test: cover plan, tool_call, ext, and protocol_level regression risks

[cursor]

Risky behavior now covered

This PR closes meaningful regression gaps in four files that previously had zero direct unit tests (only indirect coverage through v1↔v2 conversion round-trips). All four sit on hot paths in the protocol:

plan.rs (v1 + v2)

• Plan::entries resilient deserialization (introduced in feat(rust-only): better tolerate malformed optional fields in deserialization agentclientprotocol/agent-client-protocol#1006) had no direct test on the protocol-level type. New tests prove:
  • A single malformed PlanEntry does not poison a whole plan update (DefaultOnError<VecSkipError<_, SkipListener>>).
  • Outer-shape errors on entries (null / string / object / number) collapse to an empty vec.
  • The entries key is still required if missing (intentional, documented in the test).
  • PlanEntryPriority / PlanEntryStatus snake_case wire format is locked in.
  • _meta round-trips when set and is omitted when unset.

tool_call.rs (v1 + v2)

The largest blast-radius file in the schema crate (674 lines, 0 tests). New coverage:

• ToolKind::Other / ToolCallStatus::Pending defaults are skipped from the wire (is_default).
• ToolKind #[serde(other)] decodes unknown variants to Other — protects forward compatibility with future spec revisions.
• ToolCallStatus has no other catch-all (locked in via the update-fields test where DefaultOnError collapses unknown statuses to None).
• ToolCall::update() merge semantics: only present fields overwrite; collections (content/locations) are replaced, not extended — this is easy to silently break.
• TryFrom<ToolCallUpdate> for ToolCall requires title and surfaces a stable, actionable error message + InvalidParams code.
• From<ToolCall> for ToolCallUpdate preserves every field, including raw_input / raw_output — the ToolCall → ToolCallUpdate → ToolCall round-trip is asserted.
• ToolCallUpdate flattens fields next to toolCallId on the wire.
• ToolCallContent discriminator is type with snake_case content / diff / terminal.
• Resilient deserialization on content / locations (mirrors Plan::entries).

ext.rs (v1 + v2)

• ExtRequest::new preserves the method name verbatim from any Into<Arc<str>> and reuses an existing Arc<str> without reallocating. Protects the fix from fix(rpc): preserve '_' prefix for extension methods and reject empty ext agentclientprotocol/agent-client-protocol#883 at the type level even after the Side/decode_request machinery moved into the SDK in feat(rust-only): Remove unused RPC message schema types (schema.json unchanged) agentclientprotocol/agent-client-protocol#1009.
• ExtRequest, ExtNotification, and ExtResponse all serialize as the raw params payload only — the method name must never leak into the JSON, otherwise custom RPC handlers would receive unexpected fields.

protocol_level.rs (v1 + v2)

• CANCEL_REQUEST_METHOD_NAME stays "$/cancel_request" and is re-exported via PROTOCOL_LEVEL_METHOD_NAMES. Renaming this is a wire-format break.
• ProtocolLevelNotification::method() returns the right name for the cancel variant.
• CancelRequestNotification round-trips with and without _meta and for every RequestId shape (null / number / string).

Test files added/updated

• src/v1/plan.rs — new tests module (5 tests)
• src/v2/plan.rs — new tests module (5 tests)
• src/v1/tool_call.rs — new tests module (16 tests)
• src/v2/tool_call.rs — new tests module (16 tests)
• src/v1/ext.rs — new tests module (4 tests)
• src/v2/ext.rs — new tests module (4 tests)
• src/v1/protocol_level.rs — new tests module (5 tests)
• src/v2/protocol_level.rs — new tests module (5 tests)

Why these tests materially reduce regression risk

All four files contain logic that is invisible at compile time but corrupts wire-protocol behavior if broken:

1. Resilient deserialization (DefaultOnError<VecSkipError<_, SkipListener>>) is purely a serde annotation. Dropping or rewriting these attributes silently flips the schema from forgiving to strict, breaking any client that sends a slightly-malformed list element — exactly the case the pattern was introduced to handle (feat(rust-only): better tolerate malformed optional fields in deserialization agentclientprotocol/agent-client-protocol#1006).
2. ToolCall::update merge semantics are easy to "simplify" in ways that change behavior (e.g. extending vs replacing collections, treating None as clear-the-field). A regression here corrupts every tool-call stream.
3. TryFrom<ToolCallUpdate> requiring title is the contract every agent relies on when reconstructing a missed tool-call event from a later update.
4. Extension RPC encoding (#[serde(skip)] on method, #[serde(transparent)] on the struct) is invisible to type-checking. Any change that lets the method name leak into params breaks downstream extension handlers — exactly the regression fix(rpc): preserve '_' prefix for extension methods and reject empty ext agentclientprotocol/agent-client-protocol#883 fixed.
5. $/cancel_request method name is a wire-format string with no compile-time validation. The test makes accidental renames a test failure rather than a runtime mystery.

Validation

• cargo test --all-features: 320 passed (was 260 before this PR — 60 new tests).
• cargo test --lib --no-default-features: 66 passed.
• cargo test --lib --features unstable_protocol_v2: 124 passed.
• cargo clippy --all-features --all-targets -- -D warnings: clean.
• cargo fmt --check: clean.
• npm run spellcheck: clean.
• npx prettier --check: only flags AGENTS.md, which is pre-existing and unrelated to this PR (introduced in 2691825).

Tests are deterministic, hermetic (no I/O, no time, no thread-locals), and follow the existing #[cfg(test)] mod tests + serde_json::to_value / from_value round-trip convention used elsewhere in the crate.

  

---

Summary by cubic

Add unit tests for v1 and v2 protocol hot paths (plan, tool_call, ext, protocol_level) to lock in wire format and update behavior, reducing regression risk.
Covers resilient list parsing, default-skipping for ToolKind/ToolCallStatus, ToolCall::update replacing collections, TryFrom<ToolCallUpdate> requiring title with InvalidParams, From<ToolCall> preserving fields, transparent extension RPC payloads, and the stable $/cancel_request method with RequestId round-trips; also tidies AGENTS.md key commands table formatting.

^{Written for commit 3c07b86. Summary will update on new commits.}

Note

Add regression test coverage for plan, tool_call, ext, and protocol_level modules

• Adds #[cfg(test)] modules to plan, tool_call, ext, and protocol_level in both v1 and v2, covering serialization, deserialization, round-trips, and edge cases.
• Tests verify correct snake_case enum variants, omission of default/absent fields, skipping of malformed collection items, and stable error messages for TryFrom conversions.
• Protocol-level tests confirm CANCEL_REQUEST_METHOD_NAME constant value and RequestId shape handling (null, number, string).
• Tool call tests cover update semantics (selective field overwrite, collection replacement) and From/TryFrom conversion correctness.

^{Macroscope summarized 3c07b86.}

---

Note

Low Risk
Test-only changes with no runtime or schema logic edits; risk is limited to CI time and possible future test maintenance if serde attributes change intentionally.

Overview
Adds ~60 new unit tests in v1 and v2 for plan, tool_call, ext, and protocol_level—modules that previously had little or no direct coverage. Production types are unchanged; tests assert serde wire behavior and in-memory update/conversion rules.

Coverage locks in forgiving list parsing on Plan::entries and ToolCall content/locations, default omission for ToolKind/ToolCallStatus, ToolCall::update (partial field merge, collections replaced not extended), TryFrom<ToolCallUpdate> requiring title with InvalidParams, extension types serializing params only (method name must not leak), and the stable $/cancel_request method plus CancelRequestNotification / RequestId round-trips.

AGENTS.md only reformats the key-commands table for readability.

^{Reviewed by Cursor Bugbot for commit 3c07b86. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_91a5f40e-c0f7-492d-9ce3-fa8fb215f183)

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_2902aa92-200b-4b8d-bdc4-3d4cb34be7b3)

[Copilot]

Pull request overview

This PR adds direct unit-test coverage for several protocol “hot path” schema modules in both v1 and v2, primarily locking in serde wire-format behavior and update/merge semantics to reduce regression risk.

Changes:

• Add comprehensive #[cfg(test)] modules for plan, tool_call, ext, and protocol_level in both src/v1/ and src/v2/.
• Assert resilience behaviors (DefaultOnError + VecSkipError), default-skipping, flattening, and conversion/update invariants.
• Reformat the “Key commands” table in AGENTS.md.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/v1/plan.rs	Adds tests for resilient entries parsing, snake_case enums, and _meta serialization behavior.
src/v2/plan.rs	Mirrors v1 plan tests for v2 types/wire format.
src/v1/tool_call.rs	Adds extensive tests covering default omission, update semantics, flattening, resilient list parsing, and conversions.
src/v2/tool_call.rs	Mirrors v1 tool_call tests for v2 types/wire format and invariants.
src/v1/ext.rs	Adds tests ensuring extension payloads serialize transparently and method names don’t leak.
src/v2/ext.rs	Mirrors v1 ext tests for v2 types/wire format.
src/v1/protocol_level.rs	Adds cancel-request wire-format tests (currently missing required feature gating).
src/v2/protocol_level.rs	Mirrors v1 protocol_level tests (currently missing required feature gating).
AGENTS.md	Reflows the key-commands table for readability.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.