Bug: Claude Code's second User-Agent is unrecognised, so ~13% of its events label as Other #1363

What happens

Claude Code sends three distinct User-Agents on live traffic, and knownClients recognises only one. Measured from one laptop proxy's session store:

The middle row is the bug: claude-code is a value in knownClients but never a key. ParseUserAgent looks the first product token up against the map's keys, and that token is claude-code, so the lookup misses and the traffic is unrecognised.

Why it matters

Those 70 events — ~13% of this capture's Claude Code traffic — pool into agentop's catch-all Other row in the AGENTS pane instead of claude-code, and their spend goes with them. pipeline.IsKnownAgent is false for them, so they are also excluded wherever that gates (core/cost/usage). Claude Code is the reference integration, so this is the one agent whose attribution was assumed correct.

Session attribution is unaffected — AffinityName() returns claude-code for all three, because its comment-scan fallback compares against canonical names and claude-code is one. So this is a labeling/aggregation bug, not a session-grouping one.

The third row is not a bug

Claude-User (claude-code/…) staying unrecognised is intentional, and AffinityName's doc comment explains why: reading an agent's name out of a comment into Name would move those rows to a different agent in the ledger and the AGENTS pane than the history beside them. Leave it as is.

Suggested fix

One entry:

"claude-cli":  "claude-code",
"claude-code": "claude-code",   // the Agent-SDK-shaped UA

I probed this locally and verified three things:

1. claude-code/2.1.233 (…) labels as claude-code / version 2.1.233.
2. Claude-User (claude-code/…) still does NOT label — its first token is Claude-User, and trailingKnownClient cuts on / without trimming punctuation, so the later token reads as (claude-code and misses. The deliberate behaviour survives.
3. The existing core/pipeline suite stays green.

Worth adding a TestParseUserAgent_* row for each of the three live UAs above, including the negative for Claude-User, so the deliberate case is pinned alongside the fixed one.

How this was found

While verifying Codex's agent recognition for #1337, which fixed the same class of bug for Codex: an agent that sends more than one product token, with the lookup table covering only some of them. Codex needed codex-tui and codex_cli_rs added beside codex_exec for exactly this reason. Worth a look at whether bob/bob-shell/opencode have unlisted second tokens too — OpenCode reportedly uses one UA for every mode, but that is from a source comment rather than a capture.

Not fixed in #1337 deliberately: that PR is Codex support, and this is a different agent.

Unknown

Whether headless Claude Code (claude -p) is one of these two shapes or a fourth. Both recognised shapes mention claude-vscode and agent-sdk, so they look like different components (CLI binary vs Agent SDK) rather than interactive-vs-headless, but I have no headless capture to confirm it.

Assisted-By: Claude (Anthropic AI) noreply@anthropic.com