feat(coding-agent/prompts): added a model-family prompt overlay#2669
feat(coding-agent/prompts): added a model-family prompt overlay#2669metaphorics wants to merge 1 commit into
Conversation
- Added a pure prompts/model-overlay module: resolveModelOverlay(model, mode) detects the model family (GPT-5, Claude Opus, Kimi K2) and returns a short family-specific tuning section, with omp tool names; off disables and a named mode forces a family. - Wired it into buildSystemPrompt (appends the section to block 0 when a family matches and the mode is not off) and threaded the prompt.modelOverlay setting through the sdk system-prompt rebuild. - Added the prompt.modelOverlay enum setting (auto/off/gpt-5/claude-opus/kimi-k2, default auto) plus an integration test that the overlay lands in block 0 for a matching model, is omitted when off or unmatched, and can be forced.
There was a problem hiding this comment.
Pull request overview
Note
Copilot was unable to run its full agentic suite in this review.
Adds a configurable “model-family prompt overlay” feature that appends a short family-specific tuning section to the generated system prompt.
Changes:
- Introduces model-family detection + overlay section rendering (GPT‑5 / Claude Opus / Kimi K2).
- Wires overlay resolution into
buildSystemPromptand the agent session settings plumbing. - Adds unit + integration tests, docs, and changelog entry for the new
prompt.modelOverlaysetting.
Reviewed changes
Copilot reviewed 11 out of 11 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| packages/coding-agent/src/system-prompt.ts | Appends resolved model overlay section to system prompt block 0 |
| packages/coding-agent/src/system-prompt-model-overlay.test.ts | Integration test ensuring buildSystemPrompt wiring appends/omits overlay correctly |
| packages/coding-agent/src/sdk.ts | Passes prompt.modelOverlay setting into buildSystemPrompt |
| packages/coding-agent/src/prompts/model-overlay/sections.ts | Defines overlay section text per model family |
| packages/coding-agent/src/prompts/model-overlay/model-overlay.ts | Implements ModelOverlayMode and resolveModelOverlay() |
| packages/coding-agent/src/prompts/model-overlay/model-overlay.test.ts | Unit tests for overlay detection + forcing behavior |
| packages/coding-agent/src/prompts/model-overlay/index.ts | Re-exports overlay API |
| packages/coding-agent/src/prompts/model-overlay/detect.ts | Adds model string → family detection |
| packages/coding-agent/src/config/settings-schema.ts | Adds prompt.modelOverlay enum setting with UI metadata |
| packages/coding-agent/CHANGELOG.md | Documents the new overlay feature |
| docs/settings.md | Mentions prompt.* group including the overlay setting |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| export function getModelOverlaySection(family: ModelOverlayFamily): string { | ||
| if (family === "gpt-5") { | ||
| return GPT5_OVERLAY_SECTION; | ||
| } | ||
| if (family === "claude-opus") { | ||
| return CLAUDE_OPUS_OVERLAY_SECTION; | ||
| } | ||
|
|
||
| return KIMI_K2_OVERLAY_SECTION; | ||
| } |
| workspaceTree: { | ||
| rootPath: process.cwd(), |
| const rendered = prompt.render(resolvedCustomPrompt ? customSystemPromptTemplate : systemPromptTemplate, data); | ||
| const systemPrompt = [rendered]; | ||
| const overlaySection = resolveModelOverlay(model, modelOverlay); | ||
| const systemPrompt = [overlaySection ? `${rendered}\n\n${overlaySection}` : rendered]; |
![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: c98cc6c93a
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".
| model: settings.get("includeModelInPrompt") ? getActiveModelString() : undefined, | ||
| personality: agentKind === "sub" ? "none" : settings.get("personality"), | ||
| modelOverlay: settings.get("prompt.modelOverlay"), |
There was a problem hiding this comment.
When includeModelInPrompt is false, the model argument passed to buildSystemPromptInternal is undefined, so the newly added prompt.modelOverlay: auto path has no active model to inspect and silently omits the default overlay. This affects users who hide the literal model name but leave the overlay at its default/auto setting; force modes still work, but auto does not match the documented setting. Pass a separate active model value for overlay detection (and rebuild on model switches) instead of reusing the rendered model field.
Useful? React with 👍 / 👎.
| @@ -0,0 +1,31 @@ | |||
| import type { ModelOverlayFamily } from "./detect"; | |||
|
|
|||
| export const GPT5_OVERLAY_SECTION = `## Model overlay: GPT-5 | |||
There was a problem hiding this comment.
These constants are provider-facing prompt text, but the repo-level AGENTS.md requires prompts to live in static .md files and be imported with with { type: "text" }, not built as inline strings/template literals. Keeping the overlay copy in TS violates that prompt-file contract and bypasses the same prompt asset workflow used by the rest of prompts/system, so please move each section to .md and import it.
Useful? React with 👍 / 👎.
| - Preserve the full requested scope: words like every, all, and for each apply to the complete set, not the first matching item. | ||
| - Follow ordered instructions in order; if asked to do X then Y, do not invert or merge the sequence. | ||
| - Maintain precise state across long tool workflows without drifting from the original goal. | ||
| - Use task for parallelizable multi-file work, and keep edits narrow and grounded in observed files.`; |
There was a problem hiding this comment.
This overlay hard-codes task (and sibling bullets hard-code other tool names) without checking the active tool set or wire-visible names. In contexts such as subagents at task.maxRecursionDepth where task is not registered, read-only sessions with edit/bash disabled, or modes where a tool has a custom wire name, the system prompt now tells the model to call tools it cannot use; the existing Handlebars prompt avoids this by gating tool advice on availability and rendering through toolRefs.
Useful? React with 👍 / 👎.
Summary
Adds a pure model-family overlay module (
resolveModelOverlay) wired intobuildSystemPrompt: a short family-specific tuning section (GPT-5, Claude Opus, or Kimi K2) is appended to the system prompt based on the active model, controlled by theprompt.modelOverlaysetting (auto/off/forced-family, default auto).Tests
Family-detection module test + a buildSystemPrompt integration test (overlay lands in block 0 for a matching model, omitted when off/unmatched, forceable);
bun checkgreen.Note: touches
config/settings-schema.tsanddocs/settings.mdalongside #2655 and #2660; rebase if flagged.Closes #2661