[noahgift]

The apr-cli Ollama shim always ended a turn with done_reason: "stop". It now reports "length" when the turn hit num_predict:

• Coalesced chat/generate and batch-NDJSON replies read the upstream OpenAI finish_reason.
• The live APR-CPU stream ends with StreamEnd::LengthAt(max_tokens). This is approximate: a natural stop on exactly the last allowed token also reads "length". A backend error reports "stop".

Not covered: the SafeTensors upstream still hardcodes finish_reason "stop", so its Ollama routes cannot report "length" yet.

cargo test -p apr-cli --lib -- commands::serve: 405 passed.

Refs #5357
keep-open: #5357 tracks the remaining shim gaps (SafeTensors finish_reason, the compat_ollama copy, json_schema refusal text).
ont-delta: none

🤖 Generated with Claude Code

[noahgift]

seat 1, aprender-seat-o10 @c0e73ccbed APPROVE

Opus seat. be492a5 and c0e73cc are Sonnet 5.5 (la-0.71.1a). The middle commit 5e20d9b is a main-merge with no conflict hunks (git diff-tree --cc is empty). This session authored no commit on the PR.

Read and checked at the head:

• Coalesced chat/generate: openai_done_reason returns "length" only on a 2xx whose choices[0].finish_reason == "length". Anything else is "stop": non-2xx, unparseable body, missing field. That is the conservative default; it never reports a cap hit it cannot see.
• Batch NDJSON reshape: it maps the upstream reply to StreamEnd::Length/Stop, which is consistent with the coalesced path.
• Live APR-CPU stream: I checked spawn_cpu_token_text_stream at :181. It sends one channel message per token id via the generate callback, so the stream's count is a token count and LengthAt(max_tokens) compares like with like. max_tokens comes from ollama_sampling: num_predict (u32) or a default of 32, capped at 4096. So a turn that hits the default cap also correctly reads "length".
• Backend errors: an Err on the channel is forced to "stop" (ended @ (Some(Err(_)) | None)), and the row live_stream_error_at_the_cap_reports_stop_not_length pins it. A generate() failure only reaches eprintln and then closes the channel, so it can read "length" only if the cap had already been reached.
• Other callers keep their old behavior: ollama_ndjson_stream now delegates with StreamEnd::Stop, so its other callers (ollama_stream_error, the existing tests) still answer "stop". ollama_stream_final has exactly one non-test caller, and it was updated.
• Tests: they cover both reasons on every path: the coalesced chat/generate and the batch reshape for length and stop, the live stream at cap-1, at the cap and on error, and the StreamEnd boundaries at 2/3. Each assertion would fail if the old hardcoded "stop" came back. The PR body reports 405 passed for commands::serve; I did not re-run it.

Not blocking, and already stated in the body:

1. LengthAt is approximate: a natural EOS on exactly the last allowed token reads "length". If the generate callback also fires for the EOS id, a reply of cap−1 content tokens plus EOS reads "length" too. Worth one row on #5357 once the SafeTensors finish_reason follow-up lands.
2. num_predict: 0 gives LengthAt(0) → "length" on an empty turn. Arguably correct, since the cap was hit.

CI on this head was PENDING at read (authorize/x86-main/shards queued or running). This APPROVE is bound to c0e73cc; the PR is READY only once the rollup reads green on it.

[noahgift]

seat 2, aprender-seat-o9 @c0e73ccbed -- APPROVE (no blocking finding)

Reviewed the full diff (be492a5, the main merge 5e20d9b, c0e73cc): ollama.rs and handler_apr_cpu_completion.rs. It is Sonnet-authored, and this seat runs Opus.

• Coalesced chat/generate: openai_done_reason reads choices[0].finish_reason. It returns "length" only for a 2xx with finish_reason == "length". Non-2xx, bad JSON or a missing field return "stop", which is the old behaviour, so the change never invents a "length".
• Batch NDJSON: the upstream "length" maps to StreamEnd::Length. Anything else maps to Stop.
• Live APR-CPU stream: StreamEnd::LengthAt(max_tokens) gives "length" when count >= cap. spawn_cpu_token_text_stream sends one Ok per generated token, so count is a token count. The exact-cap natural-stop case the PR body states is the only false "length". A Some(Err(_)) terminal is forced to "stop" (c0e73cc), and a test covers it.
• ollama_ndjson_stream keeps its signature through a Stop wrapper, so other callers are unchanged. ollama_stream_final has one non-test caller, and it was updated.
• Tests discriminate: cap 2 vs 3 over 2 tokens, length vs stop for each reshaped route, error at the cap, and the four openai_done_reason cases.

Measured locally on a worktree at c0e73cc: cargo test -p apr-cli --lib -- commands::serve::ollama gave 25 passed and 0 failed, rc=0.

Non-blocking: as the body says, the SafeTensors upstream still hardcodes finish_reason: "stop", so those routes stay "stop". keep-open: #5357 tracks it.

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=5414 head=95dfdbd8adb3658858d972417fae293aa3f3266c verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

seat 1, aprender-seat-o10 @95dfdbd8ad APPROVE

Re-seat. The only change from c0e73cc (my seat-1 APPROVE above, whose reasoning carries over unchanged) is 95dfdbd. It is a Sonnet 5.5 commit that touches only docs/audits/surface_audit.csv, 2 lines. No code changed.

The two apr-cli .../serve/ollama.rs refs move from :650/:654 to :726/:730. That is a uniform +76 shift, which matches the 76 lines this PR adds above the route table (ollama_pull_body moves 622 → 698).

Not blocking. The bad refs come from main, and this commit faithfully carries them forward: both refs were already off on main and are still off here.

• On main, :650 is }; and :654 is "/api/show". The real lines were /api/pull :659 and /api/delete :664.
• At this head, :726 is }; and :730 is "/api/show". The correct refs are POST /api/pull → :735 and DELETE /api/delete → :740.

Worth fixing in this PR if it is cheap. Otherwise it belongs on the stale-ref ticket I recommended on #5393: whatever G2.1 checks, it does not check that a ref lands on the route it names.

CI on this head was PENDING at read: x86-main and shards 1-3 in progress. This APPROVE is bound to 95dfdbd; READY only once the rollup is green on it.

[noahgift]

seat 2, aprender-seat-o9 @95dfdbd8ad -- APPROVE (no blocking finding)

Delta reviewed: c0e73cc..95dfdbd, one commit. It changes docs/audits/surface_audit.csv only (2+/2-). git diff c0e73ccbed 95dfdbd8ad -- crates is empty, so my code review at c0e73cc carries over unchanged.

The renumber is consistent. At the merge base, ollama.rs has vec![ at 651, "/api/pull" at 659 and "/api/delete" at 664, and the ledger anchors are 650 and 654. At 95dfdbd those lines are 727, 735 and 740. The new anchors 726 and 730 keep the same offsets (+76, matching the PR's inserted lines). No ledger row is added, dropped or re-graded.