[leehack]

Summary

• Require hosted dinja ^1.3.1 to render valid Jinja \xHH literals, including Flint's embedded \x00 replacement (Refs Flint chat fails because its embedded template uses unsupported hexadecimal escapes #1335; lexer fix).
• Add the exact licensed Flint template, seven upstream prompt references, public LlamaEngine.create() regressions, typed-history checks, and malformed-escape cases.
• Update the docs, consumer skill, example and Unreleased changelogs; regenerate all three tracked dinja consumer locks and preserve exact captured fixture bytes on Windows.

Production-readiness scope

• User-facing scope: Embedded and custom templates accept two-digit hexadecimal escapes, preserve literal backslashes and Unicode, and report malformed literals as typed inference errors.
• Supported platforms/paths: Shared Dart rendering on VM and Chrome, including the actual public create/render path.
• Unsupported or intentionally unavailable paths: Flint model architecture and generated tool-output support remain unqualified on the pinned native runtime.
• Out of scope / follow-ups: Native runtime qualification stays in llamadart-native #112; runtime pins and parsing/grammar source are unchanged.

Completeness checklist

• Declared rendering scope is complete.
• Negative cases fail before backend generation; runtime limits are documented.
• README, website guide/support matrix, example, skill and both changelogs updated.
• Exact upstream prompts, public create, Unicode/NUL/marker handling, typed-history preservation and malformed escapes covered.
• Capability registry: N/A; existing LlamaEngine.chatTemplate bugfix, no new API or model/runtime support claim (Flint chat fails because its embedded template uses unsupported hexadecimal escapes #1335).
• Security/privacy reviewed: synthetic fixtures; no secrets.
• Native follow-up tracked above.

PR type guidance

• Bugfix: Earlier dinja releases reject valid hexadecimal literals before generation. The repaired hosted release fixes the shared renderer on VM/browser.

Test Plan

• dart run tool/prepare_workspace.dart; initial clean preparation and regenerated consumer locks.
• dart format --output=none --set-exit-if-changed .
• dart analyze and platform boundaries.
• dart test -p vm -j 1 --exclude-tags local-only — exact-head CI split core/runtime/tooling/native lanes and coverage.
• dart test -p chrome --exclude-tags local-only — exact-head CI.
• Final-head independent VM 28, Chrome 23, compiled grammar 4; hosted 1.3.0 control fails all 23 consumer cases, restored 1.3.1 passes.
• Three independent production mutations fail causally: typed reconstruction, unknown-tool rollback, and forced-thought grammar.
• Final-head desktop bundles and extracted archive help/list smoke pass on all three OSes in CI, preserving clean source provenance.
• Docs generation/links and all three desktop bundle jobs pass in CI.

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
static-format-analyze	Root/examples and platform boundaries	Flutter 3.47.1	PASS	CI
root-vm	Core/runtime/tooling/native regressions	Linux, macOS, Windows VM	PASS	Exact-head CI lanes
root-chrome	Shared rendering/browser contracts	Chrome	PASS	Exact-head CI; independent Flint 23/23
coverage-lib	Existing library regression coverage	Linux VM	PASS	Exact-head coverage lane
structured-output-adversarial	Rendering/history and upstream parity; conservative parser/grammar controls	Exact Flint artifact, Qwen35 fixtures, pinned compiled grammar	PASS	Independent VM28/Chrome23/compiled4 and causal controls below
high-risk-exact-head-independent-qa	Blocking adversarial review	Head 6a4451af1990d4d8b12bc9ede9a12ce9d575c607	PASS	Fresh auditor accepted; zero P1 and unresolved threads
docs-site	Site generation and internal links	Jaspr	PASS	Docs Build Check
validation-harness	Private consumer dependency and portable artifacts	Three desktop OSes, web/Android/iOS inputs	PASS	Validation bundles; all tracked consumer locks resolve 1.3.1
native Flint qualification	Model architecture/generation/tool output	Pinned runtime	N/A	Separate owner issue #112; no support claim

Review Notes

• Independent review: audit_flint_consumer accepted the exact final head/base after the validation-lock and Windows fixture corrections. It took no part in implementation.
• CI: all current-head checks passed (selected N/A jobs skipped), run.
• Initial CI exposed a stale private-harness lock and Windows CRLF conversion of exact capture files. Regenerating the lock and adding two narrow Git attributes fixed both without changing production source or the clean-source guard. Independent Windows-style checkout controls confirm exact fixture hashes.

High-risk regression review

• Classification: high-risk — artifactConsumer, structuredOutput.
• Implementation task: flint-consumer-pr; regression worker flint_consumer_regressions; validation-lock worker flint_validation_lock.
• Independent blocking QA task: audit_flint_consumer; accepted after both CI corrections and re-audit.
• Exact head / current base: 6a4451af1990d4d8b12bc9ede9a12ce9d575c607 / 09c189fab37f6683b771aea8b9211d86a3439a60.
• Production-branch deletion, bypass, or miswire proof: Hosted 1.3.0 fails all 23 consumer cases. Three isolated production mutations also fail; all restored. Public create tests prohibit native template fallback.
• Affected-family real-model/artifact evidence: Exact Flint Q4_K_M embedded template and seven official upstream prompt emissions; hosted dinja 1.3.1 archive independently verified. Fixtures include license, revision and hashes.
• Explicit unavailable-family or other N/A evidence: Native Flint architecture/generation and emitted tool calls remain outside this rendering repair.
• Known PR-caused P1 regressions: 0.
• Unresolved review threads: 0, checked against the exact revision.

Exact-head independent review and schema-v2 evidence

Fresh independent review accepted all 19 changed files after both CI corrections. The JSON preserves audit-time CI notes; all current-head checks subsequently passed, as linked above. Controls verify hosted 1.3.0 fails the 23 consumer cases; all pass after restoration. Parser and compiled-grammar mutations fail causally and are restored. Independent Git checkout controls reproduce CRLF conversion before the two exact-capture attributes, then preserve both captures byte for byte after the fix. The validation lock remains unchanged under dependency resolution.

The documented local evaluator exits 2 (unverifiedPrerequisites): repository-local evidence is internally consistent; authenticated external enforcement is intentionally unconfigured, as described in the readiness contract. Manual gates still require current-head CI and zero unresolved threads.

{
  "schema": "llamadart.high-risk-readiness-evidence",
  "schema_version": "2.0.0",
  "test_evidence": {
    "test/integration/core/template/flint_template_test.dart": {
      "test_case": "Flint keeps typed argument and result values while sanitizing text",
      "test_snippet": "'Flint keeps typed argument and result values while sanitizing text'",
      "production_refs": [
        {
          "path": "lib/src/core/template/handlers/hermes_handler.dart",
          "snippet": "final template = Template(templateSource);"
        },
        {
          "path": "lib/src/core/template/template_render_context.dart",
          "snippet": "rendered['content'] = result is String ? result : jsonEncode(result);"
        }
      ],
      "command": "TMPDIR=\"$(mktemp -d)\" dart test -p vm -j 1 test/integration/core/template/flint_template_test.dart",
      "head_result": "pass",
      "control_result": "fail",
      "control_kind": "before-fix",
      "control_description": "In an isolated copy of final 6a4451a source, pubspec_overrides.yaml selects hosted dinja1.3.0. The exact Flint template fails at lexer escape x before rendering. Remove override and offline pub restores1.3.1; source untouched.",
      "evidence_notes": "Independently repeated at final 6a4451a: consumer-independent-6a-1.3.0-red.log +1 fixture/-23 consumer cases; Unknown escape x at Hermes Template. consumer-independent-6a-1.3.1-restored.log +24. Final isolated Chrome +23 in consumer-independent-6a-chrome.log. Fixture integrity is separate +1 proof, not a causal case."
    },
    "test/integration/core/template/qwen35_tool_result_test.dart": {
      "test_case": "typed result history ${malformed ? 'rolls back undeclared tools' : 'preserves schema-directed scalar and container types'}",
      "test_snippet": "'typed result history ${malformed ? 'rolls back undeclared tools' : 'preserves schema-directed scalar and container types'}'",
      "production_refs": [
        {
          "path": "lib/src/core/engine/chat_completion_stream_parser.dart",
          "snippet": "ChatTemplateEngine.parse("
        },
        {
          "path": "lib/src/core/template/xml_tool_call_format.dart",
          "snippet": "final decoded = decodeToolSchemaText(raw, schema);"
        },
        {
          "path": "lib/src/core/template/tool_schema_utils.dart",
          "snippet": "return validateToolSchemaValue(raw, schema);"
        },
        {
          "path": "lib/src/core/template/handlers/qwen3_coder_xml_handler.dart",
          "snippet": "schemas: tools == null ? null : toolSchemas(tools),"
        }
      ],
      "command": "TMPDIR=\"$(mktemp -d)\" dart test -p vm -j 1 test/integration/core/template/qwen35_tool_result_test.dart",
      "head_result": "pass",
      "control_result": "fail",
      "control_kind": "mutation",
      "control_description": "In isolated copy only, return invalid for schema-declared strings in decodeToolSchemaText. Separate negative control removes undeclared XML tool-name rejection. Both restored.",
      "evidence_notes": "Final 6a4451a consumer-independent-6a-green.log +28 (Flint23+fixture1+parser4). Repeated consumer-independent-6a-parser-mutant.log +3/-1: typed reconstruction falls back to raw envelope. consumer-independent-6a-rollback-mutant.log +3/-1: malformed unknown tool becomes executable call. Covers auto/required with thought prefixes and none/no-grammar history branch. Mutations restored."
    },
    "test/e2e/template/specialized_tool_grammar_validation_e2e_test.dart": {
      "test_case": "Hermes required calls may follow forced-open reasoning",
      "test_snippet": "test('Hermes required calls may follow forced-open reasoning', () {",
      "production_refs": [
        {
          "path": "lib/src/core/template/tool_call_grammar_utils.dart",
          "snippet": "static String afterForcedOpenThought(String grammar, String endTag) => ["
        },
        {
          "path": "lib/src/core/template/handlers/hermes_handler.dart",
          "snippet": "grammar: requiredGrammarAfterThought("
        }
      ],
      "command": "LLAMA_CPP_GBNF_VALIDATOR=/private/tmp/flint-fix-1335/grammar-validator-build/bin/test-gbnf-validator TMPDIR=\"$(mktemp -d)\" dart test -p vm -j 1 test/e2e/template/specialized_tool_grammar_validation_e2e_test.dart --run-skipped --name 'Hermes canonical alternatives|Hermes and Mistral compile|Hermes required calls|open schema objects constrain'",
      "head_result": "pass",
      "control_result": "fail",
      "control_kind": "mutation",
      "control_description": "In isolated copy only, replace afterForcedOpenThought wrapper with bare grammar. Restored after observation.",
      "evidence_notes": "consumer-independent-6a-grammar-green.log +4 maintained compiled cases: malformed/unknown envelopes, missing/wrong-type fields, nullable nested schemas and required thought prefixes. consumer-independent-6a-grammar-mutant.log -1 forced-thought valid sequence rejected at initial N when afterForcedOpenThought returns bare grammar. Fresh validator source clean HEAD=v0.7.0 commit620d2225cc436ab0b001e3200e590e3b286ef9ab matching native wrapper v0.7.0-2; binary sha256 dbdeb6f33d062eeefcab67225329be6001aa7b7e8f49b7b805c49ae6404ce07d. Mutation restored byte-for-byte."
    }
  },
  "timestamp": "2026-10-23T11:29:03Z",
  "correlation_id": "pr1342-flint-consumer-audit-6a4451a",
  "repository": "leehack/llamadart",
  "pr_number": 1342,
  "expected_pr_head_sha": "6a4451af1990d4d8b12bc9ede9a12ce9d575c607",
  "current_base_sha": "09c189fab37f6683b771aea8b9211d86a3439a60",
  "pr_author": "leehack",
  "classification": "high-risk",
  "surfaces": [
    "artifactConsumer",
    "structuredOutput"
  ],
  "required_matrix_row_ids": [
    "high-risk-exact-head-independent-qa",
    "structured-output-adversarial"
  ],
  "matrix_row_evidence": {
    "high-risk-exact-head-independent-qa": {
      "row_id": "high-risk-exact-head-independent-qa",
      "result": "pass",
      "command": "Independent blocking-only audit of exact head 6a4451af1990d4d8b12bc9ede9a12ce9d575c607 against base09c189fab37f6683b771aea8b9211d86a3439a60",
      "evidence_notes": "Fresh independent re-audit of19 final files. New .gitattributes delta only protects two exact upstream captures. Independent old/new core.autocrlf=true checkout: old template60CRLF/capturedJSON176CRLF; final0CRLF/exact Git hashes. Old checkout integrity fails -1; fixed checkout finalVM28/Chrome23/compiled4 pass. Validation package lock remains pub-stable. Root refreshed final GraphQL with zero threads; exact-head CI still running. No earlier broad/parity counts are claimed as final6a readiness gates."
    },
    "structured-output-adversarial": {
      "row_id": "structured-output-adversarial",
      "result": "pass",
      "command": "TMPDIR=\"$(mktemp -d)\" dart test -p vm -j 1 test/integration/core/template/flint_template_test.dart test/integration/core/template/flint_fixture_integrity_test.dart test/integration/core/template/qwen35_tool_result_test.dart\nLLAMA_CPP_GBNF_VALIDATOR=/private/tmp/flint-fix-1335/grammar-validator-build/bin/test-gbnf-validator TMPDIR=\"$(mktemp -d)\" dart test -p vm -j 1 test/e2e/template/specialized_tool_grammar_validation_e2e_test.dart --run-skipped --name 'Hermes canonical alternatives|Hermes and Mistral compile|Hermes required calls|open schema objects constrain'",
      "evidence_notes": "Final6a isolatedVM28, Chrome23, compiled4. Hosted1.3.0 control +1/-23 then restored1.3.1 +24. Schema decoding +3/-1, unknown-tool guard deletion +3/-1, grammar bypass -1; all restored. Separate Windows checkout integrity control -1→pass. Actual effect rendering/history; unchanged parsing/grammar conservatively covered. No Flint model/tool-output qualification."
    }
  },
  "independent_audit": {
    "auditor_identity": "codex-auditor-flint-consumer",
    "audit_kind": "codex-adversarial",
    "audit_head_sha": "6a4451af1990d4d8b12bc9ede9a12ce9d575c607",
    "audit_base_sha": "09c189fab37f6683b771aea8b9211d86a3439a60",
    "decision": "accepted",
    "unresolved_review_threads": 0,
    "known_pr_caused_p1_regressions": 0,
    "summary": "Independent blocking re-audit accepted for6a4451a vs09c189f. All19 changed files inspected. Increment fromdcc is four .gitattributes lines: comment/blank plus two narrow -text exact-capture rules. Independent core.autocrlf=true old/new checkout reproduces and repairs byte/hash failure; integrity negative control fails and fixed bytes pass. All final-head behavior/parser/compiled tests and causal controls rerun. Zero known PR-caused P1. Root GraphQL refreshed exacthead/base and zero threads; current CI remains pending. Prior head audits are superseded; no Flint architecture/generation/output tools qualified."
  },
  "structured_output_evidence": {
    "coverage": {
      "input_rendering_history": [
        "test/integration/core/template/flint_template_test.dart",
        "test/integration/core/template/qwen35_tool_result_test.dart"
      ],
      "compiled_grammar_acceptance": [
        "test/e2e/template/specialized_tool_grammar_validation_e2e_test.dart"
      ],
      "compiled_grammar_rejection": [
        "test/e2e/template/specialized_tool_grammar_validation_e2e_test.dart"
      ],
      "schema_reconstruction": [
        "test/integration/core/template/qwen35_tool_result_test.dart"
      ],
      "streaming_rollback": [
        "test/integration/core/template/qwen35_tool_result_test.dart"
      ],
      "tool_choice_thinking": [
        "test/integration/core/template/qwen35_tool_result_test.dart",
        "test/e2e/template/specialized_tool_grammar_validation_e2e_test.dart"
      ],
      "upstream_parity": [
        "test/integration/core/template/flint_template_test.dart"
      ]
    },
    "impacts": {
      "inputRenderingHistory": {
        "applicable": true,
        "rationale": "Hosted dinja 1.3.1 is consumed by every Template constructor. Flint rendering reaches Hermes, shared context and public create. Tests compare upstream bytes and preserve typed history.",
        "production_refs": [
          {
            "path": "lib/src/core/template/handlers/hermes_handler.dart",
            "snippet": "final template = Template(templateSource);"
          },
          {
            "path": "lib/src/core/template/template_render_context.dart",
            "snippet": "rendered['content'] = result is String ? result : jsonEncode(result);"
          }
        ]
      },
      "outputParsingStreaming": {
        "applicable": true,
        "rationale": "No parser or grammar implementation bytes change in this PR. Evidence-only structured paths impose this conservative minimum impact; unchanged production boundaries were inspected and validated with maintained cases and isolated causal mutations.",
        "production_refs": [
          {
            "path": "lib/src/core/engine/chat_completion_stream_parser.dart",
            "snippet": "ChatTemplateEngine.parse("
          },
          {
            "path": "lib/src/core/template/xml_tool_call_format.dart",
            "snippet": "final decoded = decodeToolSchemaText(raw, schema);"
          }
        ]
      },
      "grammarSchema": {
        "applicable": true,
        "rationale": "No parser or grammar implementation bytes change in this PR. Evidence-only structured paths impose this conservative minimum impact; unchanged production boundaries were inspected and validated with maintained cases and isolated causal mutations.",
        "production_refs": [
          {
            "path": "lib/src/core/template/tool_call_grammar_utils.dart",
            "snippet": "static String afterForcedOpenThought(String grammar, String endTag) => ["
          },
          {
            "path": "lib/src/core/template/handlers/hermes_handler.dart",
            "snippet": "grammar: requiredGrammarAfterThought("
          }
        ]
      }
    },
    "families": [
      {
        "family": "Flint-template-artifact",
        "status": "tested",
        "evidence_test_paths": [
          "test/integration/core/template/flint_template_test.dart"
        ],
        "rationale": "Exact upstream template/captured apply-template emissions at model revision 468307a1b38259e9e4c9d4fe53d4e3b76386572c; seven prompt cases, typed history and public mocked create. Native architecture/generation/output tools are deliberately unqualified (native issue112)."
      }
    ]
  },
  "affected_test_paths": [
    "test/integration/core/template/flint_template_test.dart",
    "test/integration/core/template/qwen35_tool_result_test.dart",
    "test/e2e/template/specialized_tool_grammar_validation_e2e_test.dart"
  ],
  "evaluation": {
    "evaluated_at": "2026-10-23T11:31:10.094644Z",
    "changed_files": [
      {
        "path": ".gitattributes",
        "status": "modified"
      },
      {
        "path": "CHANGELOG.md",
        "status": "modified"
      },
      {
        "path": "README.md",
        "status": "modified"
      },
      {
        "path": "example/basic_app/README.md",
        "status": "modified"
      },
      {
        "path": "example/chat_app/pubspec.lock",
        "status": "modified"
      },
      {
        "path": "packages/llamadart_validation/pubspec.lock",
        "status": "modified"
      },
      {
        "path": "pubspec.yaml",
        "status": "modified"
      },
      {
        "path": "skills/llamadart-chat-streaming/SKILL.md",
        "status": "modified"
      },
      {
        "path": "test/fixtures/flint/Flint.jinja",
        "status": "added"
      },
      {
        "path": "test/fixtures/flint/LICENSE",
        "status": "added"
      },
      {
        "path": "test/fixtures/flint/README.md",
        "status": "added"
      },
      {
        "path": "test/fixtures/flint/render_upstream.json",
        "status": "added"
      },
      {
        "path": "test/integration/core/template/flint_fixture_integrity_test.dart",
        "status": "added"
      },
      {
        "path": "test/integration/core/template/flint_template_test.dart",
        "status": "added"
      },
      {
        "path": "test/support/flint_template_fixture.dart",
        "status": "added"
      },
      {
        "path": "website/docs/changelog/recent-releases.md",
        "status": "modified"
      },
      {
        "path": "website/docs/guides/chat-template-and-parsing.md",
        "status": "modified"
      },
      {
        "path": "website/docs/platforms/support-matrix.md",
        "status": "modified"
      },
      {
        "path": "website/examples/observability/pubspec.lock",
        "status": "modified"
      }
    ],
    "decision": "unverifiedPrerequisites",
    "failure_classification": "externalPrerequisitesUnavailable",
    "message": "Repository-local evidence is internally consistent, but auditor authentication, GitHub App publication, protected-environment provenance, and ruleset enforcement are not available. This is not operational merge readiness.",
    "external_prerequisites": {
      "app_installed": false,
      "protected_environment_configured": false,
      "independent_auditor_authenticated": false,
      "ruleset_enforced": false,
      "diagnostic_message": "No repository-local input can authenticate the dedicated GitHub App, protected environment, independent auditor, or conditional ruleset. See doc/high_risk_pre_merge_readiness.md."
    }
  }
}

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1342.