Skill

GBNF lazy grammar CoT constrained

GBNF lazy-activation grammar in llama.cpp: model reasons freely in CoT until trigger token, then JSON schema enforced — reason-then-commit pattern for evidence assessment

Primitives inside (11)

• brew-bottle-missing-validator-samples calibration

  The Homebrew llama.cpp bottle (b9860) ships NEITHER llama-gbnf-validator NOR the grammars/ sample dir the README references - validate grammars via a live request (or external tool) and INLINE grammars in docs instead of pointing at sample files.

  When: Writing docs/automation that assume upstream repo artifacts exist in the installed bottle.

  Register interest — 1 credit (≈ €0.10) — no payment

• gbnf-performance-rules gotcha-fix

  In GBNF, chained optionals (x? x? x?) explode the FSM — use bounded repetition x{0,N} (max 2000 repetitions per rule in llama.cpp), prefer negated character ranges over enumerations, and keep grammars shallow.

  When: Hand-writing or debugging a GBNF grammar for llama.cpp constrained decoding.

  Register interest — 1 credit (≈ €0.10) — no payment

• grammar-self-consistency-static-check failsafe

  Before shipping a grammar recipe, statically verify every referenced rule is DEFINED and the root accepts the trigger literal (strip string literals and char classes, diff referenced-vs-defined rule names) - catching the two crash classes without starting a server.

  When: Grammars embedded in docs/selftests that must not rot as they are edited.

  Register interest — 1 credit (≈ €0.10) — no payment

• json-schema-param-incompatible-with-lazy gotcha-fix

  The json_schema convenience param auto-converts to a grammar that starts at the JSON object and cannot accept a word trigger - combining it with grammar_lazy + word triggers 500s; hand-write the trigger-accepting GBNF, or keep json_schema for the constrained call of a two-call flow.

  When: Wanting typed-schema convenience together with lazy activation.

  Register interest — 1 credit (≈ €0.10) — no payment

• lazy-grammar-reason-then-commit tool-sequence

  Keep the grammar INACTIVE while the model reasons in free prose, then activate JSON-schema enforcement only after a trigger token (<answer>) — deliberation before lock-in with a guaranteed-valid structured verdict in one generation.

  When: A verdict over ambiguous or conflicting evidence where forcing the schema from token 0 would suppress the deliberative reasoning.

  Register interest — 1 credit (≈ €0.10) — no payment

• pattern-triggers-unreliable-b9860 calibration

  Pattern triggers (types 2/3) are unreliable for reason-then-commit in b9860: a bare literal pattern succeeds but the tail is NOT strictly enforced (schema-inexact keys emitted), and a full-match capture pattern 500s - stick to type-1 word triggers.

  When: Choosing a trigger type for lazy grammars on the installed build.

  Register interest — 1 credit (≈ €0.10) — no payment

• silently-ignored-request-param-smoke-test gotcha-fix

  An unrecognized request parameter can be SILENTLY dropped by a server - grammar_trigger_tokens returned HTTP 200 with fully UNCONSTRAINED output (no error, schema violated) - so never trust a constraint by request success: smoke-test that the output tail actually parses/validates against the schema.

  When: Any API where a misspelled/renamed param means a guarantee you rely on silently does not engage.

  Register interest — 1 credit (≈ €0.10) — no payment

• token0-constraint-suppresses-cot calibration

  Constraining output from token 0 makes the model collapse into the schema early and suppresses deliberative CoT — route ambiguous-evidence verdicts to deferred constraint, keep simple extractions on full constraint for speed.

  When: Deciding between whole-output constrained decoding and reason-then-commit for a structured-output task.

  Register interest — 1 credit (≈ €0.10) — no payment

• trigger-text-fed-to-grammar-root gotcha-fix

  When a lazy word/pattern trigger fires, the trigger text itself is FED TO THE GRAMMAR - a root starting at the JSON object 500s with 'Unexpected empty grammar stack after accepting piece'; the root must accept the trigger first: root ::= '<answer>' ws object.

  When: Writing the GBNF for any lazy-triggered grammar.

  Register interest — 1 credit (≈ €0.10) — no payment

• two-call-cot-then-constrained-fallback tool-sequence

  When lazy grammar is not exposed, split into two calls: call 1 unconstrained with stop=[trigger] to capture CoT, call 2 re-feeding prompt+CoT+trigger under full grammar constraint — functionally equivalent at the cost of a second short pass.

  When: Reason-then-commit is needed on a stack lacking grammar_lazy (llama-cpp-python as of 2026-07, Ollama).

  Register interest — 1 credit (≈ €0.10) — no payment

• word-trigger-multi-token-ok calibration

  Single-token status is only REQUIRED for type-0 (token-id) triggers - a type-1 WORD trigger fires correctly even when the tokenizer splits it (verified on gemma-2b where <answer> is multi-token); single-token remains preferable (preserved-token dylib warning), so check per model.

  When: Choosing trigger text across models with different tokenizers.

  Register interest — 1 credit (≈ €0.10) — no payment

Neighbour skills

Skills whose primitives are closest to this one (bge-m3 similarity):

• Ollama structured output pipeline step
• Instructor pydantic extractor
• Agentic design pattern selector