Agent View 0.10: Supervising a Claude Code Fleet — Triage, Evidence Gates, and Native Signals

When we launched Agent View for Claude Code in June, it answered one question: what is my fleet of agents doing right now? Six releases later, the question has changed. Running five to twenty Claude Code sessions is no longer the hard part — supervising them is. Every session asks its own questions, finishes with work you have to verify, and records state you can't see from one terminal. This is what has shipped since, from v0.4 to v0.10.

Claude Code itself now keeps a registry of running sessions in ~/.claude/sessions/ — the source behind claude agents and its built-in fleet view for background jobs. As of 0.10, Agent View reads that same registry, read-only, for busy / idle / waiting status with no hooks installed, across other windows, the Claude Code extension, and background sessions. So the two are complementary: Claude Code records the fleet; Agent View is the editor-side layer on top — spawning agents into isolated git worktrees, collapsing their questions, verifying their output, and landing their diffs without leaving VS Code, Cursor, Antigravity, Windsurf, or VSCodium.

When you supervise many sessions, the bottleneck is your attention — and agents ask the same things: package manager, branch strategy, test framework, what is off limits. Triage compresses N sessions' questions down to the distinct ones. Claude judges which questions are the same ("install deps" and "install dependencies" are one decision; "tests for auth" and "tests for billing" are two), with an offline lexical matcher as shortlist and fallback. One answer resumes every agent waiting on it. Answers go into a decision log with a scope and an expiry, and a later matching question is auto-answered only after Claude confirms the recorded answer genuinely resolves that question — the agent receives your answer verbatim.

• ·Permission asks are never auto-answered from a general decision, and file conflicts are never auto-answered at all.
• ·Past a 5% auto-answer error rate, Triage demotes itself to dry run and tells you.
• ·Hold (opt-in) suspends an asking agent inside its own AskUserQuestion call until you answer, burning no tokens — a hold can delay a prompt, never lose it.
• ·Triage is off by default since 0.10, and writes nothing into your project unless you choose repo storage.

Agents should finish with proof. When an Agent View-spawned agent settles, Evidence Gates run the project's checks — typecheck, lint, test, tsc --noEmit, cargo/go fallbacks, or your own commands — inside its worktree, and show checks ✓ or checks 1/3 ✗ on the row. Auto-detected commands need a one-time per-repo approval before they ever run, and evidence from before new edits shows as stale rather than passing for current work. Review & Land then turns each agent's work into a reviewable unit: every changed file as a native side-by-side diff against the agent's exact fork point, change requests sent back to its terminal, Open PR, or a reversible squash-merge. Nothing is auto-merged and no branch is ever deleted.

Unattended Fleet is a governed auto-pilot: spawned agents get acceptEdits, so file edits don't block but Bash and other tools still prompt; stalled agents are nudged a capped number of times, then escalated; and a per-agent cost cap pauses an agent that exceeds it, keeping its branch and diff. Turn it off and the Fleet Shift Report opens — what is waiting on you, plus a per-agent table of status, plan, diff size, evidence verdict, tokens, and cost.

The Session Board replaced the Pinboard: pick a session and its transcript materialises into swim-lanes by type and columns by episode — your prompt with its requirements as chips, the plan as one living object, verification evidence, and commit milestones. The Scroll is a session minimap, one tick per event, with your own prompts as the anchors. Both are pure transcript projection, with no LLM calls. And 0.8 made Claude Code's Agent Teams messaging visible: the native team store feeds the Teams cockpit, with a who-talked-to-whom message timeline, read-only against ~/.claude.

The cost fix matters if you use Unattended's budget cap: the old estimate missed most cache reads, so the real figure on long sessions is often several times higher — if you set mas.unattended.maxCostUsd, revisit it. The AI helpers also now run as bare text models, with no tools and no MCP servers, which closes a path where a prompt-injected transcript could have run tools.

# Open VSX-based IDEs (Antigravity / Cursor / Windsurf / VSCodium)
antigravity --install-extension inthepond.agent-view-for-claude-code

# VS Code (Microsoft build)
code --install-extension inthepond.agent-view-for-claude-code

Agent View is an unofficial community project, not affiliated with or endorsed by Anthropic; "Claude" and "Claude Code" are trademarks of Anthropic, used only to describe interoperability.