Reply honesty: five disclosure defects found dogfooding v0.38.0 (wrong unmeasured-population prose, concise rows drop resolved_by, startup ghost labels, "nearest" wording, cursor/did_you_mean diagnoses) #379

Closed
opened 2026-10-15 07:59:05 +02:00 by buildagent · 1 comment
Member

Found dogfooding v0.38.0 (4ab97367) on this repo. All five are in reply text or fields; the index itself was correct.

  1. The UNMEASURED POPULATION prose ignores the row's reason.

    • compose_evidence_semantics (mcp-server/src/server.rs ~31840) emits one fixed paragraph, written for no_use_reference_channel. It says "no reference kind captures a USE … A 0 in this reply is STRUCTURAL".
    • find_callers(all_plugins) ships that paragraph for a row whose reason is uses_may_be_generated, in a reply with total_resolved: 262.
    • search_symbols explains the same symbol correctly.
    • Fix: the paragraph must follow each row's reason.
  2. Concise find_references/find_callers rows drop resolved_by and influence. The tool descriptions say concise keeps file/line/kind/name/target_id/resolution/resolved_by/influence. Make the rows and the description agree.

  3. Deleted files are mislabelled during the startup reconcile window. 82 files deleted while no daemon ran were listed until the first reconcile removed them (about 70 s):

    • project_overview counted them as permanently_unclaimed (symbol_blind_extensions rust / permanently_unclaimed, 82).
    • list_files marked each one disk_unmeasured: not_readable.
    • A file that is not on disk is neither permanently unclaimed nor unreadable. It is gone and pending removal.
    • Disclose it as what it is, and find out why those rows read as unclaimed.
  4. The lang hint's "nearest" wording. search_text(lang:"py") says "the 12 nearest of the 20 spellings": python (an extension match) and then 11 alphabetical ones. Only the near ones may be called nearest.

  5. Two misdiagnoses.

    • search_symbols(query:"all_plugins", lang:"rs") ships did_you_mean: ["ALL","all","ALLOC","alloc"] (other symbol names), although its own empty_population shows the lang filter alone emptied the result.
    • A malformed cursor ("2@x") is refused as "belongs to a DIFFERENT index snapshot". It is malformed, not stale.
Found dogfooding v0.38.0 (4ab97367) on this repo. All five are in reply text or fields; the index itself was correct. 1. **The UNMEASURED POPULATION prose ignores the row's reason.** - `compose_evidence_semantics` (`mcp-server/src/server.rs` ~31840) emits one fixed paragraph, written for `no_use_reference_channel`. It says "no reference kind captures a USE … A `0` in this reply is STRUCTURAL". - `find_callers(all_plugins)` ships that paragraph for a row whose reason is `uses_may_be_generated`, in a reply with `total_resolved: 262`. - `search_symbols` explains the same symbol correctly. - Fix: the paragraph must follow each row's `reason`. 2. **Concise `find_references`/`find_callers` rows drop `resolved_by` and `influence`.** The tool descriptions say concise keeps `file/line/kind/name/target_id/resolution/resolved_by/influence`. Make the rows and the description agree. 3. **Deleted files are mislabelled during the startup reconcile window.** 82 files deleted while no daemon ran were listed until the first reconcile removed them (about 70 s): - `project_overview` counted them as `permanently_unclaimed` (`symbol_blind_extensions` `rust` / `permanently_unclaimed`, 82). - `list_files` marked each one `disk_unmeasured: not_readable`. - A file that is not on disk is neither permanently unclaimed nor unreadable. It is gone and pending removal. - Disclose it as what it is, and find out why those rows read as unclaimed. 4. **The `lang` hint's "nearest" wording.** `search_text(lang:"py")` says "the 12 nearest of the 20 spellings": `python` (an extension match) and then 11 alphabetical ones. Only the near ones may be called nearest. 5. **Two misdiagnoses.** - `search_symbols(query:"all_plugins", lang:"rs")` ships `did_you_mean: ["ALL","all","ALLOC","alloc"]` (other symbol names), although its own `empty_population` shows the `lang` filter alone emptied the result. - A malformed cursor (`"2@x"`) is refused as "belongs to a DIFFERENT index snapshot". It is malformed, not stale.
Author
Member

Released in v0.39.0 (PR #386, 25a45918). Every item:

  • Unmeasured explanation: it follows each symbol's own reason, so a symbol that has an attribute is no longer described as having "no reference channel".
  • list_files: reports not_on_disk for a file that is gone. A real permission error stays not_readable.
  • Deleted while no daemon ran: files deleted in that window now leave the index before the first rebuild after an upgrade.
  • lang hint: says "nearest" only for an extension, a prefix, or a typo within 1–2 edits.
  • Cursors: a cursor is refused as MALFORMED, FOREIGN or STALE, each with its own hint, for every tool.
  • did_you_mean: no longer offered when a filter, not the query, emptied the result.
  • Concise ref rows: described exactly as they ship.

Also in this release, for the indexer:

  • Deciding whether an unwalked path is gone lists each directory once per pass (20,000 ignored files took 91 s before).
  • An existing file that cannot be inspected is withdrawn, not deleted.

Two small leftovers from the dogfood of this release are in the next train: search_symbols lang:"rsut" does not say "nearest" in its "all N fit" branch, and a project-pinned reply reports the primary project's index_state.

Released in [v0.39.0](https://git.h-dv.de/h-dv/code-index/releases/tag/v0.39.0) (PR #386, 25a45918). Every item: - **Unmeasured explanation:** it follows each symbol's own reason, so a symbol that has an attribute is no longer described as having "no reference channel". - **`list_files`:** reports `not_on_disk` for a file that is gone. A real permission error stays `not_readable`. - **Deleted while no daemon ran:** files deleted in that window now leave the index before the first rebuild after an upgrade. - **`lang` hint:** says "nearest" only for an extension, a prefix, or a typo within 1–2 edits. - **Cursors:** a cursor is refused as MALFORMED, FOREIGN or STALE, each with its own hint, for every tool. - **`did_you_mean`:** no longer offered when a filter, not the query, emptied the result. - **Concise ref rows:** described exactly as they ship. Also in this release, for the indexer: - Deciding whether an unwalked path is gone lists each directory once per pass (20,000 ignored files took 91 s before). - An existing file that cannot be inspected is withdrawn, not deleted. Two small leftovers from the dogfood of this release are in the next train: `search_symbols` `lang:"rsut"` does not say "nearest" in its "all N fit" branch, and a project-pinned reply reports the primary project's `index_state`.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
h-dv/code-index#379
No description provided.