[Adam-Dalloul]

Custom agents can be missing from the Token Usage and Manage Conversations filters. Include configured agents and agents with saved conversations, retain selected agents after their registration is removed, and use their current display names in both the filters and usage breakdown. Selecting an agent sends its exact ID. Narrow conversation rows also retain their titles and metadata.

This PR now contains only the filters and agent facets. Recording and native usage recovery are separate changes so their accounting contracts can be reviewed independently.

Validation: 63 UI tests, 52 token-usage/service tests, TypeScript, scoped ESLint, server/MCP/CLI Clippy and ordinary production compilation passed. The strengthened regressions exercise rendered breakdown renames, removed-agent filter IDs, and configured names that differ from the fallback. Browser checks passed at 1365/390/320px and Arabic RTL at 390px, including exact filtering and estimate updates using fixture data.

[xintaofei]

@Adam-Dalloul thanks for this, and for the thorough description and checks. A lot of it is careful work:

• The isolated-home gate. It only accepts the official adapter with an explicit CLAUDE_CONFIG_DIR override, refuses the built-in's home and ~/.claude (canonicalized, case-insensitive on Windows), and follows spawn's empty-value rule.
• Turn keys are namespaced per session.
• The projection change keeps usage for a turn with no output without letting an orphan footer create a turn.
• The filters list configured agents by name, follow renames, and keep a selection after its registration is gone.

We're not going to merge it in its current form, though. The usage half has two problems that I think need a design decision rather than a patch:

1. §1 One usage source per session drops recorded usage after a context reset. This is the main one.
2. §2 PromptResponse.usage doesn't mean the same thing for every agent it's recorded for. Built-in Command Code and TRAE turns can be over-counted.

§3 covers smaller things, and the end has a suggestion for splitting this up.

§1 One source per session drops recorded usage after a context reset

This applies to custom claude-agent-acp agents with an isolated CLAUDE_CONFIG_DIR, the only case native recovery handles. For each session in the continuation chain, native_usage::read picks one source, in this order:

1. Claude's own <CLAUDE_CONFIG_DIR>/projects/*/<session id>.jsonl, whenever it yields any usage.
2. Saved native facts.
3. Only then, that session's ACP transcript (native_usage.rs#L153-L180).

The description puts it as "new native records replace overlapping ACP counts". That assumes the native file named after the ACP session id covers the whole ACP session. With claude-agent-acp, the file stops covering it in two everyday flows.

Approving a plan with "Yes, clear context …". This is the first option in the ExitPlanMode prompt whenever there's a plan (tools.ts#L73-L95, v0.94.0). Choosing it does the following:

• The adapter keeps the ACP session id (acp-agent.ts#L10986-L10988).
• It restarts Claude under a fresh private id: options.sessionId = creationOpts.publicSessionId ? randomUUID() : sessionId (acp-agent.ts#L11398-L11401).
• It continues the pending ACP turn in the new session with "Implement the following plan: …" (clear-context-coordinator.ts#L102-L163, exit-plan.ts#L224-L234).
• Later prompts in that ACP session go to the new private session too.

So <ACP session id>.jsonl ends at the plan. The implementation and everything after it land in a file nothing links to. codeg's built-in Claude history has the same truncation; it's noted as unsolved in registry.rs#L1425-L1430.

/clear. Claude rolls over to a new <uuid>.jsonl while the ACP session id stays the same (claude.rs#L1288-L1292). native_usage deliberately doesn't follow rollovers (native_usage.rs#L120-L124).

In both flows, the ACP transcript keeps receiving PromptResponse.usage for the turns after the reset, which is exactly what this PR starts recording. But as long as the original file yields any usage for that session, or saved native facts exist for it, those facts win and the session's ACP records are never read.

Example: you plan in plan mode, approve with clear context, let the implementation run, then continue for a few turns. The dashboard shows only what was spent before the approval. The rest was recorded and is then dropped.

The same rule also freezes a session whose native file goes away while the chat continues:

• Its saved native facts keep winning (native_usage.rs#L160-L162), so the ACP usage of turns completed afterwards in that session is ignored.
• If no other session in the chain supplies fresh facts, the result is KeepRecorded and the sync settles without rewriting anything (token_usage.rs#L1059-L1062).

The obvious fixes fall short because neither source is complete on its own. I ran claude-agent-acp 0.94.0 (the version codeg pins) against a local stub model, with one prompt that started a background sub-agent:

• The response's usage was 4,000 in / 10 out.
• _meta.quota.model_usage in the same response was 6,000 / 15. The sub-agent's 2,000 / 5 isn't in usage.
• The adapter also leaves autonomous results out of the per-prompt tally on purpose (acp-agent.ts#L6924-L6936).

Claude's own transcript does include sub-agent spend. So "native wins" loses the post-reset work, and "ACP only" loses sub-agent spend.

Shortcuts I looked at

• Take the larger of the two totals per session. This is never worse than now. But it still loses the post-reset work whenever the pre-reset part, with its sub-agents, is the larger one.
• Follow /clear with find_clear_rollover_successor.
  • It matches a sibling file that starts with a /clear record within five minutes of the old file's last record. When two sessions in the same project both clear within that window, it can pick the other session's file.
  • It also does nothing for clear-context.
• Switch to ACP from the reset onwards.
  • The ACP turn that spans the reset already includes the pre-reset main-thread usage, which the adapter carries over (clear-context-coordinator.ts#L137, acp-agent.ts#L4780-L4791). But it doesn't include that part's sub-agent spend, which only the native file has.
  • A recorded /clear prompt doesn't prove a reset happened, because codeg records the prompt before sending it (connection.rs#L13464-L13468).

So recovery needs a policy for combining the two sources per turn, or a reliable way to find the post-reset file. We'd like to settle that policy before code.

§2 PromptResponse.usage doesn't mean the same thing everywhere

The PR records response.usage for every agent whose history codeg keeps as an ACP transcript: custom agents, ZCode, and the built-in TRAE and Command Code (agent.rs#L80-L86, connection.rs#L13976-L13984). parse_usage then reads it as cache-exclusive: input + output + cache read + cache write (acp_native.rs#L621-L635). Two of those built-ins report something else.

Command Code (command-code@1.86.3, the version codeg pins; read from the published bundle):

• promptResult returns usage as a running total since the session was opened, because each turn_end is added to sessionUsage.
• The per-prompt figure is only in _meta.usage, with cacheReadTokens/cacheWriteTokens spellings that parse_usage doesn't read.
• Its counters come from the AI SDK's totalUsage. Command Code's own cost estimate treats inputTokens as including cache: estimateCatalogCostUsd prices max(0, inputTokens − cacheReadTokens − cacheWriteTokens) as fresh input.
• So three 1,000-token prompts in one opening of a session are stored as 1,000 + 2,000 + 3,000, and on top of that any cache reads are counted twice.

TRAE CLI 0.209.1. I ran a turn against a local stub model with 1,234 input tokens (1,000 of them cached) and 56 output.

• It came back as usage: {inputTokens: 1234, cachedReadTokens: 1000, outputTokens: 56, totalTokens: 1290}.
• TRAE's own _meta["_trae/tokenUsage"].turnDisplay gives the fresh input as 234.
• parse_usage would store that turn as 2,290 tokens instead of 1,290.

This affects turns recorded after upgrading:

• Command Code is overcounted when more than one prompt response is recorded during the same opening of a session and an earlier one had usage, because the running totals are summed. Reopening the session resets the total. It is also overcounted whenever input was cached.
• TRAE is overcounted whenever input was cached.

The same applies to hand-registered custom agents running those CLIs, with ids like command-code or trae that predate the built-ins (agent.rs#L176-L181).

For other custom agents, the field's meaning isn't settled yet:

• The schema describes it as usage "for a prompt turn", but its fields as "across session" / "across all turns" (agent.rs#L3433-L3445).
• The end-turn token usage RFD is kept in Draft (end-turn-token-usage.mdx#L14), and its first open question is exactly per-turn vs cumulative (#L95-L97).

claude-agent-acp reports per prompt and cache-exclusive (acp-agent.ts#L11855-L11880), so recording it as-is is right for the Claude custom agents in your description. For adapters we don't recognize, we'd want an explicit rule rather than an assumption.

§3 Smaller things

• Stopped turns record nothing.
  • On Stop, codeg deliberately doesn't write a turn end (connection.rs#L14292-L14300).
  • The late cancelled response is awaited in the background and dropped (connection.rs#L14383-L14385), so its usage never reaches the transcript.
  • This isn't a regression, since nothing was recorded before.
  • For Command Code, usage it had already counted during a stopped prompt (it adds each internal turn_end to the running total) does show up in the next response from the same open session. That is one reason to difference that total rather than read _meta.usage. It can't recover spend the adapter never reported.
• Re-registering a deleted agent can drop earlier sessions' saved usage.
  • Deleting a custom agent with its transcripts (custom_agents.rs#L140-L149) also removes the continuation headers, so a later sync only walks the current session.
  • Suppose the same id is registered again with an eligible isolated home. The next sync of that conversation (after a new turn, or a full rebuild) reads fresh native facts for the current session and replaces the conversation's rows with them.
  • Saved facts are only carried over for sessions the walk visits (native_usage.rs#L128-L134), and acp-usage: rows aren't carried over at all (native_usage.rs#L93-L95). So the earlier sessions' usage is gone.
  • While the agent stays deleted, the rows are kept.
• A few tests would still pass with the behaviour they cover broken:
  • token-usage-page.test.tsx doesn't exercise the breakdown labels: by_agent is empty and the charts are mocked. Your description reports Chromium checks covering these labels.
  • "keeps a removed agent from saved chats selectable even when its filter returns no rows" never asserts that selecting the removed agent sends agent_type: "custom:retired-reviewer". A change that dropped an unknown filter to null would still pass.
  • "keeps custom accounts reachable by name in the overflow menu" uses names identical to the humanized ids (Claude Code 2 for custom:claude-code-2), so it can't tell a configured name from the fallback.

Suggested way forward

1. The filter work. This is the management dialog, the Token Usage filter, and facet_agents. It is independent of the usage changes and looks good to me, so it could be split out and reviewed on its own.

2. Recording on its own. This is the Cargo feature, record_turn_end, and the projection change.

   • Command Code and TRAE need to be either normalized or left out for now:
     • Command Code: difference successive running totals, with a new baseline each time the session is opened, and take cache out of inputTokens. Also decide where a stopped prompt's share goes; otherwise it lands on the next recorded response, under that response's model.
     • TRAE: take cache out of inputTokens. feat: 为 TRAE CLI 会话添加原生 token 使用量的导入和统计 #1019 also changes TRAE usage, so the two will need to agree.
   • It also needs an explicit rule for custom adapters we don't recognize.

3. Native recovery as a separate PR. It should combine the sources per turn instead of picking one per session. The rule has to cover:

   • turns that predate recorded ACP usage;
   • turns the native file no longer covers;
   • turns both sources cover, keeping native's sub-agent spend without double-counting the carried pre-reset part.

   I'm happy to talk the policy through in an issue first.

Your tests and CI pass; they just don't exercise the adapter behaviour or the cleanup/re-registration case described above. If I've misread any of it, please say so and I'll take another look. That applies especially if you have a session where native-first gives the right total after a clear-context approval.

[Adam-Dalloul]

@xintaofei, this PR now contains only the filter fixes and strengthened tests; recording is separate in #1048. I reproduced the reset/resume cases on 0.97.0 and opened #1047 for the recovery policy.