[leehack]

Summary

• Seed-OSS: a reply to a prompt that opens no thought (the stock template) has one only when it starts with <seed:think> after ASCII whitespace only (C isspace, as upstream; NBSP/U+3000 do not count), kept verbatim up to the first </seed:think>, as llama.cpp 620d2225 parses it; later thinking tags stay content. Fixes #1561.
• Cohere2 MoE (#1566):
  1. With a thought a handler closed for thinking off, Plan<|START_TEXT|>Sunny. now gives reasoning Plan / content Sunny.: a reply that reaches a marker parses (and streams) as after the forced-open prompt llama.cpp renders.
  2. Thinking tags inside a text or action body are text (e.g. a literal <|END_THINKING|> in a text block).
  3. Response-schema renders (another handler renders the Cohere2 MoE template) close the thought with thinking off, as the Cohere2 MoE handler does.
• No public API change. The stream parser parses its renders through the internal parseRenderedOutput, which reads the render's internal closed-thought mark, so the engine's stream and final message match llama.cpp for marked and unmarked renders. The public ChatTemplateEngine.parse keeps its signature and cannot see the mark, so for Seed-OSS it still ends a thought at a </seed:think> in a reply that does not start with <seed:think>, as base did and as llama.cpp does after a template that opens the thought (e.g. genkit_llamadart's parse of a custom open-thought template). Trade-off: a public parse of a stock-template reply with such a stray end tag keeps base's result, not upstream's.
• #1565 items: tests kill Hold a closed Cohere2 MoE, Command R7B or Seed-OSS thought; stop holding Ministral and Solar Open #1560 audit survivors S5, S10, S11; changelog "hold it the same way" rewording.
• Audit follow-ups at 1bb9148: tests kill N20 (Cohere2 MoE end tag before a body, upstream-accepted on templates that open the thought), A13 (empty reply on a marked render), A14 (bare <seed:think> at EOF), A26 (response-schema close only for Cohere2 MoE) and A28 (its multimodal render path); changelog . ( typo fixed.

Root cause: Seed-OSS parsed any <seed:think>…</seed:think> anywhere as reasoning, while llama.cpp's autoparser only takes a leading thought when the prompt opens none. Cohere2 MoE extracted thinking over the whole reply instead of stopping at the body, and the marked (thought-closed) result parsed as unforced, so text before <|START_TEXT|> was content; the response-schema render path skipped the handler that closes the thought.

Production-readiness scope

• User-facing scope: Final messages and streams for Seed-OSS and Cohere2 MoE match llama-server on the Seed-OSS parse takes a leading thought as reasoning when the prompt never opened one #1561/Cohere2 MoE parse diverges from upstream on START_TEXT-ended thoughts and in-text END_THINKING #1566 inputs; stream equals parse at every chunking tested.
• Supported platforms/paths: All (pure Dart template parsing / stream parser; native and Web).
• Unsupported or intentionally unavailable paths: None changed.
• Out of scope / follow-ups: Pre-existing divergences, unchanged by this PR (listed under Review Notes).

Completeness checklist

• Declared scope is fully implemented, or explicitly reduced above.
• Unsupported platform/option combinations fail loudly with actionable diagnostics or are clearly documented as unavailable. (N/A: no new unsupported paths)
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant. (chat-template-and-parsing.md, template-engine-internals.md; both changelogs. README/support matrix/examples/skills unaffected.)
• Regression coverage covers the original issue plus key negative/version-skew paths where relevant.
• Public capability changes update the validation coverage registry with a pack or a reasoned, issue-linked exemption; inventory checks pass (or N/A with scope). (N/A: parse-parity fix, no new capability)
• Security/privacy review completed: no secrets, bearer tokens, signed URLs, or raw secret-bearing paths leak through logs, cache keys, metadata, errors, or snapshots.
• Follow-up work that is useful but not required for this PR is tracked in GitHub Issues, or explicitly marked "None" above.

PR type guidance

• Bugfix PR: root cause above; regression tests are upstream-oracle fixture cases in test/fixtures/closed_thought_upstream.json (new no_thought_templates group plus Cohere2 MoE open-thought cases) driven by test/integration/core/engine/closed_thought_upstream_test.dart, and stream-parser unit tests. Affected: all platforms.
• Feature PR: N/A. Docs-only PR: N/A.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze
• dart test -p vm -j 1 --exclude-tags local-only
• dart test -p chrome --exclude-tags local-only — ran the template/engine subset (test/unit/core/template test/unit/core/engine); the change is pure Dart parsing.
• Other: auditor's oracle harness (main, X and Y corpora) at the head; dart run tool/testing/check_platform_boundaries.dart; dart test -p vm test/unit/tooling/; ./tool/testing/run_template_parity_suites.sh (exit 0).

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
static-format-analyze	format, analyzer, boundaries	macOS, Flutter 3.47.1	PASS	0 changed, no issues, boundaries OK
root-vm	VM unit/integration	macOS arm64	PASS (+9639 ~126)	-j 1 --exclude-tags local-only
root-chrome	browser template/engine subset	Chrome	PASS (+2135 ~1)	test/unit/core/template test/unit/core/engine
coverage-lib	lib coverage	CI	N/A	CI lane
template-parity	vendored llama.cpp detection/render/parse parity, GBNF	macOS, llama.cpp b10549 sources	PASS	all five suites pass, exit 0
structured-output-adversarial	partial/final streaming and rollback, tools on/off, thinking on/off, pinned upstream parse parity	llama.cpp 620d2225 oracle	PASS	Oracle: 0 OK→BAD, 0 new stream≠parse, 0 public-parse OK→BAD over 229,656 rows (below)
high-risk-exact-head-independent-qa	independent blocking review of exact head	fresh auditor session, exact head 70332a4	PASS	CLEAR: CLEAR delta re-audit at 70332a4, 0 P1/P2; P3s on #1565; evidence /opt/UnitySrc/personal/llama/llamadart-evidence/pr-1569/evidence-1569-70332a4.json
gguf-chat-features-smoke / llama-cpp-chat-template-smoke	real GGUF	—	N/A	no Seed-OSS or Cohere2 MoE weights available; evidence is primary upstream parser emissions (llama.cpp 620d2225) plus durable fixtures

Upstream oracle (auditor's harness: llama.cpp 620d2225 vs llamadart engine-derived parse, public ChatTemplateEngine.parse, and whole/bytewise/every-2-split streams; base f3706ff vs head):

Corpus	Rows	OK→BAD	New stream≠parse	Public parse OK→BAD	BAD→OK	stream≠parse fixed
Main	185,160	0	0	0	2,236 (Seed-OSS 2,144, Cohere2 MoE 92)	2,160
X corpus	37,056	0	0	0	823	637
Auditor Y corpus	7,440	0	0	0	408	178

The 237 X and 45 Y X-Seed-open public-parse rows the audit flagged match upstream again. Visible-byte rows differing only by a trailing < on Seed-OSS with tools (8 main, 6 X) are unchanged from the audited head.

Mutation: first pass 34 mutants, 32 killed (S14 equivalent; N20 now killed). Audit pass at 1bb9148: N20, A13, A14, A26, A28 survived on correct code; all five are now killed, plus the new code's mutants (public Seed-OSS end-tag rule removed, ignoring a leading start tag, applied to every format, applied to the internal parse; Unicode trimLeft in the handler and the stream; narrowed ASCII whitespace set): 13 of 13 killed.

Review Notes

• Independent review status: CLEAR at exact head 70332a4 by a fresh auditor session (0 P1, 0 P2).
• CI status / head SHA: 70332a4
• Latency: unmarked Seed-OSS content streams earlier; unmarked Cohere2 MoE holds a trailing opening prefix (first content of a bare <|START_ACTION|> without tools moves from byte 9 to 16); marked Cohere2 MoE text with tools streams at EOF, as forced-open Cohere2 MoE with tools already does.
• Pre-existing, unchanged: forced-open Cohere2 MoE action whose arguments contain <|START_TEXT|>, incomplete actions, <|END_OF_TURN_TOKEN|> after a text, and a trailing < partial differ from upstream; forced Seed-OSS (custom template opening the thought) is still not verbatim; unmarked Cohere2 MoE with an end tag before any start tag streams its first characters as content when tokens split before the end tag; Seed-OSS content keeps Dart's Unicode trim() at its ends.

High-risk regression review

• Classification: high-risk (surfaces: backendRuntime, structuredOutput)
• Implementation task: llamadart 0.17.0 lane (#922) for #1561 and #1566
• Independent blocking QA task: fresh auditor session, CLEAR, 0 known PR-caused P1, 0 unresolved threads
• Exact head / current base: 70332a4 / dc8d179 (merge-tree clean)
• Production-branch deletion, bypass, or miswire proof: mutation checks above (remaining survivor S14 is equivalent); failing-first: new fixture cases fail at base (18 failures)
• Affected-family real-model/artifact evidence: unavailable (Seed-OSS, Cohere2 MoE weights); primary upstream emissions via llama.cpp 620d2225 parser plus durable fixtures
• Explicit unavailable-family or other N/A evidence: Seed-OSS, Cohere2 MoE GGUF weights unavailable
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1569.