Skip to content

Manual: answer “what is a skill” directly and pin the four capability queries #1937

Closed

Description

@santoshkumarradha

Follow-up to #1378 after checking dev at 3ba1d8f76. The original capability denial is addressed by the skills/manual work in #1410, but its exact retrieval acceptance is not fully covered.

Observed behavior

The current binary's manual command returns skills content for all four original questions. However, codeaf manual "what is a skill" begins with Which folders are read, followed by attachment and picture-exclusion sections, rather than a direct definition. The four verbatim questions are also absent from the retrieval regression table in internal/manual/chat_test.go.

The manual now correctly explains that skills work with automatic memory saving off; the old expectation that memory off disables skills must not be reintroduced.

Remaining work

  1. Make "what is a skill" retrieve a direct definition and how a person can use one.
  2. Add explicit retrieval probes for can you use skills, do you have skills, what is a skill, and why can't you use my skills.
  3. Assert that each answer reaches the appropriate capability/definition/troubleshooting section, including the current memory-independent behavior.
  4. Prove the probes catch a removed or unreachable answer section, and verify the same questions through the built binary's manual command.

This issue preserves the remaining retrieval and regression work from #1378; it does not report skills as unavailable.

Activity

  1. santoshkumarradha commented on Oct 25, 2026

    @santoshkumarradha
    MemberAuthor

    Fixed by #1938, merged into dev as c4bd2576a.

    The four original questions now have explicit retrieval regressions. “What is a skill” leads with a direct definition and how to use one; capability and troubleshooting answers explain that automatic memory saving being off does not disable skills.

    Verification:

    • Each answer section was removed separately and the regression caught every removal; pages were restored byte-for-byte.
    • The built binary's manual command passed all four questions, and make test-quick plus make pr-ready passed.
    • All eight real hosted-terminal cases passed in 88 seconds: four questions with memory saving on and off. Actual manual tool results were checked, and all 26 completed model calls returned HTTP 200 on exact deepseek/deepseek-v4.1-flash.
    • All GitHub PR checks passed. Standards and specification reviews found no blocking issues.

    The terminal test explicitly requests the manual lookup; it does not claim unprompted tool selection or arbitrary paraphrase quality. This completes the remaining retrieval work from #1378. Closing as fixed.

  2. added a commit that references this issue on Oct 25, 2026
    85b6049
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions