fix(core): tune provider generation and drift coverage#170
Open
fix(core): tune provider generation and drift coverage#170
Conversation
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
|
Warning Rate limit exceeded
To continue reviewing without waiting, purchase usage credits in the billing tab. ⌛ How to resolve this issue?After the wait time has elapsed, a review can be triggered using the 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 configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (10)
📝 WalkthroughWalkthroughRefactors provider callsites to use intent-based GenerationProfile objects, implements profile-to-provider translation in Anthropic/OpenAI/Gemini providers, updates default model identifiers (OpenAI → gpt-5 / gpt-5-mini; Anthropic → claude-haiku-4-5 / claude-sonnet-4-6), and updates schemas, tests, docs, and small config/tooling. ChangesGeneration profiles & provider refactor
Model identifier, schema, docs & env
Drift-analysis prompt & evidence formatting
Tooling, config, and small infra changes
Sequence Diagram(s)sequenceDiagram
participant Orch as Analysis Orchestrator
participant Base as BaseProvider
participant Prov as Concrete Provider
participant SDK as LLM SDK
Orch->>Base: executeStage(phase, prompt, optional generationProfile?)
Base->>Base: profile = getGenerationProfileForPhase(phase) (if absent)
Base->>Prov: callModel(model, prompt, phase, generationProfile)
Prov->>Prov: compute token/effort (max tokens / reasoning / thinkingConfig) from profile
Prov->>SDK: send request with computed config
SDK-->>Prov: response (complete or incomplete/truncated)
Prov->>Base: return normalized text or throw ErodeError (truncation/safety)
Base->>Orch: return result / propagate error
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
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. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@packages/web/public/schemas/v0/eroderc.schema.json`:
- Around line 212-218: The schema's default model IDs for the "fastModel" and
"advancedModel" properties are invalid; update the defaults in the
eroderc.schema.json to use real OpenAI model identifiers (e.g., replace
"gpt-5.3-mini" with a valid mini model like "gpt-5.4-mini" or "gpt-5-mini/nano"
and replace "gpt-5.3" with a valid full model like "gpt-5.4" or "gpt-5.5") so
API calls won't fail—ensure you only use model names listed in the OpenAI models
docs and update both "fastModel" and "advancedModel" default values accordingly.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: a9556923-6e87-4f4e-842f-dafe0044a3d6
📒 Files selected for processing (9)
.env.examplepackages/core/schemas/eroderc.schema.jsonpackages/core/src/providers/__tests__/provider-factory.test.tspackages/core/src/providers/anthropic/__tests__/provider.test.tspackages/core/src/providers/anthropic/models.tspackages/core/src/providers/openai/__tests__/provider.test.tspackages/core/src/providers/openai/models.tspackages/web/public/schemas/v0/eroderc.schema.jsonpackages/web/src/content/docs/docs/reference/ai-providers.md
parse
changed the title
Move output sizing intent into shared generation profiles. Keep provider-specific token parameters inside each AI provider. Document the provider boundary change and add web lint coverage to CI.
Disable Gemini thinking for low-reasoning phases. Raise Gemini output budgets and report max-token truncation clearly.
parse
changed the title
Greptile SummaryThis PR refactors how AI providers control generation output by replacing raw per-call
Confidence Score: 5/5Safe to merge; all provider paths are covered by tests and the generation-profile abstraction is well-bounded. The generation-profile refactor is self-contained and well-tested across all three providers. The OpenAI migration to the Responses API is the largest surface area change, but the new handleIncompleteResponse now correctly throws for every known and unknown incomplete reason. The Gemini thinking-config dispatch correctly handles 2.5, 3-dash, and 3-dot point-release model naming via a clean predicate chain verified by the new tests. No token-budget regressions were introduced: CHANGE_ANALYSIS now reaches Anthropic's large bucket (8192 tokens) and MODEL_UPDATE gets a dynamic content hint that promotes the limit well above the base bucket. No files require special attention; the two findings are non-blocking style improvements. Important Files Changed
Sequence DiagramsequenceDiagram
participant BP as BaseProvider
participant GP as generation-profile.ts
participant P as Provider (OpenAI/Gemini/Anthropic)
participant SDK as Provider SDK
BP->>GP: getGenerationProfileForPhase(phase)
GP-->>BP: GenerationProfile {outputSize, reasoningEffort, outputContentHint?}
BP->>P: callModel(model, prompt, phase, generationProfile)
P->>GP: resolveOutputTokenLimit(profile, sizeTable)
GP-->>P: maxOutputTokens (max of bucket limit vs hint/4)
alt OpenAI
P->>SDK: responses.create({input, max_output_tokens, reasoning?})
SDK-->>P: Response {status, output_text, output[]}
P->>P: handleIncompleteResponse() [throws on incomplete]
P->>P: extractText() [output_text or walk output items]
else Gemini
P->>P: getThinkingConfig(model, reasoningEffort)
P->>SDK: generateContent({config: {maxOutputTokens, thinkingConfig}})
SDK-->>P: GenerateContentResponse
else Anthropic
P->>SDK: messages.create({max_tokens, messages})
SDK-->>P: Message {stop_reason, content}
end
P-->>BP: text string
BP->>BP: unwrapModelPatchResponse(text) [strip markdown fences]
Reviews (4): Last reviewed commit: "fix(core): restore drift analysis output..." | Re-trigger Greptile |
There was a problem hiding this comment.
Actionable comments posted: 3
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
packages/core/src/providers/anthropic/provider.ts (1)
69-76:⚠️ Potential issue | 🟡 Minor | ⚡ Quick winAnthropic truncation
userMessageis less actionable than its OpenAI / Gemini counterparts.The current message
'The AI response was truncated. The output may be partial.'doesn't tell the user what "output budget" means or what to do. The OpenAI and Gemini equivalents both say "used the available output budget before completion. Try a smaller change or tune the provider output budget".✏️ Align with OpenAI / Gemini wording
- 'The AI response was truncated. The output may be partial.', + 'The Anthropic response used the available output budget before completion. Try a smaller change or tune the provider output budget.',🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/core/src/providers/anthropic/provider.ts` around lines 69 - 76, The Anthropic truncation error message is less actionable; update the ErodeError thrown when response.stop_reason === 'max_tokens' (in the provider.ts block that constructs ErodeError using model, phase, outputTokenLimit) to use the same actionable wording as OpenAI/Gemini: explain that the model used the available output budget before completion and suggest trying a smaller change or tuning the provider output budget, while keeping the same ErrorCode.PROVIDER_INVALID_RESPONSE and the existing metadata (model, phase, outputTokenLimit).
🧹 Nitpick comments (3)
packages/core/src/providers/openai/provider.ts (2)
52-114: ⚡ Quick winSilent fallthrough for unknown
incompletereasons.When
response.status === 'incomplete'with a reason other thanmax_output_tokensorcontent_filter,handleIncompleteResponsereturns without throwing, and partially-truncated text fromextractTextis treated as a successful response. The OpenAI SDK typesincomplete_details.reasonas an open string, so future or undocumented reasons would be swallowed silently and propagate corrupted data downstream into JSON parsing/validation.Consider a final
throwfor unknown reasons so the caller surfaces a clear truncation error instead of a generic schema validation failure:♻️ Proposed catch-all branch
if (response.incomplete_details?.reason === 'content_filter') { throw new ErodeError( 'OpenAI safety filters blocked the response', ErrorCode.PROVIDER_SAFETY_BLOCK, 'Content was blocked by the AI provider safety filters. Try simplifying the input.', { model: incompleteModel, phase: incompletePhase } ); } + + throw new ErodeError( + 'OpenAI returned an incomplete response', + ErrorCode.PROVIDER_INVALID_RESPONSE, + `The AI response was marked incomplete (reason: ${response.incomplete_details?.reason ?? 'unknown'}).`, + { model: incompleteModel, phase: incompletePhase } + ); }🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/core/src/providers/openai/provider.ts` around lines 52 - 114, handle the silent fallthrough in handleIncompleteResponse by adding a catch-all branch for any incomplete_details.reason not equal to 'max_output_tokens' or 'content_filter' that throws a descriptive ErodeError; locate the function handleIncompleteResponse (used when response.status === 'incomplete') and after the two existing if blocks inspect response.incomplete_details?.reason and throw an error indicating an unknown/unsupported incomplete reason (include model/phase/maxOutputTokens and the raw reason in the error context) so callers never treat truncated responses from extractText as valid.
82-158: 💤 Low valueHoist helpers out of
callModelfor readability and reuse.The five nested function declarations (
handleIncompleteResponse,extractText,supportsReasoningEffort,getOpenAIReasoningEffort,getMaxOutputTokens) sit after thetry/catch/returnthat calls them. They work due to hoisting, but reading the method top-down requires jumping past the early return to find behavior. None of them referencethis, so they can be promoted to module-private functions next toMAX_OUTPUT_TOKENS_BY_OUTPUT_SIZE. That also makes them straightforward to unit-test in isolation (e.g., reasoning-effort mapping, hint-to-token conversion) without instantiating the provider.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/core/src/providers/openai/provider.ts` around lines 82 - 158, These helper functions (handleIncompleteResponse, extractText, supportsReasoningEffort, getOpenAIReasoningEffort, getMaxOutputTokens) are declared nested after the callModel return and should be hoisted to module scope for readability and reuse; move each function out of the callModel block and place them as module-private functions adjacent to MAX_OUTPUT_TOKENS_BY_OUTPUT_SIZE, keeping their signatures and internal logic unchanged, update any local references in callModel to use the relocated functions, and ensure any used types (OpenAI.Responses.Response, AnalysisPhase, GenerationProfile, OpenAIReasoningEffort, ReasoningEffort) are available in the file scope (import or reference) so the functions compile and can be unit-tested independently.packages/core/src/providers/base-provider.ts (1)
77-92: 💤 Low valueTiming logs are inconsistent across stages and lost on error.
patchModel(line 158) wraps the timingdebugLogin.finally(), so duration is reported even when the call throws.selectComponent(line 92) andexecuteStage(line 182) instead place thedebugLogafterawait, meaning they only log on success — which is precisely the case where timing is least interesting for triaging. Aligning all three to log via.finally()(or a smalltime(label, fn)helper) would make the observability behavior uniform and surface latency on the failure path too.♻️ Example for selectComponent
- debugLog('selectComponent using model', this.fastModel); - const startedAt = process.hrtime.bigint(); - const responseText = await withRetry( - () => - this.callModel( - this.fastModel, - prompt, - AnalysisPhase.COMPONENT_RESOLUTION, - getGenerationProfileForPhase(AnalysisPhase.COMPONENT_RESOLUTION) - ), - { - retries: 2, - shouldRetry: (error) => this.isRetryableError(error), - } - ); - debugLog('selectComponent completed in', formatDuration(startedAt)); + debugLog('selectComponent using model', this.fastModel); + const startedAt = process.hrtime.bigint(); + const responseText = await withRetry( + () => + this.callModel( + this.fastModel, + prompt, + AnalysisPhase.COMPONENT_RESOLUTION, + getGenerationProfileForPhase(AnalysisPhase.COMPONENT_RESOLUTION) + ), + { + retries: 2, + shouldRetry: (error) => this.isRetryableError(error), + } + ).finally(() => { + debugLog('selectComponent completed in', formatDuration(startedAt)); + });Also applies to: 174-182
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/core/src/providers/base-provider.ts` around lines 77 - 92, selectComponent and executeStage only log duration after the await, so failures don't report timing; patchModel uses .finally() correctly. Modify selectComponent (the block that calls withRetry/this.callModel and uses startedAt = process.hrtime.bigint()) and executeStage (the analogous call) to attach the debugLog('... completed in', formatDuration(startedAt)) in a .finally() on the promise returned by withRetry (or wrap the call in a small time(label, fn) helper) so the duration is logged on both success and error—keep existing symbols: withRetry, this.callModel, AnalysisPhase.*, debugLog, formatDuration, startedAt.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@packages/core/schemas/eroderc.schema.json`:
- Around line 167-172: The schema currently uses floating Anthropic aliases;
update the defaults to pinned snapshot identifiers to ensure stable behavior:
replace the default value "claude-haiku-4-5" with the pinned snapshot
"claude-haiku-4-5-20251001" (for the fast/haiku model entry that corresponds to
the ERODE_*_FAST_MODEL env var) and replace "claude-sonnet-4-6" in the
"advancedModel" property with the pinned snapshot "claude-sonnet-4-6-20260218";
edit the default strings in the properties whose default values currently show
those alias names (e.g., the property with default "claude-haiku-4-5" and the
"advancedModel" property) so new installs use the fixed snapshot versions.
In `@packages/core/src/analysis/section-formatters.ts`:
- Around line 60-63: The evidence string in the map callback (where evidence is
built from dep.code.trim()) can contain newlines and currently only the first
line is indented, breaking bullet structure; change the construction of evidence
in that map (the anonymous function mapping dep) to normalize multiline evidence
by splitting dep.code.trim() on newlines and rejoining with a consistent
per-line prefix (e.g., two spaces plus the same "Evidence:" header on the first
line or an indented continuation for subsequent lines) so every evidence line
remains indented under the dependency bullet (ensure you update the variable
used to return the final string in that map).
In `@packages/core/src/providers/openai/provider.ts`:
- Around line 129-133: supportsReasoningEffort currently treats any model
starting with prefixes like "gpt-5-" as reasoning-capable, which incorrectly
includes chat-tuned variants (e.g., "gpt-5-chat-latest"); update the function
(supportsReasoningEffort) to also reject models that contain the 'chat'
identifier (for example check model.includes('chat') or a '-chat' segment) so it
only returns true when the model equals a prefix or starts with `${prefix}-` and
does not identify as a chat variant.
---
Outside diff comments:
In `@packages/core/src/providers/anthropic/provider.ts`:
- Around line 69-76: The Anthropic truncation error message is less actionable;
update the ErodeError thrown when response.stop_reason === 'max_tokens' (in the
provider.ts block that constructs ErodeError using model, phase,
outputTokenLimit) to use the same actionable wording as OpenAI/Gemini: explain
that the model used the available output budget before completion and suggest
trying a smaller change or tuning the provider output budget, while keeping the
same ErrorCode.PROVIDER_INVALID_RESPONSE and the existing metadata (model,
phase, outputTokenLimit).
---
Nitpick comments:
In `@packages/core/src/providers/base-provider.ts`:
- Around line 77-92: selectComponent and executeStage only log duration after
the await, so failures don't report timing; patchModel uses .finally()
correctly. Modify selectComponent (the block that calls withRetry/this.callModel
and uses startedAt = process.hrtime.bigint()) and executeStage (the analogous
call) to attach the debugLog('... completed in', formatDuration(startedAt)) in a
.finally() on the promise returned by withRetry (or wrap the call in a small
time(label, fn) helper) so the duration is logged on both success and error—keep
existing symbols: withRetry, this.callModel, AnalysisPhase.*, debugLog,
formatDuration, startedAt.
In `@packages/core/src/providers/openai/provider.ts`:
- Around line 52-114: handle the silent fallthrough in handleIncompleteResponse
by adding a catch-all branch for any incomplete_details.reason not equal to
'max_output_tokens' or 'content_filter' that throws a descriptive ErodeError;
locate the function handleIncompleteResponse (used when response.status ===
'incomplete') and after the two existing if blocks inspect
response.incomplete_details?.reason and throw an error indicating an
unknown/unsupported incomplete reason (include model/phase/maxOutputTokens and
the raw reason in the error context) so callers never treat truncated responses
from extractText as valid.
- Around line 82-158: These helper functions (handleIncompleteResponse,
extractText, supportsReasoningEffort, getOpenAIReasoningEffort,
getMaxOutputTokens) are declared nested after the callModel return and should be
hoisted to module scope for readability and reuse; move each function out of the
callModel block and place them as module-private functions adjacent to
MAX_OUTPUT_TOKENS_BY_OUTPUT_SIZE, keeping their signatures and internal logic
unchanged, update any local references in callModel to use the relocated
functions, and ensure any used types (OpenAI.Responses.Response, AnalysisPhase,
GenerationProfile, OpenAIReasoningEffort, ReasoningEffort) are available in the
file scope (import or reference) so the functions compile and can be unit-tested
independently.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 147a4c9b-dfa1-4843-83c7-9f299e3b8b0a
⛔ Files ignored due to path filters (1)
.claude-adr.logis excluded by!**/*.log
📒 Files selected for processing (26)
.githooks/post-commitpackages/core/schemas/eroderc.schema.jsonpackages/core/src/analysis/__tests__/prompt-builder.test.tspackages/core/src/analysis/__tests__/section-formatters.test.tspackages/core/src/analysis/prompts/drift-analysis.mdpackages/core/src/analysis/section-formatters.tspackages/core/src/providers/__tests__/generation-profile.test.tspackages/core/src/providers/__tests__/provider-factory.test.tspackages/core/src/providers/anthropic/__tests__/provider.test.tspackages/core/src/providers/anthropic/provider.tspackages/core/src/providers/base-provider.tspackages/core/src/providers/gemini/__tests__/provider.test.tspackages/core/src/providers/gemini/provider.tspackages/core/src/providers/generation-profile.tspackages/core/src/providers/openai/__tests__/provider.test.tspackages/core/src/providers/openai/models.tspackages/core/src/providers/openai/provider.tspackages/eslint-config/base.jspackages/eslint-config/index.d.tspackages/eslint-config/index.jspackages/web/eslint.config.mjspackages/web/package.jsonpackages/web/public/schemas/v0/eroderc.schema.jsonpackages/web/src/content/docs/docs/contributing/adrs/011-intent-based-provider-generation-profiles.mdpackages/web/src/content/docs/docs/contributing/adrs/index.mdpackages/web/src/content/docs/docs/reference/ai-providers.md
✅ Files skipped from review due to trivial changes (4)
- packages/web/src/content/docs/docs/contributing/adrs/index.md
- packages/eslint-config/index.js
- packages/web/src/content/docs/docs/contributing/adrs/011-intent-based-provider-generation-profiles.md
- packages/web/src/content/docs/docs/reference/ai-providers.md
🚧 Files skipped from review as they are similar to previous changes (3)
- packages/core/src/providers/openai/models.ts
- packages/web/public/schemas/v0/eroderc.schema.json
- packages/core/src/providers/tests/provider-factory.test.ts
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (1)
packages/core/src/providers/openai/provider.ts (1)
142-170: 💤 Low valueMap
'low'reasoning effort to OpenAI's'low'instead of silently downgrading to'minimal'.The function maps both
'low'andundefinedto'minimal', but OpenAI's Responses API accepts'low'as a distinct tier between'minimal'and'medium'. Multiple analysis phases explicitly request'low'(DEPENDENCY_SCAN, COMPONENT_RESOLUTION, CHANGE_ANALYSIS), so these requests are being silently downgraded. Either map'low'→'low'to preserve caller intent, or document this intentional downgrade with rationale ingeneration-profile.ts.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/core/src/providers/openai/provider.ts` around lines 142 - 170, The getOpenAIReasoningEffort function currently maps both 'low' and undefined to 'minimal', which silently downgrades callers requesting 'low' (e.g., DEPENDENCY_SCAN, COMPONENT_RESOLUTION, CHANGE_ANALYSIS); update the switch in getOpenAIReasoningEffort (and its ReasoningEffort -> OpenAIReasoningEffort mapping) so case 'low' returns 'low' (preserve undefined -> 'minimal' if desired), ensuring the function returns OpenAI's distinct 'low' tier rather than downgrading to 'minimal'.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@packages/core/src/providers/anthropic/provider.ts`:
- Around line 86-93: Extract the duplicated character-to-token logic into a
shared exported helper (e.g., resolveOutputTokenLimit(profile, sizeTable)) in
packages/core/src/providers/generation-profile.ts, implement it to compute
profileLimit from the provided sizeTable and hintedLimit as
Math.ceil(profile.outputContentHint.characters / 4) and return
Math.max(profileLimit, hintedLimit); then replace the body of
getOutputTokenLimit in packages/core/src/providers/anthropic/provider.ts and
getMaxOutputTokens in packages/core/src/providers/openai/provider.ts to call
resolveOutputTokenLimit(profile, MAX_TOKENS_BY_OUTPUT_SIZE) (or the OpenAI size
table) and move these calls to module scope so the helper isn’t re-declared per
call.
In `@packages/core/src/providers/gemini/provider.ts`:
- Around line 134-146: The switch handling reasoningIntent currently forces any
"pro" model detected by isGemini3ProModel(thinkingModel) to ThinkingLevel.HIGH
for the "medium" case; update the logic so Gemini 3.1+ Pro models can remain
MEDIUM while only legacy gemini-3-pro* IDs escalate to HIGH. Modify either
isGemini3ProModel to distinguish legacy IDs (e.g., match exact "gemini-3-pro" or
"gemini-3-pro-*" but not "gemini-3.1-pro*"/"gemini-3.x-pro*"), or add a new
helper (e.g., isLegacyGemini3ProModel) and use that in the medium branch inside
the switch to decide HIGH vs MEDIUM while preserving ThinkingLevel.MEDIUM for
3.1+ Pro models.
---
Nitpick comments:
In `@packages/core/src/providers/openai/provider.ts`:
- Around line 142-170: The getOpenAIReasoningEffort function currently maps both
'low' and undefined to 'minimal', which silently downgrades callers requesting
'low' (e.g., DEPENDENCY_SCAN, COMPONENT_RESOLUTION, CHANGE_ANALYSIS); update the
switch in getOpenAIReasoningEffort (and its ReasoningEffort ->
OpenAIReasoningEffort mapping) so case 'low' returns 'low' (preserve undefined
-> 'minimal' if desired), ensuring the function returns OpenAI's distinct 'low'
tier rather than downgrading to 'minimal'.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: fa0956b5-1ff8-4c0c-b80c-8000222e0273
📒 Files selected for processing (12)
.env.examplepackages/core/src/analysis/__tests__/prompt-builder.test.tspackages/core/src/analysis/__tests__/section-formatters.test.tspackages/core/src/analysis/section-formatters.tspackages/core/src/providers/__tests__/generation-profile.test.tspackages/core/src/providers/anthropic/__tests__/provider.test.tspackages/core/src/providers/anthropic/provider.tspackages/core/src/providers/gemini/__tests__/provider.test.tspackages/core/src/providers/gemini/provider.tspackages/core/src/providers/openai/__tests__/provider.test.tspackages/core/src/providers/openai/provider.tspackages/web/src/content/docs/docs/contributing/adrs/011-intent-based-provider-generation-profiles.md
✅ Files skipped from review due to trivial changes (1)
- packages/web/src/content/docs/docs/contributing/adrs/011-intent-based-provider-generation-profiles.md
🚧 Files skipped from review as they are similar to previous changes (1)
- packages/core/src/analysis/section-formatters.ts
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Refactors provider generation controls from shared raw token budgets to provider-agnostic generation intent profiles, with provider-local translation for OpenAI, Gemini, and Anthropic. Also updates provider defaults/docs, adds ADR documentation, fixes web lint coverage in CI, improves drift model-update coverage, and keeps the simplified post-commit hook cleanup because the previous ADR automation did not work.
Consumer-facing changes
gpt-4.1-mini/gpt-4.1togpt-5-mini/gpt-5.claude-haiku-4-5/claude-sonnet-4-6.max_output_tokensand reasoning effort instead of Chat Completionsmax_tokens.gemini-2.5-flash/gemini-2.5-pro; generation controls are model-aware so Gemini 2.5 usesthinkingBudgetand Gemini 3-style models usethinkingLevel.ERODE_*_FAST_MODELandERODE_*_ADVANCED_MODELoverrides.Validation
npm run typecheck --workspace=packages/corenpm run lintnpm run testnpm run buildnpm run check:ciSummary by CodeRabbit
New Features
Refactor
Tests
Chores
Style