[devopsfortunato]

What this changes

docs/api.md's Reasoning section said a no-thinking GLM-5.3 prompt has "no reasoning-effort line". Native renders the checkpoint's template unchanged and appends </think> once to close its dangling think block (zig/src/server/prompt.zig), so the template's Reasoning Effort head stays and the prompt is the render plus one token. The released Python 0.6 engine instead rewrites GLM's render (thinking_off strips the head and closes the block inline), which is what its smaller usage.prompt_tokens reports. The section now states both engines' behavior and what follows: a prompt count, a cache key or a kept state from one engine does not match the other's for enable_thinking: false. The CHANGELOG carries one bullet.

• Issue: Refs glm: the released Python engine renders the --no-thinking prompt 6-7 tokens shorter than the checkpoint template; native renders it verbatim #1911 — the issue asked for a docs line if the engines' prompts differ deliberately; @ashhart's measurement there (5 Nov) settled native's shape (the render unchanged plus one appended </think>, id 154842, the 22-token vector pinned in the host suite) and confirmed the Python lane's 6-7-token-shorter prompts. This PR records that in the docs; the issue's quality question (whether the kept head degrades greedy no-thinking output) stays open there, and nothing in this change touches code.
• Stacked on: main, 8993a3cca.
• atc overlap --json: run after the final rebase on 8993a3cca. 384 entries; complete is false because ATC reports pullFilesLimit, so 374 entries are unknown for lack of those pull requests' file lists. Every open pull request overlaps only CHANGELOG.md; five report conflicts (glm cuda: TF_GLM_GRAPHS=1 replays verify windows as CUDA graphs (#1490) #2279, contract: an XPU device runner (gates 1 to 4 on a B70, gate 5 live, planted controls) #2269, server: an explicit thinking off that a template can't honour is refused #2170, CUDA: #1764's BF16 prompt tile for Qwen ModelOpt, Nemotron and Dory, plus two exact Qwen prompt kernels (#1762) #2008, qwen CUDA: a faster opt-in paged path (TF_QWEN_KV_PAGES=256), --parallel a cap with pages set #2011) and each is that one file — every one prepends a bullet under ## Unreleased, resolved at landing by keeping all bullets. docs/api.md overlaps with no open pull request; one local worktree holding this pull request's previous head shares both files and conflicts on CHANGELOG.md alone.

Checks on this head

• Head commit: 3648d035b (rebased on 8993a3cca; the approved head was dd28d3360 on 2e81d5df1). The rebase changed no content: docs/api.md is byte-identical to that approved head, and the only conflict was the CHANGELOG.md top bullet, resolved by putting this change's bullet at the top of ## Unreleased.
• Lean checker, python3 tools/zig/lean_check.py: base 8993a3cca 425 problems, this head 425 problems; lean_new.py reports 425 problems, 425 on the base, 0 new. The nonzero count is inherited from main.
• Tests run, with the command and pass, fail and skip counts: no code changed, so no test counts apply. On this head zig fmt --check build.zig zig exit 0 and git diff --check exit 0.
• CI: all seven checks pass on this head — zig 15m03s, zig-windows 22m52s, zig-macos-dev-tools 9m19s, release-paths 1m17s, fmt 13s, benchmark-clients 13s and 22s.
• Not run, and why: zig build test and the other suites locally — the diff is two Markdown files (docs/api.md, CHANGELOG.md); the build graph is unchanged, and CI builds and runs the unchanged tree on this head.

Receipt

n/a — documentation only. The change cannot reach inference: no source, build-graph or configuration file is touched, so prompts, tokens and caches are byte for byte as main's.

Checklist

• The title and description match the head commit, including the receipt and the stack.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. — n/a, no new tests; no code changed.
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

[ashhart]

Thanks @devopsfortunato. This matches what we measured on #1911, and what zig/src/server/prompt.zig does: native renders the template unchanged and appends </think> once. It also replaces the old sentence, which was wrong for native. It's queued to land, and #1911 closes with it.

[ashhart]

Thanks @devopsfortunato. The push still says what we measured on #1911 and what prompt.zig does: the render is unchanged, </think> is appended once when the render ends in <think>, and the template's own head stays. It lands from the queue. At landing we'll make the CHANGELOG line cite "(#2297, #1911)" with your thanks. The body still names ccb635cfd, so please point it at the current head if you touch it again.

The quality question you left open on #1911 (whether the kept head changes greedy no-thinking output) stays there: we'll keep #1911 open for it and note that the docs half is done.

[devopsfortunato]

Done — the body now names head 3648d035b (rebased on 8993a3cca), not ccb635cfd: head commit, base, the lean checker on both (425 and 425, lean_new.py 0 new), the post-rebase atc overlap --json and the CI run are all recorded under "Checks on this head". CI is 7/7 green on that head — zig 15m03s, zig-windows 22m52s, zig-macos-dev-tools 9m19s, release-paths 1m17s, fmt 13s, benchmark-clients 13s and 22s. Content is unchanged from dd28d3360; the only resolution was moving this bullet to the top of ## Unreleased. Main has moved again since, so the changelog line reads CONFLICTING again — leaving that to your landing flow as you described, and no further pushes unless you want one. Thanks for the landing note and the "(#2297, #1911)" citation.

[ashhart]

Thank you @devopsfortunato. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.