- remove() drops the session's slot once the file is gone: a session that comes
back with the same id (reboot restore) starts afresh instead of answering
`removed` until the next restart. A run in flight stays on its own slot's store,
and the routes start a sync in the same tick as their session check, so a
deletion right after it still stops that sync. A store that is not loaded is
deleted without being parsed (deleteStore).
- A new conversation (/clear, resume) first reads the previous transcript, next
to the new one, to its end with its subagents, so edits it got after the last
read are not lost with its cursors. "Load older" reads the stored conversation
while the switch is pending.
- At most MAX_LOADED_STORES (3) stores stay in memory; the least recently used one
without a run is written if it changed (settle) and loads again on demand.
- Tmp files are named per process start, so a reused pid (a container's pid 1)
no longer keeps a stale one through the startup sweep.
- load() keeps only a plain transcript id (the previous transcript's path is built
from it) and no cursors without one; a pending Bash call with an id, cwd or
agent id over 4,096 characters is not kept at parse time either.
- nextRange() is gone: production reads forward through the reader. The store
tests mirror the reader's range choice, and a new end-to-end test runs the
real reader, sync and store (/clear cases, 40 seeded random transcripts).
- The wiring guard scans all of src for holders of the sync.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Forward: after a line longer than the budget, the next line is read from
its start instead of being skipped with it when it crosses the chunk end.
- Backward: the line before a skipped long line is read whole; a floor that
turned out to be a line start no longer stalls.
- A chunk the shared forward budget cut short is not applied (it read as one
over-long line) but read in a later pass, so a first open of a long main
transcript reaches the subagents in the same sync.
- Load older reads a full budget block of up to three transcripts per click,
taking turns, instead of splitting one budget (a normal Write longer than
the share was skipped for good).
- A subagent transcript that vanished loses its cursor, so Load older does not
stay on forever.
- POST /api/cleanup-state forgets the history of every record it drops; a
startup sweep deletes edit-history files no session or record owns and
stale tmp files of other processes.
- load() rebuilds entries from known fields (string diff lines, per-edit cap
re-applied, ids counted against the byte cap), bounds pending Bash fields
and agent ids, and does not parse a file over 32 MiB.
- No re-read after Codeman's own keep-alive turn.
- Tests: regression tests for each of the above (they fail on the previous
commit); the remove-during-write test now meets the write under way; the
sync and route tests remove their stores before the temp dir, so no late
save recreates it in /tmp.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- The history is deleted when the session's persisted record goes (forgetEditHistory):
kill via cleanupSession (a pinned record is only demoted and keeps it), the
persisted-only DELETE, and stale cleanup at startup. After remove() neither a sync in
flight nor a later request writes the file again; writes are serialized.
- load() re-bounds a damaged file: cursors, pending Bash calls (MAX_PENDING_BASH) and
subagent labels (MAX_AGENTS, oldest unused dropped first).
- Load older reads one block of every transcript, budget split between them, and only
the files it asked for; a block short of budget reads its end. A line longer than the
budget is skipped going back as well (floorMidLine).
- Concurrent first callers share one store; a forward request folded with an older one
still reads forward.
- Images outside the workspace get no thumbnail (one attachment id per render); a click
opens the shared preview, which registers once with notify: false.
- Docs: per-edit diff cap, deletion rule, load-time bounds, thumbnails, follow triggers.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Part 1 of #643 (local sessions only; remote follows as its own PR). An
opt-in, per-device header button opens a panel with the active session's
edits, newest first: every Edit/Write/MultiEdit/NotebookEdit of the main
agent and its subagents with its diff, plus images a Bash call rendered or
the agent sent. Remote, Docker and non-claude sessions answer `unsupported`.
- Reads only on request: `?sync=1` when the panel opens; while it is open on
that session it re-reads after an edit-capable `transcript:tool_end` or a
`transcript:complete` from the existing transcript watcher. No hook, timer,
second watcher or new SSE event; a wiring guard pins it.
- Main transcript: the path the transcript watcher follows (hook-reported),
else `<id>.jsonl` under the session's own CLAUDE_CONFIG_DIR override or
the server's config dir. A config dir set outside Codeman is found with
the next hook; until then the panel says so.
- Store: dataPath('edit-history')/<id>.json, 0600 via tmp+rename, capped at
500 entries and 4 MiB of diff text per session, deleted with the session,
blocked in isSensitivePath, never fed to /api/search. Every route goes
through findSessionOrFail with req.
- Images only through existing file surfaces: file-raw inside the
workspace, the attachment route with notify:false outside it.
- textContent rendering, bounded byte-offset reads, marker-class header
button, hidden on phones (locked in the mobile policy test).
Docs: Working-With-Files, Settings-Reference, architecture invariants, API
reference, CLAUDE.md key pattern.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Second merge of #674: the author pushed the thumb-arc handback from the review themselves (e8fc08be), byte-identical in code to the merge-time fix already here, so this brings their round in and keeps the browser test and docs added at merge.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
# Conflicts:
# docs/wiki/Mobile-Guide.md
# src/web/public/terminal-ui.js
# test/session-swipe-gesture.test.ts
Conflict in CLAUDE.md's State Files line resolved by keeping #661's new
config-backup entries and this PR's "case default groups" in state.json.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Conflicts in CLAUDE.md, README.md, docs/api-reference.md and
src/web/server.ts were additions beside #662's Copilot quota read and
#666's keep-warm sections; both sides kept.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A scroll whose first move past the tap slop leaned sideways (a thumb arc) locked as 'swipe' for the whole contact, and SwipeHandler drops a swipe at 50 px of vertical drift, so the gesture did nothing. Past that limit the drag is now a 'y' scroll. Also brings the mobile-handlers fileoverview to the current 500 ms / 50 px limits and notes in the Mobile Guide that a selection pauses the swipe and a wide terminal pans instead.
Squash of PR #658. Discovery records each llama.cpp model's parallel slot
count (from --parallel/-np, or /props total_slots when the launch command
states none), and a Claude launch warns when the context split across
those slots leaves less than 40K tokens per request. An explicit
--parallel beside a /props context records a slot count of 1, so a later
round overwrites an earlier count, and a non-2xx /props is remembered per
launch command while a network error is retried.
Merge-time fix by the maintainer: the architecture-invariants sentence on
the explicit --parallel case now says it records a slot count of 1.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merge-time fix for #676 (review finding 3 and nit 9). The State Files entry
for state.json names the new caseDefaultGroups key, and the Owner tab layouts
paragraph says where a case's default session group sits in the placement
order of sessionCreated() and that only TabLayoutService reads or writes it.
The wiki endpoint map counts the two new case routes (37 to 39).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Squash of PR #662. The header plan-usage chip gains a Copilot row: the
calendar-month premium-request quota, with the request counts and the
reset in the tooltip, read every 10 minutes from GitHub's
copilot_internal/user endpoint with the Copilot CLI's own sign-in token.
It rides showPlanUsageLimits and is on by default while the CLI is
installed and signed in; any failure leaves the chip without the row.
Merge-time fixes by the maintainer: the gates (setting, then token, then
the CLI probe) live in readCopilotPlanUsageIfEnabled() so a host with no
Copilot sign-in never runs the login-shell probe each tick, with a test
that Plan Usage Limits off means no request; the X-GitHub-Api-Version
header the Copilot CLI itself does not send is dropped; and CLAUDE.md,
architecture-invariants, README and copilot-integration describe the read.
The hooks-config import in server.ts was merged with #666's.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The remote-history PATH bootstrap that rode along in this PR was left out
at merge (it belongs in its own PR that routes every remote tmux call
through one wrapper), so 242bbfcd's title overstates what lands here.
Conflicts in CLAUDE.md and docs/architecture-invariants.md resolved by
keeping this PR's rewrite of the wheel/touch paragraphs and re-adding
#664's touch-forwarding sentence to both.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Squash of PR #669. Browser terminals could be moved onto another session
by tmux's prefix keys: Ctrl+B then ), (, s or w switched the tab's tmux
client to a different session while the tab kept the original's name.
Codeman's tmux socket now keeps only Ctrl+B Ctrl+B (send-prefix) in its
prefix table, and codeman tui and scripts/tmux-manager.sh get F1 as their
way out instead of Ctrl+B d.
Merge-time fixes by the maintainer: on an empty prefix table the
unbind-key that fails no longer skips binding send-prefix; a key tmux lists
escaped (prefix C-\ shows as C-\\) compares equal, so the reset does not rerun on every create;
and reconcileSessions() reduces the table at boot when the socket has live
sessions, so sessions that survive an upgrade are covered too.
Co-authored-by: christianhaberl <116439789+christianhaberl@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merge-time fix for #665.
- A markdown link converted to a preview link became `a.rv-path`, so its
label ("Our incident record") took the path style: monospace, 0.92em
and `word-break: break-all`, which split prose mid-word on a phone. The
anchor now also carries `rv-path-link`, and
`.rv-text a.rv-path.rv-path-link` resets font and word-break, as the
File Viewer does for its labelled document links.
- The first new test adds `/tmp/a.json?raw=true` (pins the `?query`
strip) and `/tmp/` (pins the trailing-slash folder check, since
findFilePathLinks finds no folder in a bare `/tmp/`), and checks the
new class.
- The `_linkifyFilePaths` JSDoc and docs/architecture-invariants.md say
the function also rewrites anchors whose href is a scheme-less,
non-`//` path, on purpose without FILE_PATH_LINK_PATTERN's root and
extension filter.
- Mobile Guide: "a web link" in the response viewer opens a tab.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merge-time fix for #661. The pre-restore copy was forced, so a restore made
while the live files equalled the newest backup spent a retention slot on a
duplicate. It now goes through the normal change check: when nothing changed
since the newest backup, no copy is made and safetySnapshot names that
backup. Two tests that relied on the forced copy now change a file before the
restore; a new one pins the reuse.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merge-time fix for #674: the mobile-handlers.js @fileoverview still
gave the old 300 ms and 100 px limits (now 500 ms and 50 px), and the
Mobile Guide now says a visible text selection pauses the swipe and a
terminal wider than the screen pans instead of switching.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merge-time fix for #659.
- Alt+Left/Right (File Viewer back/forward) is a display-only entry in
DEFAULT_SHORTCUTS, so the Ctrl+? overlay lists it, and a row in the
help modal's Panels group, pinned in help-modal-shortcuts.test.ts.
zh-CN entry for the new label; the "Restore size" entry is dropped,
since nothing shows that string any more.
- The wiki row moves from "Sessions and tabs" to "Everything else".
- The grip comment in styles.css says only phones hide the grip;
tablets show and use it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merge-time fix for #661. A restore that could not write one file (a full
disk, a directory where the file should be) used to stop there and answer a
bare 500, naming neither the files it had already put back nor the
pre-restore copy, and the Settings form was not re-read, so the next Save
wrote the old settings back over a restored settings.json.
restoreConfigBackup now writes every verified file it can, then throws
ConfigBackupPartialRestoreError with restored, failed (name and error),
untouched and safetySnapshot. The route runs the same reload hooks and audit
for the files written as a full restore does, and answers 500 with the error
envelope plus those fields (and restartRequired) in data, as the tab-layout
409 carries its layout. Settings refills the form when settings.json is among
the restored files, also on that path. Tests: a unit case where a later file
is still restored, the clis.json-as-a-directory route case, and two vm cases
for the Settings handler.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merge-time fix for #666. `recordHookEvent()` ran for every `stop`, the run
summary files a `stop` as a warning, and the Away digest lists warnings under
Needs Attention, so each keep-alive nudge added a row there and to the
session's Summary timeline. The hook route now skips the run summary while
`keepAliveReason` is set (the keep-alive turn's `stop` and `idle_prompt`).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merge-time fix for #666. A nudge only goes out inside the last five minutes
before `expiresAt`, so the check armed a whole `KEEP_WARM_LEAD_MS` after a
failed write landed at or past the expiry and stopped with "too close to
expiry": the retry the status promised never happened. The window
`beginKeepAliveTurn()` opened also stayed open for its ten-minute ceiling
with no turn running.
A failed write now ends the window (`endKeepAliveTurn()` joins
`KeepWarmSessionView`; `Session` already has it) and re-arms at
`min(now + KEEP_WARM_RETRY_MS, expiresAt - KEEP_WARM_LAST_CHANCE_MS)`. The
test advances the clock and sees the second write, and a second one pins
that the retry never lands past the last-chance line.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Case settings gains a "Default session group" selector (Ungrouped plus every
named tab group of the owner). A new session of the case lands in that group
in the same versioned layout write that registers it, so it is drawn in place
with no Ungrouped-to-group jump.
- Stored per owner and case in state.json (caseDefaultGroups), as the group's
stable id or null for Ungrouped; read and written only through
TabLayoutService under the owner lock.
- Precedence in sessionCreated: explicit tabGroupId, then the parent
placement, then the case default, else Ungrouped. A deleted or foreign id
never fails a spawn and is forgotten only after the creating write lands.
- POST /api/quick-start (local, linked, remote, Docker) and
/api/ralph-loop/start pass their case; POST /api/sessions infers it
lexically from workingDir. The create response carries tabLayout when a
case default placed the session.
- GET/PUT /api/cases/:name/default-group; deleting a case (or an admin
deleting a user's case folder) forgets the value.
- The selector mirrors desktop and phone, reverts on a refused save, and is
redrawn from every adopted tab layout; zh-CN strings added.
The tab layout's 512 refs also count pinned sessions kept as stopped records, so
the ceiling holds at the default MAX_WEBVIEWS and with no pinned closed sessions.
SKILL.md, verbs.md (both copies) and the cron guide no longer give a fixed 50/25.
Co-Authored-By: Christian Haberl <116439789+christianhaberl@users.noreply.github.com>
Co-Authored-By: Claude <noreply@anthropic.com>
An owner's tab layout holds at most MAX_TAB_REFS (512) refs, counting saved web
tabs (up to MAX_WEBVIEWS, 50 by default) as well as sessions. With a ceiling of
500, session creation could still fail with HTTP 500 on the layout limit before
the capacity message. 512 - 50 = 462; a test pins the ceiling plus MAX_WEBVIEWS
inside MAX_TAB_REFS so the two limits cannot drift apart again.
Co-Authored-By: Christian Haberl <116439789+christianhaberl@users.noreply.github.com>
Co-Authored-By: Claude <noreply@anthropic.com>
It does match one Claude shows inline when the first line has 24+
non-space characters; it misses a short first line and the
[Pasted text #N +k lines] placeholder. Say "cannot reliably" instead of
"cannot".
Co-Authored-By: Christian Haberl <116439789+christianhaberl@users.noreply.github.com>
Co-Authored-By: Claude <noreply@anthropic.com>
Review follow-up. Other bounds assume a small session count: an owner's
tab layout holds at most MAX_TAB_REFS (512) refs and PUT
/api/session-order accepts at most 500 ids, so with
CODEMAN_MAX_SESSIONS=600 the 513th session creation answered HTTP 500
instead of the capacity message. The value is now clamped to a named
MAX_CONCURRENT_SESSIONS_CEILING of 500; any digit string above it
(including one past the safe-integer range) resolves to 500 instead of
falling back to 50. The resolver reads process.env directly, the unused
env parameter is gone.
Docs: a CODEMAN_MAX_SESSIONS row in the Settings Reference environment
table, the env name next to the cap in the cron guide and the
architecture invariants, and both copies of the skill's endpoints
reference now give the per-user cap as half the global cap (25 by
default). The changeset says "1 to 500". The test drops the
`as unknown as string` casts and adds '1000', '501' and an
out-of-range value, each refusing at 500.
Co-Authored-By: Christian Haberl <116439789+christianhaberl@users.noreply.github.com>
Co-Authored-By: Claude <noreply@anthropic.com>
The first commit left the old rule standing in places: the gotchas list in
extending-codeman.md still said newlines were stripped and that
"echo A\necho B\r" ran `echo Aecho B`, and so did both READMEs and the
skill's big-input section. Those now describe the paste.
Newlines are only kept on the mux path, so the docs that describe the paste
say "with useMux": without it a multi-line body is not a plain prompt and is
written straight into the pane.
Cron keeps rejecting multi-line prompts, but its docs and the code comments
in cron-service.ts, schemas.ts and respawn-controller.ts gave a reason this
change removes (writeViaMux joining lines, a newline submitting in Ink). The
reason that remains is the submit check, which cannot recognise a
multi-line paste in the composer, and for cron's paste mode the newlines
reaching the CLI as keys.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
TmuxManager.sendInput() removed every `\n` before the text reached tmux, so a
multi-line prompt sent with `useMux` (or a multi-line respawn prompt) arrived as
one line with its lines run together: `"echo A\necho B\r"` ran `echo Aecho B`.
The text half of a submit has gone in as one bracketed paste
(buildPasteTextCommands) since the Codex paste-burst fix, and inside the paste
markers Claude Code and Codex insert a pasted line break into the composer
instead of submitting. So the `\n` is now kept (`\r\n` loses only its `\r`) and
the existing paste carries the lines; the Enter stays a separate tmux call.
Multi-line text without an Enter is pasted too rather than typed with
`send-keys -l`, where every line break would reach the pane as a key.
Single-line input is unchanged.
The test-mode gate moves from the top of sendInput() into inputRunner(), as
readPaneRows() does, so a test subclass can record the commands sendInput()
would run instead of pinning the branch shapes on the source.
SubmitVerifier still compares the composer's last glyph line with the prompt
minus its newlines, so it does not reliably recognise a multi-line paste and
may not re-press an Enter that Claude Code ignored. `agent send` and the TUI
therefore keep refusing or joining multi-line input; the docs that described
the stripping now describe the paste and that limit.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Review round 3 on #661: retention aged each snapshot by its own createdAt, so the first change after a quiet month deleted the last good copy; it now ages from when the next snapshot was made. users.json lastLoginAt, settings.json lastUsedCase and intents.json recentPrompts/updatedAt are volatile keys for the change check. The keep count has a minimum of 2. Wiki restore paragraph aligned with RESTART_REQUIRED.
Review round 2 on #661: the hostname in the origin orphaned every snapshot when a Docker container was recreated, so the origin is a random id kept in <data dir>/config-backup-id; after restoring settings.json the Settings form is re-read from /api/settings so Save cannot write the old values back; the restore confirm names the files and the credential ones; clis.json is reloaded (reloadCliRegistry) and users.json needs no restart; API reference and route docs corrected.
Review round on #661: webhook.json and intents.json are redacted in compare and URL values lose userinfo/query; a custom folder is probed (bounded, pastCap) before each pass and in each route and only created when its parent exists; manifests record an origin and change check, prune, list, restore and diff are scoped to it; a restore reloads IntentStore, lists untouched files and treats settings.json as restart-required; remote-cases.json is tracked; stale .tmp-cfg-* folders are swept and a failed restore unlinks its temp; the group is hidden from non-admins; an empty days field saves 30.
Settings -> System -> Config backups gains a Compare view: a backup against the
running config, a backup against the one before it, or any two backups. Each
differing file lists its changed keys as before/after, plus files that appeared
or vanished. GET /api/config-backups/diff?from=&to= serves it (`to` defaults to
`current`; admin only in multi-user mode).
Credential values are never returned: .env, users.json and push-keys.json report
names only, and any other key that looks like a token, secret, password,
credential or key is reported as changed without its value. A non-JSON file
reports only that it differs, a file is cut at 200 changes, and a self-refreshed
discovery timestamp does not count as a difference.
custom-model-hosts.json carries a lastDiscoveredAt per host that the 5-minute model
discovery poll rewrites, so the content hash never matched the latest snapshot and a
snapshot was written every tick. At the default keep count of 20 that left about 100
minutes of history and evicted every useful older copy.
The change check now looks through volatile keys per tracked file. The snapshot still
stores the file as written and restore still verifies against its own digest.
- New wiki page Config-Backups: what is backed up (and what is not, and why), when a backup is made,
where and under what name (cfg-YYYYMMDD-HHMMSS, manifest.json, reasons), retention and its
guarantees, the Settings controls, restore (verify, pre-restore copy, atomic write, which files
need a restart), security, the HTTP API and troubleshooting.
- Linked from the Home index and sidebar; Settings Reference, Troubleshooting ("my settings were
reset"), Security and Core Concepts updated; README and CLAUDE.md entries; the api-reference
section now states the naming convention and the full tracked list.
- Found while documenting: remote-hosts.json, docker-hosts.json, docker-cases.json and
webviews.json are user configuration too and were not tracked; they are now. state.json stays
out on purpose (rewritten per session event, also holds cron jobs and respawn presets) and the
docs say so.
The files that hold a user's customization (settings.json, clis.json, custom-model-hosts.json,
webhook.json, intents.json, linked-cases.json, push-keys.json, users.json, .env) are plain JSON
in the data dir, rewritten by the server, the Settings UI and the CLI, with no copy to go back to.
A snapshot is taken at startup and whenever the tracked files' content changes (checked every
five minutes, and only written when a digest of the file set differs from the latest snapshot's),
into `<data dir>/backups/config` by default or any absolute folder. Retention keeps the newest
configBackupKeepCount (1-500) and, when configBackupKeepDays > 0, drops older ones; it only ever
deletes folders it made (name pattern and a parseable manifest) and never the newest.
GET/POST /api/config-backups list and back up now; POST .../:id/restore verifies every file against
the manifest's SHA-256, takes a pre-restore snapshot, then replaces files atomically with mode
0600, writing only allowlisted names. Snapshots are 0700 folders of 0600 files and all routes are
admin only in multi-user mode. Settings -> System gets the controls and a Restore list.
Claude Code draws its prompt suggestion in an empty composer, the first letter in the reverse-video cursor cell and the rest faint, so a plain capture read it as typed text and keep-warm never nudged those sessions. The keep-warm pane read now uses capture-pane -e and drops the suggestion before checking the composer.
Also from review: an off status carries no window, automation and an unreadable composer retry instead of stopping, a failed evaluation re-arms a retry, a trigger during an evaluation queues one follow-up that remove() cancels, and the cap's clock is seeded from lastSubmitAt for pre-upgrade state.
Only a human spends an idle alert, but the keep-alive turn resolved an unseen one three ways: its working transition, its stop hook, and its own idle prompt replacing the item with a pre-acknowledged one. While the turn runs, a pending idle item is now left exactly as found, the page's stop handler returns on the keep-alive reason before clearing hooks or raising Response Complete, and session:working/session:idle carry keepAlive so the stuck-detection timer neither restarts nor re-arms.
The window that marks the turn closes on its own idle prompt or any other submit instead of a fixed three minutes, which a slow turn's idle prompt (sent a minute after the turn ends) could outlast. The service keeps the default and cap in memory, read through a strict reader so a mid-write read is retried rather than kept, and the cap is clamped before the strict settings save.
Claude Code lets a subscription's one-hour prompt cache expire after an hour idle, and the first turn back re-writes the whole conversation at the cache-write rate; nearly every gap over an hour goes cold, while the saving is concentrated in gaps of one to four hours. Codeman now sends one short keep-alive turn five minutes before the expiry Claude Code reports on its statusline, so the next real prompt still reads the conversation from cache, and stops two hours (configurable) after the last real prompt.
Codeman owns the idle policy for these sessions, as agreed in Discussion #554: a launch with keep-warm on also carries idleCompaction: false in its one --settings JSON, so Claude Code's own idle compaction cannot replace the conversation with a summary a minute before the nudge. It is decided once per launch (keep-warm on, the statusLine exporter injected, Claude Code 2.1.290 or newer, an unknown version fails closed like --name), recorded on the mux record, and picked up by any relaunch; a launch with keep-warm off is byte-identical. A session switched on after launch is kept warm best-effort and the readout says so.
Every guard lives in one pure rulebook: the opt-in (an agent-spawned session does not inherit the app default), a launch that reports cache state, a warm 1h cache with enough tokens at stake and no compaction this gap, a live pane, no usage-limit pause, no respawn, Ralph, task or waiting agent, idle, no open dialog, an empty composer and no recent keystroke. The nudge is a submit and moves lastSubmitAt, which the response viewer's conversation correlation needs; the cap counts from its own persisted lastRealPromptAt instead. Read My Mind drops the exact nudge text, and the turn's stop hook and idle prompt open pre-acknowledged so nothing alerts. The last cache report and the clock survive a restart, so the timer re-arms.
