[yan-ad]

Summary

Split src/acp/service.rs into a private service/ module tree without changing behavior. The original file is 5,549 lines at the updated #64 head; service/mod.rs is now 190 lines and every extracted file is now at most 374 lines, including adjacent test files.

Depends on #64. This branch starts from #64's current head, including the compaction-mode tooltip follow-up. Please merge #64 first; until then GitHub's diff against main also includes that prerequisite. The refactor commits are above #64; the latest follow-up further narrows module responsibilities.

Closes #65.

Module boundaries

• mod.rs: service/session state, capabilities, shared cancellation and error helpers.
• sessions.rs: new/load/resume/fork/list/close/delete/cancel and persisted attachment handling.
• prompt.rs and stream.rs: prompt preparation, execution, stream events, transcript recording, and turn finalization.
• compaction.rs: explicit/automatic compaction, summary generation, transcript replacement, and notifications.
• usage.rs: provider context measurements, live usage, cumulative billing, and ACP usage updates.
• config.rs: MCP config, modes, model/reasoning selection, limits, and picker descriptions.
• commands.rs: command advertisement, expansion, and side questions.
• attachments.rs: input conversion, image/audio storage, and rollback guard.
• permissions.rs, questions.rs, and terminals.rs: permission previews/grants, elicitation, and client-terminal bridging/cleanup.
• notifications.rs: text/tool projections, locations/diffs/media, and transcript replay.
• test_support.rs: shared test fixtures, compiled only under cfg(test).

Tests move alongside their owners; stdio integration tests stay in tests/acp_stdio.rs. Production imports are explicit, child modules remain private, sibling-used helpers use pub(super), and the server-facing service::AcpService API is unchanged.

Behavior preservation

No changes to function bodies, public signatures, lock scopes, persistence ordering, cancellation handling, compaction thresholds, context-versus-billing semantics, or terminal cleanup. Compared all 202 functions/methods/types/fixtures/tests against #64 using Rust syntax parsing: production function and method bodies are token-identical; moved test bodies differ only in formatting-only trailing commas. Only internal type/field visibility and import/module placement change.

Validation

• cargo test acp:: --bin crabcode: 76 passed.
• cargo test session::compaction:: --bin crabcode: 21 passed.
• cargo test --test acp_stdio: 8 passed, including automatic compaction and serialized mode descriptions.
• cargo fmt --all -- --check and git diff --check: passed.
• Serial full-unit run before the tooltip follow-up: 1,850 passed, 3 ignored, one existing unrelated failure in ui::components::chat::tests::test_edit_tool_renders_codex_style_diff_summary. It embeds /Users/carlo/Desktop/Projects/crabcode and expects a basename on other checkouts; the same failure is documented in fix(acp): trigger automatic compaction from provider context usage #64.

Synced to yan-ad/crabcode:yan-flavor (refactor commit ac52d96, latest smaller-module sync head ca801fe), preserving that branch's older lifecycle implementation and concurrent install/upgrade changes. Flavor validation: 61 ACP tests, 21 compaction tests, and 4 stdio tests pass. Five consecutive parallel stdio runs pass after test-only fixtures eliminated models.dev and Ollama discovery dependencies. All 181 flavor functions/types/fixtures/tests are preserved, with function and test bodies token-identical to the flavor baseline.

Follow-up: smaller source chunks

The smaller-module follow-up (3189129) separates models, modes, and MCP config; tool content, tool locations, and replay; attachment storage; command expansion; permission previews; context accounting; transcript recording; and compaction summary construction. Long lifecycle tests are in adjacent service/tests/ files, still compiled as child modules of their owners.

This is source organization, not Rust minification or a change to model context size. Related execution flows remain intact rather than splitting functions to satisfy an arbitrary line count. Upstream's largest file is now 374 lines (previously 622); the shared core remains 202 lines. The adapted flavor retains its existing 521-line prompt flow, with all other files at most 349 lines.

Validation after this follow-up: upstream 76 ACP tests, 21 compaction tests, and 8 stdio tests; flavor 61 ACP tests, 21 compaction tests, and 4 stdio tests. Both stdio suites pass in three consecutive complete runs. Formatting and whitespace checks pass. A separate upstream test-only follow-up (bbea752) isolates mock compaction from runtime-provider discovery without altering authentication tests.

[Blankeos]

refactor/acp-service-modules → main
Model: claude-sonnet-5-5, effort unknown

Confidence: 4/5

Pure file split of src/acp/service.rs into a private service/ tree. I could not find any behaviour change, and build, fmt and the ACP tests are green. Held at 4 because it is stacked on open #64 and will collide with three other open PRs.
Checks run:

• cargo fmt --check: clean.
• cargo build: ok, 0 warnings.
• cargo test --bin crabcode acp::: 76 passed (3 runs).
• cargo test --test acp_stdio: 8 passed (2 runs). One run under heavy parallel load failed login_with_piped_stdin_fails_promptly..., a timing test this PR does not touch. It passed on rerun.
• Whole cargo test --bin crabcode, serial (--test-threads=1): 1851 passed, 1 failed, 3 ignored. The one failure is ui::components::chat::tests::test_edit_tool_renders_codex_style_diff_summary, a hard-coded /Users/carlo/... path that fails on any other checkout. It is not caused by this PR and the PR notes it.
• Whole cargo test in parallel: 3 to 6 flaky failures per run, varying run to run. They are model::discovery, model::catalog, command::handlers and ui::chat finalize-TPS tests, all in files this PR does not touch, and they look like shared global state or timing. One run also failed acp::service::stream::...cancelled_permission..., which passed alone and in 3 later acp:: runs. This was on a loaded machine, so I am not calling it a regression, but the new service/tests/stream.rs is timing-sensitive.
• Behaviour-preservation check of my own: I compared the old service.rs (at fix(acp): trigger automatic compaction from provider context usage #64 head 4a4f25a) against the concatenated new files, ignoring whitespace, pub(super)/pub visibility and trailing commas. Differences are use/mod lines and re-indentation of test code. No production logic change found.
• Ran the built binary for real: crabcode acp over stdio with a temp XDG_* and CRABCODE_TEST_MODE=1. initialize returned the same capabilities (loadSession, list/delete/fork/resume/close), session/new returned modes, config options and available_commands_update, and session/list returned [].
• Clippy: 9 findings fall in src/acp/service/*, all in moved code (sort_by_key, MutexGuard across await, collapsible if, 9 args, assert_eq! with a literal bool). None are new logic. The repo-wide 176 errors are the known pre-existing set.
• No UI or TUI change, so no screenshots.
  Confidence bumps: none left that I can do. Re-checking after fix(acp): trigger automatic compaction from provider context usage #64 lands would be cheap, because the diff against main then shrinks to the real refactor (no score change expected).

Summary

Splits the 5.5k-line src/acp/service.rs into 25 private modules plus 3 test files, with the largest file at 374 lines. The public service::AcpService API is unchanged. The commits on top of #64 only move code, narrow visibility to pub(super), and add test fixtures (test_support.rs, a mock auth.json in tests/acp_stdio.rs).

Notes:

• The PR is stacked on fix(acp): trigger automatic compaction from provider context usage #64 (d1ae121, 4a4f25a). Against main the diff also contains fix(acp): trigger automatic compaction from provider context usage #64's behaviour changes in session/compaction.rs, tests/acp_stdio.rs and the compaction code. Review those in fix(acp): trigger automatic compaction from provider context usage #64, and merge fix(acp): trigger automatic compaction from provider context usage #64 first.
• The PR text says service/mod.rs is 190 lines. It is 202. Cosmetic.
• This is a big reshuffle (+6370/-4993) for a file that three other open PRs are editing (see below).

Map

flowchart TD
  M[service/mod.rs: AcpService, AcpSessions] --> S[sessions.rs]
  M --> P[prompt.rs] --> ST[stream.rs] --> T[transcript.rs]
  P --> A[attachments.rs + attachment_storage.rs]
  M --> C[compaction.rs + compaction_summary.rs] --> U[usage.rs + context.rs]
  M --> CF[config.rs + models.rs + modes.rs + mcp_config.rs]
  M --> CM[commands.rs + command_expansion.rs]
  M --> PQ[permissions.rs + permission_preview.rs + questions.rs + terminals.rs]
  M --> N[notifications.rs + tool_content.rs + tool_locations.rs + replay.rs]
  M --> TS[test_support.rs + tests/*]

Loading

Start at service/mod.rs, then prompt.rs and stream.rs.

Regressions

• None found in production code.
• Merge-conflict risk with other open PRs: fix(acp): keep sessions usable after provider failures #72, feat(auth): add Google Antigravity browser OAuth #70 and feat(cli): add --auto permission mode #69 all edit src/acp/service.rs, which this PR deletes. Each will hit a modify/delete conflict and need its hunks re-applied by hand into the new files. fix(acp): trigger automatic compaction from provider context usage #64 is the base of this PR. git merge-tree of this branch against current main is clean, but that is only because fix(acp): trigger automatic compaction from provider context usage #64's changes ride along.
• Timing-sensitive tests: see Checks run.

Migrations

None. No schema, config or wire change.

Before merging

• Merge fix(acp): trigger automatic compaction from provider context usage #64 first (this PR contains its commits).
• Decide merge order against fix(acp): keep sessions usable after provider failures #72, feat(auth): add Google Antigravity browser OAuth #70 and feat(cli): add --auto permission mode #69. Merge this one last or first, and rebase the others onto it.
• Re-run cargo test --test acp_stdio and cargo test --bin crabcode acp:: after the rebase onto main.
• Optional: fix the PR text (mod.rs is 202 lines, not 190).