Subagent cards became non-expandable

Date: 2026-08-02 Status: fixing Reported: user, live panel (“いつの間にか、サブエージェントのカードが展開できなくなってます”)

1. Symptom

Completed subagent cards show no chevron and do not react to clicks.

2. Root cause A (proven): wire vs disk field-name mismatch

SubagentCard decides expandability up front:

bool hasDetails = _subagent.blocks.Count > 0
    || !string.IsNullOrEmpty(_subagent.summaryMarkdown)
    || _subagent.status == "running";

When false the chevron is hidden AND the ClickEvent handler is never registered — the card is inert for the rest of its life.

Live measurement of the user’s session (3fd4e6ad, 5 subagent records): blocks=0 summary=0 tok=0 for every one, so hasDetails is false.

Why the data is empty: TranscriptLoader.RestoreToolResult reads the Agent tool_result envelope as userNode["tool_use_result"], but that is the wire (stdout stream-json) spelling. The on-disk transcript JSONL spells the same envelope toolUseResult. Both verified today:

source	spelling	evidence
stdout wire	tool_use_result	real captures Tests/Editor/Fixtures/askuser_inbound.jsonl etc.
disk transcript	toolUseResult	real session file: top-level keys include toolUseResult, has snake tool_use_result at top? false

So on every restore (history switch, editor restart, domain reload, --resume rebuild) ApplySubagentCompletion never runs: status, token totals and the summary are all dropped. R02c section 1 point 4 documented the wire shape; the loader adopted that name for a disk-format reader, and TranscriptLoaderTests pinned it with a HAND-WRITTEN line using the wire spelling — so the suite stayed green while the real path never worked.

This is not a new regression; it has been wrong since Phase 4. It became visible now because recent releases forced many reloads/restarts, making restored (rather than live-streamed) cards the common case.

3. Root cause B: hasDetails ignores lazily-loadable content

SubagentCard.TryLazyLoadNestedBlocks can pull a completed subagent’s whole nested transcript from subagents/agent-<taskId>.jsonl on first expand. But hasDetails only looks at what is ALREADY in memory, so a card whose sidechain is sitting on disk still renders inert. Fixing A alone would leave this trap for any record whose summary happens to be empty (e.g. a subagent that returned only tool output).

4. Decisions

1. TranscriptLoader: read toolUseResult first, fall back to tool_use_result. Tolerating both costs one lookup and makes the reader correct for either serialization (the panel feeds it disk files today; the wire spelling stays valid if that ever changes).
2. SubagentCard.hasDetails: also true when the record has a toolUseId (a sidechain MAY be loadable). Expanding a card that turns out to have nothing shows an explicit “no details” line rather than an empty box, so the affordance never lies in either direction.
3. Regression guards must use REAL captured shapes, not hand-written JSON: pin the camelCase disk spelling against a fixture copied from an actual transcript line, and keep a wire-spelling case for the fallback.
4. Correct R02c: state explicitly that the envelope is tool_use_result on the wire and toolUseResult on disk.

Rejected: making hasDetails unconditionally true — a genuinely detail-free record (no toolUseId, e.g. a synthetic/legacy row) would then offer an expander that can never show anything.

5. Process note

The hand-written fixture is the failure that matters here: a test authored from the same wrong assumption as the code cannot catch that assumption. Transcript-format tests must be pinned to captured lines.