Repository navigation
Conversation
|
The following comment was made by an LLM, it may be inaccurate: |
![opencode-agent[bot]](https://avatars.githubusercontent.com/in/1549082?s=60&v=4){kind=small}
There was a problem hiding this comment.
Requesting changes. Agent headers and generation keys such as temperature now reach model requests, but agent body works differently from the model/variant/provider body it is documented alongside.
How I checked: on the base, nothing passes the agent's request into SessionModelRequest (from reading the code). With this PR I ran a throwaway test that sent agent body { temperature: 0.1, top_p: 0.9, service_tier: "flex", maxTokens: 5 } through requests.primary and encoded it with OpenAIChat.fromRequest. temperature and maxTokens reached the request; top_p and service_tier were silently dropped, and maxTokens: 5 replaced the computed output limit. The three test files you listed pass (18 tests), and bun run typecheck in packages/core passes.
Public API: unchanged in code. The config key agents.*.request.body/headers already existed in the schema and now takes effect. Existing configs whose agent body uses names the provider expects (e.g. top_p) will still be ignored, without any warning.
Main concern: model, variant and provider body are raw JSON fields merged into the request the provider receives (the http.body overlay, see models.mdx/providers.mdx). This PR puts agent body into Session options instead, so only keys with the internal option names (temperature, topP, maxTokens, …) are sent. Keys like top_p, service_tier or reasoning_effort go to providerOptions and are dropped. The same block that works on a model does nothing on an agent. The new docs text also says agent values override model and variant defaults; because model body is merged in last, a model's temperature would most likely still win (I didn't test this combination). Passing agent body through http.body, the way model body is passed, would give the same rules and the intended precedence. Details are inline.
| kind === "primary" || kind === "compaction" | ||
| ? { maxTokens: outputLimit(model.limit, kind, input.inputTokens) } | ||
| : {}, | ||
| ? { maxTokens: outputLimit(model.limit, kind, input.inputTokens), ...input.agentRequest?.body } |
There was a problem hiding this comment.
This makes agent body a set of Session options rather than a JSON body overlay like model/variant/provider body. In a quick check, { top_p: 0.9, service_tier: "flex" } ended up in providerOptions and was missing from the encoded OpenAI Chat request. Only keys with the internal names (temperature, topP) were sent. Also, spreading the agent body after maxTokens lets an agent maxTokens replace the computed output limit for primary/compaction. Could agent body go through http.body instead (merged like the model's body), so the same config works on agents and models?
| model: model.model, | ||
| http: { | ||
| headers: { | ||
| ...input.agentRequest?.headers, |
There was a problem hiding this comment.
Headers are spread as an object, so an agent header that differs from a runtime header only in case (e.g. X-Session-Affinity) survives next to the lowercase runtime key. I saw both in the prepared headers. If session identity should never be replaced by agent config, drop the agent headers that match runtime-owned names case-insensitively before merging.
| <Callout type="warning"> | ||
| The V2 session runner preserves these values but does not yet send them with model requests. Configure active request | ||
| settings on the provider, model, or model variant instead. | ||
| Agent request values override model and variant defaults. Session hooks may override the agent body options, and |
There was a problem hiding this comment.
"Agent request values override model and variant defaults" doesn't seem to hold for body. Model/variant body is merged into the request the provider receives after encoding, so a model temperature would most likely win over the agent's. Body keys written the way the provider expects them (top_p, service_tier) are dropped entirely. The docs should match whichever behaviour is kept.
|
Addressed in a7f1fca. Agent body now uses http.body rather than Session generation options. The regression checks the transport JSON for temperature, top_p and service_tier, nested overlay precedence, and that maxTokens in raw body does not replace the internal output budget. Runtime-owned agent headers are filtered case-insensitively. Documentation now describes raw overlay semantics and explicitly notes request.settings remains unsupported. The 18 focused tests and Core typecheck pass with Bun 1.4.2; no live provider call or full local check was run. |
|
Automated PR Cleanup Thank you for contributing to opencode. Due to the high volume of PRs from users and AI agents, we periodically close older PRs using automated criteria so maintainers can focus review time on the most active and community-supported contributions. This PR was closed because it matched the following cleanup criteria:
PRs created within the last month are not affected by this cleanup. If you believe this PR was closed incorrectly, or if you are still actively working on it, please leave a comment explaining why it should be reopened. A maintainer can review and reopen it if appropriate. Thanks again for taking the time to contribute. |
Issue for this PR
Related to #49550. This PR addresses body and header overlays only; agent-level
request.settingsremains unimplemented.#55009 reports the same omission and has been identified as a duplicate of #49550.
Type of change
What does this PR do?
Agent config populated
request.bodyandrequest.headers, but Session request preparation received only the agent ID. The configured values never reached model requests.Pass the selected agent's request overlays into primary, compaction, generate, and title request preparation. Body is passed through
http.body, matching provider/model/variant raw JSON overlays. Provider field names such astop_pandservice_tierare preserved, and agent body no longer changes the internal computed output budget. Headers matching runtime-owned names are filtered case-insensitively before merging;model.requesthooks can still override headers.Update the Agents documentation to describe the implemented behavior. Agent request settings such as transport and timeouts remain outside this change.
How did you verify your code works?
Using Bun 1.4.2, from
packages/core:bun test test/session-model-request-hooks.test.ts test/session-model-request.test.ts test/session-native-compaction.test.ts: 18 tests passed.bun typecheck: passed.temperature,top_p,service_tier), nested body merging, preserved internal output limits, and case-insensitive runtime header protection.git diff --check: passed.No live provider request was made.
Screenshots / recordings
Not applicable; this changes request construction, not UI.
Checklist