Compile time: replace serde Content buffering with JSON-value deserializers

[vorporeal]

Description

This is PR 3 of a stack that reduces the compile time of the warp crate. It is stacked on #15454.

Why

For enums that are internally tagged, adjacently tagged, or that mix tagged and untagged variants, serde's derive cannot deserialize in one pass. It generates buffering machinery (Content, ContentRefDeserializer, ContentDeserializer) that copies the input into a generic tree and replays it for each candidate variant. Fields with deserialize_with also each get a private __DeserializeWith wrapper struct. This machinery is generic over the input lifetime and the error type, so it monomorphizes widely and it pulls large dependent trees (for example BlockContext and its field types) into the instantiations.

The affected types are only deserialized from JSON (serde_json::from_str / from_slice / from_value). I verified each call site before this change. Because the input format is known, a hand-written Deserialize can buffer into serde_json::Value once and then parse the payload directly, with no generic replay machinery.

What changed

Serialization derives are unchanged everywhere. Only deserialization is hand-written, and it accepts the exact same wire format as before:

• DProtoHook (app/src/terminal/model/ansi/dcs_hooks.rs): a small raw struct captures the hook tag and the value payload, and a 13-arm match parses the payload. The SourcedRcFileForWarp shape is frozen because the payload literal ships inside user RC files; a test pins it.
• BootstrappedValue (same file): a raw twin struct replaces the per-field deserialize_with wrappers (empty_string_is_none, float parsing, shell-option splitting). The conversions now run in one non-generic place. A RawBootstrappedField helper keeps the old missing-versus-present semantics for optional fields.
• Artifact (app/src/ai/artifacts/mod.rs): an ArtifactEnvelope struct replaces the ArtifactHelper mirror enum for the adjacently tagged (artifact_type / data) format. The PULL_REQUEST arm still derives repo and number from the URL.
• AIAgentContext and AIAgentAttachment (app/src/ai/agent/mod.rs): these mix externally tagged variants with an untagged Block(BlockContext) variant, which is the most expensive derive pattern. A private derived twin enum handles the tagged variants, and the impl falls back to BlockContext when the tagged parse fails — the same fallthrough order the derive used. This also removes BlockContext's large deserializer tree from the ContentRefDeserializer instantiations.

Alternatives considered

• Re-tagging the enums (for example, making Block a tagged variant) would remove the machinery with less code, but it changes the wire format. These payloads live in persisted conversations, RC-file snippets, and server messages, so the format must not change.
• Keeping the derives was the status quo; the buffering machinery accounted for ~167K IR lines at the base of this branch.
• CodeSource was on the candidate list but is a plain externally tagged enum. Its derive does not generate the buffering machinery, so it was left alone.
• PrecmdValue/PromptMetadata use #[serde(flatten)], which generates similar buffering. That is left as follow-up work because the flatten structure is more intricate.

Measured effect (vs. PR 2, Apple M5 Pro, rustc 1.92.0, dev profile)

• cargo llvm-lines -p warp --lib: 17,022,929 → 16,966,243 lines (−0.33%); 573,201 → 571,576 copies.
• Serde buffering machinery (Content* / __DeserializeWith symbols): 167,090 → 89,078 IR lines (−47%). The remainder comes from other crates' types that still use these patterns.
• Per type (IR lines): DProtoHook 24,664 → 6,773; BootstrappedValue 13,299 → 7,158; Artifact symbols 39,443 → 18,148 (ArtifactHelper 23,101 → 0); AIAgentContext 21,473 → 18,172; AIAgentAttachment 15,841 → 13,418.
• Macro expansion (-Zmacro-stats): 32,172,116 → 31,907,058 bytes (−0.8%).
• Clean warp lib rebuild (deps cached, CARGO_INCREMENTAL=0, two runs): 71.4 s / 71.1 s → 69.2 s / 69.1 s.
• Incremental touch rebuild (two runs): 15.4 s / 14.8 s → 14.7 s / 14.2 s.

The wall-clock gain is small (~2 s clean); the main value is removing the widest remaining serde instantiations in the crate and establishing the pattern for the flatten-based types.

Linked Issue

No linked issue. This PR comes from a compile-time investigation; see the conversation link below.

Testing

• New round-trip and frozen-format unit tests: app/src/terminal/model/ansi/dcs_hooks_tests.rs (new file), plus new tests in app/src/ai/artifacts/mod_tests.rs and app/src/ai/agent/mod_tests.rs. They serialize each variant with the unchanged derived Serialize and assert the hand-written Deserialize returns an equal value, and they pin raw JSON literals for the frozen formats (RC-file hook, artifact envelope, pull-request number coercion).
• cargo nextest run -p warp -E 'test(/ai::|conversation/)': 2372 passed.
• ./script/format and all three presubmit clippy passes are clean.
• Not manually run with ./script/run; behavior is covered by the deserialization tests above, and the wire formats are unchanged.

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

Conversation: https://staging.warp.dev/conversation/024e797d-1d35-4c62-8394-abf93f7ddb0e

Co-Authored-By: Warp agent@warp.dev

CHANGELOG-NONE

[warp-for-oss]

@vorporeal

I'm starting a first review of this pull request.

You can view the conversation on Warp.

I completed the review and no human review was requested for this pull request.

Comment /oz-review on this pull request to retrigger a review (up to 3 times on the same pull request).

Powered by Oz

[warp-for-oss]

Overview

This PR replaces several serde derive-generated deserializers with handwritten JSON-value-based deserializers while preserving the existing serialization derives and wire formats. The implementation is broadly scoped to DCS hooks, bootstrapped shell payloads, AI agent context/attachments, and artifacts, with unit tests added around the main formats.

Concerns

• The new handwritten dispatch paths are not fully covered by the added tests. In particular, the DProtoHook tag match omits round-trip coverage for several variants, and the AIAgentContext tagged-variant table omits ExecutionEnvironment even though both paths are now maintained manually.

Verdict

Found: 0 critical, 2 important, 0 suggestions

Request changes

Comment /oz-review on this pull request to retrigger a review (up to 3 times on the same pull request).

Powered by Oz

[warp-for-oss]

⚠️ [IMPORTANT] This new handwritten tag match needs coverage for every DProtoHook variant; the added tests only exercise a subset, so tag/value-mapping regressions for the remaining variants would no longer be caught by serde derive.

[warp-for-oss]

⚠️ [IMPORTANT] This table omits ExecutionEnvironment, even though the PR adds a handwritten AIAgentContextTagged::ExecutionEnvironment arm; add that case so every tagged variant is covered.