Skip to content

rt-observe & the Toolkit Dashboard

Chapter 12 of the ResearchTools manual. Back to table of contents. Skill row: 06-skills.md. Command summary: 04-commands.md.

What it answers

Answers one question: is this toolkit correctly deployed to every harness in use, and which empty mirror cells are deliberate rather than lost. Serves the same snapshot as a loopback dashboard (/rt-dashboard, rt-dashboard.ps1/.sh/.bat) - mirror matrix, fan-out wiring diagram, Real-Time Process tab, sessions strip, and the rail's seven receipt-bearing panels (repository, plan, services, code graph, hooks, journal, actions - the last a closed, tested action whitelist, dry-run by default). The journal panel (2026-09-24) reads an OPTIONAL, absent-by-default persistence layer - PostgreSQL identity/account-mapping plus an OpenObserve trace/audit stream - that changes nothing about a clone with neither configured.

Full reference (states, dashboard layout, actions panel, session messaging, adapters): docs/rt-observe.md.

Screenshot

rt-observe dashboard, Real-Time Process tab

The Real-Time Process tab (one of the four left-side views, alongside Mirror Matrix, Fan-out, and Sessions), captured on branch feat/rt-dashboard, profile engineering. The top strip walks the harness state machine for the current turn — SessionStart → waiting → UserPromptSubmit → reasoning + tool → tool call → PostToolUse → security audit — with the hooks firing at each step named above their box (install-junctions, memory-upkeep, caveman-mode-tr..., betterleaks-hoo..., pip-audit-hook...). Below it, the fan-out diagram draws one turn's actual calls: the tool node branching into Write, prompt, and Bash counts, out to Obsidian vault and Graphify graph as the two memories, and down to a spawned local-writer subagent — this is the "runner" the tests describe travelling a live edge. The right rail shows the Repository panel (branch, profile, suite status green . 69 passed, stamp age, definitions count, registry health, and run-offline-tests.ps1 as the prover), with Plan, Services, Graph, and Hooks as sibling tabs. The 2 STALE badge top-right and the footer strip (drawn by rt_state.py, generation timestamp, policy hash, last install, repo root) are the receipts every panel here carries per R9 — a claim with no stamp is not trusted.

The voice panel (2026-09-26)

A fifth tab, Voice, next to Real-Time Process. Hold P to record, release to send; a dropdown next to the button picks the language (Auto-detect, English, Francais), which drives both the speech-to-text decode and the answer's own language. The transcript updates live while you speak, the same progressive re-transcription Devoir2 uses, and the view draws the voice pipeline in the Real-Time Process tab's own idiom: one agent node whose ring is your microphone level, four stage boxes (listening, transcribing, thinking, replying) that light up in turn with their measured durations, and a live spectrum strip.

Speech-to-text runs locally on the GPU: faster-whisper large-v3-turbo in int8_float16 on CUDA. It won a scored comparison of four Whisper variants run beside the resident writer-role model: 2.5% word error on the operator's own recordings, a 19.5 s recording transcribed in 0.74 s, and 945 MiB of VRAM (see docs/rt-observe.md for the full table). Install scripts/requirements-voice.txt, plus scripts/requirements-voice-cuda.txt for the NVIDIA runtime on a machine without a system CUDA Toolkit. A machine without them sees the panel name the missing install command, never a broken dashboard. The model is loaded and warmed when the dashboard starts, so the first press already has a live transcript.

The panel never opens the Obsidian vault or the graphify graph itself: every question is relayed to the vault daemon's own read-only ask queue (.claude/skills/obsidian-cli's daemon_ask.py), the same boundary this page's vault/graph panels already hold themselves to. Since 2026-10-02 that queue answers in two spoken parts. It searches the vault first and speaks that answer right away; then, if the matched vault note's project folder names a repo: property, it asks a short English-keyword question and queries that project's own graphify graph (daemon_graph.py), read-only, and speaks a short code-structure note on top - labelled as code structure, not progress, since the graph is AST-only and holds no decision log. No repo:, or no graph for that repository, says so plainly instead of guessing. POST /api/voice/ask returns at once; the panel polls GET /api/voice/answer until a part arrives or the per-part wait (timeouts_seconds.voice_ask_wait) runs out. Every spoken part is read back through the browser's own speechSynthesis, not a new server-side voice.

Local calls and plan stages. Two snapshot sections follow local coding. local_calls, read by collect_local.py, reads a bounded tail of the bridge log and shows each role with its model, window, call and token totals and whether it is active. rounds, read by collect_rounds.py, lists each plan of a round folder with its executor, tier and task counts, and the review loop of every worktree as rounds used of the maximum. The Real-Time tab draws a lane and a token ring for each active bridge role, and shows an aider run's fixed cost and task budget in the hover. See chapter 11.

The /rt-dashboard command

Starts the rt-observe dashboard on loopback and reports the URL. With no argument it dry-runs first, which names the interpreter it resolved, the bind address, the page and every per-section TTL, and starts nothing. With --json it skips the server entirely and dumps the snapshot, which is the right form inside a script or when the only question is whether anything is lost.

Its three refusals are read back to you rather than worked around: no interpreter found names every path tried and the rebuild command and exits 2 (it looks in .venv-skills of the clone, then of the main clone from a worktree, and uses a python from PATH only with -AllowPathFallback, or --allow-path-fallback for the .sh launcher); the Repository panel shows the same interpreter, or the reason there is none; a port held by something else names the holding PID and exits 1, because two dashboards showing two different snapshots is worse than none; and a dashboard already running is reported with its URL rather than started twice.

The command is a convenience, never the only way in. rt-dashboard.ps1, rt-dashboard.sh, rt-dashboard.bat and the VS Code task all reach the same launcher without Claude Code.

File: .claude/commands/rt-dashboard.md


← 09 Harnesses and mirrors | Table of contents | 11 Local coding and the cloud →