feat(conversation): Auto phase routing with aionrs model hot-swap

[JeffC0628]

Summary

• Phase 2 for [Feature]: Auto planner/worker model routing for long-running aionrs agents #4143: before each Auto send, decide planner vs worker, resolve the slot model, and hot-swap via config-options/model when the provider matches.
• Pill shows Auto · {phase}/{model}; Guid first create seeds planner.
• Depends on AionCore hot-swap: feat(aionrs): hot-swap model without agent rebuild AionCore#923 (graceful fallback if Core lacks the option).

Test plan

• With Core feat: fixed the issue where the agent icon data sources of the three … #923: Auto conversation first message uses planner slot; follow-ups use worker without agent rebuild
• Replan-style user text escalates to planner
• Cross-provider slot bindings still rebuild (expected)
• bun run test -- tests/unit/renderer/autoModel
• Without Core feat: fixed the issue where the agent icon data sources of the three … #923: send still works (hot-swap fails soft, persist-only)

Related: Phase 1 PR #4144

[codecov]

Codecov Report

❌ Patch coverage is 48.14815% with 182 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...tings/SettingsModal/contents/ModelModalContent.tsx	53.01%	23 Missing and 16 partials ⚠️
...s/desktop/src/renderer/utils/autoModel/settings.ts	20.00%	20 Missing ⚠️
...es/conversation/platforms/aionrs/AionrsSendBox.tsx	17.39%	18 Missing and 1 partial ⚠️
...p/src/renderer/hooks/agent/useAutoModelSettings.ts	0.00%	17 Missing ⚠️
...sation/platforms/aionrs/useAionrsModelSelection.ts	40.00%	15 Missing ⚠️
...renderer/pages/guid/hooks/useGuidModelSelection.ts	0.00%	14 Missing ⚠️
...p/src/renderer/utils/autoModel/resolveAutoModel.ts	74.46%	2 Missing and 10 partials ⚠️
...versation/platforms/aionrs/AionrsModelSelector.tsx	15.38%	5 Missing and 6 partials ⚠️
...c/renderer/pages/guid/components/GuidActionRow.tsx	8.33%	11 Missing ⚠️
...nderer/pages/guid/components/GuidModelSelector.tsx	18.18%	4 Missing and 5 partials ⚠️
... and 5 more

📢 Thoughts on this report? Let us know!

[IceyLiu]

Thanks for this — the direction is right and the engineering hygiene is good (self-contained autoModel/ module, PRD included, all 13 locales complete, unit tests present).

I ran the full gate locally on feat/auto-phase2-routing:

Check	Result
bunx tsc --noEmit	✅ 0 errors
bun run test -- tests/unit/renderer/autoModel	✅ 9/9 passed
bun run i18n:types + scripts/check-i18n.js	✅ in sync, all locales
bun run lint --quiet	✅ 0 errors

Note for reviewers: the feat/auto-phase2-routing branch already contains all commits from #4144, so merging this PR covers both.

Below are the blocking and should-fix items before merge.

---

🔴 Blocking: unmerged AionCore dependency + a regression on the existing model-switch path

This PR depends on AionCore#923, which is still open. The Auto path itself degrades gracefully (setConfigOption is wrapped in try/catch), so that part is fine.

The problem is ChatConversation.tsx — onSelectModel no longer stops the active turn when the provider is unchanged:

const sameProvider = conversation.model?.id === selected.id;
if (runtimeView.activeTurnId && !sameProvider) { /* stop */ }

This changes behaviour for all manual model switches, not just Auto. Without Core#923, Core does not hot-swap either, so a user switching models mid-turn within the same provider gets a UI that says the model changed while the turn keeps running on the old one — a silent no-op regression on a path everyone uses.

Please either land and release Core#923 first, or split this sameProvider stop-skip out of the PR so it ships together with the Core support.

🟠 Vision filtering is dead code — image sends can pick a text-only model

requireVision is declared in ResolveAutoModelInput and honoured inside resolveAutoModel, but no caller ever passes it:

resolveAutoModel.ts:34   if (input.requireVision && ...)
resolveAutoModel.ts:53   if (input.requireVision && ...)
types.ts:48              requireVision?: boolean;

AionrsSendBox has files in scope at the applyAutoModelForTurn call site, and handleSelectAuto has no vision context at all. Sending an image in an Auto conversation can therefore resolve a non-vision worker model and fail the request. Please thread requireVision through from the send path (and consider re-resolving when attachments are present).

🟠 The tier heuristics rank the wrong models, and the preference switch is a no-op

I ran scoreModelForSlots / preferenceWeight over a realistic catalogue (claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5, gpt-5, gpt-4o-mini, deepseek-chat, kimi-k2, glm-4.6, qwen3-coder-plus, gemini-3-pro-preview):

cost/planner    -> gemini-3-pro-preview:5   opus:4  gpt-5:4  sonnet:3
balance/planner -> gemini-3-pro-preview:7   opus:4  gpt-5:4  sonnet:3
quality/planner -> gemini-3-pro-preview:14  opus:8  gpt-5:8  sonnet:6

cost/worker     -> qwen3-coder-plus:12  gemini-3-pro-preview:10  sonnet:8  haiku:8
balance/worker  -> qwen3-coder-plus:12  sonnet:8  deepseek-chat:8
quality/worker  -> qwen3-coder-plus:12  sonnet:8.9  gemini-3-pro-preview:8.1

cost/utility    -> haiku:6  gpt-4o-mini:6  gemini-3-pro-preview:6
balance/utility -> haiku:4  gpt-4o-mini:4  gemini-3-pro-preview:4

Four concrete problems:

1. gemini-3-pro-preview always wins planner, ahead of Opus and GPT-5, because it double-scores on both /pro-preview/i (+4) and /gemini.*pro/i (+3) while /opus/i only yields +4.
2. qwen3-coder-plus always wins worker (12) by double-scoring /qwen.*plus/i and /coder/i. Any model with coder in its name captures the worker seat regardless of size or price.
3. cost / balance / quality produce identical winners in every seat. The preference selector in Settings has no observable effect on this catalogue.
4. tierHeuristics.ts:41 is an identity expression:

   utility: Math.max(utility, planner > 6 ? 0 : utility)

   Math.max(u, 0) === u for all u >= 0, and Math.max(u, u) === u. The intended "zero out utility for strong planners" never happens — which is why gemini-3-pro-preview also ties for the utility seat.

Also, kimi-k2 and glm-4.6 score {planner:0, worker:4, utility:0} — treated as unknown. Regex-guessing model names is going to keep drifting as new models ship. Consider an explicit capability/tier table, or requiring explicit slot bindings and treating automatic as a last resort.

🟠 hasPriorUserTurns resets on every conversation open, so reopening re-runs planner

useEffect(() => { autoUserTurnsRef.current = 0; }, [conversation_id]);
...
hasPriorUserTurns: autoUserTurnsRef.current > 0

The counter is in-memory only. Reopen a 50-turn Auto conversation after a restart and the next message is classified as the first turn → planner. That defeats the cost-saving goal in the most common long-running scenario, which is exactly the one #4143 targets. Derive this from message history or from the persisted extra.auto_model state instead.

🟡 Two features are wired to nothing

• consecutiveWorkerFailures is read in decideAutoModelPhase and forwarded by applyAutoModelForTurn, but no production caller ever supplies it. The "≥2 worker failures escalate to planner" rule in the PRD is unreachable outside the unit test.
• The utility slot is configurable in Settings but never consumed by any resolve path (PRD defers it to F-AUTO-07). Users can bind a model to a seat that does nothing — please hide it until it is wired, or wire it.

🟡 AutoModelSettingsBlock renders above the empty state

In ModelModalContent.tsx the block is now rendered unconditionally at the top of the scroll area, ahead of the !data || data.length === 0 branch. A first-run user with zero providers sees three empty slot dropdowns before the "no configured models" onboarding copy. Suggest moving it below the provider list and gating it on data?.length.

🟡 Auto is silently skipped for team conversations

The send path is guarded by if (autoEnabled && !teamSendMessage), but the pill still renders Auto · {phase}/{model}. The UI claims routing that isn't happening.

🟡 Per-turn latency: 3 extra round trips before each send

useAcpConfigOptions.setConfigOption does prepareSetRuntime() → fetchConfigOptionsOnce() → PUT config-options/model, and then applyAutoModelForTurn issues a PATCH /conversations/:id. That's 3–4 sequential round trips added in front of every Auto send, on the critical path to first token. Worth short-circuiting when the resolved model equals the current one (the alreadyOnModel check covers the PUT but not the PATCH), and consider skipping the pre-fetch.

🟡 Possible bypass of a locked model option

handleSheetModelSelect guards on runtimeConfig.isConfigOptionBlocked?.('model'), but the Auto send path does not. If setConfigOption throws config_update_in_progress (or the option is blocked by an assistant that pins the model), the error is swallowed by the try/catch and the code still persists conversation.model. Please confirm Auto respects the same block.

---

Housekeeping

The branch is 12 commits behind main — please rebase so the merge diff can be re-checked.

---

Happy to re-review once the Core dependency is resolved and the vision / first-turn / heuristics items are addressed. The module boundaries and test setup are solid, so most of this is contained work.

[JeffC0628]

Addressed @IceyLiu review feedback in latest push (rebased on main):

• Blocking: Restored stop-on-model-change for manual model switches during active turns (same-provider hot-swap only on Auto per-turn path).
• Should-fix:
  equireVision\ wired when send includes image attachments; tier heuristics/preference scoring fixed; Auto settings hidden when no providers configured.

Superseded #4144 (closed). Auto hot-swap at send time still depends on AionCore #923.

[JeffC0628]

Addressed @IceyLiu review feedback in latest push (rebased on main):

• Blocking: Restored stop-on-model-change for manual model switches during active turns (same-provider hot-swap only on Auto per-turn path).
• Should-fix:
  equireVision\ wired when send includes image attachments; tier heuristics/preference scoring fixed; Auto settings hidden when no providers configured.

Superseded #4144 (closed). Auto hot-swap at send time still depends on AionCore #923.

[JeffC0628]

Addressed @IceyLiu review feedback in latest push (rebased on main):

• Blocking: Restored stop-on-model-change for manual model switches during active turns (same-provider hot-swap only on Auto per-turn path).
• Should-fix:
  equireVision\ wired when send includes image attachments; tier heuristics/preference scoring fixed; Auto settings hidden when no providers configured.

Superseded #4144 (closed). Auto hot-swap at send time still depends on AionCore #923.

[JeffC0628]

Addressed @IceyLiu review feedback (rebased on main, CI green):

• Blocking: Restored stop-on-model-change for manual model switches during active turns.
• Should-fix: requireVision when send includes images; tier heuristics/preference scoring fixed; Auto settings hidden when no providers configured.

#4144 closed (superseded). Auto per-turn hot-swap still depends on AionCore #923.

[JeffC0628]

Additional hardening after self-review:

• Fix: \hasPriorUserTurns\ now derives from loaded message history (\conversationHasUserTurns) instead of a session-local counter — reopening an existing Auto conversation no longer mis-routes every send to planner.
• Tests: added \�pplyAutoModelForTurn, vision/history helpers, requireVision filtering, and balance worker selection coverage (22 autoModel tests total).