Per tests/test_issue6067_artifact_paths.py, this key is intentionally
English-fallback-owned (same convention as profile_concept_*) and must
be absent from every non-English locale block. CI caught this
(test_issue6067_session_source_key_is_english_fallback_owned) — the key
now falls through to LOCALES.en[key] as intended, same as before.
master was ~1,070 commits behind nesquena/hermes-webui:master, and roughly
40 of those touched static/i18n.js: new features (workspace panel, session
sharing, extensions gallery, cron sentence-builder UI, gateway controls),
plus a community PR that bulk-translated the old TODO-placeholder set
across all locales. Rebasing the old ja branch directly conflicted because
the English source text itself had moved since those translations were
written.
Rebuilt the ja block against current upstream/master instead: reused the
existing translation wherever the paired English source is unchanged, and
wrote fresh translations for everything new, reworded, or still sitting in
English. ja is back to full key parity with en per this repo's own
tests/test_japanese_locale.py contract (key coverage, duplicate-key
parity, placeholder preservation, arrow-function parity).
Left brand/acronym terms (MCP, YOLO, JSON, HTML, URL, sk-..., Hermes
Agent) and profile_concept_* (explicitly English-fallback-owned per
tests/test_issue2147_profile_concept_help.py) untouched, matching this
repo's existing convention across all locales.
Codex gate on #6312 found the extracted helper caught only (TypeError, ValueError)
while the inline code it replaced caught Exception; float(10**400) raises
OverflowError, escaping the helper and aborting terminal-error cleanup before the
error message is persisted / apperror is emitted. Widen _terminal_turn_duration's
float() guard to (TypeError, ValueError, OverflowError) — the exact exceptions
float() can raise. Adds 10**400 to the existing invalid-origin parametrize
(red-before-green verified).
The suite failed 6 tests (test_sprint12 default-model, test_issue1538 Nous
catalog, test_issue1567 picker capacity/symmetry) whenever it ran on a box that
exports HERMES_MODEL (e.g. inside a live Hermes agent session where HERMES_MODEL
is the agent's runtime model). Root cause: get_effective_default_model() in
api/config.py treats HERMES_MODEL / OPENAI_MODEL / LLM_MODEL as the
highest-priority default-model source — above the isolated test config — so a
runner-exported value overrode the sandbox config and polluted the model
picker/catalog (e.g. injecting "opus-4.8" and breaking the @nous: prefix
invariant / default-model readback).
Fix: scrub the three model-selection env vars at conftest module level (before
any product module imports, covering in-process tests) and again in the
out-of-process test_server subprocess env, and add them to the per-test
_scrub_provider_env helpers in the two Nous test modules. The isolated config /
HERMES_WEBUI_DEFAULT_MODEL is now authoritative regardless of the runner env.
Verified: with HERMES_MODEL=opus-4.8 set, the full suite goes from 6 failed to
0 failed (13550 passed); non-vacuity confirmed (disabling the scrub reproduces
the failures).
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* fix: Fix height of consolidated kanban board
There was a lot of empty space below the kanban board columns
in the condolidated view (particularly on desktop). Modify the
CSS such that the consolidated view always fills the viewport
vertical space.
* Release: consolidated Kanban board fills vertical space (#6308)
---------
Co-authored-by: Joe Palazzolo <joe@joepalazzolo.net>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* feat(extensions): token-v1 proxy->sidecar auth boundary
Mint a per-extension secret core injects (X-Hermes-Sidecar-Token) on every
proxied request; sidecars validate it. Closes the hole where a loopback sidecar
port is reachable by any local process and cannot distinguish a proxied request
from a direct one.
- api/extension_sidecar_auth.py: per-extension token lifecycle (atomic mint,
re-read-verified so an unpersisted token is never injected, per-request
mtime-cached read for live rotation, path-escape-safe)
- manifest proxy_auth negotiation: absent=legacy, token-v1=enforce, unknown=fail-closed
- proxy: inject token, strip inbound + response x-hermes-*, auth-off posture
(loopback-only local_unprotected, else 503), fail-closed when token unavailable
- consent-time auth_required in status payload; mint-on-consent
- docs/EXTENSIONS.md proxy_auth section
- 7 new tests (26/26 green)
* fix(extensions): address Codex+Fable gate on token-v1 (6 findings)
- align token-module extension-id grammar with core _EXTENSION_ID_RE (was
narrower -> legally-named ext consented then 503'd forever)
- resolve token dir dynamically (mirror _extension_state_dir) -> real test
isolation, no import-time STATE_DIR cache
- cross-process mint: O_CREAT|O_EXCL no-clobber claim (was os.replace clobber)
- rotation cache keyed on full fingerprint (ino/dev/mtime/ctime/size) +
re-fingerprint after read -> same-size/mtime replacement no longer stale
- validate token format on read (url-safe, 16-256) -> malformed file can't
leak via a ValueError echoed in a 502
- consent fails 503 when token can't be provisioned (was silent-swallow ->
persisted consent then 503 forever)
- rename status auth_required -> posture enum (protected|local_unprotected):
nothing is blocked for loopback, so 'required' was misleading
- panels.js: render local_unprotected warning on the consent row (+ CSS)
- docs: token-path resolution order, 401-vs-503, explicit 'legacy' acceptance
- tests: route-level token-injection+response-strip test; fix illusory isolation
in token-module test; 27/27 green
* fix(extensions): close 2 token-mint races (Codex re-gate round 2)
- mint via temp-file + atomic os.replace (not O_CREAT|O_EXCL) so the final
path is never observed empty/half-written — a concurrent loser can no longer
read an empty token file and 503
- single _stable_read helper (fingerprint-read-refingerprint, bounded retry on
mid-read change) used by BOTH ensure_token and current_token — a token that
changes during the read is never returned or cached stale
- stress-verified: 20 concurrent first-mints converge on 1 persisted token; 27/27
* fix(extensions): atomic no-clobber token publish via os.link (Codex re-gate round 3)
os.replace fixed empty-file exposure but still clobbered cross-process: two
processes could both write+replace and a reader between them got a token no
longer on disk -> 401. Switch to the repo's TOCTOU-safe os.link create-or-fail
idiom (session_recovery.py:627): write temp -> link into place (fails if a
winner already published) -> loser drops its temp and reads the winner via
_stable_read. Proven: 16 concurrent PROCESSES converge on 1 persisted token,
all matching disk. 27/27 tests green.
* fix(extensions): resolve Frank+Greptile #6331 review — token-v1 fail-closed when auth off (consent+resolution), token-v1-only proxy_auth/posture status fields, fullmatch ext-id validator
* fix(extensions): finish Frank and Greptile sidecar review
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* fix: Fix kanban scrolling in mobile viewports
Scrolling vertically (particularly in expanded view) in kanban on
a mobile viewport is difficult. The columns have overscroll
disabled, meaning tap and drag will only scroll within the column
and will not continue to the next section. On desktop it's much
easier to get the mouse outside the column div, on mobile you have
to deliberately try to tap very close to the edge of the viewport.
Disabling that behavior makes the experience much better
Author: Joe Palazzolo <joe@joepalazzolo.net>
* docs(changelog): stamp #6306 mobile Kanban column scroll fix
---------
Co-authored-by: Joe Palazzolo <joe@joepalazzolo.net>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* fix: preserve live anchor run identity
* Validate envelope run ids before snapshot cursor use
* docs(changelog): stamp #6201 stable Anchor run identity
---------
Co-authored-by: Frank Song <franksong2702@gmail.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* fix(webui): keep the content search from evicting the user's working set
/api/sessions/search?content=1 walks EVERY session and pulls each one through
get_session(), which inserts it into the SESSIONS LRU and marks it
recently-used. On any install with more sessions than sessions_cache_max
(default 300), a single search therefore flushes the whole cache and refills it
with sessions the user is not looking at — the classic buffer-pool
scan-pollution problem. The sessions actually open in the UI are exactly the
ones evicted, and the search is keystroke-debounced, so it repeats while typing.
A scan reads each session exactly once, so nothing it touches has earned
"recently used". get_session_for_scan() reuses a resident session without
promoting it, and reads a cold one straight from disk without caching it. It
returns None rather than raising, since a scan skips what it cannot open.
This is a correctness fix for cache behaviour, not a latency fix. The
multi-second searches that led here were contention, not scan cost: a trivial
/api/profiles took 9.2s in the same window, and a full read+parse of ~1700 real
sessions measures ~4s total.
test_sessions_search_depth_validation patched api.routes.get_session. With the
search reading through the scan accessor it now patches get_session_for_scan —
left unfixed, two of its cases fail and the third passes vacuously against an
empty result set.
Validation:
pytest tests/test_issue4765_sessions_lru_eviction.py
tests/test_sessions_search_depth_validation.py -> 13 passed
Both added eviction tests fail on the pre-fix accessor (verified by revert):
the working set drops from 4/4 to 0/4 resident after one scan.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* docs(changelog): stamp #6084 content-search working-set preservation
---------
Co-authored-by: ai-ag2026 <261867348+ai-ag2026@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* feat(terminal): broadcast output to every viewer instead of one shared queue
TerminalSession.output was a single queue.Queue read destructively by the SSE
handler. Two tabs/windows viewing the SAME session each open their own
EventSource, so two _handle_terminal_output handlers competed on that one queue:
every PTY chunk was delivered to exactly one of them. Each tab saw a disjoint
half of the byte stream, and only one ever received terminal_closed.
Output now fans out, mirroring StreamChannel/SessionChannel: each SSE consumer
subscribe()s its own queue (seeded with a bounded backlog so a first/late attach
still replays the recent scrollback, preserving the old buffer-until-first-
consumer behaviour), and put_output broadcasts to all subscribers. A slow
viewer's queue drops its own oldest chunk (drop-oldest, isolated per subscriber)
so one lagging tab can't starve another. The handler unsubscribes in a finally
so the subscriber list can't grow.
Stacked on the terminal fd-leak fix (same file).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: cover terminal broadcast lifecycle
* fix: serialize terminal subscriber fanout
* test: cover terminal unsubscribe publication race
* restore timing-flaky pytest.skip on test_terminal_survives_short_lived_request_thread (keep API-updated body); changelog #5836
---------
Co-authored-by: ai-ag2026 <261867348+ai-ag2026@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* fix: catch unhandled exception in POST /api/updates/check (defensive hardening) (#6180)
* fix: generate non-empty approval_id when gateway approval.request omits it (#6008) (#6168)
* chore: mark update-check try/except as defensive-only guard, drop #6086 linkage
Per maintainer review, the try/except wrapper is defense-in-depth only —
it does NOT fix#6086 (root cause is signal/process-group reaping).
Updated log message and added inline comment to make this explicit.
Leave #6086 open.
* test: isolate state-dir probes from user state
* docs(changelog): stamp #6168 approval_id, #6180 update-check guard, #6305 test isolation
---------
Co-authored-by: webtecnica <75556242+webtecnica@users.noreply.github.com>
Co-authored-by: webtecnica <webtecnica@users.noreply.github.com>
Co-authored-by: pxxD1998 <214340659+pxxD1998@users.noreply.github.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
Add an opt-in setting to hide Transparent Stream's per-event timestamp chips while keeping the response footer time visible, for users who found the per-event chips noisy. Thanks @rodboev. Closes#6099.
Recover checkpointed core background processes and rebuild PROCESS_SESSION_INDEX on WebUI startup, so an ordinary terminal(background=True, notify_on_complete=True) proc_* completion that outlives a WebUI restart can still wake its original session. Complements #6283 (which covered async_delegation completions). Thanks @allenliang2022.
Fall back to a copytree (with read-only handling) when the test-server fixture can't create the skills symlink on native Windows without SeCreateSymbolicLinkPrivilege (WinError 1314). Test-infra only. Thanks @rodboev. Closes#6240.
Route async-delegation completions by the immutable origin_ui_session_id (exact origin tab) and deliver them through a durable claim/complete/release lifecycle so they arrive exactly once, restart-safe, on both the background wakeup and next-turn drain paths. Combines #6185 (@carlotestor) + #6159 (@sysophelper-droid); supersedes #6002/#6225.
Co-authored-by: carlotestor <carlotestor@users.noreply.github.com>
Co-authored-by: sysophelper-droid <sysophelper-droid@users.noreply.github.com>
Prevent bare-id model picker revert when the provider hint is empty: an ambiguous bare id that collides across provider groups no longer snaps to the default group on re-render. Adds a revert-sensitive regression test and fixes three cross-file test-isolation leaks found while gating. Thanks @webtecnica.
Co-authored-by: webtecnica <webtecnica@users.noreply.github.com>
* ci: docs-only fast path + minimal docs CI
Skip the full pytest matrix + browser smoke on docs/CHANGELOG/README-only PRs
(required checks still report green fast via a fail-safe 'changes' gate job), and
add a lightweight Docs CI: critical_markdown_check.py (rendering-breaks only, not
style) + lychee broken-link check. Detection fails safe (any uncertainty or any
non-doc path -> full suite runs).
* ci: tighten docs-only detection — extension/type wins over name
A code file whose NAME contains README/CHANGELOG (scripts/CHANGELOG_stamp.py,
static/README_renderer.js) was wrongly classified docs-only, which would SKIP the
test suite on a real code change. Now a path is docs only by doc extension, exact
doc basename, or non-code file under docs/. Verified against 16 cases incl. every
code-with-docs-name trap.
* ci: fix all Fable+Codex gate findings on docs-only fast path
- BLOCKER: drop *.txt from is_docs (requirements.txt is a dep manifest — a bump
would have skipped the whole test matrix). docs = *.md/.markdown/.rst + bare
doc basenames only; strict allowlist, no docs/** denylist.
- Rename hole: use 'git diff --name-only --no-renames' so a rename src/app.py ->
docs.md reveals the code-side deletion instead of collapsing to the doc dest.
- SECURITY: docs-ci.yml no longer interpolates untrusted fork-PR filenames into
run: via ${{ }} (which executes $() in a crafted name). File list flows through
a file + mapfile-as-args; lychee gets a fixed glob, not the attacker list.
Added permissions: contents: read.
- Wedge belt-and-suspenders: 'if: ${{ always() }}' on the required test +
browser-smoke jobs so a failed 'changes' job can't skip them; step guards treat
missing/empty docs_only as run-full.
- critical_markdown_check.py: corrected the core rule — a newline in the whitespace
AROUND a link destination is valid CommonMark (was a false positive); only a
newline INSIDE the destination token, or an unclosed inline link, breaks
rendering. Verified full agreement with the markdown-it-py reference parser +
0 false positives on all 43 repo docs. Also blank 4-space indented code + multi
backtick spans. Reworded 'block the merge' -> 'break rendering' (non-required).
* ci: address Codex re-gate — lint always() + markdown title/unclosed cases
- Add if: always() to the lint job too (not currently required, but future-proof
against the wedge class if it's ever promoted).
- critical_markdown_check.py: handle two more CommonMark cases Codex found —
a newline inside a "title" string is legal (skip), and a newline-terminated
unclosed dest ([x](url\n at EOF/EOL with no close) is broken (flag). After the
destination token ends, valid continuations are ')' or a title opener (" ' ();
bare text after the newline is the real break. Verified full agreement with
markdown-it-py across 13 cases + 0 FP on all 43 repo docs.
* ci: model the CommonMark inline-dest grammar (root-cause fix for markdown checker)
Round-3 gate found the title heuristic caused sibling regressions: a parenthesized
multi-line title (url (a\nb)) was falsely flagged, and a quote glued into the URL
(exa"part) was wrongly treated as a title-start and skipped. Rather than patch more
heuristics, replace the ad-hoc newline logic with _scan_inline_dest(), which walks
the actual grammar: skip leading ws -> bare dest (balanced parens, ends at ws or the
depth-0 ')') or <angle> dest -> after ws the next char must be ')' or a real title
opener (" ' () -> else the destination is split across the line (broken). Verified
FULL agreement with markdown-it-py across 20 adversarial cases (incl. both round-3
regressions, balanced parens, angle dests, multiline titles) + 0 FP on all 43 docs.
* test: pytest suite for critical_markdown_check (42 cases)
Durable, repeatable verification for the docs-CI markdown checker: 19 verdict cases
+ 19 cross-checked against the markdown-it-py CommonMark reference (skips cleanly if
the lib is absent) + 3 code-span-safety cases + empty/no-link inputs. Covers both
round-3 regressions (parenthesized multi-line title valid; quote-glued-in-URL broken),
balanced parens, angle destinations, and multiline titles. 42 passed.
* ci: fix 2 grammar edges from Codex round-4 (escaped-> in angle dest, unbalanced bare-dest parens)
- Angle dest <...> now honors backslash escapes: [x](<foo\>bar>) renders (the \> is
escaped), was a false positive.
- Bare dest must have BALANCED parens: [x](foo(\n)) does not render (a '(' stays open
when whitespace ends the token) — now returns split, was a false negative.
Both verified against markdown-it-py + added as pytest cases. 46 passed, 0 FP on 43 docs.
* ci: escaped-newline in angle dest is still a raw newline (Codex round-5)
<...> escape handling skipped the char after backslash including a newline, so
[x](<foo\<nl>>) returned ok but doesn't render (blockquote on line 2). An escaped
newline inside an angle destination is still a raw newline -> split. Preserves
[x](<foo\>bar>). Added regression case. 48 pytest cases pass, 0 FP on 43 docs.
* ci: mirror escaped-newline guard to bare dest (Codex round-6, GFM-correct)
The angle branch already treated backslash-newline as split; the bare branch skipped
it, so [x](foo\<nl>bar) returned ok. GFM/CommonMark forbid line endings in bare
destinations (GitHub's cmark-gfm won't render it), though markdown-it-py permissively
does. Since these docs are GitHub-rendered we follow GFM: flag it. Added as a
SPEC_DIVERGENT test case (verdict-asserted, excluded from the permissive parser
cross-check). 49 pytest pass, escaped paren/space still valid, 0 FP on 43 docs.
* ci: delimiter-aware title scan (Codex round-7 unclosed-link false negatives)
Naive text.find(')') matched the TITLE's own ')' not the link's outer ')', so
[x](foo (title), [x](<foo> (title), [x](foo "title" all returned ok despite being
unclosed links. Now parse the title to its actual closing delimiter ("..", '..',
or a (..) that forbids nested unescaped '(' per CommonMark), then require the link's
own ')' after optional whitespace. Also catches the nested-paren-title break
[x](foo (a (b) c)). Verified vs markdown-it-py; escaped parens + single-quoted
titles containing parens still valid. 61 pytest pass, 0 FP on 43 docs.
* ci: remove unused variable (ruff F841 in critical_markdown_check)
The CI lint gate (ruff forward E9+F+B on new lines) caught a dead 'stripped =
line.lstrip()' leftover from an earlier refactor in _blank_code — the fence
detection matches on the raw line. Removed. No behavior change (61 pytest pass).
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Render data:image URIs as inline images (raster + base64 SVG) instead of raw base64 text, route file:// images through the media pipeline, with a strict allowlist + 2MB cap and img-only data: sanitizer. Thanks @ai-ag2026.
Co-authored-by: ai-ag2026 <ai-ag2026@users.noreply.github.com>
Deduplicate configured model badges so one configured model shows a single picker entry, with provider-collision + colon-bearing-id routing correctness. Thanks @happy5318.
Co-authored-by: happy5318 <happy5318@users.noreply.github.com>
* perf: optimize large session reconnect by adding file-size threshold to tail optimization (#6241)
When a sidecar JSON file exceeds 500 KB, the display-path tail optimization
now fires even if the message count is within the raw_budget. This prevents
sessions with few messages but large tool outputs (multi-MB JSON) from
forcing a full-scan merge of all messages on reconnect.
Changes:
- Added _sidecar_file_exceeds_threshold() helper
- Added _SIDECAR_BYTE_TAIL_THRESHOLD = 500_000 constant
- Fall-through to truncation in _state_db_since_timestamp_for_limited_display
when the sidecar file exceeds the threshold, regardless of message count
* CHANGELOG: byte-size reconnect tail optimization (#6260)
---------
Co-authored-by: webtecnica <webtecnica@users.noreply.github.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* fix(transparent-stream): remove dead rowIndex param from pushRow (#6189 follow-up)
In #6189 / #6217 the final-segment eligibility was migrated from an
index comparison (rowIndex > lastNonTerminalWorkRowIndex) to a WeakSet
lookup (finalSegmentLiveProseRows.has(row)). The rowIndex parameter on
pushRow became dead code — it's declared but never read, and the
call-site still passes idx from forEach. Remove the unused parameter
and simplify the call-site.
This addresses the greptile review feedback on the original PR.
Closes#6189
* CHANGELOG: dead rowIndex param removal (#6258)
---------
Co-authored-by: webtecnica <webtecnica@users.noreply.github.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* fix: split OIDC allowlist on commas only, preserve scope whitespace-split (#6244)
_normalize_text_list is shared with _normalize_scopes — OAuth scopes
are space-delimited per RFC 6749 §3.3. Created _normalize_allow_values
that splits on commas/newlines only, keeping multi-word group names
like 'Hermes Users' intact.
* fix(oidc): filter blank allow_values list elements + add parser-split test (gate follow-up)
The new comma/newline-only _normalize_allow_values() list-path retained empty
strings that the shared _normalize_text_list() had filtered, so a YAML
allow_values: [""] would brick an OIDC-only deployment (every callback 403s).
Filter stripped-empty collection elements. Adds a regression test asserting
allowlist multi-word preservation, comma/newline splitting, blank filtering,
and that scopes stay space-delimited (RFC 6749 §3.3).
Co-authored-by: webtecnica <webtecnica@users.noreply.github.com>
* CHANGELOG: OIDC allowlist whitespace fix (#6244)
---------
Co-authored-by: webtecnica <webtecnica@users.noreply.github.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* fix: intercept /sessions and /resume slash commands in WebUI (#6224)
Add a native-intercept branch in the block of
send() alongside the /pet special-case. When the user types /sessions
or /resume, expand the sidebar and refresh the session list instead of
sending the raw slash text to the agent.
Root cause: the agent command registry exposes sessions/resume as
non-CLI-only commands, so the autocomplete popup shows them, but the
WebUI send-time dispatch had no branch to catch them, causing the
literal text to be sent as a prompt.
* fix(commands): use mobile-aware session-browser opener for /sessions /resume (gate follow-up)
The intercept called expandSidebar() directly, which is a no-op on phone-width
layouts, so /sessions and /resume silently did nothing on mobile (composer
cleared, nothing shown). Use the mobile-aware _openProfileSwitchSessionBrowser()
first, falling back to expandSidebar(). Reproduced + specified by the pre-release
Codex gate.
Co-authored-by: webtecnica <webtecnica@users.noreply.github.com>
* CHANGELOG: intercept /sessions /resume slash commands (#6245)
---------
Co-authored-by: webtecnica <webtecnica@users.noreply.github.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
* fix(reasoning): gate reasoning_effort ladder to GLM-5.2+ on native zai
Z.AI's API (docs.z.ai) defines reasoning_effort as GLM-5.2+ exclusive, but
hemes-webui advertised the full 6-level ladder for all 7 GLM models because
_candidate_supports_reasoning has an unconditional glm token match and
_filter_reasoning_efforts_for_provider had no ZAI branch. Six of seven catalog
models (glm-5.1, glm-5, glm-5-turbo, glm-4.7, glm-4.5, glm-4.5-flash) showed a
selector whose values the endpoint silently ignores, and GLM-4.7 (forced thinking
that cannot be disabled per Z.AI docs) showed a 'none' option with no effect.
Add a ZAI branch to _filter_reasoning_efforts_for_provider mirroring the existing
OpenAI/Gemini/Anthropic ceiling pattern: strip the whole ladder for pre-5.2 GLM
models and for the forced-thinking GLM-4.7 family; preserve the full ladder for
GLM-5.2+ (whose accepted values match VALID_REASONING_EFFORTS exactly). The gate
is scoped to the native zai provider only (aliases glm/z-ai/z.ai/zhipu all
resolve to zai); aggregator providers are untouched because they route through
their own routers, not Z.AI's native endpoint.
The glm family-detection heuristic in _candidate_supports_reasoning is unchanged
— GLM models DO support the thinking on/off toggle at the family level; this fix
is specifically about the reasoning_effort intensity ladder.
State layer: agent.reasoning_effort config + UI dropdown options derived from
resolve_model_reasoning_efforts. Invariant: UI options and coercion now agree
and match Z.AI's per-model docs (max offered only for GLM-5.2+, none never
offered for forced-thinking models). Out of scope: the thinking:{type:...}
request-field translation lives in the external agent/gateway layer.
* fix(reasoning): close ZAI coercion gap + harden test assertions
Address Greptile review on #6219 (round 1):
1. Vacuous test assertion (test_glm_5_2_preserves_none_sentinel): the 'or'
fallback made the assertion always-true, so a regression stripping 'none'
for GLM-5.2 would go undetected. Rewrote to inject 'none' via the raw source
(mocking _resolve_model_reasoning_efforts_impl) and assert it survives — the
test now genuinely exercises the preservation branch.
2. Coercion gap for non-max stored levels on pre-5.2 GLM: the existing
'if ceiling and raw not in ceiling' guard treats an empty ceiling as 'no
rule' (preserving the configured effort verbatim per #3505), so a stored
'high'/'medium'/'low' for glm-5.1/glm-4.5/glm-4.7 on native zai was forwarded
to Z.AI unchanged and silently ignored — contradicting the PR's stated
UI/coercion agreement invariant.
Root cause: the ZAI gate returns [] to mean 'known-empty' (no
reasoning_effort at all), but the coercion path treated all [] as
'ambiguous/unknown, preserve verbatim'. Fixed by extracting the ZAI decision
into _zai_glm_reasoning_efforts_supported (True/False/None sentinel) shared by
both the filter and coercion, then special-casing the known-False result in
coerce_reasoning_effort_for_model to return '' (send no field). The #3505
preserve-verbatim behavior for genuinely-unknown models on non-zai providers
is unchanged.
Added 10 regression tests (all fail before the coercion fix, pass after):
- All 6 levels (max..minimal) coerce to '' for each pre-5.2 GLM + glm-4.7
- All 4 aliases (glm/z-ai/z.ai/zhipu) resolve through the same coercion gate
- GLM-5.2 preserves all 6 levels verbatim
- Regression guard: unknown model on custom: provider STILL preserves verbatim
State layer: agent.reasoning_effort config + the value forwarded to Z.AI.
Invariant now fully holds: UI offers no options for pre-5.2 GLM AND coercion
sends no reasoning_effort field for any stored level on those models.
* fix(reasoning): preserve ZAI GLM 4.5-5.1 thinking toggle when effort ladder empty
Address nesquena-hermes round-2 review on #6219: returning [] for the effort
ladder on sub-5.2 GLM models hid the entire reasoning chip in the composer
(static/ui.js:4932 treats empty supported_efforts as 'no reasoning control at
all'), silently regressing the working thinking on/off toggle for GLM-4.5/4.6/
5.0/5.1 users. Per Z.AI's own docs, those models accept the thinking
{type:enabled|disabled} toggle even though they do not accept the
reasoning_effort intensity ladder.
Fix decouples thinking-toggle capability from the effort ladder:
Backend (api/config.py):
- Refactor _zai_glm_classification returns one of 'effort' (GLM-5.2+), 'thinking'
(GLM-4.5 up to but not including 5.2), 'forced' (GLM-4.7 family), or None
(non-zai / non-GLM). Single source of truth shared by all three consumers.
- _zai_glm_reasoning_efforts_supported now wraps classification for the coercion
contract (unchanged behavior).
- _zai_glm_thinking_toggle_supported returns True for 'effort' or 'thinking',
False for 'forced', None otherwise.
- get_reasoning_status gains a supports_thinking_toggle field = bool(supported)
OR (zai_thinking is True). Non-zai providers default to bool(supported_efforts)
so their chip-visibility behavior is unchanged.
Frontend (static/ui.js):
- New _currentReasoningToggleSupported state var (default undefined = treat as
true so legacy responses without the field do not newly hide the chip).
- _applyReasoningChip shows the chip when hasEffortLadder OR toggleSupported.
Empty efforts + toggle=True keeps the chip visible with just the None/On
control (the existing _applyReasoningOptions already shows 'none' when the
ladder is empty). Empty efforts + toggle=False (GLM-4.7 forced) hides it.
- Profile-transition and fetch-failure resets now pass
supports_thinking_toggle:false alongside the empty efforts so the chip hides
during the unknown-state window, matching the prior reset contract.
Tests (30 new):
- _zai_glm_classification parametrized across all three tiers + aliases + defer
- get_reasoning_status supports_thinking_toggle per tier (GLM-5.2 both, GLM-4.6
toggle-only, GLM-4.7 neither, non-zai defaults to effort capability)
- Frontend _applyReasoningChip behavior via node driver: empty efforts +
toggle=True stays visible, toggle=False hides, effort ladder alone is
sufficient, absent field keeps prior behavior
State layer: agent.reasoning_effort config + supports_thinking_toggle field in
/api/reasoning + composer chip visibility. Invariant: the chip is hidden ONLY
when the model supports neither the effort ladder nor the thinking toggle
(GLM-4.7 forced, or genuinely non-reasoning models); GLM-4.5-5.1 retain the
working On/None control they had before the round-1 effort gate.
* fix(reasoning): make ZAI thinking toggle two-way + force GLM-4.7 stored none
Address nesquena-hermes round-3 review on #6219 — two SILENT gaps in the
thinking-toggle path, plus a click-handler sibling I found while auditing.
Gap #1 — ONE-WAY toggle for GLM-4.5/4.6/5.0/5.1 (api/config.py:4247,
static/ui.js:4902, static/index.html:757):
The round-2 fix kept the chip visible for thinking-tier models but the only
rendered dropdown option was 'None' (the HTML had no Default option, and
set_reasoning_effort rejected empty effort with 400). So a GLM-4.6 user could
turn thinking OFF but never back ON — worse than the original bug.
Fix:
- set_reasoning_effort now accepts empty effort as 'clear the override' (removes
agent.reasoning_effort so the provider default takes effect). Invalid values
still raise ValueError.
- static/index.html gains a <div data-effort=''>Default</div> option.
- _applyReasoningOptions always shows both Default ('') and None alongside the
effort ladder, so a thinking-tier model (empty ladder + toggle=true) renders
an operable Default+None two-state control.
- Click handler (ui.js:5106) checks option presence (if(opt)) not truthiness
(if(effort)) — the old check silently ignored data-effort='' clicks, which
would have left the Default button dead even after the HTML/backend changes.
Gap #2 — GLM-4.7 not forced when 'none' stored (api/config.py:3984, :3857):
When GLM-4.7 had agent.reasoning_effort=none configured, coercion preserved
'none' via the early return at line 3985, so streaming built disabled reasoning
for a model that forces thinking on regardless. Separately, when the raw
capability source listed 'none', resolve_model_reasoning_efforts reattached it
to GLM-4.7's supported_efforts (['none']), leaking an 'off' option to the UI
for a forced-thinking model.
Fix:
- coerce_reasoning_effort_for_model checks _zai_glm_classification == 'forced'
BEFORE the generic 'none' early-return, coercing stored 'none' to '' (default
= thinking on) for forced models.
- resolve_model_reasoning_efforts returns [] early for forced-tier models,
skipping the 'none' reattachment entirely.
Tests (17 new):
- Gap #2: coerce('none', glm-4.7) -> '', regression guard that non-forced GLM
still accepts 'none'; resolve does not reattach 'none' for forced but DOES
for thinking-tier; end-to-end get_reasoning_status for forced+stored-none.
- Gap #1 backend: set('') clears the key (no raise), set('garbage') still
raises, all 7 valid levels still save.
- Gap #1 frontend: three new node-driver tests asserting the dropdown exposes
both Default and None for thinking-tier (two-state), Default+None+ladder for
effort-tier, and the off->on->off round trip keeps both visible throughout.
- Updated test_reasoning_show_hide.test_set_reasoning_effort_rejects_invalid to
reflect the new contract (empty accepted, garbage rejected).
Regression gate: gap#2 coerce tests fail without the forced-tier check (3
failures); gap#1 backend test fails with 'ValueError: effort is required'
without the empty-acceptance change. 219 passed, 1 pre-existing skip across
the full reasoning + chip + config-cache surface.
State layer: agent.reasoning_effort config + supports_thinking_toggle field +
composer dropdown options. Invariant: the thinking toggle is now genuinely
two-way for GLM-4.5-5.1 (Default=on, None=off, both always visible) and
GLM-4.7 forced-thinking never offers 'none' in any path (coerce, resolve,
status, or UI).
* CHANGELOG: GLM per-version reasoning controls (#6219)
---------
Co-authored-by: Ruby Hartono <58564005+rh-id@users.noreply.github.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
