feat(chat): make MCP and tool rows legible end to end (parser + UI, flag-gated)

[ElbertePlinio]

The backend and frontend halves of #362 and #365, consolidated. #390 merged into this branch, so it ships as one gated unit rather than leaving the parser emitting events nothing renders.

Nothing non-Bash ever completed

if self.tools.get(item_id) != Some(&RememberedTool::Bash) {
    return None;
}

Every other tool was emitted InProgress and never updated — permanently "working" long after it finished.

Lifting that guard alone would not have helped: RememberedTool was { Bash, Other }, discarding the name, so there was nothing to route a completion to. It now retains identity (Bash, Mcp { server, tool }, Other(name)). Both issues named this type as the thing whichever PR landed first must own.

MCP rows were starved on purpose

The MCP branch emitted detail: None while the non-MCP branch immediately below it attached compact_input_summary(&input). That asymmetry is why an MCP call rendered as one dead grey line with a disabled chevron.

The generic row starved on a four-key allowlist

file_path, notebook_path, pattern, description. AskUserQuestion, Task, Skill, Glob, ToolSearch and Read-with-offset carry none of them. Known keys still win; anything else falls back to a capped rendering of the whole input.

The frontend threw away what did arrive

lib/agentChat.ts has always declared status and detail on both events. The timeline item type did not, so the reducer read them and dropped them, and ChatTimeline rendered <McpCard server tool /> with no detail — which is precisely why canExpand() was false.

Now: both item types carry status and detail (optional, so rows persisted before this still render); a completion carrying no result keeps the arg summary already on the row; expansion moves into the hoisted RowExpansion map so the toggle survives scroll-back; status renders through StatusPill, not new chrome.

reduceMcpToolCall gained the itemId match path — it appended unconditionally, unlike its reduceToolUse sibling. Without it these new completions turn "the row never finishes" into "the row finishes twice". That is why these commits cannot be separated.

Capping

In Rust, not the view, because rows are persisted. Output tails keep the 2000-char tail; argument summaries take the head at 300 — the meaningful part of an argument list is at its front, and it rides on one collapsed row.

Flag

mcpToolDetail, default off. A persisted row's shape changed and the parser only emits completions on newer builds, so main stays releasable while this is dark — the same precedent as changesReview. flags.test.ts enumerates the registry, so the flag is asserted rather than silently appearing.

Tests

Every one verified load-bearing.

Rust (8): MCP arg summary + InProgress → Completed; failed MCP resolves Failed; a generic tool completes under its own name; Bash still maps to CommandDone; an orphan tool_result still ignored; summary falls back to whole input, prefers a known key, is capped, and yields None on empty. Restoring the Bash-only guard fails exactly the three completion tests.

Frontend (5): MCP status/detail carried and resolving in place; a detail: null completion keeping the arg summary; a generic tool reaching failed; no duplicate row; distinct calls still appending separately.

Validation

• cargo test -p pickforge-core — 738 passed
• cargo clippy -p pickforge-core --all-targets -- -D warnings — clean
• bun run test:unit — 1606 passed
• bunx tsc --noEmit clean; bun run lint clean

Still open on these two issues

Stated rather than implied — #365's headline complaint is not fixed here:

• tool_progress elapsed time and system/status compacting/requesting need a new event variant in event.rs. Without them, WorkingRow still names nothing and shows no elapsed time.
• MCP tool rows in chat are dead ends: no args, no result, no completion — and pickforge-lanes should show lane cards #362 PR 3: pickforge-lanes rendering lane cards inline. Open decision (live vs snapshot) still unresolved.
• Row-height estimates unchanged — these rows gain a pill but stay one line, so COMPACT_LINE_PX holds. Re-measure when elapsed lands.
• No VRT fixture for a populated-detail MCP row.

Refs #362, #365