History: usage restore + browsing at scale (2026-08-03)

Two requests from one session:

1. Bug — 「このセッションの使用状況」 shows nothing after restoring a session from History.
2. Feature — the History list becomes unusable as sessions pile up. Rework it, modelled on Claude Desktop.

1. Bug: usage is discarded on restore

1.1 Root cause

AgentHub.SwitchToSession builds a brand new ChatSession and copies only the messages (AgentHub.cs:870):

var session = new ChatSession { sessionId = sessionId };
if (restoredMessages != null) { session.messages = restoredMessages; }
...
_lastModelUsage = new Dictionary<string, ModelUsage>();

so totalInputTokens / totalOutputTokens / totalCostUsd / completedTurns are all zero and the per-model map is empty. Two surfaces go blank:

surface	source	after restore
status bar “N tok”	session.total*	0 tok
“Usage this session” popover	AgentHub.LastModelUsage	“No completed turn yet.”
context meter	LastModelUsage[..].ContextWindow	hidden

Nothing reads the transcript for usage; AccumulateTurn is only ever called from the live result handler (AgentHub.cs:2494). The live session survives a domain reload only because SessionCacheFile persists the totals — a restore has no such cache.

1.2 What the transcript actually contains (measured)

18 real transcripts in ~/.claude/projects/C--Unity-UnityProjects-DevelopmentProject:

• No result lines and no cost field of any kind. total_cost_usd is stream-only, so cost cannot be restored. (Moot for subscription auth, where the CLI reports 0 anyway, and showCostUsd gates the display.)
• No contextWindow. The context meter’s denominator is stream-only too, so the meter stays hidden until the next turn completes. That is already its “no data” behaviour, so nothing regresses; it just does not come back early.
• assistant lines carry message.usage with input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens — everything the status-bar total needs.

1.3 The trap: usage lines are duplicated

One API response is written to the JSONL as several lines (one per content block), each repeating the same usage object:

file	assistant lines	distinct message.id	ids whose usage differs
62fbbf7d	23	16	0
b1228ae1	5	3	0
e233e417	101	68	0

Naive summing over-counts badly — 62fbbf7d reads 6679 output tokens naively vs 3147 deduplicated. Aggregation must dedupe by message.id.

1.4 Second trap: the <synthetic> model

e233e417 reports two models: claude-opus-5 and <synthetic> (the placeholder on isApiErrorMessage lines, 0 tokens). It must be excluded or the popover grows a meaningless row.

1.5 Design

Add a TranscriptUsage result to TranscriptLoader, filled during the scan Load already performs over every line (no extra I/O):

• TotalInput/Output/CacheRead/CacheCreation — deduped by message.id over the whole file (not the maxMessages-trimmed view: the trim happens after the scan, and a truncated total would be wrong, not just partial).
• LastTurnModelUsage — per-model usage for the assistant messages after the last real user prompt, i.e. exactly what AgentHub.LastModelUsage means (the last turn), not a session-cumulative aggregate that the context meter would misread. ContextWindow stays 0 (§1.2).
• UserPromptCount → completedTurns; last line timestamp → lastActivityTimestamp.

SwitchToSession applies them. Load’s existing signature is kept and delegates to a new overload, so no caller churns.

Known gap, stated rather than papered over: subagent tokens live in separate transcript files and are not summed. None of the 18 sampled transcripts spawned a subagent (isSidechain 0, Task/Agent tool_use 0 even in the 7 MB one), so the divergence from a live result total is unmeasured — it is not claimed to be zero.

2. Feature: browsing at scale

Requested: search, date grouping, incremental paging, pin, archive, delete, rename, and grouping by scene / project / custom group.

2.1 Data availability (measured)

need	available?	evidence
Conversation title	yes	ai-title lines in 18/18 files, always at line index 6-8, exactly 1 distinct title per file
Working directory	yes	cwd present, always by line index 2
Unity scene	no	zero scene references in any transcript; ContextBlockFormatter attaches none either

The title/cwd scan folds into SessionIndexEntry’s existing bounded 500-line preview scan — same single pass, no extra I/O, and both keys are found ~60x inside the cap. Scene grouping has no historical data: it requires new recording, so pre-upgrade sessions group under “(未記録)”.

2.2 Storage: SessionMetaFile

Pin / archive / rename / custom-group membership / recorded scene are our own state keyed by session id. Persisted as a JSON sidecar at UserSettings/AgentPanel/SessionMeta.json, following SessionCacheFile’s established contract exactly (atomic tmp+replace, never throws, tolerant load, pure System.IO + Core/Json).

Not the State.asset ScriptableSingleton: renamed titles are raw user text, and UnityYAML round-trips such strings unreliably — the exact defect that forced the transcript cache out of the asset (ARCHITECTURE.md D5).

2.3 Two decisions that need calling out

Delete. The transcript is the CLI’s own canonical user data under ~/.claude. Rather than unlink it, delete moves the file to UserSettings/AgentPanel/DeletedSessions/<sessionId>.jsonl behind the existing confirm-bar pattern, and the confirm text says where it went. Moving it out of ~/.claude/projects/<dir> entirely (rather than into a subfolder there) avoids assuming anything about how the CLI globs its own session directory. Reversible by hand; escalating to a hard delete is a one-line follow-up if that is what is wanted.

Project grouping. Every session the panel lists already belongs to one Unity project (the projects directory is keyed by cwd), so the only reading that adds anything is cross-project browsing. Implemented as: on demand, enumerate ~/.claude/projects/*, read each session’s cwd, and keep only directories that are real Unity projects (<cwd>/ProjectSettings/ProjectVersion.txt exists) — this machine has 30+ throwaway probe directories that would otherwise flood the list. Foreign sessions are browse-only: resuming another project’s session against this cwd is meaningless, so their rows explain that instead of switching.

2.4 UI shape (Claude Desktop as the reference)

Toolbar: search field, group-by dropdown (日付 / シーン / プロジェクト / カスタム), and an archived-visibility toggle. Rows show the ai-title (or the rename override) as the title with the first user message as a dimmer sub-line, plus the existing relative-time/size caption. Pinned sessions sort into a leading group regardless of the grouping mode. A per-row ”…” button opens pin / rename / group / archive / delete. Initial render caps at 50 rows with a 「もっと見る」 button, so list cost stays flat.

Search matches title, rename override, first-user preview and session id, case-insensitively; date groups are 今日 / 昨日 / 過去7日間 / 過去30日間 / それ以前.

3. Regression guards

Pure, Unity-free logic is the test surface (this is why grouping, search, sorting and paging are static methods taking explicit inputs):

• TranscriptUsage dedupe: a fixture with the same message.id on multiple lines must count once — pinned against captured real lines, not hand-written ones (the v0.12.2 lesson: a fixture written from the same wrong assumption as the code cannot catch that assumption).
• <synthetic> excluded from the per-model map.
• Totals computed over the whole file even when maxMessages trims.
• SwitchToSession populates totals and LastModelUsage.
• SessionMetaFile round-trip, tolerant load, atomic save (mirrors SessionCacheFileTests).
• Search/group/sort/page: pure-function tests including pinned-first ordering, archived hidden by default, and empty-query passthrough.

4. Verification

Sandbox AITemp: compile clean with no warnings from this package, full EditMode suite 1541 total / 1536 passed / 0 failed / 5 skipped (1466 before this change; +75 tests).

One failure was caught and fixed on the way: UssHygieneTests rejected three new rules that used var() inside a padding/margin shorthand. UI Toolkit does not expand custom properties inside multi-value shorthands, so the repo has a standing scan for exactly that; the declarations were split into longhand.

Live editor (DevelopmentProject), against the real 18-session history:

sessions on disk: 18   with ai-title: 18   with cwd: 18
largest transcript (7.1 MB): in=127 out=59781 cacheRead=5484673
                             cacheCreate=190248 prompts=4
status bar would show: '59.9k tok'   (before this fix: '0 tok')
date groups: today 12 / yesterday 2 / last-7-days 4
unity projects discovered: 2 (DevelopmentProject, AITemp)

Three things this confirms beyond the unit tests:

1. The reconstructed totals match the ground truth computed independently from the same file, on a transcript 160x larger than the fixture.
2. ai-title and cwd extraction succeeded for 18/18 real sessions, so the single-scan approach holds outside the fixtures.
3. The Unity-project filter kept 2 directories out of the 30+ under ~/.claude/projects on this machine — without it, “group by project” would have listed 30 throwaway probe directories.

Note lastTurn models=0 for that particular transcript: its final user prompt was never answered (an interrupted turn), so there is genuinely no last-turn usage to show and the popover correctly reads “No completed turn yet”. The populated case is pinned by the fixture test instead.