Commit Graph
2 Commits
Author SHA1 Message Date
8ad7d78a70 fix: parameter coercion and validation for oneOf/anyOf/allOf schemas (#1397)
* fix: parameter coercion and validation for oneOf/anyOf/allOf schemas

WASM extension tools with multi-action schemas (e.g. github extension)
fail when the LLM passes numeric parameters as strings because the
coercion layer skips JSON Schema combinators. This causes serde
deserialization errors like `invalid type: string "100", expected u32`.

Add discriminated-union resolution to the coercion layer: for oneOf/anyOf,
match the active variant by const or single-element enum discriminators;
for allOf, merge all variants' properties. Also propagate combinator
awareness to schema validators, WASM wrapper helpers, and tool discovery
so they no longer reject or ignore valid combinator-based schemas.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* test: add e2e tests for oneOf discriminated union parameter coercion

Add three end-to-end tests using a fixture tool that mirrors the github
WASM tool's oneOf schema with #[serde(tag = "action")] deserialization.
Each test sends string-typed numeric/boolean params through the full
agent loop, verifying that coercion resolves them before serde runs:

- list_issues: limit "100" → 100 (integer in oneOf variant)
- get_issue: issue_number "42" → 42 (integer in different variant)
- create_pull_request: draft "true" → true (boolean in variant)

Without the coercion fix these fail with:
  invalid type: string "100", expected u32

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* test: add real WASM github tool e2e tests with HTTP interception

Load the actual compiled github WASM binary, send params with string-typed
numbers through the coercion layer, and verify the WASM tool constructs
correct HTTP API calls via a new HTTP interceptor in the WASM wrapper.

Changes:
- Add `http_interceptor` field to `StoreData` and `WasmToolWrapper` so
  WASM tool HTTP requests can be captured/mocked in tests
- Make `prepare_tool_params` and `coercion` module public for integration tests
- Add 3 e2e tests loading the real github WASM binary:
  - list_issues: `limit: "50"` → URL contains `per_page=50`
  - get_issue: `issue_number: "42"` → URL contains `/issues/42`
  - list_pull_requests: `limit: "25"` → URL contains `per_page=25`

Tests gracefully skip if the WASM binary isn't compiled.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* refactor: simplify WASM e2e tests to use TestRig with with_wasm_tool()

Replace the manual WasmToolWrapper construction with TestRig integration:

- Add `with_wasm_tool(name, wasm_path, capabilities_path)` to TestRigBuilder
  that loads real WASM binaries and wires the shared HTTP interceptor
- Build the HTTP interceptor before tool registration so it can be shared
  between AgentDeps and WASM tool wrappers
- Rewrite github WASM e2e tests to use the standard trace pattern:
  TraceLlm sends tool calls with string params, http_exchanges specify
  expected outgoing requests and canned responses

The test code is now identical to other trace-based e2e tests — no custom
interceptors or manual WASM construction needed.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* fix: address review comments on combinator schema support

- Validate `has_combinators` checks array type (`.as_array().is_some()`)
  instead of bare `.is_some()` to reject malformed `{ "oneOf": {} }`
- Validate top-level `required` keys against merged combinator variant
  properties when no top-level `properties` exists (both validators)
- Deduplicate oneOf/anyOf handling into single loop in coercion.rs
- Revert `pub mod coercion` to private; only re-export `prepare_tool_params`
- Call `after_response` on interceptor after real HTTP when `before_request`
  returns None (recording mode correctness)
- Fix formatting (CI failure)

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* fix: address second round of review comments

- Fix headers deserialization bug: deserialize resp.headers_json as
  HashMap<String, String> then convert to Vec, not directly as Vec
- Sort interceptor headers for deterministic trace fixtures
- Update after_response comment: RecordingHttpInterceptor does exercise
  this path (returns None from before_request)
- Mark WASM tests #[ignore] instead of silent skip — avoids false-green
  CI while keeping them runnable with --ignored
- Fix with_wasm_tool signature: Option<PathBuf> instead of
  Option<impl Into<PathBuf>> which doesn't compile in nested position
- Fix with_wasm_tool doc comment to match actual behavior
- Revert prepare_tool_params to pub(crate) — no longer needed publicly

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* fix: coerce empty strings to null for optional tool parameters

LLMs often send "" instead of null/omitting optional parameters, causing
parse errors in tools that expect typed values (e.g., timezone, schedule).

PR #1127 fixed this per-field in the time tool. This commit adds
dispatcher-level coercion so all tools benefit:

- Non-required properties with value "" are coerced to null at the
  object level (based on the schema's `required` array)
- Explicitly nullable schemas (`type: ["string", "null"]`) coerce ""
  to null in the per-value coercion path
- Required string-only fields keep "" unchanged

Closes #755

Co-Authored-By: spiritj <[email protected]>
Co-Authored-By: Xing Ji <[email protected]>
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* feat: complete coercion coverage for $ref, nested combinators, and additionalProperties

Close remaining coercion gaps so 3rd-party tools (MCP servers, complex
WASM tools) work correctly:

- $ref resolution: inline all #/definitions/<name> and #/$defs/<name>
  references in a pre-pass before coercion, with depth limit (16) for
  circular ref safety
- Nested combinators: resolve_effective_properties now recurses into
  variants that themselves contain allOf/oneOf/anyOf (depth limit 4)
- additionalProperties inheritance: check allOf variants and matched
  oneOf/anyOf variant for additionalProperties schemas

New tests:
- resolves_ref_and_coerces_referenced_properties
- resolves_nested_refs_in_oneof_variants
- coerces_nested_combinators_allof_containing_oneof
- coerces_array_items_with_oneof_discriminator
- circular_ref_does_not_infinite_loop

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* fix: address third round of review comments

- Validators: tighten has_combinators to require at least one object-typed
  variant (has type:"object" or properties), rejecting non-object combinator
  schemas like { "oneOf": [{"type":"integer"}] }
- Empty-string coercion: only coerce "" → null when schema allows null or
  doesn't allow string; pure type:"string" fields keep "" as meaningful
- Fix comment: "coerce to null" → "return unchanged" for empty strings
  with no type match (code returns None, not null)
- Redact credentials before passing to after_response interceptor to
  prevent secret leakage into recorded trace files
- Switch to tokio::fs::read for async WASM binary loading in test rig
- Add doc comment explaining soft URL check in WASM e2e tests

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* ci: retrigger after staging merge [skip-regression-check]

* fix: merge staging, report non-array combinator values as errors

Merge latest staging to fix CI (missing fallback_deliverable field).
Add explicit error reporting when oneOf/anyOf/allOf values are not
arrays in both strict and lenient validators.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* fix: recurse into combinator variants that have properties but no explicit type

Both validators only recursed into variants with `type: "object"`,
missing variants that define `properties` without an explicit type
(common in allOf patterns). Now recurse when variant has either.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>
Co-authored-by: spiritj <[email protected]>
Co-authored-by: Xing Ji <[email protected]>
2026-03-21 12:41:46 -07:00
Henry ParkandGitHub c79754df28 Fix schema-guided tool parameter coercion (#1143)
* Fix schema-guided tool parameter coercion

* Fix CI checks for coercion regression tests

* Finish panic-scan annotations

* Avoid redundant worker param preparation

* Keep panic-scan annotations rustfmt-stable

* Handle nullable WASM schema review feedback

* Address param coercion review notes
2026-03-14 16:27:18 -07:00