feat(chat): name tool calls by what they did

[incognitojam]

Note

Tool rows in the timeline said "Tool call" and printed raw JSON. They now say what the tool did, with a readable argument, consistently across Claude, Codex and ACP providers.

The problem

Every tool row's title came from a 7-value item-type table, one per adapter, so a Read, a Skill, a ToolSearch and a SendMessage all rendered as Tool call followed by their serialized input. The three adapters also disagreed with each other: the same action read "Command run" from Claude and "Ran command" from Codex in the same thread. TaskCreate classified as a file write, so it rendered as File change with an edit pencil.

The clients could not do better on their own: projectActivityPayload rebuilt payload.data from an allowlist that kept toolName only for mcp_tool_call, so for every other call the row genuinely did not know which tool ran.

None of this is a regression — titleForTool and summarizeToolRequest have been unchanged since the Claude adapter landed.

Before / after

Before	After

How it works

packages/shared/src/toolRowPresentation.ts resolves a row in layers — tool name, then item type, then the provider's own label — and returns a heading plus a typed argument that the caller formats. The timeline, the agents panel, the approval card and mobile all read from it, so the vocabulary converges instead of gaining a fourth dialect.

• Well-typed actions share one verb across providers: Ran command, Edited file / Edited 3 files, Viewed image.
• The unknown-tool bucket names itself: Read, Skill, ToolSearch, SendMessage, Monitor. New tools stay correct with no table entry.
• Three overrides where the tool name would mislead: Created task / Updated task (they arrive typed as file changes) and Saved memory / Updated memory.
• A title the provider chose for a specific call (an ACP tool's own name, Codex's server · tool) outranks anything derived here.

The projection now ships data.toolName plus a short, length-capped argument allowlist. Because the projection runs at read time over persisted payloads, existing threads rename retroactively.

Notes for review

• Titles, labels and details are never rewritten. They are the identity tool rows collapse on, and the server's toolLifecycleIdentity drops superseded rows using the same triple; changing one side without the other would silently stop update rows folding into their completions. Naming happens at render time only.
• Wire size. The argument allowlist is keyed and value-capped at 200 characters, with content, new_string, old_string, prompt and message bodies deliberately excluded; a test asserts a 50 KB edit cannot survive projection. MCP keeps its arbitrary argument names (an allowlist would blank every server's expanded row) and caps values instead — that path previously shipped input uncapped.
• Icon ordering fix. image_view is now tested before the file-change branch; a viewed image discovers a path, which was stamping a read-only row with an edit pencil.
• Deliberately untouched: classifyToolItemType. Its misclassification is corrected at display time so the adapter stays byte-identical to upstream. Filed Approval routing misclassifies TaskCreate as a file change #115 for the approval-routing half of that bug, which is not cosmetic, and Upstreamable: MCP tool arguments ship uncapped in thread payloads #116 for the uncapped MCP input, which is upstream's.

Verification

• vp test run across the touched packages: 1,035 passing, plus the full apps/server suite at 2,563.
• vp run -r typecheck clean; targeted lint and vp fmt --check clean.
• Replayed 4,000 real persisted rows through the new naming to confirm every category renames and Codex's already-good titles do not move.
• Screenshots above are a live client against seeded fixtures.

---

Written by an agent (Claude Code, claude-opus-5).