Skip to content

fix(chat): model-dropdown switches actually stick - #73

Merged
KrasimirKralev merged 4 commits into
ID-Robots:betafrom
KrasimirKralev:fix/chat-model-switcher-sticks
Apr 18, 2026
Merged

fix(chat): model-dropdown switches actually stick#73
KrasimirKralev merged 4 commits into
ID-Robots:betafrom
KrasimirKralev:fix/chat-model-switcher-sticks

Conversation

@KrasimirKralev

@KrasimirKralev KrasimirKralev commented Apr 18, 2026

Copy link
Copy Markdown
Contributor

Summary

fix(chat): green confirmation pill on model switch, keep red for errors

The chat panel's ChatMessage type uses the same role: 'system'
for both error banners (WebSocket dropped, model switch failed) and
the "Switched chat to X." confirmation we emit after a successful
dropdown pick. The shared style renders both in a red pill with red
text, so a successful model switch looks like a failure in the
stream. In the screenshot the user showed, "Switched chat to
llamacpp/gemma4-e2b-it-q4_0." appeared directly above a real
Error: banner in identical styling — no visual distinction.

Changes in src/components/ChatPopup.tsx:

  • Add an optional variant: 'success' | 'error' to the ChatMessage
    interface. Defaults to 'error' (no behaviour change for existing
    call sites: all the error paths keep their red styling).
  • Tag the Switched chat to X. message with variant: 'success'.
  • Update the inline message styles in the renderer to branch on
    msg.variant for system-role bubbles: emerald (#22c55e fg,
    rgba(34,197,94,0.15) bg) for success, keep the existing red
    (#ef4444 fg, rgba(239,68,68,0.15) bg) for errors.

Other system-role call sites — WS error banners, connection
failures, "provider not configured" hints — all keep the default
'error' variant and render red, so nothing else regresses.

fix(chat): model-dropdown switches actually stick to the active session

The ClawBox chat-panel model dropdown (ChatPopup.tsx) POSTs to
/setup-api/chat/model when the user picks a model. That route
previously did only two things:

await exec(openclaw, ["config", "set", "agents.defaults.model.primary", target]);
await restartGateway();

Both writes are correct, and new sessions opened after the restart
inherit the chosen model. But they left the currently open chat
session's per-session override untouched. That override lives in
<agents-dir>/<agent>/sessions/sessions.json as:

"providerOverride": "deepseek",
"modelOverride": "deepseek-chat",
"modelOverrideSource": "auto",
"authProfileOverride": "deepseek:default",
...

and it's what OpenClaw actually consults at chat time to decide which
provider to dial. Because the source was "auto", OpenClaw's own
auto-picker happily rewrote it to whatever it preferred on every new
message — usually the local provider when reachable — so the chat

Test plan

  • Deployed and verified live on Jetson Orin during development session (2026-04-18)
  • CodeRabbit bot review (triggered automatically on draft open)

@coderabbitai

coderabbitai Bot commented Apr 18, 2026 •
edited
Loading

Copy link
Copy Markdown

Warning

Rate limit exceeded

@KrasimirKralev has exceeded the limit for the number of commits that can be reviewed per hour. Please wait 5 minutes and 31 seconds before requesting another review.

Your organization is not enrolled in usage-based pricing. Contact your admin to enable usage-based pricing to continue reviews beyond the rate limit, or try again in 5 minutes and 31 seconds.

⌛ How to resolve this issue?

After the wait time has elapsed, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

We recommend that you space out your commits to avoid hitting the rate limit.

🚦 How do rate limits work?

CodeRabbit enforces hourly rate limits for each developer per organization.

Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout.

Please see our FAQ for further information.

ℹ️ Review info
⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 0e72b08d-5df6-4872-9fe1-163cb0d1876a

📥 Commits

Reviewing files that changed from the base of the PR and between f4260b2 and cdbee87.

📒 Files selected for processing (4)
  • src/app/setup-api/chat/model/route.ts
  • src/components/ChatPopup.tsx
  • src/lib/openclaw-config.ts
  • src/tests/routes/chat-model.test.ts
✨ Finishing Touches
🧪 Generate unit tests (beta)
  • Create PR with unit tests

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

❤️ Share

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

@KrasimirKralev
KrasimirKralev force-pushed the fix/chat-model-switcher-sticks branch from 360e50f to 8e1c657 Compare April 18, 2026 14:10
@KrasimirKralev

KrasimirKralev commented Apr 18, 2026

Copy link
Copy Markdown
Contributor Author

@coderabbitai review

@coderabbitai

coderabbitai Bot commented Apr 18, 2026

Copy link
Copy Markdown
✅ Actions performed

Review triggered.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

@KrasimirKralev
KrasimirKralev force-pushed the fix/chat-model-switcher-sticks branch from 8e1c657 to ccec7d6 Compare April 18, 2026 17:15
@KrasimirKralev
KrasimirKralev marked this pull request as ready for review April 18, 2026 17:48
@KrasimirKralev
KrasimirKralev requested a review from a team as a code owner April 18, 2026 17:48
@KrasimirKralev

KrasimirKralev commented Apr 18, 2026

Copy link
Copy Markdown
Contributor Author

@coderabbitai review

@coderabbitai

coderabbitai Bot commented Apr 18, 2026

Copy link
Copy Markdown
✅ Actions performed

Review triggered.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

@KrasimirKralev
fix(chat): model-dropdown switches actually stick to the active session
3b7c174 
The ClawBox chat-panel model dropdown (`ChatPopup.tsx`) POSTs to
`/setup-api/chat/model` when the user picks a model. That route
previously did only two things:

  await exec(openclaw, ["config", "set", "agents.defaults.model.primary", target]);
  await restartGateway();

Both writes are correct, and new sessions opened after the restart
inherit the chosen model. But they left the *currently open chat
session's per-session override* untouched. That override lives in
`<agents-dir>/<agent>/sessions/sessions.json` as:

  "providerOverride":       "deepseek",
  "modelOverride":          "deepseek-chat",
  "modelOverrideSource":    "auto",
  "authProfileOverride":    "deepseek:default",
  ...

and it's what OpenClaw actually consults at chat time to decide which
provider to dial. Because the source was `"auto"`, OpenClaw's own
auto-picker happily rewrote it to whatever it preferred on every new
message — usually the local provider when reachable — so the chat
panel dropdown *appeared* to have no effect from the user's point
of view. They would click "ClawBox AI", the pill would update, the
next answer would still come from llama.cpp. Reproduced live on a
Jetson: ran the CPU watch monitor on llama-server while the user
picked ClawBox AI in the dropdown and sent messages — llama-server
CPU stayed flat, the gateway's outbound TLS to openclawhardware.dev
never opened, the dropdown was effectively decorative.

Fix:

- src/lib/openclaw-config.ts: new `applyModelOverrideToAllAgentSessions`
  helper that walks every sessions.json under the agents directory
  and rewrites the seven override fields to the given target, tagged
  with the given source (default `"manual"`). Atomic temp+rename
  writes so a crash mid-sweep can't corrupt sessions. Individual
  bad files are logged and skipped rather than aborting the whole
  sweep. Exported alongside a small `parseFullyQualifiedModel`
  helper for the `"<provider>/<modelId>"` format the rest of the
  codebase already uses.

- src/app/setup-api/chat/model/route.ts: after the existing
  `agents.defaults.model.primary` write, call the new helper with
  `source: "manual"` so OpenClaw's auto-picker leaves the chosen
  override alone on subsequent messages. Sweep failures are logged
  but don't fail the request — worst case the default write still
  takes effect for newly-opened sessions. Also bump the per-call
  timeout on the underlying `openclaw config set` invocation from
  10 s to 30 s: baseline CLI startup on Jetson Orin is 10-12 s
  (measured with three sequential idle runs), so the previous bound
  was guaranteed to trip and surface a spurious "command failed …
  timed out after 10000ms" to the browser on every legitimate
  dropdown click. The helper alias is renamed `exec` → `runOpenclaw`
  inline to make it clear at the call site that we're invoking the
  `openclaw` CLI via `execFile` (no shell), not the `exec` family.

Result: picking a model in the chat panel dropdown now changes the
running session immediately and the change survives further messages.
The auto-picker respects explicit user choices (since those are now
tagged `manual`) and only fires on untouched sessions.

The session-sweep helper is exported from `src/lib/openclaw-config.ts`
so the pending Local-only-mode session-overrides fix (separate
branch) can share the same implementation rather than duplicating
the traversal and write logic.
@KrasimirKralev
fix(chat): green confirmation pill on model switch, keep red for errors
e2cf68f 
The chat panel's `ChatMessage` type uses the same `role: 'system'`
for both error banners (WebSocket dropped, model switch failed) and
the "Switched chat to X." confirmation we emit after a successful
dropdown pick. The shared style renders both in a red pill with red
text, so a successful model switch looks like a failure in the
stream. In the screenshot the user showed, "Switched chat to
llamacpp/gemma4-e2b-it-q4_0." appeared directly above a real
`Error:` banner in identical styling — no visual distinction.

Changes in src/components/ChatPopup.tsx:

- Add an optional `variant: 'success' | 'error'` to the `ChatMessage`
  interface. Defaults to `'error'` (no behaviour change for existing
  call sites: all the error paths keep their red styling).
- Tag the `Switched chat to X.` message with `variant: 'success'`.
- Update the inline message styles in the renderer to branch on
  `msg.variant` for `system`-role bubbles: emerald (`#22c55e` fg,
  `rgba(34,197,94,0.15)` bg) for success, keep the existing red
  (`#ef4444` fg, `rgba(239,68,68,0.15)` bg) for errors.

Other system-role call sites — WS error banners, connection
failures, "provider not configured" hints — all keep the default
`'error'` variant and render red, so nothing else regresses.
@KrasimirKralev
test(chat-model): mock applyModelOverrideToAllAgentSessions + parseFu…
6c6660f 
…llyQualifiedModel

PR ID-Robots#67's merge into beta brought runOpenclawConfigSet and the test
mocks for it. This branch additionally adds two more openclaw-config
helpers that the route now imports — extend the mock factory and the
beforeEach with sensible defaults so the test file compiles and the
route code paths that call them don't throw.
@KrasimirKralev
refactor(chat): post-simplify cleanups from /simplify pass
cdbee87 
Applied code-simplifier to the branch diff before pushing for CodeRabbit review.
Two substantive improvements surfaced:

- openclaw-config.ts: collapsed the dual-name `SESSION_OVERRIDE_FIELDS` +
  `SESSION_OVERRIDE_FIELD_LIST` alias into a single direct export. The
  re-export was also throwing away the `as const` literal types by
  retyping to `readonly string[]`.
- openclaw-config.ts: fixed a latent counter bug in
  applyModelOverrideToAllAgentSessions — `sessionsUpdated` was previously
  incremented inside the mutation loop before the atomic write, so a
  write failure would inflate the returned count. Now only accrues on
  successful write via a `touchedInFile` local.
- ChatPopup.tsx: flattened the new triple-nested ternary for
  success-vs-error banner coloring by lifting `isSuccess`,
  `systemBg`, `systemColor` into locals in the message map callback.

No behavior change aside from the counter-accuracy fix, which only
matters when a sessions.json write fails mid-sweep.
@KrasimirKralev
KrasimirKralev force-pushed the fix/chat-model-switcher-sticks branch from ccec7d6 to cdbee87 Compare April 18, 2026 18:05
@KrasimirKralev

KrasimirKralev commented Apr 18, 2026

Copy link
Copy Markdown
Contributor Author

@coderabbitai review

@coderabbitai

coderabbitai Bot commented Apr 18, 2026

Copy link
Copy Markdown
✅ Actions performed

Review triggered.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

@KrasimirKralev
KrasimirKralev merged commit 7966e11 into ID-Robots:beta Apr 18, 2026
5 checks passed
KrasimirKralev added a commit to KrasimirKralev/clawbox that referenced this pull request Apr 18, 2026
@KrasimirKralev
fix(ai-provider): sweep sessions, show progress + green banner on switch
bbccfd9 
Changing the primary AI provider in Settings (e.g. OpenAI → ClawBox AI)
had three user-visible problems:

1. Chat kept responding as the OLD provider after the switch.
   The configure route updated agents.defaults.model.primary but
   left every existing session's per-session override pointing at
   the previous model. OpenClaw's resolver respected those sticky
   overrides on the next turn, so the chat stayed bound to the old
   provider until the user cleared the session — or refreshed the
   page and noticed it still answered wrong.

2. Chat 'froze' during the gateway restart. The chat popup only
   had skill-install-aware retry logic, so a provider-change gateway
   restart dropped the WS and the user saw no progress; only the
   bare retry loop slowly reconnected.

3. No affirmative 'switched to X' feedback — when reconnect finally
   landed, the chat looked identical to before. The dropdown already
   shows a green 'Switched chat to X' banner when switching models
   from within the chat pane (PR ID-Robots#73); provider changes via Settings
   deserved the same.

Changes:

- src/app/setup-api/ai-models/configure/route.ts: after
  'agents.defaults.model.primary' is written, sweep per-session
  model overrides via applyModelOverrideToAllAgentSessions with
  source: 'user' so OpenClaw's resolver keeps the new primary
  bound across existing sessions. Mirrors the sweep in
  /setup-api/chat/model from PR ID-Robots#73. Only runs when this call
  actually set a new primary (skips local-only scopes).

- src/components/ChatPopup.tsx: listens for the existing
  clawbox:primary-ai-configured event (dispatched by SettingsApp
  after provider configure succeeds) alongside
  clawbox-skill-installed. Reuses the reloadingSkill overlay +
  progress bar + quadrupled retry budget — makes the gateway
  restart visible instead of letting the chat look frozen. Banner
  label adapts ('Switching AI provider...' vs 'Reloading skills...')
  via a new reloadReason state; a matching reloadReasonRef escapes
  the frozen closure captured by the one-time WebSocket hello
  resolve callback. After reconnect on a provider change, the
  auto-'my skills were just updated' prompt is skipped (no skill
  changed, nothing to confirm) and a green 'Switched chat to
  <label>' banner is pushed using the refreshed chat/model state.

- src/app/setup-api/chat/model/route.ts: PROVIDER_LABELS map was
  missing 'openai-codex', so labelForProvider fell back to the
  generic 'AI Provider' placeholder — which made the new banner
  read 'Switched chat to AI Provider' when switching to or from
  openai-codex. Added 'openai-codex: "OpenAI Codex"'.

Post-fix /simplify pass converted the inline fetch().then().then()
chain in ChatPopup's reload branch to async/await to match the
existing refreshChatModelState pattern in the same file. Semantics
unchanged.
@KrasimirKralev KrasimirKralev mentioned this pull request Apr 18, 2026
Merged
4 tasks
KrasimirKralev added a commit that referenced this pull request Apr 19, 2026
@KrasimirKralev
fix(ai-provider): sweep sessions + show switch progress + green banner (
4c8e4c5 
#83)

* fix(ai-provider): sweep sessions, show progress + green banner on switch

Changing the primary AI provider in Settings (e.g. OpenAI → ClawBox AI)
had three user-visible problems:

1. Chat kept responding as the OLD provider after the switch.
   The configure route updated agents.defaults.model.primary but
   left every existing session's per-session override pointing at
   the previous model. OpenClaw's resolver respected those sticky
   overrides on the next turn, so the chat stayed bound to the old
   provider until the user cleared the session — or refreshed the
   page and noticed it still answered wrong.

2. Chat 'froze' during the gateway restart. The chat popup only
   had skill-install-aware retry logic, so a provider-change gateway
   restart dropped the WS and the user saw no progress; only the
   bare retry loop slowly reconnected.

3. No affirmative 'switched to X' feedback — when reconnect finally
   landed, the chat looked identical to before. The dropdown already
   shows a green 'Switched chat to X' banner when switching models
   from within the chat pane (PR #73); provider changes via Settings
   deserved the same.

Changes:

- src/app/setup-api/ai-models/configure/route.ts: after
  'agents.defaults.model.primary' is written, sweep per-session
  model overrides via applyModelOverrideToAllAgentSessions with
  source: 'user' so OpenClaw's resolver keeps the new primary
  bound across existing sessions. Mirrors the sweep in
  /setup-api/chat/model from PR #73. Only runs when this call
  actually set a new primary (skips local-only scopes).

- src/components/ChatPopup.tsx: listens for the existing
  clawbox:primary-ai-configured event (dispatched by SettingsApp
  after provider configure succeeds) alongside
  clawbox-skill-installed. Reuses the reloadingSkill overlay +
  progress bar + quadrupled retry budget — makes the gateway
  restart visible instead of letting the chat look frozen. Banner
  label adapts ('Switching AI provider...' vs 'Reloading skills...')
  via a new reloadReason state; a matching reloadReasonRef escapes
  the frozen closure captured by the one-time WebSocket hello
  resolve callback. After reconnect on a provider change, the
  auto-'my skills were just updated' prompt is skipped (no skill
  changed, nothing to confirm) and a green 'Switched chat to
  <label>' banner is pushed using the refreshed chat/model state.

- src/app/setup-api/chat/model/route.ts: PROVIDER_LABELS map was
  missing 'openai-codex', so labelForProvider fell back to the
  generic 'AI Provider' placeholder — which made the new banner
  read 'Switched chat to AI Provider' when switching to or from
  openai-codex. Added 'openai-codex: "OpenAI Codex"'.

Post-fix /simplify pass converted the inline fetch().then().then()
chain in ChatPopup's reload branch to async/await to match the
existing refreshChatModelState pattern in the same file. Semantics
unchanged.

* fix(chat): address CodeRabbit #83 — preserve transcript + force hello on provider change

Two Major findings on PR #83:

1. Provider switches were wiping the chat transcript. setMessages([])
   ran unconditionally in the skillInstalledRef branch before the
   wasProviderChange check, so a Settings provider change cleared
   the user's visible history. Skill installs still want the reset
   (new tools may change the assistant's behavior), but provider
   changes don't — the conversation semantics are unchanged.
   Guarded the reset + greetedRef bump behind !wasProviderChange.

2. Provider-reload overlay could get stuck on fast gateway restarts.
   The configure route restarts the gateway before returning its
   response, and the primary-ai-configured event fires only after
   that response lands. If the WS retry loop has already
   reconnected by then, no future hello triggers the reload branch
   and the overlay stays up indefinitely.

   The provider handler now, after seeding state:
   - closes any currently-open WebSocket (so the resolve-branch
     fires on a fresh connect rather than relying on a stale hello
     that missed skillInstalledRef=true)
   - resets retryCountRef and calls connect() directly

   This makes the completion independent of any in-flight hello's
   timing relative to when the event is dispatched.
@KrasimirKralev
KrasimirKralev deleted the fix/chat-model-switcher-sticks branch July 4, 2026 17:42
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@KrasimirKralev