Captain Claw v0.6.5

Captain Claw v0.6.5 Release Notes

Release title: Teams that wait for each other — Vatra rendezvous, specialist tiers, and a tidier deck
Release date: 2026-06-29

A polish-and-plumbing release on top of 0.6.4's Frontier Horizon. The headline is a
wait rendezvous for Vatra: a specialist that genuinely depends on a teammate's
artifact now blocks until it's ready instead of improvising a guess. Around it: the
Library gains dedicated coding and vision tiers (so code and image work route to the
right model), Basna/Vatra/Plan runs get a sticky "what phase is running now" pill plus
wait events in the timeline, Plan mode's live agent cards finally show per-agent token
counts, and a new Agent Folders admin page lets you inspect and clean up fd-data
subfolders. Fully additive and backward compatible with 0.6.4 — nothing changes for existing
runs unless you reach for it.

---

What's new

Vatra wait — agents block on a teammate's artifact instead of guessing

Agents on a Vatra team often look for a piece a teammate is supposed to produce — a VFS file
or a board post — don't find it yet, and improvise a guess. A new blocking wait
action lets a specialist that truly depends on a specific artifact pause until it's ready.

• Two conditions. wait takes a path (vfs:<project>/<file>) or a query (board
  keywords). The /agent/wait endpoint long-polls the VFS file or the blackboard every
  second.
• Bounded, never hangs a run. Capped at 90s (tool timeout raised to 120s), which is
  below the dispatch timeout, so a wait can't stall the run. On timeout it returns a board
  digest plus a "proceed deliberately" nudge — so the agent does the right something else
  rather than erroring out. Sibling owners keep working while one awaits (they're concurrent
  coroutines).
• Visible in the timeline. A wait emits a wait stage event when it starts (⏳ … ≤Ns),
  when it resolves (✓ got <file> / board match), and on timeout (⌛ not ready) — so a
  run shows exactly who is blocked on what instead of a waiting agent just looking slow.
• Steered, not forced. Worker prompts now point agents to wait when genuinely blocked,
  and no longer say "don't stall waiting" unconditionally — don't busy-poll, but wait once
  when you really depend on something. There's no forced depends_on ordering; the tool is
  the whole mechanism. Owner+project file resolution is sandbox-checked (resolve_under), so
  the multi-user Flight Deck process serves the right run's files and rejects escapes.

Coding & Vision tiers in the Library

Two new model tiers join the Library so work routes to a model suited for it:

• coding (default opus-4-8, 65 536 output context) for multi-file diffs and debugging.
• vision (default sonnet-4-6) for image / screenshot / document understanding.

Both are wired through every touchpoint — the archetypes.json tier table, the backend
_VALID_TIERS, Dubina's TIER_ORDER, consciousness's _TIER_RANK, and the frontend
TIER_ORDER. A backfillTierMap makes existing saved tier sets gain the new cards
(inheriting their provider/key) instead of rendering empty. Code-producing archetypes
(code-implementer, debugger, simplifier) are reassigned from balanced to coding,
and a new "Vision & Multimodal" family (visual-extractor, ui-reviewer,
brand-design-reviewer) plus vision routing lines in the project-coordinator and concierge
rosters round it out.

Sticky phase pill + per-agent tokens in Plan mode

Runs emitted plenty of detail (action / usage / narration) but no clear "what phase is
running now" signal — under a long event stream the stage markers just scrolled away.

• Explicit phase events are now emitted at every transition and surfaced stickily in
  the Progress header (with a spinner while running), and phase rows are accented in the log:
  • Vatra: Planning → Spawning → Intro → Main → Review → Verifying slices → Synthesizing →
    Verifying deliverable → Learning → Done.
  • Basna: Routing → Spawning → Generating → Merging → Verifying → Done.
  • Plan: Step x/y: <goal> per step, plus Planning / Synthesizing / Done.
• Per-agent token counts in Plan mode. The plan-step progress mirror previously dropped
  usage events, so Plan-mode agent cards showed activity but no running token count. Usage
  events (prompt / completion / total) are now forwarded into the parent plan log, so the
  cards render the per-agent token header like a normal Basna run.

Agent Folders — inspect and clean up fd-data

A new admin-gated Flight Deck page to inspect and clean up agent subfolders under
fd-data.

• What it shows. Each folder with its on-disk size, file counts, last-modified,
  and desktop presence — running / on-desktop / orphaned, where orphaned means a
  folder whose slug is no longer in the process registry.
• Backend (/fd/agentfs, admin-gated): list folders, list and view/download workspace
  files, and delete a folder (refused if the agent is still running). Path-sandboxed to
  direct children of fd-data.
• Frontend: AgentFoldersPage + agentFsStore, wired into routing and the sidebar.
  Light-theme-friendly colors, filters (running / on-desktop / orphaned / name search), and a
  rich file viewer that renders HTML (sandboxed iframe) and Markdown with GFM tables.

---

Upgrade notes

• No schema changes and no new required configuration. Everything above is additive.
• Library tiers. Existing saved tier configs are auto-backfilled with the new
  coding / vision cards (they inherit the provider/key of your existing tiers) the next
  time the Library loads — review them if you want code/vision work pointed at different
  models.
• New endpoints: POST /fd/<run>/agent/wait (Vatra) and the /fd/agentfs/* router
  (admin-gated). The wait action is available to Vatra workers automatically.
• Frontend bundle rebuilt for the phase pill, Plan token cards, tier UI, and Agent
  Folders page.

Backward compatible with 0.6.4.

Full Changelog: v0.6.3...v0.6.5