Parity with the gateway /reasoning handler and the new /model default:
a bare /reasoning <level> now applies to the current session only;
--global persists agent.reasoning_effort to config.yaml. --session is
still accepted as an explicit alias for the default. Display toggles
(show/hide/full/clamp) remain persistent as before — they are user
preferences, not conversation state.
Builds on YAMAGUCHI Seiji's #51158 (session-scope plumbing + /new reset,
cherry-picked as the previous commit) with the default flipped to match
the session-first policy. Fixes the CLI half of #54084.
Flip the resolve_persist_behavior() fallback from persist-to-config to
session-only. A plain /model <name> (typed or via any picker — CLI,
TUI/Desktop, gateway) now affects only the current session; --global
persists explicitly, and model.persist_switch_by_default: true restores
the old opt-out behavior for users who want switches to stick.
This is the root cause behind the recurring 'session switch applied
globally' bug class (#61458, #63083, #58290, #61190): every surface
funnels its no-flag default through this one function, so per-surface
patches kept missing paths. Fixing the default fixes all surfaces at
once: CLI typed + picker, TUI/Desktop config.set + slash, gateway typed
+ inline picker.
Builds on liuhao1024's #58371 (--provider session scoping, cherry-picked
as the previous commit) and supersedes the per-surface #61488.
When /model is called with --provider but without --global or --session,
the switch now defaults to session-only instead of persisting to
config.yaml. Provider switches are typically exploratory — the user is
trying a different backend for this conversation, not reconfiguring the
default. --global can still force persist when desired.
This addresses a regression from fad4b40d9 where /model switched to
persist-by-default, causing /model xxx --provider xxx to overwrite the
global config when the user only intended a temporary switch.
Fixes#58290
Deduplicates the SimpleNamespace stub-response construction that was
copy-pasted between the tool-call-drop guard and the text-only-drop
guard (both in interruptible_streaming_api_call). The third site
(error-handler at ~L3775) has a structurally different shape (different
role/reasoning/model/usage sources + _content_filter_terminated tag)
and is left inline with an existing comment.
When a streaming response ends cleanly (HTTP 200) with no finish_reason
after delivering text but no tool calls, the chunk collector silently
stamps finish_reason='stop' and the conversation loop presents truncated
text as a complete response.
Three stream-drop paths now exist after chunk collection:
1. Zero-chunk → EmptyStreamError, retried (existing)
2. Tool-call in progress, no finish_reason → partial-stream-stub (existing)
3. Text-only, no finish_reason → partial-stream-stub (NEW — this fix)
Path 3 routes through the same PARTIAL_STREAM_STUB_ID + FINISH_REASON_LENGTH
machinery as path 2. The conversation loop shows 'Stream interrupted —
requesting continuation' and injects a continue prompt, giving the model
a chance to resume where the stream dropped.
Observed with DeepSeek provider where CloudFront drops SSE streams
mid-response after delivering partial text.
developers.cloudflare.com/cloudflare-one/connections/connect-networks/
returns 301 Moved Permanently to
/cloudflare-one/networks/connectors/cloudflare-tunnel/ after Cloudflare
reorganized the Cloudflare One docs. The new page is the named-tunnel
workflow this paragraph points readers at.
The salvaged retention check compared the built-in memory snapshot
before vs after the disk reload. That holds for a long-lived CLI agent,
but on fresh-agent surfaces (gateway per-turn agents, TUI) the cached
prompt is restored from the session DB and can predate mid-session
memory writes that the fresh MemoryStore already absorbed at init: the
snapshot is then identical on both sides of the reload while the prompt
itself is stale, so compression would retain (and re-persist via
update_system_prompt) a prompt missing the new memory for the life of
the session.
Replace the equality check with a containment check
(_cached_prompt_reflects_builtin_memory): retain the cached prompt only
when the freshly-reloaded rendered blocks appear verbatim inside it,
and rebuild when a leftover block header remains for a target whose
entries have since been emptied or disabled. Block headers are shared
via MEMORY_BLOCK_HEADERS in tools/memory_tool.py so the check stays in
lockstep with MemoryStore._render_block.
Adds regression guards for the gateway stale-restore path and the
emptied-memory leftover-block path; verified with a real-MemoryStore
E2E matrix (9 scenarios) against a temp HERMES_HOME.
- extra_headers participates in the section-3 group identity (mirrors
section 4 — header-routed tenants behind one proxy URL stay distinct)
- model declarations go through _declared_model_ids() so
models: [{id: ...}] rows keep working
- gateway model-switch handler moved to gateway/slash_commands.py since
the PR branched — re-applied the display-form edits there (both the
legacy picker closure and the current typed path)
- regression tests: same-endpoint fold, api_mode separation,
header-routed separation, list-of-dict models, RID display stripping
Address review on PR #36998: the inline ri.<service>..<ns>. stripper in
_get_status_bar_snapshot was a one-off heuristic that:
* lived in cli.py with no shared call site, so the switch-confirmation
banner ("✓ Model switched: ri.language-model-service..…") and the
[Note: model was just switched from … to …] system-prompt nudge still
printed the full opaque RID — exactly what the screenshot reported;
* split on '..' and re-split on '.', which would mis-handle any RID
whose namespace token isn't a single dotted segment.
Refactor:
* New module-level helper hermes_cli.model_switch.format_model_for_display
matches on a startswith() allow-list (_OPAQUE_MODEL_PREFIXES) and
returns the trailing slug. Falls through to the original string for
every non-Palantir id, so HF paths (meta-llama/Llama-3.3-70B-Instruct),
plain Claude/GPT names, .gguf paths, and aliased ids are untouched.
Allow-list is extensible — add a prefix tuple entry for future
proxies that wrap real names in a namespace (Bedrock ARNs are
already covered by the slash-split fallback and have a different shape).
* _get_status_bar_snapshot() now delegates to the shared helper after
the reverse-alias miss (so configured aliases still win over the
helper output).
* cli.py::_handle_model_command — both confirmation-print blocks
(~7720 and ~7975) now run result.new_model AND old_model through
the formatter before they hit _cprint() and the
_pending_model_switch_note text.
* gateway/run.py model-switch handler (~10915) — same treatment for
_pending_model_notes[_session_key] and the
t('gateway.model.switched', model=…) confirmation line returned to
the gateway client.
The formatter is DISPLAY-ONLY. The session_model_overrides map,
ModelSwitchResult.new_model, persistence to config.yaml, alias lookups,
and every wire call still carry the full opaque RID — Palantir's API
requires it.
Verification: unit reproducer covers (a) all four Palantir model RIDs
from this user's config stripped to the trailing slug, (b) plain
model names (claude-4-7-opus-20260101, gpt-5.4, HF paths, empty
string) passed through unchanged, (c) prefix-only edge preserved
(no infinite-loop / empty-output regression).
Refs: PR #36998 review feedback; screenshot showed model banner still
printing the long RID after the original status-bar-only fix landed.
Two related TUI quality-of-life fixes for users running multiple models
behind a single proxy/aggregator (e.g. Palantir Foundry, Bedrock,
self-hosted vLLM behind a single key).
1. _get_status_bar_snapshot() — friendlier model name in the status bar.
Long catalog IDs (Palantir RIDs like
``ri.language-model-service..language-model.anthropic-claude-4-7-opus``)
were truncated to ``ri.language-model-ser...`` by the existing 26-char
slash-split, leaving the user with no way to tell which model is active.
The status bar now:
* Reverse-looks up the model id in config.yaml ``model_aliases:`` /
``model.aliases:`` and shows the shortest configured alias when one
exists (so users who set up a friendly alias get it for free).
* Falls back to stripping Palantir's ``ri.<service>..<ns>.`` RID prefix
before length-truncation, so the truncated label carries the actual
model identity (``anthropic-claude-4-7-opus``) instead of the URN
scheme.
* Reverse-alias map is cached at module level (config is loaded once
per session; no need to re-resolve on every status-bar refresh).
2. list_authenticated_providers() section 3 — group ``providers:`` entries
by (api_url, key_env, api_mode), mirroring section 4's existing grouping
for ``custom_providers:`` lists.
Before: a Palantir Foundry config with two Anthropic-proxy entries
(``palantir-claude46`` + ``palantir-claude47``) produced two near-
duplicate picker rows labelled ``Palantir Claude 4.6 Opus`` and
``Palantir Claude 4.7 Opus`` — same endpoint, same PALANTIR_TOKEN,
same anthropic_messages wire protocol, differing only by model id.
After: those entries collapse into a single ``Palantir Claude`` row
with both models in the dropdown. Same-host entries with a different
``api_mode`` (e.g. an OpenAI-compat ``palantir-gpt54`` alongside the
Anthropic claude rows on the same host) keep distinct rows since
the wire protocol differs — same safety invariant section 4 already
enforced for ``custom_providers:``.
Group display name strips per-version trailing tokens (``Palantir
Claude 4.7 Opus`` → ``Palantir Claude``) only when the prefix has
≥2 words, so single-word names aren't over-trimmed.
The new code records (raw_display_name, api_url) into
_section3_emitted_pairs for every raw entry that joined the group, so
section 4's compatibility-merged ``custom_providers`` view (built by
``get_compatible_custom_providers()`` which calls
``providers_dict_to_custom_providers()`` to convert ``providers:``
into custom-provider shape) still dedupes against this grouped row.
Manual smoke test on a config with three Palantir entries
(claude-4.6, claude-4.7, gpt-5.4): before — 3 picker rows; after — 2
picker rows (1 row "Palantir Claude" with 2 models, 1 row
"Palantir GPT-5.4" with 1 model).
The collapsed-providers atom (`hermes.desktop.collapsed-providers`) is a
global presentation-layer preference, but the catalog the picker renders is
profile-scoped (`getGlobalModelOptions` routes through `profileScoped()`,
model-menu-panel.tsx:87-93). The previous code called `pruneStaleCollapsed`
on every render against `pickerProviders`, which silently deleted any
slug not present in the active catalog.
Bug class (review from @teknium1):
- Profile switch to a catalog that lacks a previously-collapsed provider
→ that collapse preference is permanently lost.
- Refresh Models that drops a provider (revoked key, plugin disabled,
backend policy change) → same loss.
- Any transient empty catalog (loading → []) → guarded by `length > 0`,
but still loses state once the new (smaller) catalog resolves.
Fix:
- Remove the prune `useEffect` from model-menu-panel.tsx.
- Delete `pruneStaleCollapsed` (no other caller; the AGENTS.md "no
speculative infrastructure" rule applies — keeping it exported with a
plausible-sounding docstring is a foot-gun for future contributors who
would re-call it against a single active catalog and reintroduce the bug).
- Document on the atom why we deliberately don't prune: provider slugs
come from a bounded configured set (not user input); the render loop
only visits providers in the active `groups`; dead slugs have no
observable effect (`collapsedProviders.includes(slug)` against an
absent slug is a no-op).
Tests:
- `preserves the collapsed set across a profile switch whose catalog
lacks the slug` — regression pin for the profile-switch case.
- `preserves the collapsed set when Refresh Models drops a provider` —
regression pin for the refresh-models case.
Verified: `tsc -p . --noEmit` is clean on the changed files. CI is the
source of truth for the vitest run (the local React-detection env
failure pre-dates this PR; CI passes).
Closes the review thread on #64690.
Click a provider header (DEEPSEEK, GOOGLE, etc.) in the model picker dropdown
to collapse/expand its model list. State persists across sessions via localStorage.
- New provider-collapse store with persistentAtom<string[]> + stale-key pruning
- Provider headers become clickable DropdownMenuItems with chevron indicators
- textValue="" excludes headers from Radix typeahead (both reviewers flagged)
- Auto-expands the active provider so the checkmark is always visible
- Search bypasses collapse (typing shows all matching models)
- Keyboard accessible via Enter/Space on the header row
- Store double-read eliminated per cross-vendor review feedback
Reviewed-by: Flash + GPT-OSS cross-vendor review (all SHOULD-FIXs addressed)
Related: #60966 (different interaction model — hover-to-expand)
The new top-level mcp: config section surfaces exactly one field
(auto_reload_on_config_change) in the dashboard settings schema, which
tripped the no-single-field-categories invariant. Merge it into the
agent tab like onboarding/computer_use.
Follow-up on the salvaged #67449: auxiliary.mcp is the side-LLM task
provider block (provider/model/timeout for MCP aux calls) — a watcher
behavior toggle doesn't belong there. Move it to a new top-level mcp:
runtime section and read it from the same freshly-parsed config.yaml the
watcher already diffs (no second load_config() per tick, and flipping the
toggle + editing mcp_servers in one edit behaves correctly).
Also adds a regression test for the salvaged #55701 false-positive fix:
${VAR} templates in mcp_servers made the raw-yaml-vs-expanded-snapshot
comparison permanently unequal, so ANY save_config_value() rewrite (e.g.
/reasoning changing agent.reasoning_effort) fired a full MCP reconnect.
Credits: @OYLFLMH (#55701 env-expand fix), @TurgutKural (#67449 opt-out).
The opt-out default was declared in DEFAULT_CONFIG["auxiliary"]["mcp"][...]
but the watcher in _check_config_mcp_changes() read top-level
load_config().get("mcp") — a key that does not exist in the loaded
config shape. Consequently the declared default was never observed and
the fallback stayed True at runtime: setting auto_reload_on_config_change
to false in config.yaml silently did nothing.
Resolve through the same path the default is declared on:
cfg["auxiliary"]["mcp"]["auto_reload_on_config_change"]
Tests:
- test_optout_disables_auto_reload: mocked config now mirrors the real
DEFAULT_CONFIG shape (auxiliary.mcp), so the test exercises the actual
lookup path instead of a separately mocked shape.
- test_optout_path_is_auxiliary_mcp_not_top_level: regression guard — a
config that sets ONLY top-level mcp.auto_reload_on_config_change=false
must NOT disable the reload. This pins the config-path contract so a
future regression to _cfg.get("mcp") is caught.
Addresses sweeper review: the declared default was never observed at
runtime because the watcher read a different config path than the one
where the default was defined.
Co-authored-by: Turgut Kural <turgut.kural@gmail.com>
The automatic MCP reload added in #1474 watches config.yaml's mcp_servers
section every 5s and reloads on any change. Every reload rebuilds the agent
tool surface and INVALIDATES the provider prompt cache — the next message
re-sends the full input prefix, which is expensive on long-context /
high-reasoning models. When config.yaml is rewritten frequently (external
tooling, multiple Hermes instances, or a flapping MCP server that rewrites
config), this causes silent, repeated cache-breaking reloads.
Add `mcp.auto_reload_on_config_change` (default: true, backward compatible).
When set to false:
- The config change is still DETECTED (watcher keeps running).
- No automatic reload happens.
- The user is told the config changed, that new settings are NOT yet
applied, and how to apply them on their own terms with /reload-mcp —
including the explicit warning that /reload-mcp invalidates the prompt
cache.
Manual /reload-mcp is unaffected and still works for users who want to
apply changes deliberately.
Tests: extend TestMCPConfigWatch with test_optout_disables_auto_reload.
Co-authored-by: Turgut Kural <turgut.kural@gmail.com>
_check_config_mcp_changes compared mcp_servers from two inconsistent
sources:
- init: self.config.get('mcp_servers') -> from load_config() + _expand_env_vars -> expanded values
- watcher: yaml.safe_load(cfg_path) -> raw templates
When mcp_servers uses env-var templates like ${POWERMEM_API_KEY},
every save_config_value() that rewrites config.yaml (even for unrelated
keys) triggers a false-positive MCP reload, reconnecting all servers.
Apply _expand_env_vars() to the raw watcher value before comparison
so both sides use the same expanded representation.
Test plan: tests/cli/test_cli_mcp_config_watch.py (6/6 pass)
Follow-up to the salvaged #66083/#42792 commits:
- alibaba (Qwen Cloud coding-intl) gets qwen3.7-plus too — same platform
allowlist as alibaba-coding-plan (issue #44662 comment by @coder-movers)
- qwen3-max substring context entry (262144) so the newly-listed
qwen3-max-2026-01-23 snapshot doesn't fall to the generic 131072 qwen
fallback
Add qwen3.7-plus to DEFAULT_CONTEXT_LENGTHS with 1M context window.
Without this entry, the model falls back to the generic 'qwen' entry
(128K), causing premature context compression at 50% (64K tokens)
instead of the correct 500K threshold.
Official docs: https://help.aliyun.com/zh/model-studio/developer-reference/
The model list for alibaba coding plan is currently out of sync with
the actually supported models. See the official documentation[1].
Per the docs, alibaba coding plan does not support qwen3.7-max;
it supports qwen3.7-plus instead. Additionally, qwen3-max-2026-01-23
was missing from the model list.
Changes to the alibaba-coding-plan model list:
- Replace qwen3.7-max with qwen3.7-plus
- Add qwen3-max-2026-01-23
[1] https://www.alibabacloud.com/help/en/model-studio/coding-plan
The supermemory SDK already honors SUPERMEMORY_BASE_URL, but the raw
urllib call used for session-end conversation ingest hardcoded
https://api.supermemory.ai/v4/conversations, so ingest always hit the
cloud even when pointing at a self-hosted server (e.g.
http://localhost:6767).
Resolve the base URL as config (supermemory.json base_url) >
SUPERMEMORY_BASE_URL env var > https://api.supermemory.ai, strip any
trailing slash, and use it for both the SDK client and the
/v4/conversations ingest endpoint.
Phase 2c found 3 tests that called _run_and_exit_oneshot without
mocking _cleanup_oneshot_runtime, causing real cleanup (terminal,
browser, MCP, auxiliary) to run in the pytest worker. Add the mock
to all three for test isolation.
Also remove redundant 'import logging' inside _exit_after_oneshot
(already imported at module level, line 729).
The initial salvage from #43698 only shut down MCP servers and cached
auxiliary clients. The interactive CLI's _run_cleanup() also closes
terminal environments, browser sessions, and interrupts async
delegations — all of which can hold native-extension-backed resources
(aiohttp connectors, websocket clients) that SIGABRT during
Py_FinalizeEx.
Add the missing three sites to _cleanup_oneshot_runtime(), matching the
order in cli.py:_run_cleanup(). Update tests to cover the expanded
cleanup chain.
Credit: @konsisumer (#67768) identified the full cleanup surface.
Follow-up fixes for salvaged PR #67686 (issue #67639):
1. Windows regression: bare 'import fcntl' at module level in entry.py and
slash_worker.py would crash on Windows (fcntl is POSIX-only). entry.py
explicitly aims to 'import cleanly on Windows'. Extracted all fcntl/socket
logic to tui_gateway/_stdin_recovery.py with try/except import guards.
2. fd leak: socket.fromfd() dups fd 0; s.detach() returned the fd number
without closing it, leaking one fd per recovery call. Changed to s.close()
(safe — fromfd duped the fd, closing won't close stdin).
3. Code duplication: ~80 lines of recovery loop + diagnostic were copy-pasted
between entry.py and slash_worker.py. Extracted to shared
tui_gateway/_stdin_recovery.py (handle_spurious_eof + diagnose_stdin_state).
4. Profile-unsafe path: scanner used Path.home() / '.hermes' / 'plugins' but
canonical plugin discovery uses get_hermes_home() / 'plugins'. With
HERMES_HOME pointing elsewhere, scanner missed the actual plugin dir.
Also gated project plugins behind HERMES_ENABLE_PROJECT_PLUGINS to match
hermes_cli/plugins.py.
5. SO_RCVTIMEO not cleared: recovery only called os.set_blocking(0, True)
but a child that set SO_RCVTIMEO would cause the next readline to time
out and loop. Now clears SO_RCVTIMEO alongside O_NONBLOCK.
The copy-paste config sample had reasoning_effort: low active, which
would silently downshift effort for anyone pasting the block. Keep it
commented like other optional keys. Also add the contributor email
mapping for the salvage.
The dict-form guard from PR #67878 only covered the mapping shape
({model: {context_length: ...}}). The list-of-dicts shape
([{id: model, context_length: ...}]) is also a supported config form
(per _declared_model_ids) and was still being replaced with a flat
list of strings, destroying per-model metadata.
Sibling site for #67841.
When custom_providers[].models uses the mapping form to store
per-model metadata (e.g. context_length), _save_discovered_models_to_config
must not replace it with a flat list of strings. Add a guard that skips
entries whose models value is a dict, preserving the user's curated
metadata.
The regression was introduced by PR #65652, which added the auto-save
helper without considering the dict form.
Consolidates the three Qwen provider slugs (alibaba / Qwen Cloud,
alibaba-coding-plan / Alibaba Cloud Coding Plan, qwen-oauth / Qwen CLI
OAuth) under a single 'Qwen' group row in the interactive provider
pickers, matching the existing OpenAI / Kimi / MiniMax / xAI groups.
Display-only via PROVIDER_GROUPS — slug identity, --provider, and
/model <provider:model> paths are unchanged. Because group_providers()
is the shared fold, the CLI 'hermes model' picker, the setup wizard,
and the Telegram /model keyboard all pick up the grouping with no
per-surface changes.
#60194 flipped SessionResetPolicy's default to mode: none, but
cli-config.yaml.example still shipped session_reset.mode: both. Every
install path (install.sh, install.ps1, docker stage2-hook, hermes
doctor) copies the template verbatim to ~/.hermes/config.yaml, so fresh
installs got an EXPLICIT mode: both that overrides the code default —
users hit 24h-idle resets with 'nothing' in their config enabling it.
- cli-config.yaml.example: session_reset.mode both -> none, comments
rewritten to describe auto-reset as opt-in
- docs/session-lifecycle.md: appendix example updated to match
- tests/gateway/test_config.py: invariant tests — template seed, absent
config, and mode-less session_reset block all resolve to mode none;
explicit opt-in still honored
_find_tail_cut_by_tokens aligns cut_idx away from tool-call/result
boundaries (_align_boundary_backward), and both tail anchors re-align after
moving it. The final statement then raised the result to head_end + 1 so
compression always claims at least one message — without that floor the
caller's compress_start >= compress_end guard turns the pass into a no-op
that re-runs forever.
That raise discarded the alignment. When the floor landed inside a tool
group, the parent assistant(tool_calls) fell in the summarised region while
its tool results started the tail, and _sanitize_tool_pairs dropped those
orphans outright — so the tool output was neither summarised nor kept. It
vanished. That is exactly the silent loss _align_boundary_backward's own
docstring says the alignment exists to prevent.
Two back-to-back tool calls are enough to trigger it on default settings
(protect_first_n=3):
system, assistant(call_1), tool, tool, assistant(call_2), tool
aligned cut = 4 (keeps call_2's group together)
returned cut = 5 (floor overrode it)
summarised region = [assistant(call_2)]
tail = [tool(call_2)] -> orphan -> dropped
Sweeping every well-formed block layout up to length 6 (21840 transcripts),
5623 of them — 26% — split a call/result pair this way.
Re-align FORWARD after applying the floor. Forward, never backward: pulling
back would hand return the message the floor just claimed and reopen the
no-op loop. Sliding forward instead moves the cut past the end of the group,
so the whole call/result pair is summarised together and nothing is
orphaned. The same sweep reports 0 violations after the change, and the
progress guarantee is pinned by its own test.
Salvage of PR #67447 — the original PR fixed 3 of 7 missing keys.
gateway/config.py reads 4 more top-level keys (stt_echo_transcripts,
reset_triggers, always_log_local, filter_silence_narration) that
produced the same false 'Unknown top-level config key' warning.
Add all 4 and extend the regression test to cover them.
Regression for known_plugin_toolsets / group_sessions_per_user /
thread_sessions_per_user so validate_config_structure no longer
false-positives on keys Hermes owns.
