Skip to content

feat(search): abstain names the exact call that continues (option C) - #109

Merged
ManSio merged 1 commit into
mainfrom
feat/retry-hint
Nov 10, 2026
Merged

ManSio merged 1 commit into
mainfrom
feat/retry-hint

Conversation

@ManSio

@ManSio ManSio commented Nov 10, 2026

Copy link
Copy Markdown
Owner

feat(search): abstain now names the exact call that continues (option C)

The owner chose C: never present an unverified row as an answer, but stop
leaving the continuation in prose that a single-shot caller does not act on.

On a withheld answer the tool now ends with:

Neighbours withheld: 8 (unverified).
next: investigate(query="...", include_neighbours=true, limit=8)

and ask_rewrite gets the same shape with hints=["<identifier>"]. The verdict
stays abstain and no row is added; the JSON-quoted query round-trips verbatim,
so a host can call it without re-parsing prose.

Measured on the B1 dev split (71 solvable tasks):

hint emitted on abstains   36/36
default hit@5              19/71   (E68 baseline: 19 - no rows leaked)
recovered by following it  18/36
default + hint total       37/71   [0.41, 0.63]   (neighbours arm: 38)

The first run of that measurement compared "18 recovered" against a panel-level
threshold and failed: a metric bug, not a product bug. The threshold was not
moved; the metric was fixed and an assertion now enforces that every hinted task
has no default hit, so the sum cannot double count.

The owner chose C: never present an unverified row as an answer, but stop
leaving the continuation in prose that a single-shot caller does not act on.

On a withheld answer the tool now ends with:

    Neighbours withheld: 8 (unverified).
    next: investigate(query="...", include_neighbours=true, limit=8)

and `ask_rewrite` gets the same shape with `hints=["<identifier>"]`. The verdict
stays `abstain` and no row is added; the JSON-quoted query round-trips verbatim,
so a host can call it without re-parsing prose.

Measured on the B1 dev split (71 solvable tasks):

    hint emitted on abstains   36/36
    default hit@5              19/71   (E68 baseline: 19 - no rows leaked)
    recovered by following it  18/36
    default + hint total       37/71   [0.41, 0.63]   (neighbours arm: 38)

The first run of that measurement compared "18 recovered" against a panel-level
threshold and failed: a metric bug, not a product bug. The threshold was not
moved; the metric was fixed and an assertion now enforces that every hinted task
has no default hit, so the sum cannot double count.
@coderabbitai

coderabbitai Bot commented Nov 10, 2026

Copy link
Copy Markdown

Important

  • Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: Organization UI
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: cef0e7d0-ac32-4698-b025-e55269dd23ec
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

@ManSio
ManSio merged commit 79c9297 into main Nov 10, 2026
13 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant