
bubble, from evidence only:
accepted, waiting for your answer (with a way to the card), done with
what showed it, runs in
the terminal (with Open Terminal), or no result seen. A command the
agent keeps no transcript
  turn for still shows its sent row. A terminal screen that
changed after the send folds under the chip as an unconfirmed preview,
never proof that the
command finished. Terminal-only commands instead leave one quiet row
with Ran in terminal and
Open Terminal, without outcome chips or unconfirmed previews. Hermes
informational result cards
  keep showing what the agent printed.
  ([#1450](#1450))
- Settings → Terminal → Follow herdr's focus: when herdr moves to
another pane (a click in its
window, `herdr tab focus`, a script), this device opens that pane too.
Off by default, and set per
  device, so a phone open on another pane stays there.
  ([#1432](#1432))
- Hermes `/help`, `/status`, `/version`, `/profile` and `/reload-skills`
sent from Chat show what
Hermes printed in a card closed with Dismiss, which sends nothing to the
pane.
  ([#1450](#1450))
- Chat lists the commands and skills an OpenCode v2 pane's own
background service reports for its
folder, refreshed when they change, and settles a command it sent from
that session's own events.
The service is only read, never started, and only when it is the process
its service.json names.
  ([#1489](#1489))
- An `@skill` sent from Chat to OpenCode v2, at once or queued, is typed
through OpenCode's own
autocomplete so the skill is attached; when the row does not show, the
message says the skill
  went as plain text and points to the terminal.
  ([#1489](#1489))
- An OpenCode 1.x question is a card in Chat with each option and its
description, answered from
  the card.
  ([#1489](#1489))
- Chat lists a Codex pane's skills as the running Codex daemon of that
pane's own Codex home reports
them (Codex 0.157 or later; on Linux, or when the server's Codex home is
set), and a gjc pane's
skills as the SDK broker it is bound to reports them, asked through the
gjc CLI in the background.
The daemon and the broker are never started for this; without one, the
skills read from disk stay.
  ([#1489](#1489))
- New workspace and New tab start an agent with its model, effort and
first message where the
agent's command line takes them (Claude Code and Codex: `--model` and
the effort level; Codex also
takes the message as its own argument, behind `--` so a message like
`logout` is never a command).
For Claude Code and every other agent the message is queued on the PC as
the pane's first pending
message and sent once when the agent is ready, through the page that
started it; it shows as a
pending message that can be discarded or sent now. If that page
disconnects, reloads or leaves the tab
first, the message is held, never sent on its own, until someone sends
it now or discards it. A message the PC
could not queue, or one an older PC ignores, stays in the new tab's
message box and the dialog says
so. Pending first messages are kept in memory: restarting the web UI's
server drops them. A Codex
message with a tab or a line break is sent as the first message rather
than on its command line,
and the PC refuses launch arguments with control characters, an unknown
effort level or a model
  that is not one word. Under the message, the agent's skills in the
folder are listed before it runs, each a chip that puts the skill as the
agent types it into the
message (`$name`, `/name`, or `@name` for OpenCode 2). A first message
to OpenCode attaches its
`@skill` as any Chat send does; before OpenCode has a session nothing
tells its version, so the `@`
  is sent as the plain text it is.
In a desktop window the chat's composer can show a New tab with an
agent… button that opens the
dialog on the pane's workspace; it is off until turned on in Settings →
Chat → Composer (New tab
  button), and the tab menu opens the dialog either way.
  ([#1512](#1512))
- omp's `/handoff` runs from Chat, with or without focus instructions:
while omp writes the
handoff document, Chat shows a card that cancels it, and afterwards Chat
stays on the same
  conversation, which omp compacts in place.
  ([#1450](#1450))
- omp's model picker shows in the chat as a card, as Claude Code's,
Codex's and OmO's lists do: pick
a model and it is picked for the session, the model in use named on the
card. `/model` sent from
an omp chat opens it: the chat sends it as `/switch`, because omp's own
`/model` opens its role
  hub, whose picks are saved as the default for later sessions.
([#1377](#1377) by
@evanwy0223)
- OpenCode 2.0 or later: its questions are answered from the chat's
prompt card through OpenCode's
own background service instead of keystrokes, one question at a time
with Back to change an earlier
answer, and sent together after the last. Permissions, OpenCode 1 and
other prompts keep the
  existing generic fallback card and Terminal.
([#1257](#1257) by
@bardiz12)
- HTTP requests and WebSocket handshakes now carry request IDs and write
bounded completion
records in the app's state directory, so the owner can correlate a
response with its request.
  ([#1558](#1558))
- A workspace's row menu has Earlier sessions…: the sessions each agent
kept for the workspace's
folder, read from the agent's own store on the PC, newest first with the
title, the first prompt
and the time. Resume starts the agent on that session in a new tab of
the workspace; a session a
pane already shows is opened by selecting that pane instead. Picking
`/resume` in the chat's menu
opens the same list, and the terminal command's row offers it. A PC
whose bridge cannot list
  sessions yet has no such item.
  ([#1494](#1494))
- Agent rows show what herdr plugins report on a pane, such as the Quota
plugin's usage left or a
task's progress, in a small dim line under the row. They follow the
`$name` tokens of your herdr
`[ui.sidebar.agents]` rows (every token where none are set), change
live, and say a long value
whole in the tooltip and in a phone's ⋯ sheet. A line a plugin reports
for the agent now updates
  as soon as herdr says it.
  ([#1341](#1341))
- `HERDR_WEB_TAILSCALE_HTTPS_PORT` names the HTTPS port the installer
publishes the app on with
`tailscale serve`, so 443 can stay free for something else. Settings →
Phone & devices suggests
that port too. When another service already uses it, the installer stops
and says so instead of
taking another port, which would move the address an installed app
opens.
  ([#1433](#1433))

### Changed
- The chat's `/` menu lists the commands that run in the terminal apart,
last, under Terminal
commands, each with a Terminal badge, as the PC's command catalog
describes them. Pick or type
one and send it normally: once accepted, the pane's own terminal opens
over the chat, in this
browser only, while the chat stays where you were reading and the pane
stays on Chat. A head
says what ran, with Back to Chat and Keep Terminal; the header's Chat,
Ctrl/Cmd+Shift+J and a
phone's Back return too, and Escape goes to the agent. On a desktop the
message box stays under
the terminal, dimmed; on a phone the key bar takes its place. Claude
Code's panels (`/config`,
`/permissions`, `/theme` and the like) can bring the chat back by
themselves when you close them,
once Claude records it; other agents' recorded screens do once their
input row is back, and the
  rest return by hand. A Ran in terminal row's
Open Terminal shows the terminal over the chat again. Busy commands stay
queued; only an
explicit Send now opens the terminal, never automatic delivery. Refused
sends keep their draft
and note. Watch-only devices never send or open it, and /resume still
opens Earlier sessions.
Commands the agent's own menu hides are not listed, but still go when
typed by hand.
  ([#1441](#1441))
- For an OpenCode pane whose skills are `@` attachments, the `@` menu
lists them under Skills
  before the files, and picking one inserts the skill's id.
  ([#1441](#1441))
- Growing OMP and Claude Code tasks now parse appended transcript
records incrementally, reducing
repeated work when long chats refresh while preserving tool results,
notices and history resets.
Live parser state shares Codex's bounded cache; earlier responses stay
unchanged as new records arrive.
([#1561](#1561) by
@Yoonwoo-Ha)
- The app is now called Ovis across the browser, install messages and
documentation, matching
the phone apps. Existing installs, settings, plugin identifiers and
update channels stay the same.
  ([#1454](#1454))
- A question card whose screen reads keep failing now says it may be out
of date ("Not updated: the
question on screen may have changed.") until a read answers again.
Answering stays safe: the
server reads the screen again first and refuses a changed question
(#1335).
([#1360](#1360) by @Noobear)
- The Earlier sessions dialog picks the agent from the same list as the
New tab and New worktree
dialogs, each agent with its icon and the chevron spaced as the app's
other controls, instead of
  a plain select whose arrow sat at the far edge.
  ([#1573](#1573))
- A finished agent opened in one browser is read in every browser of the
same server: opening it
on the phone clears its DONE dot on the desktop too, and the other way
around.
  ([#1463](#1463))
- Remove PC has a restrained red background and border before and after
its confirmation step,
retaining visible keyboard focus and the explicit second click to remove
the registration.
([#1394](#1394) by @snowykr)
- After a terminal command sent from Chat, Codex, omp, pi, gjc, OmO,
Hermes (classic CLI), Devin,
agy and Grok bring the chat back by themselves when you close the picker
or panel it opened,
once the agent's own empty input row is back on the screen. Only the
commands each agent was
recorded doing that for count (Codex's `/permissions`, omp's
`/settings`, Grok's `/resume` and
the like); print forms, toggles, editor handoffs, Codex's `/side` and
`/btw`, OpenCode and
  anything unrecorded still return by hand.
  ([#1637](#1637))

### Fixed
- The New session and History pickers show Hermes, Grok, Antigravity,
Devin and Command Code by
  their own names instead of their lowercase ids.
  ([#1543](#1543))
- The terminal attach process and its Node host no longer inherit the
server's `HERDR_WEB_TOKEN`
environment variable; they keep the socket path and the rest of their
environment.
  ([#1562](#1562))
- Choosing Jump to latest remembers bottom-follow before the virtual
list scroll completes.
Switching to Terminal immediately after that choice and returning to
Chat no longer restores
the older reading position. This fixes explicit Jump-to-latest intent,
not arbitrary upward
  restoration on long, active or reasoning-heavy tabs.
([#1346](#1346) by
@Yoonwoo-Ha)
- Remember bottom-follow without measuring rendered rows or looking up a
turn anchor; readers away
  from the end retain the existing paragraph-and-offset restoration.
([#1346](#1346) by
@Yoonwoo-Ha)
- Selecting and copying chat text now preserves the text as shown
instead of adding Markdown
syntax. Rich-text paste and the whole-answer MD and TXT copy actions
remain available.
  ([#1409](#1409))
- Chat no longer treats a command that leaves the agent waiting in a
screen Chat cannot show as an
ordinary send. Claude Code's `/usage`, `/cost`, `/status`, `/help`,
`/memory`, `/mcp` and
`/add-dir`, Codex's `/new` and `/fork` in a repository, `/cd`, `/usage`,
`/copy` and `/side`,
and the panels, pickers, confirmations and editors of omp, gjc, OmO,
Hermes, Devin,
Antigravity, OpenCode and Command Code are now Terminal commands: sent,
they open Terminal with
Back to Chat. A value typed with a picker command, such as `/fast off`
or `/export notes.md`,
  stays an ordinary send.
  ([#1596](#1596))
- A Claude Code skill or custom command named like a built-in command
(`/compact`, say) is listed
once in the composer's command menu, as the skill Claude runs, instead
of beside the built-in.
  ([#1441](#1441))
- Reconnects and unconfirmed sends can leave a bounded browser-side
diagnostic trail on this
server, without recording messages, credentials, addresses or terminal
output.
([#1357](#1357) by
@christi4nity)
- The chat and the session history now apply one rule to Codex rollouts:
a subagent's rollout (one
naming a parent thread or an agent role), an exec run's, or one whose
ids are null or contradict
its file name is no longer read as a pane's conversation or offered to
resume, while older
rollouts without those fields still are. A rollout whose first line is
longer than the chat or
  the history reads is left out instead of read unchecked.
  ([#1592](#1592))
- Two Codex panes in one folder whose answers are short each show their
own chat again with Codex
0.163, which starts a thread in the same second as its process; before,
both said the
  conversation was unavailable about as often as not.
([#1522](#1522) by
@jiaoew1991)
- A Codex pane with only short answers still finds its chat after
another pane in the same folder
runs `/new`: a thread that pane shows a long answer of under its first
prompt, is bound to, or
  that herdr reports for it no longer leaves this pane unsure.
([#1522](#1522) by
@jiaoew1991)
- Of two Codex panes in one folder with short answers, the one started
first finds its chat too
once the other shows its own first exchange, where it said the
conversation was unavailable.
([#1522](#1522) by
@jiaoew1991)
- On a phone, a composer notice's actions (such as the `/resume`
notice's Earlier sessions, Open
Terminal and Run in Terminal) wrap onto their own line, and its dismiss
button stays visible
  instead of being cut off at the edge.
  ([#1543](#1543))
- Wrapped and multiline message drafts are sized with an out-of-flow
measurement field instead of
collapsing the live input, preserving transcript spacing and the
paragraph being read (#1337).
([#1347](#1347) by
@Yoonwoo-Ha)
- The measurement field shares the live input's Korean, Japanese and
Chinese line-breaking rules,
including Korean placeholder wrapping, so localized drafts keep their
natural height.
([#1347](#1347) by
@Yoonwoo-Ha)
- Local Herdr timeouts, machine connection failures and recoveries, and
abnormal WebSocket closes
now leave timestamped, rate-limited diagnostic records without logging
input or credentials.
([#1356](#1356) by
@christi4nity)
- The browser connection trace no longer re-posts a batch the server
refuses with a 4xx status,
resumes reporting after you sign in again on the same page, and
malformed reports can no longer
  use up the server's shared report quota.
  ([#1423](#1423))
- Pairing a device, renaming or revoking one, and the web push key and
subscriptions are now
flushed to disk before they replace the previous file, so a crash or
power cut right after the
change is much less likely to bring back a revoked device or make every
phone subscribe to
alerts again. How much survives a power cut still depends on the disk
and filesystem honouring
  the flush; on Windows the folder itself is not flushed.
  ([#1568](#1568))
- File search and path lookup no longer run a repository's fsmonitor
hook, follow Git variables
the server was started with, or wait on a Git listing that never ends: a
listing past 16 MiB
  or 3 seconds, or one cut off mid-name, falls back to the folder walk.
  ([#1570](#1570))
- **Files → Changes** no longer runs a program named `false` found on
`PATH` with Git 2.35 or
older, and reads the repository asked about even when the server was
started with
  `GIT_COMMON_DIR` set.
  ([#1570](#1570))
- Keep sidebar status updates responsive after many completed turns by
caching read finishes and
bounding their in-memory history instead of repeatedly scanning browser
storage.
  ([#1536](#1536))
- A first message from New workspace or New tab no longer waits forever
when the agent started for
it quits before the message goes: it is held for Send now or Discard.
The same happens when an
agent of the same kind that the user starts later comes up instead, and
when the page that sent it
leaves the pane while it is being typed, so the pane's other queued
messages are not stuck behind it.
On native Windows the PC sees only the pane's shell, so it cannot tell
that the agent quit: there the
  message keeps waiting until you discard it.
  ([#1532](#1532))
- GJC Chat hides internal agent-to-agent IRC messages and automatic
replies, while keeping user
  messages that quote them visible.
  ([#1535](#1535))
- GJC questions using the Nerd Font selection cursor now show their
actual choices in chat
instead of only Up, Down, Enter and Esc controls. Moving the cursor
preserves the question's
  identity and uses the current selection when answering.
([#1553](#1553) by
@Kinetic27)
- GJC chats on a busy machine give cold session lookups up to five
seconds per command and
ten seconds overall instead of timing out after two seconds per command.
Session snapshots
  retain the shorter status lookup budget; failed status lookups
  are held for 30 seconds without preventing chat reads from retrying.
([#1364](#1364) by @sskys18)
- GJC chat and status reuse their verified session binding when other
sessions write heartbeats,
instead of restarting the SDK lookup for every pane. Session changes
still require SDK validation.
  ([#1343](#1343))
- Grok keeps Chat with older sessions open and uses Terminal when the
session is uncertain.
([#1352](#1352) by
@infowolfe)
- One failed health check no longer shows herdr offline or switches a
Windows PC's panes to the
terminal lens: the header says Status unknown with the last herdr it
saw, and herdr unavailable only
  after a second failure in a row.
  ([#1581](#1581))
- A malformed, oversized or misaddressed answer from herdr now fails
only the request it answers,
with a clear error, instead of hanging until the timeout, growing the
server's memory or being
taken as another request's answer; event subscriptions report such a
frame and close.
  ([#1584](#1584))
- Hermes `/status` sent from Chat now shows its result card on the first
send in a new pane too, even
when the computer is busy. A send whose result was missed no longer
leaves every later `/status` in
that pane without a card for ten minutes. The same holds when another
device has the terminal
  scrolled up.
  ([#1543](#1543))
- Sending Hermes `/status` (or another informational command) from Chat
again now shows that send's
own result card each time, instead of "Runs in the terminal" once
earlier identical results fill
  the terminal.
  ([#1543](#1543))
- Slow API requests can return their result or timeout error instead of
losing the HTTP connection
  while bounded Herdr calls are still running.
([#1355](#1355) by
@christi4nity)
- The app can be installed behind a reverse proxy that signs in every
request (Authelia,
oauth2-proxy and the like): the browser now sends its cookies with the
manifest request.
([#1332](#1332) by
@ascandroli)
- Mouse wheel and phone swipes scroll terminal history on Windows and
other mirrored panes;
  read-only connections still only pan their local grid.
([#1446](#1446) by
@Davesaggio)
- A Codex started in a terminal shows its model and reasoning level
before the first message,
  from Codex's live footer, while the chat shows the terminal output.
([#1374](#1374) by
@evanwy0223)
- Before the first message, a model switched in the terminal shows in
the message box once the
supported agent draws it. omp still waits for its transcript to record
the model; the bridge
never runs `omp models`, which can refresh the CLI's OAuth credentials.
([#1374](#1374) by
@evanwy0223)
- A pane that runs OmO is now treated as OmO by the composer even while
herdr labels it Claude
(OmO's Claude child is running): sending a message no longer stops on a
Claude input-box draft
  refusal or tries to clear a Claude draft there.
  ([#1441](#1441))
- A multi-line message sent to an OmO pane that herdr has no agent label
for is now pasted as one
bracketed paste instead of line by line, so its first line is no longer
sent alone.
  ([#1441](#1441))
- An OmO pick list (an extension's select, such as the init-deep
advisor's "Run now / Skip this
time / Never…") shows in the chat as a card with its own options again,
instead of a "Waiting for
input" card offering only ↑, ↓, Enter and Esc. A pick list that counts
down in its title keeps
the same card while it ticks, so an answer tapped a second later is not
refused as stale.
([#1495](#1495) by
@nahwan-kim)
- An omp pane started with an absolute `--cwd` shows its chat when herdr
names no session file
and the directory herdr reports is not omp's own, such as the shell's or
a symlink's real path:
the session is still looked up in the directory herdr reports first,
then under that `--cwd`.
  ([#1594](#1594))
- A message or command sent from the chat to an omp that is still
starting no longer stays typed
in its input box: omp shows the box before it takes an Enter, so the
send waits until herdr's omp
integration reports the session (omp does so once it takes input), for
an omp started moments ago.
  ([#1475](#1475))
- A new OpenCode 2 pane offers your project and user skills as `@`
attachments before its first
session starts, using the OpenCode version the PC's install reports,
instead of listing them as
OpenCode 1 `/` commands. Once the pane has a session, that session's
version decides as before.
  ([#1543](#1543))
- Native question answers stay bound to the current pane, session and
form. Local service discovery
rejects stale or non-loopback endpoints, bypasses environment proxies,
and bounds slow requests.
([#1257](#1257) by
@bardiz12)
- Polling does not invalidate an unchanged question or turn an
acknowledged answer into an error.
A temporary service outage keeps an already displayed native card and
its unsent edits.
([#1257](#1257) by
@bardiz12)
- Back and question tabs preserve unsent edits, and multiselect
submissions include both checks
and custom text. Abandoned drafts are bounded and expire after 30
minutes.
([#1257](#1257) by
@bardiz12)
- OpenCode question cards keep edits and allow immediate retries after a
failed reply, and optional
questions can be skipped. Unrelated service metadata no longer resets a
form's progress.
  ([#1421](#1421))
- Answers sent from the message box show the translated warning when
OpenCode has not confirmed
delivery. Native service connections accept only the loopback addresses
127.0.0.1 and [::1].
  ([#1421](#1421))
- Subscription usage reads OpenCode 2's own credential store before the
`auth.json` OpenCode 1
wrote. OpenCode 2 keeps its provider sign-ins in `opencode.db` and
leaves that file standing with
the key it has since replaced, so the OpenCode card reported "Sign-in
expired" for a CLI that was
signed in and working. The store's active key answers first, and a store
without one (signed out of
OpenCode 2) shows no card; the file still answers beside OpenCode 1's
store, or with none.
([#1329](#1329) by
@krish-kant)
- `GET /api/pane/read` takes `lines` only as plain decimal digits from 0
to 4294967295 and refuses
any other value (blank, signed, fractional, exponent or too large) with
400 `invalid_lines`
instead of passing it to herdr; a blank value is no longer read as zero.
Every count herdr
  takes still goes through unchanged.
  ([#1571](#1571))
- A pasted or dropped PNG, JPEG, GIF or WebP image is saved with the
extension of what its bytes
are, not the one its file name suggests, and a file that only claims to
be one of those images
is refused with its reason instead of being saved for the agent to read.
  ([#1599](#1599))
- The phone header's Chat/Terminal switch keeps clear of the bar's top
and bottom edges in both
densities: its options take the density's segmented height, so its track
is 34px tall in the 52px bar
and 32px in the compact density's 46px bar, which it used to nearly
touch.
([#1519](#1519) by
@adnankassem)
- On a phone, the header's Chat/Terminal switch centers its icons and no
longer touches the header's
bottom rule (in compact density it was as tall as the bar). The buttons
still take a full
  44 px tap.
([#1367](#1367) by @Noobear)
- A pi chat now shows its context ring for models pi lists in its
commandcode cache or model
  store, not only for providers defined in `models.json`.
([#1391](#1391) by
@diegolhambi)
- Model and effort pickers no longer confirm a stale choice when the
agent stops waiting for the
  answer during the final screen read.
  ([#1435](#1435))
- A Portal that stops on its own after its address was up is started
again once, on the same relay
and address, and Settings → Phone & devices says Recovering… meanwhile;
a second stop shows the
error with Portal's output instead of leaving the address down without a
word or restarting it
  again.
  ([#1587](#1587))
- When the server cannot save a shared setting or pin, Settings now says
the change is not on the
server yet, that reloading or closing the page can undo it, and offers
Retry. After three failed
saves in a row the browser resends less often (after a minute, then up
to every five minutes)
until you press Retry or the connection reopens, and edits made
meanwhile are kept. A late resend
of a refused change no longer overwrites a newer value another device
saved meanwhile, while a
setting you pick again after the refusal still wins. While the
connection is lost it waits for it
to come back instead of resending, and a connection the browser gives up
on (for example a proxy
error while the server restarts) is opened again, so alerts and other
devices' changes no longer
  stop arriving until a reload.
  ([#1595](#1595))
- A damaged or unreadable shared preferences or read-finish file is no
longer lost to the next
change. Reading never touches it; the first change moves the damaged
file, byte for byte, to
`preferences.json.bad` or `seen.json.bad` (private to you, the latest
damaged copy only), a file
repaired by hand while the server runs is kept with the change added to
it, and a file that
cannot be read or set aside is not written over, so the change fails
instead.
  ([#1564](#1564))
- A server or remote bridge that crashed no longer blocks its restart
when its old process number
now belongs to another program: startup locks and the bridge registry
record the owner's start
  identity, and that other program is never signalled.
  ([#1567](#1567))
- Attaching to a terminal once again waits for input to be ready before
signalling completion,
preserving queued keys, clicks and text and the order of Stop followed
by a message.
  ([#1342](#1342))
- Opening a finished agent's terminal while its connection to herdr is
being restored no longer
clears its DONE dot from the screen left over from before: it clears
once the terminal shows
  current output.
  ([#1472](#1472))
- An oversized or unreadable answer from another PC's bridge is now
refused at a fixed size
instead of being held whole in memory: connecting, refreshing and the PC
proxy stop reading it,
report a short error in the user's language that repeats none of it, and
a watching device
  never gets a prompt that could not be redacted.
  ([#1588](#1588))
- Opening a pane in another tab or on a phone, or reconnecting, no
longer copies again to the
clipboard what a program in the pane copied before; its new copies still
reach the clipboard.
  ([#1600](#1600))
- Update and restart, and the header's Update, install exactly the
release they offered: a release
published in the meantime installs nothing and is shown to be checked
first, and a server whose
supervisor cannot install by revision asks for a restart instead of
installing.
  ([#1590](#1590))
- In macOS Safari, Backspace in a terminal no longer skips a press after
deleting the last letter of
a Korean syllable: 한 then four Backspaces now also deletes the character
before it.
  ([#1363](#1363))
- Chat shows only the current agent session's command results: after an
agent restarts in the same
pane, the earlier session's commands no longer appear. A `/clear`,
`/new` or resume sent from Chat
keeps its own row in the conversation it starts, without claiming that
it finished. Commands sent
before an agent's first reply stay with that agent while it keeps
running, also through a moment
in the background (Ctrl+Z on Linux and macOS); once a different agent
takes over the pane, they
are hidden, even if Chat was not open when it happened. A command sent
to a restarted agent while
the pane still shows the previous session stays once the agent's own
session appears within a
minute, unless Chat started another session after it. Within that
minute, a session change typed
in the terminal of an agent that resumed the same session can still keep
such a command. Where
Chat cannot find the agent's program at all (native Windows, or an agent
started under another
program name), such a command stays with the session that the same agent
starts in the same
terminal; a later session, another agent or terminal still hides it.
Outside Linux, an agent
restarted under the same process number cannot be told apart from the
one before it.
  ([#1543](#1543))
- SSH PC management controls now stay within narrow sidebars: the rename
field and action buttons
  reflow in place, and long labels wrap without clipping.
([#1392](#1392) by @snowykr)
- Opening a finished pane clears its DONE dot only after its answer or
terminal output is
actually visible. Covered panes, loading or failed chats, and terminal
reconnections keep
  unread finishes marked.
  ([#1342](#1342))
- Reading older chat keeps the finish unread. Jumping to the latest
answer clears it only
after the jump arrives; small scrolls above the end show the way back.
Resizing and restoring
  a short chat work without an extra scroll.
  ([#1342](#1342))
- An older answer cannot clear a newer finish. Agents whose chat cannot
prove completion keep
the dot conservatively; opening their Terminal view can acknowledge the
output.
  ([#1342](#1342))
- Read finishes stay read across pane switches and browser tabs on the
same device. Every
web-server restart or update brings read DONE dots back; read records
are not carried across
  server restarts.
  ([#1342](#1342))
- The sidebar's PC control buttons are the standard control height
again, Rename shares the name
field's row in a wide sidebar, Remove PC dims like its neighbours while
a rename or disconnect
runs, and danger buttons keep their red tint on hover everywhere in the
app.
  ([#1422](#1422))
- Documents opened through copied read links keep their own scroll area
at desktop widths,
rather than expanding beyond the window and leaving scrolling to the
page.
([#1586](#1586) by
@Kinetic27)
- A Markdown table whose separator has a one- or two-dash cell
(`|---|:-:|`) shows as a table in the
  chat, as on GitHub, instead of as its pipe-separated source.
([#1361](#1361) by @Noobear)
- Going back to the chat while a terminal command's screen is still open
(Claude Code's `/btw`
answer, a picker left up) no longer lets the next message from the
message box be typed into
that screen and do nothing, for Claude Code and the agents above (not
OpenCode): the composer
says the command is still open in the terminal, with Show terminal, and
Send (or a queued
message's Send now) keeps the message until the screen closes or you
choose Send anyway.
  ([#1637](#1637))
- Opening the app no longer flashes the dark amber default before your
saved theme: the saved light,
dark or single-mode theme (Dracula on the light side stays dark),
including System and older saved
palettes, is drawn from the first frame, and the installed app's title
bar color matches it at once.
  ([#1569](#1569))
- Settings → Updates no longer tells you to switch to main when the
server was started from an
installed update's folder instead of its source checkout: it now says to
restart the app from
  the source checkout, which is what lets updates run again.
  ([#1417](#1417))
- Typing or composing text in a wide mirrored terminal no longer pulls
the view sideways.
Windows terminal wheel panning stays in place when output arrives, and
tapping the terminal
  without scrolling no longer changes shared history.
  ([#1534](#1534))
- Loading the page no longer sends about 33 requests per workspace
folder to find its project icon:
each folder is asked once (`GET /api/fs/favicon`), and a folder without
an icon is an answer, not
a 404. A reverse proxy with a rate limit no longer answers 429 to the
burst, and the browser
console no longer fills with red 404 lines on every load. A 429 is no
longer remembered as "no
  icon" either: the folder is asked again when its row next appears.
  ([#1339](#1339))
- A dragged workspace order no longer jumps back while its moves are
still landing: a roster read
before the move, a cached roster or another PC's update cannot undo it,
and the first roster
read after the move (another browser's later order included) replaces
it. A remote PC does this
once its bridge is updated to the next release; until then it keeps the
earlier behaviour.
  ([#1593](#1593))

### Maintenance
- Composer geometry regressions freeze the demo conversation while
keeping it idle, so turn
completion cannot change the transcript during measurements on phone and
desktop.
([#1347](#1347) by
@Yoonwoo-Ha)
- Settings and PWA draft checks target the visible input explicitly;
locale regressions cover
  native draft and placeholder heights on phone and desktop.
([#1347](#1347) by
@Yoonwoo-Ha)


</details>

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
devswha added a commit that referenced this pull request Oct 28, 2026
#2197)

## Change

Part of #1468 (follow-ups of #1463). The server's shared read finishes
(`seen.json`) kept the
records of herdr runs that had ended (an old `finish_epoch`) until the
500-record cap pushed
them out. Every events-stream reconnect sent them to each browser, which
wrote them back into
its localStorage, and every change rewrote them into `seen.json`.

- `server/seen.ts`: `SeenStore` takes each PC's current finish epoch.
`GET /api/seen` and the
stream's first `seen` frame leave out a PC's records of another epoch. A
`POST` of one is
answered as sent (200, the client only reads the status) and neither
kept nor told. The next
write drops them from the file, before the cap is applied. A PC whose
epoch the server does not
  know keeps every record.
- `server/machines.ts`: `MachineManager.finishEpochs()` names the local
PC's epoch, and a remote
PC's only once a snapshot or status of the PC itself reached this server
since it started
(`runtime.heard`). A snapshot restored from `machines.json` is not
evidence: it is saved about a
second after a status frame moves the epoch, and `stop()` cancels a
pending save, so it can name
the run the PC has since left (Codex review, MEDIUM, fixed in the second
commit). A setup move
that is rolled back restores that mark with the snapshot (third commit,
Codex LOW).
- DESIGN.md's read-state entry and a changelog fragment.

No HTTP/WS shape changes: the same routes and `seen` event, with fewer
rows.

### Left in #1468, with reasons (comment on the issue)
- The MEDIUM resend after a transient error was fixed by #1747.
- Batching `seen.json` writes: a change is held and told only once it is
on disk (#1564's
contract). The file now holds only runs still going, which bounds the
rewrite much lower than
  the 1.5 MB worst case of 500 records.
- A POST with the largest finish ID: it comes from an authenticated
browser of the same server
and only marks its own read state until that herdr run ends. A cap would
need the server's own
per-pane finish view, which lags status frames, so it would risk
refusing real finishes.
- An older cached client dropping a snapshot poll on a `seen` event: an
already-cached client
  cannot be changed from here; a reload ends it.

## Validation

- `server/seen.test.ts` "leaves out and drops the finishes of a PC's
herdr run that ended, and
  keeps a PC's whose epoch it does not know": fails on main, passes now.
- `server/machines-refresh.test.ts` "names a remote PC's finish epoch
for read finishes only once
the PC told it, never from a saved snapshot": fails before the second
commit, and fails again
  with the `heard` check removed (mutation), passes now.
- Real HTTP check on a server against the isolated `ovis-test` herdr
session (mkdtemp stateDir,
port 0): `seen.json` seeded with a local record of an ended epoch and
one of a PC not in the
roster. `curl GET /api/seen` left out the ended one and kept the other;
a POST for the ended
epoch answered 200 and was not stored; a POST for the current epoch
rewrote `seen.json` without
  the ended record. (Scratch script, not committed.)
- `HERDR_TEST_MODE=unit bun test server/machine-own-update.test.ts
server/machines-refresh.test.ts
  server/seen.test.ts`: 101 pass, 0 fail.
- `bun run check fast` on the pushed head: exit 0 (7969 unit pass, 0
fail; typecheck, build).
- Reviews: omo-native-code-reviewer APPROVE (its LOW-1 is the same
restored-snapshot window,
fixed); Codex Sol 6.1 Fast xhigh REQUEST CHANGES (MEDIUM, fixed in
26da453), re-review APPROVE
with one LOW (fixed in 9589b7c). The rollback LOW has no new test: it
needs a PC that never
  connected plus a save fault; the existing #2075 rollback tests pass.

No release-guard change: `REMOTE_BUNDLE_VERSION` is raised only by
release PRs.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant