improve(core): require recurrence evidence before extracting skills

[SandyTao520]

Summary

Tighten the skill extraction agent's signal gate so it requires evidence of recurrence and durability before writing a skill. Previously, sessions with similar summaries (e.g., two one-off incidents about login debugging) could trigger skill creation. Now the agent defaults to "no skill" and must prove future reuse before proceeding.

Details

Prompt changes (skill-extraction-agent.ts):

• Add two new gate questions: "Is there strong evidence this will recur?" and "Is this broader than a single incident?"
• Default to NO SKILL — the agent must justify creation, not justify skipping.
• Explicitly reject one-off incidents, single-session preferences, and output-style preferences.
• Raise the bar on confidence tiers: medium confidence now means "usually do NOT create."
• Warn the agent that session-summary similarity alone is never sufficient evidence.
• Require naming the future trigger, evidence sessions, recurrence signal, and validation signal before writing.

Unit test (skill-extraction-agent.test.ts):

• Verifies new prompt guardrails are present (default-to-no-skill, incident rejection, summary-vs-evidence warnings).

Behavioral eval (skill_extraction.eval.ts):

• ignores one-off incidents: seeds two incident-response sessions with similar summaries and asserts zero skills are created.
• extracts a repeated project-specific workflow: seeds two sessions describing the same settings-doc regeneration workflow and asserts the extracted skill contains the correct commands and sections.

Related Issues

Closes #18007

How to Validate

1. Run the unit test:

   npm test -w @google/gemini-cli-core -- src/agents/skill-extraction-agent.test.ts

2. Run the behavioral evals (requires API key):

   npm run test -w evals -- skill_extraction.eval.ts

3. Review the prompt diff to confirm the new gate questions and rejection categories are present.

Pre-Merge Checklist

• Updated relevant documentation and README (if needed)
• Noted breaking changes (if any)
• Validated on required platforms/methods:
  • MacOS
    • npm run
    • npx
    • Docker
    • Podman
    • Seatbelt
  • Windows
    • npm run
    • npx
    • Docker
  • Linux
    • npm run
    • npx
    • Docker

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request improves the precision of the skill extraction agent by raising the bar for what constitutes a reusable skill. By requiring evidence of recurrence and durability, the agent is now less likely to create skills from one-off incidents or user-specific preferences, ensuring that only high-value, repeatable workflows are captured.

Highlights

• Stricter Skill Extraction Criteria: Updated the skill extraction agent's prompt to require explicit evidence of recurrence and durability, defaulting to 'no skill' unless justified.
• New Guardrails: Added specific rejection categories for one-off incidents, single-session preferences, and output-style preferences to prevent unnecessary skill creation.
• Enhanced Validation: Introduced new behavioral evaluations to verify that the agent correctly ignores one-off incidents while successfully extracting repeated project-specific workflows.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[github-actions]

Size Change: +2.42 kB (+0.01%)

Total Size: 33.6 MB

Filename	Size	Change
./bundle/chunk-3C5HMBV5.js	0 B	-3.8 kB (removed)	🏆
./bundle/chunk-GWNCRQFK.js	0 B	-3.42 MB (removed)	🏆
./bundle/chunk-RYLKYKIA.js	0 B	-14.5 MB (removed)	🏆
./bundle/core-PQZMUSEF.js	0 B	-46.6 kB (removed)	🏆
./bundle/devtoolsService-FR3DNE3S.js	0 B	-28.4 kB (removed)	🏆
./bundle/gemini-WOWVBYIX.js	0 B	-553 kB (removed)	🏆
./bundle/interactiveCli-6J5254PO.js	0 B	-1.29 MB (removed)	🏆
./bundle/oauth2-provider-46MXKLRP.js	0 B	-9.16 kB (removed)	🏆
./bundle/chunk-72E4X26G.js	14.5 MB	+14.5 MB (new file)	🆕
./bundle/chunk-V4H5YYIL.js	3.8 kB	+3.8 kB (new file)	🆕
./bundle/chunk-WHPGIKXF.js	3.42 MB	+3.42 MB (new file)	🆕
./bundle/core-TQMVEXLQ.js	46.6 kB	+46.6 kB (new file)	🆕
./bundle/devtoolsService-HO4JYN7I.js	28.4 kB	+28.4 kB (new file)	🆕
./bundle/gemini-2646R5JZ.js	553 kB	+553 kB (new file)	🆕
./bundle/interactiveCli-UEYEGSVA.js	1.29 MB	+1.29 MB (new file)	🆕
./bundle/oauth2-provider-6BBV3A7J.js	9.16 kB	+9.16 kB (new file)	🆕

ℹ️ View Unchanged

Filename	Size	Change
./bundle/bundled/third_party/index.js	8 MB	0 B
./bundle/chunk-34MYV7JD.js	2.45 kB	0 B
./bundle/chunk-5AUYMPVF.js	858 B	0 B
./bundle/chunk-5PS3AYFU.js	1.18 kB	0 B
./bundle/chunk-664ZODQF.js	124 kB	0 B
./bundle/chunk-DAHVX5MI.js	206 kB	0 B
./bundle/chunk-IUUIT4SU.js	56.5 kB	0 B
./bundle/chunk-PC3Y4R7E.js	1.97 MB	0 B
./bundle/chunk-RJTRUG2J.js	39.8 kB	0 B
./bundle/cleanup-MF3K2DAE.js	0 B	-932 B (removed)	🏆
./bundle/devtools-36NN55EP.js	696 kB	0 B
./bundle/dist-T73EYRDX.js	356 B	0 B
./bundle/events-XB7DADIJ.js	418 B	0 B
./bundle/gemini.js	4.97 kB	0 B
./bundle/getMachineId-bsd-TXG52NKR.js	1.55 kB	0 B
./bundle/getMachineId-darwin-7OE4DDZ6.js	1.55 kB	0 B
./bundle/getMachineId-linux-SHIFKOOX.js	1.34 kB	0 B
./bundle/getMachineId-unsupported-5U5DOEYY.js	1.06 kB	0 B
./bundle/getMachineId-win-6KLLGOI4.js	1.72 kB	0 B
./bundle/memoryDiscovery-R4AKKXHF.js	980 B	0 B
./bundle/multipart-parser-KPBZEGQU.js	11.7 kB	0 B
./bundle/node_modules/@google/gemini-cli-devtools/dist/client/main.js	222 kB	0 B
./bundle/node_modules/@google/gemini-cli-devtools/dist/src/_client-assets.js	229 kB	0 B
./bundle/node_modules/@google/gemini-cli-devtools/dist/src/index.js	13.4 kB	0 B
./bundle/node_modules/@google/gemini-cli-devtools/dist/src/types.js	132 B	0 B
./bundle/sandbox-macos-permissive-open.sb	890 B	0 B
./bundle/sandbox-macos-permissive-proxied.sb	1.31 kB	0 B
./bundle/sandbox-macos-restrictive-open.sb	3.36 kB	0 B
./bundle/sandbox-macos-restrictive-proxied.sb	3.56 kB	0 B
./bundle/sandbox-macos-strict-open.sb	4.82 kB	0 B
./bundle/sandbox-macos-strict-proxied.sb	5.02 kB	0 B
./bundle/src-QVCVGIUX.js	47 kB	0 B
./bundle/tree-sitter-7U6MW5PS.js	274 kB	0 B
./bundle/tree-sitter-bash-34ZGLXVX.js	1.84 MB	0 B
./bundle/cleanup-NMOP5TMJ.js	932 B	+932 B (new file)	🆕

_{compressed-size-action}

[gemini-code-assist]

Code Review

This pull request introduces a new evaluation suite and unit tests for the skill extraction system, while refining the agent's system prompt to prioritize durable, recurring workflows over one-off incidents. Feedback focuses on aligning the test code with repository style guides, specifically regarding the proper management of environment variables using vi.stubEnv and avoiding global state modifications like process.chdir.

[gemini-code-assist]

Modifying the process working directory with process.chdir() is a risky practice in tests as it affects the global state of the process and can cause flakiness in concurrent test runs. Since loadSettings and loadCliConfig both accept a projectRoot or cwd parameter, this global change is redundant and should be avoided.

    let config: Awaited<ReturnType<typeof loadCliConfig>> | undefined;

    try {
      resetSettingsCacheForTesting();

[github-actions]

✅ 59 tests passed successfully on gemini-3-flash-preview.

---

This is an automated guidance message triggered by steering logic signatures.

[SandyTao520]

Re: process.chdir feedback — Intentionally keeping process.chdir here. startMemoryService and its transitive callchain (e.g. skill extraction, memory consolidation) rely on process.cwd() internally, so removing the chdir would break the eval. The current code restores the original cwd in a finally block, which is safe.

The other two comments (use vi.stubEnv instead of direct process.env mutation) are addressed in 42a4dcb.

[gundermanc]

This looks pretty good! Main suggestions are that we:

• Use componentEval() test type instead so we can eliminate a lot of the E2E related boilerplate and indirection.
• Consider following up with more cases over time.
• Ensure we have telemetry and/or logging that can help us detect and produce more evals for cases that don't work well.

[gundermanc]

At some point I think it'd help to be able to furnish the agent with quantitative data, like a sqlite DB or index of past sessions vs. making it count.

[gundermanc]

💡 General thought on the system prompt -- this seems very long and very mechanical. Is it LLM generated? I find the LLM generates very structured, sometimes redundant prompts, though it's pretty good at reflecting on pre-written prompts.

I think you might be able to accomplish the same outcome with 1/4th the text and make it a bit easier to reason about. Ideally we have enough evals that we can sort of experiment and see which lines are required.

I'd suggest in a separate PR trying to ask the model to condense this, then tweaking the follow up by hand. If we have enough evals it should be possible to test how the behavior changes.

[gundermanc]

Approved with suggestion. There seems to be a compile error.