feat(desktop): attribute a sub-agent's work to it, and show what it did (R5)

[physercoe]

Closes the R5 wedge of docs/plans/desktop-companion-vision-parity.md. W4 is complete with R6 (#573).

What the director gets

The Sub-agents chip listed delegated agents but could not say what any of them did. Clicking one now opens what it was asked to do and everything it has done — its tool calls, its notes, and what it is running right now. The drill-down lives in the dock's existing panel with a back button, so no new region.

The premise was true of one engine, and not one we drive

The plan said "events already carry subagent marking". The only producer of that marking is the kimi M4 wire-tail (kimi_code/mapper.go:235). Built as written, the panel would have been permanently empty for claude and codex.

No session log on the dev box had a sidechain in it, and the field appears in no vendor doc we hold — so I made the observation: a real claude --print --output-format stream-json run making one Agent call. claude-code 2.1.220 carries more than the plan assumed:

• parent_tool_use_id on every assistant/user frame — null for the main agent, the spawning Agent call's tool_use_id for a sub-agent's work. This is the correlation the panel needs.
• subagent_type + task_description alongside it.
• Four system subtypes: task_started (full prompt), task_progress (last_tool_name + running usage), task_updated, task_notification (summary).
• The tool is named Agent, not Task — the dock's name set already matches it.

The bug that fell out

Our translator dropped all of it. A sub-agent's tool calls and prose landed in the transcript indistinguishable from the main agent's own work — read as the agent you're talking to having run them. And #374's guard, which keeps a sub-agent's token usage out of the session's turn counts, checks subagent:true and therefore skipped nothing on claude, quietly inflating them.

So R5 ships the producer half first. The frame profile stamps parent_tool_use_id plus a derived subagent boolean — the same one kimi already sets, so one vocabulary answers for both engines and the two existing consumers (digest_fold.go, transcriptStats.ts) start working with no change. Deriving it at the producer rather than in each consumer is deliberate: a predicate re-derived per consumer is the duplication this repo has been bitten by. task_progress regains tool_use_id + subagent_type, which #374 dropped along with the noise it meant to drop.

Honesty for the engines that report nothing

codex and gemini publish no per-sub-agent correlation. The panel still opens for them — the header comes off the spawning call, which every engine has — and then says "this engine does not report which events are a sub-agent's". That is a different sentence from "this sub-agent has not done anything yet", which is what an empty stream means when the engine does report. The panel picks between them by asking the data, not the engine name, so an engine that starts reporting tomorrow needs no change here.

Verification

• 8 new tests, three mutation-checked. The separating input is two sub-agents running at once — a filter keyed on the subagent boolean alone puts both agents' work under each row and passes every single-agent test. Mutations caught: subagent-bool-only join (2 fail), toolId = event id (7 fail), reported dropping the provenance probe (1 fail).
• Both translators agree: the legacy translate() mirrors the profile field-for-field (the parity test compares key sets, so fields are set unconditionally, nil included).
• The 15-frame recording is now in the parity corpus — its first sidechain coverage — scrubbed of the recording box's scratch paths and session id. The desktop's TS interpreter is pinned to the same regenerated fixture, so the local Companion gets the same events; the electron suite verifies it.
• go test ./internal/hostrunner ./internal/agentfamilies ./internal/server ✅ · desktop node --test 883 pass · electron 715 pass, 6 skipped · tsc --noEmit ✅ · full vite build ✅ · token ratchet clean, no phantom tokens · CI's whole lint set ✅ including openapi.

🤖 Generated with Claude Code