fix(v2): show read and search tool activity accurately

[Yash-Singh1]

V2 loses or mislabels some provider tool calls. Read and search activity can appear as a generic --- file, a write, or a tool with no useful path. Expanded read entries can show file contents or projection metadata instead of the request and result.

Preserve structured tool inputs through the ACP, Claude, Cursor, and OpenCode adapters. Classify read and search calls from their tool identity and arguments, then show clear titles, paths, and expanded details in web and mobile. Keep grouped tool rows aligned with the corrected labels. OpenCode codesearch now emits a code search turn item, while websearch remains a web search.

Before and after

These screenshots use a seeded local Cursor activity thread. The web and mobile captures cover the visible group summary, tool classification, and expanded read and search details.

Web group summary

Before	After

Web expanded tool details

Before	After

Mobile tool rows

Before	After

Mobile expanded read

Before	After

Verification

• 645 focused tests passed across the shared tool classifier, client presentation, web and mobile activity, and provider adapters. The follow-up OpenCode adapter test run passed all 45 tests, including an emitted codesearch and websearch pair.
• Targeted server, web, and mobile typechecks and lint passed. Formatting and git diff --check passed.
• Web and iOS UI flows were inspected against the seeded thread after the icon change. The replacement after screenshots show the eye icon on read rows and expanded details.

Model: GPT-6. Harness: Codex.

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Codex	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Codex	Live turn WebSocket wire	—	1.1 KiB	—	2.0 KiB	✅
Codex	Live turn WebSocket decoded	—	20.4 KiB	—	29.3 KiB	✅
Codex	Live turn messages	—	1	—	8	✅
Claude	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Claude	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Claude	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Claude	Live turn WebSocket decoded	—	20.8 KiB	—	29.3 KiB	✅
Claude	Live turn messages	—	2	—	8	✅

Baseline: unavailable · PR result: 6478aa4 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 106.1 KiB
• Claude decoded thread snapshot: 106.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a broad cross-layer change to provider projections, shared activity classification, and web/mobile rendering, including a changed internal representation for read operations. Although well covered by tests and focused on correcting activity display, the breadth and existing-path behavior changes warrant human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[juliusmarminge]

Thanks, the read/search labels are a real improvement, and the change is still needed on the current V2 branch. None of the ~90 commits since your base touched read/search classification: Claude Grep still renders as "Grep" with a wrench, and Cursor/Grok reads still show as "Searched files".

I merged the current t3code/codex-turn-mapping head into a scratch copy of this branch. It merges cleanly, and I found no semantic conflicts with what landed since (#13725 folds, #13790 stopped rows, #13806 Cursor stopped status, the per-turn item-ordinal maps). On that merge, tsc is clean in all six packages, the PR's own tests pass (web 180, mobile 75, client-runtime 93, shared 8, adapters 339), the replay suite is 96/96, and knip is clean. The only textual conflict is with the open #13140, in workEntryDisplayLabel.

Two things need fixing before merge (inline):

1. Mobile approval rows regress. A file-read approval shows "Read file" instead of its prompt and can't be expanded, and command/file-change approvals summarize as "Ran 1 command, changed 1 file" although nothing ran.
2. OpenCode codesearch is not a file search. It was a web code-context tool (query + tokens, no path), OpenCode removed it upstream in May, and 1.18.32 doesn't ship it. The test invents a path for it.

Smaller, non-blocking notes are inline too: two pieces of dead code, a lone web search losing its "Web search" label, multi-line approval prompts, and MCP tool names.

On tests: nothing in the replay suite asserts the new "Read …" / "Searched …" titles. Adding title assertions to the tool_call_read_only Claude and Cursor replay outputs would prove the change better than the new adapter unit tests, and OpenCode read/grep have no replay coverage at all. The ACP mock-agent frames for late locations and empty rawInput resends don't appear in any of the 23 recorded Grok/registry transcripts, so I'd drop them unless a real transcript shows them.

Not verified: live providers and a real web/mobile UI pass.