#4771 added an explicit relay-failure 409 (code: gateway_run_unavailable)
to /api/approval/respond so a gateway approval whose run is gone keeps its
card actionable instead of silently failing. But the 409 was emitted
regardless of backend.
On the DEFAULT local in-process backend, every guarded command parks an
entry in tools.approval._gateway_queues (via _await_gateway_decision), and
the local streaming notify callback mirrors it into _pending tagged with
_GATEWAY_MIRROR_FLAG. There is no gateway run and no _STREAM_RUN_IDS entry
by design. _gateway_pending_approval_without_run_id() therefore returned
True for that purely local approval, so the handler 409'd
("active run unavailable") and refused to resolve an approval that resolves
perfectly well locally. Local-backend users hit this on every approval
click (reported by b3nw on Discord; independently filed as #4948 by
claw-io, which also notes the local-runtime trigger).
Fix: gate the 409 on the WebUI actually running the gateway chat backend
(webui_gateway_chat_enabled). Local approvals fall through to the local
resolution path; the legitimate #4771 gateway behaviour is unchanged.
Tests:
- New tests/test_issue4771_local_approval_regression.py drives the real
_handle_approval_respond handler and asserts: local backend -> 200 +
agent thread unblocked; gateway backend -> 409 preserved. These fail
before the fix (409) and pass after.
- The two existing #4771 tests that asserted the 409 are updated to pin
HERMES_WEBUI_CHAT_BACKEND=gateway at respond time, since the 409 is now
gateway-deployment behaviour.
Reported-by: b3nw
Reported-by: claw-io (#4948)
/api/sessions re-ran the expensive CLI/cron session projection on every
poll while a turn was streaming — a major cause of the multi-second sidebar
latency + 100% CPU on cron-heavy installs (continuing #4672/#4808/#4889).
Root cause: the CLI/cron projection is gated by _CLI_SESSIONS_CACHE, whose
key folds in _sqlite_file_stat_cache_key -> _sqlite_content_fingerprint
(MAX(rowid) FROM messages). During a live turn the gateway writes a message
row per streamed delta, so that fingerprint advances on essentially every
~5s poll — busting the cache and re-running the full candidate-join +
projection (and the lineage-metadata pass) on every poll, contending for the
same SQLite/global lock the streaming worker holds.
The route-level session-list cache already froze its key during streaming
(#4808 _session_list_cache_streaming_freeze_marker), but that freeze never
reached this inner CLI-sessions cache, so the heavy CLI/cron query still
re-ran whenever the outer cache validated.
Fix: while any stream is active, the CLI-sessions cache key folds in the same
stable streaming-freeze marker (keyed only on the set of active stream ids)
instead of the volatile content fingerprint, and the cache TTL widens to a
streaming window. The projection is reused across polls mid-stream and
rebuilt at most once per window. The instant a stream starts/stops the marker
changes, so freshly-finished rows surface promptly. Structural mutations
(cron completion, new/renamed/archived sessions, attention) now also clear
the CLI cache via the existing session-list-changed listener — those signals
never fire per streamed token, which is what makes the freeze safe.
Idle behavior is byte-identical (freeze only engages with active streams).
Applies to both the active-profile and all-profiles cache paths.
Adds tests/test_issue4842_cli_sessions_streaming_freeze.py (6 tests): marker
idle/stable/changes semantics, the core 'key stable across message writes
while streaming' guarantee, idle key still advances with the fingerprint,
streaming TTL > idle TTL, and the structural-mutation listener clears the
CLI cache.
#4932 added a second 'const resultsByTid={}' (in _collectToolResultSnippetsByTid)
EARLIER in ui.js than the renderMessages derived rebuild, so the #4927 test's
UI_JS.index('const resultsByTid={};') sliced the wrong region and 4 assertions
failed. Re-anchor on 'const fallbackToolSources=[];' which is unique to the
derived-rebuild block. Pure test fix, no behavior change.
Codex+Opus gate (both reproduced): the new mixed-content ordered render path
resolved tool snippets only from the S.messages join (_collectToolResultSnippetsByTid)
and never consulted persistedSnippetByTid (S.session.tool_calls) — so on a
cold/paginated load where the tool_result join misses (#4927 scenario) the inline
card rendered an EMPTY body, AND its tid was added to transparentOrderedToolIds,
suppressing the post-loop derived card that WOULD have recovered via #4927. Build
a transparentPersistedSnippetByTid map from S.session.tool_calls and thread it as a
5th arg to _transparentOrderedToolCall, falling back in both the live and non-live
branches. Also gate the transparent-only scans (S.toolCalls + _collectToolResultSnippetsByTid)
behind isTransparentStream() so non-transparent renders skip the O(messages×parts)
work (Opus perf finding). Redaction (#4926/#4928) was already safe (shared buildToolCard).
+ cold-load persisted-fallback regression test.
A settled turn that recorded no live duration fell back to
(Date.now()/1000) - session.pending_started_at. But pending_started_at is the
start of an IN-FLIGHT turn, so for a settled turn it is either stale (a session
that sat idle) or about a different pending turn — rendering a bogus
'Processed 15h 32m' on fresh conversations (#4930). Gate the fallback on a turn
actually being in flight (active_stream_id || pending_user_message); otherwise
return undefined (no duration shown) rather than a fabricated one. The recorded
live duration (_turnDuration / base.turn_duration) still takes priority, so this
only removes bogus durations, never a correct one.
Codex gate (the key one): the persistedSnippetByTid fallback read S.session.tool_calls,
but loadSession() keeps the messages=0 session object (tool_calls []), and the
real messages=1 summary arrives only as the sessionToolCalls argument to
_syncToolCallsForLoadedMessages() — never copied onto S.session. So the fallback
source was empty on exactly the cold-load path it repairs. Persist
sessionToolCalls onto S.session.tool_calls in _syncToolCallsForLoadedMessages
(covering _ensureMessagesLoaded, older-message paging, and full-history loaders)
so renderMessages builds a populated persistedSnippetByTid. + guard test.
Codex gate: the third derived-push site (Anthropic tool_use content-array branch,
ui.js) still read resultsByTid[tid]||'' without the persisted fallback, so
Anthropic-format sessions could still cold-reload with an empty tool body. Add
the persistedSnippetByTid fallback there too; now all derived result-snippet
reads (OpenAI top-level, Anthropic tool_use, partial) chain to the durable
persisted snippet. Added a guard test asserting EVERY derived resultSnippet read
includes the fallback.
In the transparent stream, terminal output and patch diffs intermittently
vanished after a cold/paginated reload. A settled tool row renders its body from
`snippet`, but the raw-assistant-envelope 'derived' rebuild only joined the
result by tool id via resultsByTid (role:tool messages); when that join missed
(id mismatch, recovery-rebuilt turn) the snippet was empty, and #4622's live
enrichment only helps when S.toolCalls still matches (empty on cold load).
Build a persistedSnippetByTid lookup from the durable session.tool_calls compact
summary (which carries a bounded result snippet per tid) and use it as a
fallback in all three derived-push sites, so the persisted record is the
reliable source and the live join becomes a fast-path, not the only path that
shows the body. Part of #4925.
Settings -> Extensions offered an Install button even when extensions were
unconfigured, then dead-ended on 'Extensions not configured' because
install_extension required HERMES_WEBUI_EXTENSION_DIR to be set to an existing
directory. Default to a WebUI-managed root under STATE_DIR/extensions and
bootstrap it on first install so install just works with zero config; the env
var remains an optional override (never auto-created). Gallery-installed
extensions already auto-load via the runtime manifest, so the extension loads
on the next app-shell render with no further setup.
- api/extensions.py: _default_extension_root + _writable_extension_root;
_extension_root/_extension_root_status fall back to the managed default;
install_extension bootstraps via _writable_extension_root; dir-unavailable
warning only fires for an explicitly-set-but-broken env path.
- tests: update disabled-by-default expectations (configured=True, valid=False
pre-install) + new regression proving install bootstraps + auto-loads with
no env.
- docs/EXTENSIONS.md: document zero-config one-click install + env now optional.
- CHANGELOG [Unreleased].
Codex gate: raising content args from 120->4000 chars newly surfaces secrets on
non-first lines / past char 120 in render paths that print tc.args directly and
bypass the #4926 label redaction. Apply _redactToolTargetLabel to arg values at
every such site: the transparent Full-tab args render (_transparentToolDetailHtml),
the non-transparent tool-card args block (buildToolCard visibleArgs), the
clipboard copy path (_copyEventToClipboard), and the _toolArgsSnapshot itself.
+ node regression proving a secret past char 120 in a long command is masked in
the Full tab while the non-secret command structure is retained.
Tool-call args were truncated to 120 chars at three sites (_truncate_tool_args,
the live _tool_args_snapshot in api/streaming.py, and _toolArgsSnapshot in
static/ui.js). For incidental args that's fine, but content/diff-bearing keys
(command/cmd/script/code/patch/diff/old_string/new_string/content/path/
file_path) feed the rendered tool card AND the diff reconstruction in
recovery-rebuilt sessions, so the cap silently corrupted long commands/paths and
produced broken/empty diffs. Exempt those keys with a much larger cap
(_TOOL_ARG_CONTENT_CAP = _TOOL_RESULT_SNIPPET_MAX = 4000) at every site, keyed by
key name; incidental args keep the 120 cap; storage stays bounded. Part of #4925.
Codex re-gate: the env/AUTH '=' redaction regex matched 'token=' INSIDE a URL
(?token=...&mode=debug&signature=...) before the dedicated URL-query redactor,
and its \S+ value swallowed the rest of the URL incl later params + closing
quote. Anchor the env-assignment match to a shell-token boundary (^|[\s;|(]) so
it can't match keys preceded by ? or & — URL query params are left to the
dedicated query-param redactor, which masks the secret-looking ones while
preserving the command tail. + regression tests (URL-tail-preserved,
benign-PATH-not-over-redacted).
Codex gate: exposing the full multi-line command in the expanded shell card
newly surfaces secrets on non-first lines (the collapsed header only ever showed
line 1). The old redactor masked only password-shaped fields. Broaden it to mask
common secret forms across the WHOLE command: env/colon assignments for
*TOKEN/*API_KEY/*SECRET/*ACCESS_KEY/*PRIVATE_KEY/*CREDENTIAL/*CLIENT_SECRET/
*SESSION_KEY and *AUTH*=, --token/--api-key/--secret style flags,
Authorization: Bearer/Bot/Token headers, and secret-looking URL query params.
Test now drives the REAL redactor (not an identity stub) and asserts non-first-line
secrets are masked while the collapsed header stays first-line.
In the transparent activity stream, a shell tool card's collapsed header
correctly shows only the first line of the command, but the EXPANDED detail
lead (`$ <command>`) reused that same first-line-only value via
_toolTargetLabel, so a multi-line script still rendered as line 1 when expanded.
The full command is available frontend-side; the truncation was purely in the
label derivation.
Adds _toolFullCommandLabel(tc) (full multi-line, redacted, no first-line
collapse) and uses it in _toolDetailLeadText for the shell kind, while the
collapsed header keeps the first-line _toolTargetLabel. Part of #4925.
The first cut asserted against the REAL tools.process_registry.format_process_notification,
which isn't importable in WebUI-only CI shards (shard 2 failed: returned None).
Inject a stub agent module via sys.modules so the test validates the delegation
WIRING independent of agent-module availability, plus a graceful-degradation
test confirming async_delegation returns None (not raise) when the agent
formatter can't be imported. Production behavior is unchanged: async_delegation
events only occur when the agent is running, so the import resolves where the
events actually exist.
format_wakeup_prompt() in api/background_process.py returned None for any event
type other than 'completion'/watch_*, so async_delegation events (produced when
a background delegate_task subagent finishes) were silently dropped — the
subagent showed as perpetually 'running' and its consolidated result never
re-entered the parent conversation.
Add an async_delegation branch that delegates to the agent-side
tools.process_registry.format_process_notification (which already knows the
async_delegation shape), so the subagent result is rendered into the wakeup
prompt and re-enters chat. Scoped to async_delegation specifically (not all
unknown types) so genuinely-unrecognized events are still skipped — preserving
the existing skip-unknown-event contract.
Reported + root-caused by @mydelren.
Co-authored-by: mydelren <mydelren@users.noreply.github.com>
On Windows a reserved-device-name file (nul/con/prn/aux/com1-9/lpt1-9) can land
in the working tree (e.g. a shell command redirecting to `> nul` under Git Bash,
which treats it as a literal filename). git can't delete it via the normal Win32
path, so `git clean -fd` exits non-zero — and force_apply_update treated that as
fatal, aborting the whole force update with 'Failed to remove untracked files
before force reset' and leaving the user stuck.
The clean is now best-effort: a failure is logged (stderr captured for
diagnostics) but non-fatal, because the subsequent `git reset --hard` overwrites
any tracked-file collisions regardless, and residual untracked files git can't
delete are harmless. Updates the prior test that asserted the (buggy) fatal
behavior to the corrected non-fatal contract, plus a new regression test.
Reported + root-caused by @rodboev.
Co-authored-by: Rod Boev <rod.boev@gmail.com>
Codex gate caught that the enrich only restored when the settled snippet was
fully EMPTY — but the backend persists a 4000-char capped preview
(_TOOL_RESULT_SNIPPET_MAX), so long terminal/tool output settles to a truncated
4000-char prefix, not empty (the actual #4622 symptom). Now also restore when
the settled snippet is a bounded preview of the live one (live longer + settled
is a prefix + settled length >= the 4000 cap), without clobbering a genuinely
short non-preview value. Adds capped-preview restore + short-prefix-not-clobbered
tests (proven non-vacuous).
In the Transparent Stream / Worklog view, terminal tool cards lost their stdout
and patch/edit cards lost their diff AFTER a turn settled (live cards were fine).
Root cause (proven via a node harness driving the real settled chain
_anchorSceneRowsByMessageIndex -> _anchorSceneToolRowFromCall ->
_anchorSceneToolCallFromRow): the settled rebuild builds each tool row from the
persisted messages[].tool_calls (state.db/sidecar), which can carry only a short
preview — or, on a cold/paginated load, nothing — for the result body. The full
output lives on the live S.toolCalls entry at settle time, but the dedup loop in
_anchorSceneRowsByMessageIndex SKIPPED the matching live entry (continue) instead
of restoring the missing body onto the surviving settled row. So buildToolCard
got tc.snippet='' -> empty output region, and the patch/edit diff (detected from
the snippet) never rendered. This also blanks the transparent 'Output' tab and
the 'Full' (detailed-input) tab.
Fix: new _enrichSettledToolRowBodyFromLive() restores snippet/result/output,
command, and args onto a settled row from the matched live call, ONLY when the
settled value is empty (never clobbers a genuinely persisted body). Wired into
the dedup loop: on a tid match that would have been skipped, enrich the existing
pooled row instead of dropping the live data.
The prior #4625 guard only exercised the LIVE buildToolCard path, which is why it
stayed green while the settled path regressed. New
tests/test_issue4622_settled_toolcard_body.py drives the full SETTLED chain and
asserts: terminal stdout restored + full output in data-full + transparent Output
tab non-empty; patch diff body restored + renders as a diff + Full-tab args kept;
genuine persisted body not clobbered; args restored when dropped; no-live-match
leaves the row untouched; no duplicate row after enrich. Verified to FAIL on clean
master (assertions) and pass with the fix. Also registered the new helper in the
and now references the new symbol).
Reusable proof harness saved to ~/WebUI/qa/harnesses/.
Closes#4622.
Gate (Codex) caught two SILENT regressions in the original fix: explicitly
hiding the header in the 'else'/'empty' branches also hid populated read views.
- _setMemoryHeaderButtons('read') now shows the header for ALL read views
(incl. read-only Project Context / External Notes), gating only the Edit btn.
- _renderProfileConceptHelp() now uses a new 'help' header mode (shows header,
hides all action buttons) instead of 'empty' (which hid the help title).
Adds tests/test_issue4891_detail_header_visibility.py (proven non-vacuous).
Co-authored-by: luandnh <luandnh@users.noreply.github.com>
#4856 (mobile transcript completely unscrollable) survived multiple fixes
because regressions were guarded only by grepping one exact CSS string. This
adds an INVARIANT guard so the whole class can't recur: .messages-inner (base +
every media query) must never become a Y-axis scroll container — neither via an
explicit overflow-y scroll value nor via overflow-x:hidden/scroll/auto (which
the CSS spec coerces overflow-y->auto, collapsing min-height:auto to 0 inside
the flex column). overflow-x:clip is the only allowed horizontal clip. Also
asserts .messages stays the real scroller (overflow-y:auto + min-height:0).
Proven non-vacuous: reintroducing overflow-x:hidden on the mobile rule fails it.
For a true runtime check at a real viewport, the private workspace has
scripts/layout_scroll_probe.py.
The mobile .messages-inner rule used overflow-x:hidden, which per the CSS
overflow spec coerces overflow-y from visible to auto — turning the inner
into a scroll container. A scroll container's min-height:auto resolves to 0,
so the inner (a flex-shrink:1 item inside the .messages column flexbox)
collapsed to the viewport height instead of growing to its content. The
transcript then overflowed the clipped inner rather than the .messages
scroller, leaving nothing to scroll and pinning scrollTop at 0 — every
open/send/receive stranded mobile users at the top with no way to scroll down.
Fix: overflow-x:clip suppresses horizontal overflow (the original #4583/#4553
intent) WITHOUT coercing overflow-y or creating a scroll container, so
min-height:auto stays min-content and the inner grows to full height.
Verified empirically via headless Chrome mobile emulation (390x844, mobile=true):
scrollableDelta 0 -> 7746; open/send/receive all land at bottom; mid-scroll
position preserved across re-render. Desktop unaffected.
Updated #4553 test (clip satisfies its horizontal-pan goal) + added dedicated
Opus gate finding (BLOCKING): the two #4385 archive-reappears regression tests
mocked Session.load_metadata_only to return an archived stub WITHOUT writing a
real sidecar file, and never set SESSION_DIR. The new stat-gate in
_state_projection_sidecar_metadata short-circuits on SESSION_DIR/{sid}.json
before consulting metadata, so those mocks were never reached -> archived=False.
This is NOT a production regression (in prod, load_metadata_only reads exactly
SESSION_DIR/{sid}.json and returns None when absent, so archived metadata only
exists when the file does — in which case the stat() succeeds and behavior is
identical). The divergence existed only at the tests' mock boundary.
Fix: both tests now write a real archived sidecar JSON under a monkeypatched
SESSION_DIR, exercising the actual production read path (strictly better
coverage than mocking load_metadata_only). Also documented the
SESSION_DIR-as-sole-source assumption on the cache function per Opus's
forward-looking note.
GET /api/sessions pinned CPU to 100% and took multiple seconds on profiles with
many cron sessions. The state.db -> sidebar projection did three pieces of
redundant per-row file I/O, in BOTH the visible pass and the higher-capped
(CRON_PROJECT_CHIP_LIMIT=200) cron-only second pass:
1. _state_projection_sidecar_metadata(sid): uncached open() + 64KB prefix read +
JSON-key scan, once per row.
2. get_last_workspace(): up to two file reads + an is_dir()/remote probe,
returning the SAME workspace for every row.
3. cron/jobs.json: full read + parse per untitled cron row.
On a cron-heavy install that is hundreds of file reads per build, and because
the enclosing _CLI_SESSIONS_CACHE is keyed on a state.db content fingerprint
that advances on every streamed message row, the whole scan was re-paid on
essentially every 5s poll during a live turn (the recurring 100% CPU / slow
/api/sessions from #4672 -> #4808 -> #4842).
Fix:
- Memoize _state_projection_sidecar_metadata by the sidecar's
(path, mtime_ns, size, ctime_ns) stat signature in a bounded module cache, so
a warm build re-stats instead of re-reads. A rename/archive/edit bumps the
signature and invalidates just that one entry (no staleness). This survives
the 5s CLI-cache busting during streaming, which is the actual trigger.
Mirrors the Claude Code parse cache (#4718). Folded into
clear_cli_sessions_cache() for explicit-reset/test parity.
- Build-scoped memos for cron jobs.json (parse once -> job_id->name map) and
get_last_workspace() (resolve once), matching the existing _cron_pid() pattern.
Output is byte-identical. Adds tests/test_issue4842_cron_projection_perf.py
(7 tests: warm-rebuild no re-read, stat-change invalidation, dict-copy isolation,
missing-sidecar default, jobs.json once, workspace once, lifecycle clear).
