feat(core): note an identical call with an identical result to the model

[Max17190]

Why

A model that has finished can fail to stop: its reasoning says the task is complete, its reply is the same verification call again, the output is the same, and the turn runs to the iteration cap. On one full run of the release binary a fifth of all model turns repeated the previous turn's calls and results byte for byte, and one trial in eight ran to the cap that way. #225 removed a veto keyed on identical arguments because arguments say nothing about results and a stuck model is bounded by budgets; that reasoning stands, and this change vetoes nothing.

Summary

• RepeatTracker hashes each iteration's tool calls and the tool replies they drew. Every call still executes. From the third consecutive byte-identical pair, the tool message the model reads next carries an advisory note with the count that names both ways out: keep waiting, or stop calling tools and answer. A reply with no calls resets the count.
• The note is appended after the pair is hashed, so it never enters the next comparison, and it lands at the tail of the prompt, so no cached prefix is invalidated.
• The effect on a model's behavior is not measured here; the tests pin the mechanism (when the note appears, with what count, and that nothing is vetoed), and the next full run measures the rest.

Test Plan

• New identical_calls_with_identical_results_draw_a_note_from_the_third: five identical read_file calls; all five execute, replies one and two carry no note, three through five carry the note with the right count. Fails on main at the third reply.
• New identical_calls_with_changing_results_draw_no_note: one bash command with fixed arguments whose output moves; no note on any reply.
• cargo test --workspace --locked green; cargo +1.97.0 clippy --workspace --all-targets --locked -- -D warnings clean.

Greptile Summary

Summary

• Adds an advisory message when consecutive tool-call iterations have identical calls and identical results.
• Keeps execution unchanged and avoids the advisory when repeated calls return different output.
• Focused and multi-tool behavior checks passed; safe to merge.

Confidence Score: 5/5

Safe to merge: the repeated-call advisory behaves as intended for identical single-tool and multi-tool iterations, while changing results remain unannotated.

No issues were found. Executed checks covered the third-and-later advisory threshold, continued detection after prior advisory text was appended, multi-tool iterations, and changing-output polling.

Files Needing Attention: None.

T-Rex Logs

What T-Rex did

• Verified that repeated single-tool call/result sequences trigger an advisory starting on the third consecutive iteration.
• Verified with a local streaming endpoint that byte-identical multi-tool call/result iterations receive the advisory on iterations 3 through 5.
• Confirmed that detecting later iterations remains effective after an earlier advisory has been appended to the transcript.
• Observed that RepeatTracker::observe runs after tool execution and before append_and_emit_note, preserving detection across later iterations despite the appended note.
• Validated that the contract behavior aligns with expectations at crates/core/src/agent.rs:1657-1671 and crates/core/src/agent.rs:3489-3491.

_{Ran code and verified through T-Rex}

_{Reviews (1): Last reviewed commit: "feat(core): note an identical call with ..." | Re-trigger Greptile}