feat(agent-visualizer): show agent model names

[mirland]

What does this PR do?

• Shows the active model name in the agent detail card.
• Tracks the model per agent in the extension (Map keyed by agent name), re-emitting model_detected when the model changes (e.g. /model switch), and replays per-agent models when the webview reconnects.
• Formats Claude and GPT model IDs into readable labels, including legacy Claude 3.x, minor-version IDs, and Bedrock/Vertex provider-prefixed IDs.

How to test

1. Run pnpm run dev and open the visualizer.
2. Select an agent and verify the model label shows under the agent name in the detail card.
3. Confirm examples like claude-3-5-sonnet-20241022 and us.anthropic.claude-opus-4-1-20250805-v1:0 render as Sonnet 3.5 and Opus 4.1.
4. Close and reopen the visualizer panel mid-session: subagent model labels persist.
5. pnpm test and pnpm run build:web.

Preview

Checklist

• I have read the CONTRIBUTING guide
• I have signed the CLA

[cla-assistant]

All committers have signed the CLA.

[cla-assistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[mirland]

Hi @patoles, thanks for this repository, it has been very useful for understanding how agents move through a flow.

I added a small improvement to show which model is running for each subagent. I think this can make it easier to understand and debug agent executions.

Curious to hear what you think. Any feedback is welcome!

[patoles]

Thank you very much for this! It's a feature I'm really glad to see land, and you found exactly the right seam: the frontend was already prepared for it (Agent.model exists and handleModelDetected is keyed by agent), so per-agent emission is the correct missing piece. The detail card wiring works, the tests are appreciated, and they're picked up by the root pnpm test glob. I did find a few things worth addressing before merge. It looks like a long list, but the first three all share one root cause, so it's smaller than it appears.

1. The once-ever Set dedup loses/freezes models (root cause of three issues)

modelDetectedAgents: Set<string> records that an agent's model was emitted, but not which model, and it's never cleared or replayed. Three consequences:

• Panel reopen permanently loses subagent models. replaySessionStart re-emits the orchestrator spawn with session.model and resets sub.spawnEmitted so subagent spawns replay, but subagent spawn payloads carry no model, and the Set keeps suppressing model_detected re-emission. Close/reopen the visualizer mid-session and every subagent's model row is blank (with tokensMax falling back) for the rest of the session.
• Mid-session model switches never show. After an agent's first emission, /model switches (or a re-dispatched same-named subagent with a different model: override) are suppressed forever, leaving a stale label and a wrong context-window %.
• The one-shot can be consumed before the frontend ever sees it. On re-attach, prescan adds subagent names to spawnedSubagents without emitting agent_spawn (transcript-parser.ts:449-452). If model_detected then fires before the frontend knows the agent, handleModelDetected drops it silently, and the Set entry means it never retries.

The codebase already has the right pattern: the Codex parser tracks lastEmittedModel and re-emits on change (codex-rollout-parser.ts:80, with test coverage). Suggest replacing the Set with Map<agentName, model> and emitting when the value differs. Same footprint, subsumes the once-per-agent behavior, keeps the two watchers consistent, and gives replaySessionStart the data to re-deliver per-agent models on reconnect (mirroring what it already does for session.model).

2. scripts/relay.ts constructs WatchedSession too and wasn't updated

The new field is required in protocol.ts, but relay.ts:195 builds its own WatchedSession literal (it already initializes spawnedSubagents: new Set() right there). Nothing typechecks scripts/ (the extension's tsc --noEmit only covers extension/src), so this ships as a silent interface violation, and the lazy-init in the parser (session?.modelDetectedAgents ?? (session ? (session.modelDetectedAgents = new Set()) : undefined)) is what keeps the standalone dashboard from crashing at runtime. Easy to miss; I only caught it because I knew the relay had its own copy. Adding modelDetectedAgents: new Set() to the relay's literal lets the parser gate collapse to a plain !session.modelDetectedAgents.has(agentName), which also stops the field from looking optional when it isn't.

3. formatModelName: Bedrock IDs render the date as a minor version

The date strip /-\d{8}$/ is end-anchored, so any suffix after the date defeats it, and CLAUDE_NEW's optional (?:-(\d+))? swallows the date:

formatModelName('us.anthropic.claude-sonnet-4-20250514-v1:0')  // → "Sonnet 4.20250514"

(IDs with a minor version, like ...opus-4-1-20250805-v1:0 → "Opus 4.1", happen to work, which hides the broken case.) Making the date strip non-anchored (e.g. /-\d{8}(?=$|[-@:.])/) or stripping known provider prefixes/suffixes first would cover Bedrock/Vertex IDs. Worth a test case too.

Smaller points in the same function:

• The GPT regex is unanchored and matched against raw model instead of base, so chatgpt-4o-latest becomes "GPT-4o-latest". Anchor it (word boundary or ^) and match against base like the Claude branches.
• The two Claude branches are copy-paste with capture indices swapped. A tiny shared helper (claudeLabel(family, major, minor?)) removes the drift risk.
• web/lib/canvas-constants.ts already maintains Claude family patterns for context size and cost, both matched against lower-cased IDs. This adds a third lockstep-edit family list, matched case-sensitively. Consider deriving all three from one shared family constant and lowercasing before matching.

4. PR description vs. diff

The description says the model shows "in the agent detail card and chat panel" and that the model is "passed through the visualizer panel props", but chat-panel.tsx isn't touched (its props have no model) and no props were added anywhere; the detail card works because selectedAgent already carries model. Was a commit missed, or should the description be trimmed to the detail card?

---

The Map-on-change and the relay init are the two load-bearing ones; the formatModelName items are quick tweaks on top. Happy to talk through any of it if you see it differently, and thanks again for the contribution!

[patoles]

Nice work on the revision, this addressed everything I was worried about, and the replay handling is better than what I sketched. I took the liberty of updating the PR description to match the final diff (the chat panel part didn't make it into the change set, which is fine, the detail card is the useful surface anyway). The shared family-list idea I mentioned is small enough that I'll do it myself in a follow-up. Thanks for the thorough turnaround!