feat(config): object-style fallback_models with per-model settings#2622
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d873abb88a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
There was a problem hiding this comment.
1 issue found across 8 files
Confidence score: 3/5
- There is a concrete regression risk in
src/hooks/runtime-fallback/auto-retry.ts: object-style per-model settings can be lost during auto-retry because config is flattened to a string and only provider/model(/variant) is reconstructed. - Given the severity (6/10) and high confidence (9/10), this is more than a minor edge case and could cause user-facing behavior changes when retries occur with custom model options.
- Pay close attention to
src/hooks/runtime-fallback/auto-retry.ts- ensure auto-retry preserves full per-model object settings instead of dropping non-provider/model fields.
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/hooks/runtime-fallback/auto-retry.ts">
<violation number="1" location="src/hooks/runtime-fallback/auto-retry.ts:118">
P2: Auto-retry fallback drops object-style per-model settings by flattening to string and reconstructing only provider/model(/variant).</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2d82727956
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
RaviTharuma
force-pushed
the
feat/object-style-fallback-models
branch
from
2c3092e to
69e0f70
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 69e0f70910
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
2 issues found across 3 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/tools/delegate-task/subagent-resolver.ts">
<violation number="1" location="src/tools/delegate-task/subagent-resolver.ts:120">
P2: Object-style fallback model settings are discarded during static fallback resolution because fallback objects are flattened to strings before selection.</violation>
</file>
<file name="src/tools/delegate-task/category-resolver.ts">
<violation number="1" location="src/tools/delegate-task/category-resolver.ts:105">
P2: Flattening fallback objects to strings drops per-model settings when a fallback model is promoted to the initial execution model.</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
|
Follow-up: the newly reported delegate-task issues are fixed in |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 31ba06ce66
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
4 issues found across 17 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/shared/model-resolver.ts">
<violation number="1" location="src/shared/model-resolver.ts:96">
P1: Custom agent: **Opencode Compatibility**
Add `"minimal"` to the list of supported inline variants, as it is a valid variant supported by Opencode (e.g., for OpenAI models).</violation>
</file>
<file name="src/tools/delegate-task/sync-prompt-sender.ts">
<violation number="1" location="src/tools/delegate-task/sync-prompt-sender.ts:80">
P2: `categoryModel.maxTokens` and `categoryModel.thinking` are added and populated upstream but never included in the prompt payload, so these fallback model settings are silently ignored at runtime.</violation>
</file>
<file name="src/features/background-agent/manager.ts">
<violation number="1" location="src/features/background-agent/manager.ts:488">
P2: `resume()` drops new per-model settings (`temperature`, `top_p`, `reasoningEffort`), so resumed tasks can run with different model behavior than initial launch.</violation>
</file>
<file name="src/tools/delegate-task/category-resolver.ts">
<violation number="1" location="src/tools/delegate-task/category-resolver.ts:23">
P2: `getFallbackModelConfig` uses exact string equality to match `actualModel` against fallback chain entries, but the model resolver can return a fuzzy match (e.g., `openai/gpt-5.4-preview` when the user configured `openai/gpt-5.4`). In that case the lookup misses and all per-entry settings (`reasoningEffort`, `temperature`, `top_p`, etc.) are silently dropped. Consider using a prefix or `startsWith` match on the model ID portion to handle resolved model names that extend the configured base name.</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 809b1a2750
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
5 issues found across 14 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/features/background-agent/manager.ts">
<violation number="1" location="src/features/background-agent/manager.ts:768">
P1: Custom agent: **Opencode Compatibility**
The Opencode SDK and `PromptInput` schema do not support `temperature` or `top_p` in the `model` object.</violation>
<violation number="2" location="src/features/background-agent/manager.ts:787">
P1: Custom agent: **Opencode Compatibility**
The Opencode SDK and `PromptInput` schema do not support passing an `options` object (such as `reasoningEffort`, `thinking`, or `maxTokens`) in a prompt request.</violation>
</file>
<file name="src/tools/delegate-task/sync-prompt-sender.ts">
<violation number="1" location="src/tools/delegate-task/sync-prompt-sender.ts:87">
P1: Custom agent: **Opencode Compatibility**
The `options` object is not supported by the Opencode SDK's session prompt payload.</violation>
</file>
<file name="src/tools/delegate-task/subagent-resolver.ts">
<violation number="1" location="src/tools/delegate-task/subagent-resolver.ts:37">
P2: Combined exact/prefix matching inside a single `.find()` lets broad base-model entries shadow later exact specific-model configs.</violation>
</file>
<file name="src/tools/delegate-task/category-resolver.ts">
<violation number="1" location="src/tools/delegate-task/category-resolver.ts:25">
P2: The `startsWith(`${configuredModel}-`)` check is stricter than the upstream `fuzzyMatchModel` substring matching. If fuzzy resolution produces a model ID that extends the configured prefix without a hyphen (e.g., `gpt-5.4` → `gpt-5.4o`), this predicate misses and per-entry settings are silently dropped. Consider using a plain `startsWith(configuredModel)` check (without the trailing hyphen) to stay consistent with fuzzy matching, or use the same fuzzy matcher here.</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
Fixed and pushed in What changed:
Also fixed fallback-promotion matching:
Follow-up suite cleanup included in the same push:
Validation:
|
There was a problem hiding this comment.
1 issue found across 22 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/tools/delegate-task/category-resolver.ts">
<violation number="1" location="src/tools/delegate-task/category-resolver.ts:33">
P2: Unconstrained prefix fuzzy matching can select the wrong fallback entry for models with shared prefixes (e.g., gpt-4 vs gpt-4o).</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
️✅ There are no secrets present in this pull request anymore.If these secrets were true positive and are still valid, we highly recommend you to revoke them. 🦉 GitGuardian detects secrets in your source code to help developers and security teams secure the modern development process. You are seeing this because you or someone else with access to this repository has authorized GitGuardian to scan your pull request. |
|
Follow-up: I removed the secret-like test fixture strings that triggered the GitGuardian generic-entropy detector in |
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
@cubic-dev-ai review |
@acamq I have started the AI code review. It will take a few minutes to complete. |
There was a problem hiding this comment.
6 issues found across 38 files
Confidence score: 3/5
- There is moderate regression risk from concrete behavior gaps:
src/config/schema/fallback-models.tscurrently rejects valid Opencode/OpenAIreasoningEffortvalues ("none","minimal"), which can break compatible configs at runtime. src/tools/delegate-task/subagent-resolver.tshas a likely fallback resolution bug (high confidence) where built-infallbackChaincan be skipped when no user fallback models are set, so delegation may choose the wrong model path.- Other findings are mostly maintainability or lower-confidence concerns (duplicated parsing/type definitions in
src/shared/fallback-chain-from-models.tsandsrc/tools/delegate-task/sync-task.ts, plus uncertain stale session params insrc/tools/delegate-task/sync-prompt-sender.ts), which keeps this from being high risk but not fully low risk. - Pay close attention to
src/config/schema/fallback-models.ts,src/tools/delegate-task/subagent-resolver.ts, andsrc/tools/delegate-task/sync-prompt-sender.ts- config compatibility, fallback-chain correctness, and possible stale session tuning state.
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/config/schema/fallback-models.ts">
<violation number="1" location="src/config/schema/fallback-models.ts:6">
P1: Custom agent: **Opencode Compatibility**
Update the `reasoningEffort` enum to include `"none"` and `"minimal"`. Opencode explicitly supports these values for OpenAI models (as seen in `OPENAI_EFFORTS` inside the provider transform logic). Omitting them from the schema prevents users from configuring these legitimate fallback settings.</violation>
</file>
<file name="src/tools/delegate-task/subagent-resolver.test.ts">
<violation number="1" location="src/tools/delegate-task/subagent-resolver.test.ts:375">
P3: Test doesn’t exercise prefix-specific matching: available model exactly matches a fallback entry, so exact-match logic wins and prefix selection isn’t tested.</violation>
</file>
<file name="src/shared/fallback-chain-from-models.ts">
<violation number="1" location="src/shared/fallback-chain-from-models.ts:66">
P2: Duplicate model string parsing logic - `parseFallbackModelObjectEntry` duplicates ~10 lines of parsing code from `parseFallbackModelEntry` (trimming, splitting by `/`, provider fallback resolution, variant extraction). Refactor to call `parseFallbackModelEntry(obj.model, ...)` and merge additional properties.</violation>
</file>
<file name="src/tools/delegate-task/sync-prompt-sender.ts">
<violation number="1" location="src/tools/delegate-task/sync-prompt-sender.ts:73">
P2: Session-level prompt params are only set when tuning values exist, so later prompts without tuning for the same session leave stale temperature/topP/options in the session map and can incorrectly apply them.</violation>
</file>
<file name="src/tools/delegate-task/sync-task.ts">
<violation number="1" location="src/tools/delegate-task/sync-task.ts:20">
P2: Expanded inline `categoryModel` object type is duplicated across multiple delegate-task files, creating avoidable drift and maintainability risk; extract to a shared type/interface.</violation>
</file>
<file name="src/tools/delegate-task/subagent-resolver.ts">
<violation number="1" location="src/tools/delegate-task/subagent-resolver.ts:198">
P2: Promotion of fallback model config ignores built-in agent fallbackChain when no user fallback models are configured, because getFallbackModelConfig is called with configuredFallbackChain instead of the resolved fallbackChain.</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
Follow-up on the red |
|
@cubic-dev-ai All active review threads on this PR are resolved and the branch head is updated ( |
@RaviTharuma I have started the AI code review. It will take a few minutes to complete. |
There was a problem hiding this comment.
4 issues found across 38 files
Confidence score: 3/5
- There is a concrete medium-risk behavior in
src/features/opencode-skill-loader/config-source-discovery.ts: negation globs like!secret.mdcan be bypassed by the current OR logic between directory and file matches, so excluded files may still be discovered. src/tools/delegate-task/subagent-resolver.tshas a specificity bug in fallback selection (sorting by longest provider string instead of matching-prefix length), which can route to a less-appropriate fallback in some provider-name combinations.src/tools/delegate-task/sync-prompt-sender.tsmay carry stale session prompt params into fallback models, but confidence is low; together with duplication insrc/features/background-agent/manager.ts, this points to maintainability/regression risk rather than an immediate blocker.- Pay close attention to
src/features/opencode-skill-loader/config-source-discovery.tsandsrc/tools/delegate-task/subagent-resolver.ts- exclusion correctness and fallback-target selection have the highest user-facing impact.
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/tools/delegate-task/subagent-resolver.ts">
<violation number="1" location="src/tools/delegate-task/subagent-resolver.ts:50">
P2: Fallback matching sorts by the longest provider string in the entry rather than the length of the matching prefix, which can select a less-specific fallback when an unrelated provider name is longer.</violation>
</file>
<file name="src/features/background-agent/manager.ts">
<violation number="1" location="src/features/background-agent/manager.ts:515">
P2: Prompt-parameter normalization/mapping logic is duplicated in both launch and resume paths, increasing risk of drift when model params evolve.</violation>
</file>
<file name="src/features/opencode-skill-loader/config-source-discovery.ts">
<violation number="1" location="src/features/opencode-skill-loader/config-source-discovery.ts:43">
P2: Negation globs (e.g., `!secret.md`) are effectively bypassed because directory matches are OR’d with file matches; the negation matches non‑matching directory strings and short‑circuits to true, so explicitly excluded files are loaded.</violation>
</file>
<file name="src/tools/delegate-task/sync-prompt-sender.ts">
<violation number="1" location="src/tools/delegate-task/sync-prompt-sender.ts:64">
P2: Fallback models without explicit tuning overrides will inherit stale session prompt parameters because session params are only set when overrides exist and never cleared for a new model without overrides.</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
There was a problem hiding this comment.
3 issues found across 9 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/features/opencode-skill-loader/config-source-discovery.ts">
<violation number="1" location="src/features/opencode-skill-loader/config-source-discovery.ts:35">
P2: Extglob patterns starting with `!(...)` are misclassified as global negations, causing valid skills to be incorrectly filtered out.</violation>
</file>
<file name="src/hooks/thinking-block-validator/hook.ts">
<violation number="1" location="src/hooks/thinking-block-validator/hook.ts:119">
P1: Shared closure state (`thinkingStripped`) is used across async hook phases, creating cross-request race conditions and incorrect `chat.params` behavior under concurrency.</violation>
<violation number="2" location="src/hooks/thinking-block-validator/hook.ts:169">
P1: Removing empty assistant messages after stripping thinking blocks can create consecutive user messages, violating Anthropic API's alternating roles requirement</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
…blocks & drop empty messages When the validator strips thinking blocks from history, it must also disable the `thinking` request parameter — otherwise the API still enforces that every assistant message starts with a thinking block, causing 'Expected thinking but found tool_use' rejections. Added a shared closure flag between the two hook points: - messages.transform sets the flag when it strips - chat.params reads the flag and sets thinking=undefined Also filter out assistant messages left with empty parts arrays after stripping (e.g. thinking-only responses), which would cause separate API validation failures. Fixes cubic-dev-ai P0 and P2 findings on PR code-yeongyu#2622.
RaviTharuma
force-pushed
the
feat/object-style-fallback-models
branch
from
5137ddf to
8a9807d
Compare
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
Rebased onto latest
|
| Group | Commits | Files | Description |
|---|---|---|---|
| Core feature | d8cfe3a..3f6b6c3 |
5 | Object-style fallback_models config schema + model resolver |
| Fallback chain | 3c045fc..6cc155b |
6 | Prefix-match lookup, chain construction, promotion logic |
| Delegate-task | 8d480c2..051934f |
13 | Prompt payload alignment, subagent resolver, model selection |
| Model compat | 75cb051..d7fb439 |
3 | MODEL_FAMILY_REGISTRY, provider-agnostic detection, reasoning effort |
| Thinking blocks | 83c3464..a074115 |
5 | Strip-instead-of-inject, chat.params integration |
| Skill loader | b1d7cfb |
1 | Negated glob every() fix |
| GPT-legacy fix | e136d46 |
2 | Remove reasoningEffort from legacy GPT models |
| Misc fixes | 3df74f9..8a9807d |
4 | Review follow-ups, glob guard, test cleanup |
Each commit message has a conventional-commits scope that maps to these groups.
RaviTharuma
force-pushed
the
feat/object-style-fallback-models
branch
from
8a9807d to
36d7949
Compare
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
@cubic-dev-ai please review |
@RaviTharuma I have started the AI code review. It will take a few minutes to complete. |
There was a problem hiding this comment.
1 issue found across 34 files
Confidence score: 4/5
- This PR looks safe to merge overall; the reported issue is moderate (5/10) and describes a maintainability/regression risk rather than an immediate functional break.
- In
src/features/background-agent/spawner.ts, duplicated prompt-parameter mapping acrossstartTaskandresumeTaskcould drift over time, causing inconsistent model behavior when settings are updated. - Pay close attention to
src/features/background-agent/spawner.ts- duplicated mapping logic should be centralized to reduce future desync risk.
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="src/features/background-agent/spawner.ts">
<violation number="1" location="src/features/background-agent/spawner.ts:139">
P2: Duplicated model prompt-parameter mapping in `startTask` and `resumeTask` creates a desync risk when model settings evolve.</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
RaviTharuma
left a comment
RaviTharuma
left a comment
There was a problem hiding this comment.
Re: cubic-dev-ai review — Violation 1 (P2): Duplicated model prompt-parameter mapping
This is already addressed in the current code. The applyModelPromptParams helper (lines 21–42 of spawner.ts) was extracted exactly to centralize the mapping logic. Both startTask (line 164) and resumeTask (line 256) call the same shared function — no duplication.
The remaining parallel structure (launchModel/resumeModel and launchVariant/resumeVariant) is inherent to the two call sites having different input types (LaunchInput vs BackgroundTask). These are simple field extractions, not duplicated business logic, and they can't drift independently from the shared applyModelPromptParams.
No changes needed.
|
@acamq please check it again :) |
RaviTharuma
force-pushed
the
feat/object-style-fallback-models
branch
from
36d7949 to
1de5b66
Compare
|
@code-yeongyu I restacked this branch onto current |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1de5b66a8a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
…tings Add support for object-style entries in fallback_models arrays, enabling per-model configuration of variant, reasoningEffort, temperature, top_p, maxTokens, and thinking settings. - Zod schema for FallbackModelObject with full validation - normalizeFallbackModels() and flattenToFallbackModelStrings() utilities - Provider-agnostic model resolution pipeline with fallback chain - Session prompt params state management - Fallback chain construction with prefix-match lookup - Integration across delegate-task, background-agent, and plugin layers
There was a problem hiding this comment.
0 issues found across 8 files (changes from recent commits).
Requires human review: Large change (2800+ lines) affecting core model resolution, prompt payloads, and state management. High complexity makes 100% certainty against regressions difficult without manual verification.
RaviTharuma
force-pushed
the
feat/object-style-fallback-models
branch
from
a841eac to
fb08553
Compare
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
@code-yeongyu restacked this branch onto the current The rebase exposed one real failure on this branch: Local verification on
Fresh CI run for this push: |
|
@code-yeongyu final status on this stack after the latest restack:
From my side this PR is ready for your normal review/merge flow as stack item 1 of 3. |
3 participants
Summary
This PR makes
fallback_modelsactually usable across different model families.The real problem was that a fallback chain is not just a list of model IDs. Different fallback models often need different settings to work properly:
variantreasoningEfforttemperature,top_p,maxTokens, orthinkingsettingsWithout per-model settings, fallback was fragile:
A concrete example is falling back from Opus with
variant: "max"to Sonnet: the fallback inherits"max", but Sonnet does not support that level, so the fallback fails instead of rescuing the task.This PR fixes that by letting each fallback entry carry the settings that belong to that specific model, so the fallback chain behaves like a real fallback system instead of a best-effort model swap.
What changed
fallback_models, so each fallback model can define its own:variantreasoningEfforttemperaturetop_pmaxTokensthinking{ model: { providerID, modelID }, variant }chat.paramsgpt-5.4->gpt-5.4-previewgpt-5.4->gpt-5.4oExample config
Why this matters
This is important because fallback quality depends on more than swapping the model name.
If the fallback model does not receive the settings it expects, one of two bad things happens:
The point of a fallback chain is reliability. This PR makes fallback entries model-aware, so each fallback can run with the settings that actually fit that model.
Test plan
bun run typecheckbun test-> 4150 pass, 0 failCloses #2621
Summary by cubic
Adds object-style
fallback_modelswith per-model settings and promotes them only when a real fallback is selected across delegate tasks, subagents, background agents, and runtime-fallback (closes #2621). Prompt payloads now send only provider/model IDs and a top-level variant; per-model tuning is stored in session-scopedchat.params, applied on prompts, restored on resume, and cleared onsession.deleted.New Features
fallback_modelswithvariant,reasoningEffort(addsminimal),temperature,top_p,maxTokens,thinking; mixed arrays allowed; Zod-validated schema.findMostSpecificFallbackEntry; fuzzy model resolution returns the matched entry so per-model settings carry through;flattenToFallbackModelStrings()feeds runtime-fallback.session.deleted.normalizeFallbackModels,flattenToFallbackModelStrings(), session prompt params store, and sharedKNOWN_VARIANTS.Bug Fixes
minimal.Written for commit fb08553. Summary will update on new commits.