Skip to content

Three landed spec anchors point at the wrong lines, and one cites a string that exists nowhere #1001

Description

@localai-bot

Found by the agent merging #996, which flagged it and handed it up rather than filing silently or fixing outside its dispatch. I verified all three at origin/main before filing.

The three

.agents/specs/ltx25-retake.md:63   -> ltx2_video.cpp:1566 for an emptiness test
   actual :1566  =  context_tokens = encoded.seq;
   the test is at :1420 / :1681

.agents/specs/ltx25-retake.md:251  -> ltx2_video.cpp:1995 for
   `video.mask.assign(video.tokens, 1.0F)`
   occurrences of that string on main: 0.  It exists in NO tree checked
   (the row head, origin/main, or the merge).

.agents/specs/ltx25-dfr-pipeline.md:307 -> ltx2_video.cpp:1379-1380 for the
   stale-refusal tally
   actual :1379 = a comment fragment; the tally is at :1626

None was caused by the merge. Measured at three trees — the row head, origin/main, and the merged tree — the retake pair was already wrong when retake landed, and the DFR one was stale pre-merge (it sits in §8 Risks, authored before implementation, which is exactly when a line number is least knowable).

A fourth citation, ltx25-dfr-pipeline.md:61docs/USAGE.md:791, is correct and should not be touched: §1 is explicitly headed measured at 0e1bee42f, so it is a dated historical claim, not a live one. That is the pattern the others should follow.

Why this is worth an issue rather than a quiet fix

The second one is the interesting case. A citation whose text exists in no tree cannot be repaired by re-deriving a line number — someone has to decide what the sentence meant. That is a job for whoever owns the row, not for a passing merge.

And the first two sit in a landed spec belonging to a different row. AGENTS.md's fix-in-flow rule covers a defect you find in work you are doing; it does not license one row rewriting another row's records. The merging agent said exactly that and declined — correctly, and it is the reason this issue exists at all.

Disposition

This is a fourth instance of #911 (line-carrying citations no checker examines) and a sibling of #959 (seven files citing latent_cond.py:38-39 for writes at :40-41). The remedy is the one this repo already uses and #911 will need to enforce:

  • SHA-anchor a historical claim (path:NN @ <sha>), as ltx25-dfr-pipeline.md:61 does correctly by dating its whole section;
  • re-derive a live one with the needle taken from the claim, never by reading the span out of the cited file — that check is circular and once reported 27/27 FRESH while five anchors pointed at unrelated code;
  • and cover bare :NN continuations, the class that has now slipped seven separate audits.

No behaviour depends on any of this. The cost is that a reader who follows a citation lands on real code doing something else and is misled rather than alerted.

Related: #911, #959, #924, #986.

FOLLOWING_AGENTS_PROTOCOL

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions