chore(ticket): Track fixture review findings

docs/ticket/063sw1z-enforce-resolved-tool-argument-constraints-before-every-disp.md

@@ -0,0 +1,38 @@
+# Enforce resolved tool argument constraints before every dispatch
+
+- **Status**: Todo
+- **Kind**: Bug
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+Tool-call validation checks missing and unknown names, but does not enforce
+parameter types, enums, array item schemas, or nested value constraints.
+It runs only for local tools; MCP and built-in tools bypass it.
+
+The recorded provider fixtures already contain invalid calls that pass through
+the test harness, including Cerebras arrays with numeric items, Google values
+outside the declared enum, and llama.cpp JSON-looking strings where an array was
+intended.
+
+`conversation.tools.<name>.parameters` documents enums as allowed-value
+constraints and supports forcing a value.
+Those constraints must be enforced by JP rather than treated only as model
+guidance.
+
+Open PR \#998 validates resolved schema definitions but does not validate
+argument instances at dispatch time.
+
+Acceptance criteria:
+
+- Validate required fields, unknown fields, JSON types, complete-value enums,
+  `items`, and nested `properties` against the resolved tool schema.
+- Apply the same validation before local, MCP, and built-in dispatch.
+- Normalize strict-provider `null` placeholders back to omission or the
+  configured default before validation.
+  Optional nullable values must not leak into a tool whose source schema does
+  not accept `null`.
+- Return the existing invalid-arguments tool response without invoking the
+  target.
+- Add tests using invalid calls taken from the Cerebras, Google, and llama.cpp
+  fixtures.
+- Add tests proving MCP and built-in parameter overrides are enforced.

docs/ticket/063sw3q-exercise-recorded-provider-tool-rounds-through-a-production.md

@@ -0,0 +1,27 @@
+# Exercise recorded provider tool rounds through a production-shaped loop
+
+- **Status**: Todo
+- **Kind**: Chore
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+`TestRequest::tool_call_response` creates a fresh `ChatQuery` with no tools.
+Production keeps the full tool list on every streaming cycle and resets a forced
+choice to `Auto` after execution.
+
+The mismatch is visible in the fixtures: Ollama says no tool is defined after JP
+sends a result, and some Google forced-tool follow-ups contain no assistant
+message.
+These tests do not cover repeated tool calls or continued tool availability.
+
+Acceptance criteria:
+
+- Add a recorded-provider test path that retains tool definitions across the
+  post-result request, matching `run_turn_loop`.
+- Use a fake executor so the recorded response is produced by the same request,
+  execute, append-result, request cycle used in production.
+- Cover a second tool call after the first result.
+- Cover parallel calls where the provider supports them.
+- Assert that forced choice becomes `Auto` while tools remain declared.
+- Keep lower-level provider serialization tests where useful, but name them as
+  such.

docs/ticket/063swa4-make-provider-cassette-tests-assert-semantic-outcomes-and-co.md

@@ -0,0 +1,30 @@
+# Make provider cassette tests assert semantic outcomes and consumed interactions
+
+- **Status**: Todo
+- **Kind**: Chore
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+The provider VCR tests snapshot whatever happened but usually have no semantic
+assertion.
+`Vcr::cassette` also discards the `MockSet` returned by `playback_async`, so
+recorded interactions are not checked for exactly one use.
+Optional tool follow-ups can be silently skipped and leave stale cassette
+entries without failing.
+
+Current green examples include empty llama.cpp model details, malformed Ollama
+structured output, invalid tool arguments, and empty Google post-tool responses.
+
+Acceptance criteria:
+
+- Assert every cassette interaction is consumed exactly once during playback.
+- Fail on unexpected requests and unused recorded responses.
+- Require request-kind invariants: forced tool name, tool argument conformance,
+  structured schema conformance, nonempty or explicitly empty post-tool outcome,
+  requested model ID, and nonempty model lists.
+- Remove no-op assertion defaults where the request kind has a meaningful
+  contract.
+- Make skipped `ToolCallResponse` requests explicit in the test declaration
+  rather than controlled by a boolean that silently returns `None`.
+- Add a harness regression test with an intentionally unused second cassette
+  interaction.

docs/ticket/063swjq-wire-xct2cli-xml-fixtures-into-tests-and-anonymize-them.md

@@ -0,0 +1,21 @@
+# Wire xct2cli XML fixtures into tests and anonymize them
+
+- **Status**: Todo
+- **Kind**: Chore
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+`crates/contrib/xct2cli/tests/fixtures/sample-toc.xml` and `time-sample.xml`
+have no test consumer; the parser tests use inline XML.
+The files also retain a device name, device UUID, username-bearing absolute
+path, and process IDs while the README calls the exported fixtures safe to keep.
+
+Acceptance criteria:
+
+- Add tests that parse both checked-in fixture files through the production
+  entry points, or delete the files if the inline cases fully replace them.
+- Assert the meaningful TOC and time-sample business outcomes exactly.
+- Replace device names, UUIDs, user paths, process IDs, and addresses with
+  stable synthetic values unless a raw value is needed by the parser case.
+- Clarify the README: exported XML omits the trace bundle's environment secrets
+  but still requires anonymization before commit.

docs/ticket/063swjr-snapshot-the-full-agentic-shepherd-issue-rendering-fixture.md

@@ -0,0 +1,20 @@
+# Snapshot the full Agentic Shepherd issue rendering fixture
+
+- **Status**: Todo
+- **Kind**: Chore
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+`renders_full_issue_fixture` renders a complete user-visible Markdown document
+but checks selected substrings.
+Duplicated sections, broken ordering, leaked content, and formatting changes can
+pass.
+
+Acceptance criteria:
+
+- Replace the substring checks with one exact static output assertion or an
+  accepted snapshot.
+- Keep focused unit tests for individual rendering rules.
+- Normalize only genuinely unstable fields before comparison.
+- Verify the test fails when a section is duplicated, reordered, or rendered
+  with incorrect indentation.

docs/ticket/063swpx-make-provider-fixtures-deterministic-and-sanitize-recorded-m.md

@@ -0,0 +1,29 @@
+# Make provider fixtures deterministic and sanitize recorded machine data
+
+- **Status**: Todo
+- **Kind**: Chore
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+Several fixtures contain avoidable nondeterminism or local machine data:
+
+- The common tool prompt asks the model to provide arbitrary arguments.
+- `banana.jpg` depicts an apple and the test uses a substring assertion.
+- Every generic OpenAI fixture uses the same conversation timestamp and
+  prompt-cache key.
+- llama.cpp cassettes record `/Users/jean/...`.
+- Other recorded fixtures contain host names, UUIDs, process IDs, and absolute
+  paths.
+
+Acceptance criteria:
+
+- Replace the arbitrary tool prompt with separate deterministic scalar, array,
+  default, and nested-value cases.
+- Rename the image fixture to match its content and compare the normalized
+  answer exactly.
+- Give each cassette a fixed distinct conversation identity while keeping all
+  turns in one cassette on the same identity.
+- Extend fixture post-processing to normalize user paths, host names, device
+  UUIDs, process IDs, and timestamps when those values are not under test.
+- Add a fixture lint test that rejects common absolute home-directory patterns
+  and known secret header/query fields.

docs/ticket/063swpy-resolve-llama-cpp-model-details-when-request-aliases-differ.md

@@ -0,0 +1,26 @@
+# Resolve llama.cpp model details when request aliases differ from catalog IDs
+
+- **Status**: Todo
+- **Kind**: Bug
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+The llama.cpp fixture requests `llamacpp/qwen3.5:9b`, while `/v1/models` reports
+`unsloth/Qwen3.5-9B-GGUF`.
+`map_model` strips the vendor prefix to `Qwen3.5-9B-GGUF`, so
+`model_details("qwen3.5:9b")` misses the loaded model and returns empty details.
+The accepted snapshot loses the `/props` context window of 8192.
+
+The chat cassettes are also inconsistent with model discovery: chat responses
+identify a 35B model while `/v1/models` identifies a 9B model.
+
+Acceptance criteria:
+
+- Define and implement model identity matching for llama.cpp request aliases and
+  loaded catalog IDs.
+- Preserve the served context window when the configured request name differs
+  from the server model ID.
+- Avoid ambiguous fallback when a server exposes more than one model.
+- Re-record a self-consistent llama.cpp fixture corpus using one loaded model.
+- Make the model-details fixture assert the expected ID and context window
+  instead of accepting empty details.

docs/ticket/063swpz-reject-malformed-and-schema-invalid-structured-responses.md

@@ -0,0 +1,33 @@
+# Reject malformed and schema-invalid structured responses
+
+- **Status**: Todo
+- **Kind**: Bug
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+The Ollama structured-output cassette returns plain text despite receiving the
+title JSON schema.
+`EventBuilder` wraps the parse failure in `Value::String` and emits
+`ChatResponse::Structured`, so the background title task treats it as structured
+data, extracts no title, and returns success without updating the title.
+
+Evidence:
+
+- `crates/jp_llm/tests/fixtures/ollama/test_structured_output.yml`
+- `crates/jp_llm/tests/fixtures/ollama/test_structured_output.snap`
+- `crates/jp_llm/src/event_builder.rs`
+- `crates/jp_task/src/task/title_generator.rs`
+
+Acceptance criteria:
+
+- Malformed JSON cannot be represented as a successful
+  `ChatResponse::Structured`.
+- Parsed JSON is checked against the schema attached to the request.
+- A malformed or nonconforming response produces a typed failure that the retry
+  layer or caller can handle.
+- Background title generation reports the failure rather than silently
+  succeeding without a title.
+- Provider fixture tests assert schema conformance, including the existing
+  Ollama cassette.
+- Cover valid JSON with the wrong shape, wrong field types, and wrong array
+  cardinality.

docs/ticket/063swq0-assert-openai-prompt-cache-behavior-automatically.md

@@ -0,0 +1,23 @@
+# Assert OpenAI prompt-cache behavior automatically
+
+- **Status**: Todo
+- **Kind**: Chore
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+`test_gpt_5_6_prompt_cache_read_after_write` asks the person re-recording
+fixtures to inspect usage fields manually.
+The cassette currently contains nonzero cache writes followed by nonzero cached
+tokens, but the test cannot fail if either disappears because provider usage is
+discarded.
+
+Acceptance criteria:
+
+- Capture the relevant cache usage from recorded Responses API events in the
+  test path.
+- Assert nonzero cache writes on the first request and nonzero cached tokens on
+  the second.
+- Assert the stable prompt-cache key and expected breakpoints on both requests.
+- Keep usage out of the persisted conversation unless product behavior requires
+  it; a test-only observer is sufficient.
+- Add a negative harness test proving zero cache activity fails the cache test.

docs/ticket/063swq1-reject-malformed-streamed-tool-call-argument-json.md

@@ -0,0 +1,21 @@
+# Reject malformed streamed tool-call argument JSON
+
+- **Status**: Todo
+- **Kind**: Bug
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+`EventBuilder` converts malformed tool-call argument JSON into an empty map and
+emits a normal `ToolCallRequest` (`crates/jp_llm/src/event_builder.rs`).
+A truncated provider stream can therefore execute a no-argument tool or a tool
+whose defaults fill the missing fields.
+
+Acceptance criteria:
+
+- A tool-call buffer containing malformed JSON must not become a valid empty
+  argument map.
+- Surface a specific invalid-call response or stream error that can be recorded
+  and returned to the model.
+- Prove that the tool executor is not invoked.
+- Cover truncated JSON, a non-object top-level value, and valid `{}` separately.
+- Add a test that fails against the current fallback behavior.

docs/ticket/063swq2-preserve-ollama-provider-tool-call-ids.md

@@ -0,0 +1,24 @@
+# Preserve Ollama provider tool-call IDs
+
+- **Status**: Todo
+- **Kind**: Bug
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+Recorded Ollama responses contain provider-generated call IDs such as
+`call_77bh2121`, but JP discards them and persists synthetic IDs such as
+`run_me_2`.
+The pinned `ollama-rs` response type exposes only the function payload, so the
+ID is lost during deserialization.
+
+Acceptance criteria:
+
+- Preserve the provider ID when Ollama sends one.
+- Use a deterministic synthetic ID only for responses from older servers that
+  omit the field.
+- Keep parallel calls and repeated calls across turns distinct.
+- Round-trip the same ID through `ToolCallRequest`, `ToolCallResponse`, and the
+  next Ollama request.
+- Update the Ollama fixtures to assert the recorded provider IDs.
+- If this requires an `ollama-rs` change, pin the fixed revision and add an
+  upstream regression test.

docs/ticket/063swq3-replace-duplicated-full-provider-catalogs-with-focused-mappi.md

@@ -0,0 +1,26 @@
+# Replace duplicated full provider catalogs with focused mapping fixtures
+
+- **Status**: Todo
+- **Kind**: Chore
+- **Authors**: jp
+- **Date**: 2026-08-21
+
+Provider model-list fixtures are too large to review and several model-details
+cassettes duplicate the full catalog byte for byte.
+OpenRouter has two identical 29,699-line cassettes plus a 10,111-line model
+snapshot.
+
+Google, llama.cpp, and Ollama also duplicate their model-list cassette for
+`models` and `model_details` tests.
+
+Acceptance criteria:
+
+- Keep one recorded catalog smoke cassette per provider where live format
+  coverage is useful.
+- Test model mapping, sorting, deduplication, deprecation, capability inference,
+  and lookup with small curated static responses.
+- Reuse one cassette response for `models` and `model_details` when both
+  production methods call the same endpoint.
+- Replace full-catalog snapshots with exact assertions over selected
+  representative records and list invariants.
+- Keep a test proving an unknown requested model returns the intended fallback.