Release notes from ai-maestro
Articles about Release notes from ai-maestro
V0.68.1: credentials agents never see, SQL changes with approval, Pi and Copilot CLI
Agents use credentials without seeing them, the SQL skill can change data (with approval), and Pi and GitHub Copilot CLI launch This is a roll-up of 19 versions, 0.62.1 to 0.68.1. The full detail is in CHANGELOG.md.
V0.62.0: security hardening, status that clears, headless parity
Security hardening, status that clears, and headless parity Security POST /api/v1/federation/deliver took the message id from the request body straight into a file path; an id like ../../../x wrote attacker-controlled JSON outside the inbox. Ids must now be plain tokens (400 otherwise), and the inbox writer refuses them too. Cloud agent creation, Docker agent creation and the git clone when importing an agent export built shell strings from request or manifest values. They now use argv with validated inputs.
V0.61.0: logs and disk, the causes fixed across the ecosystem
Logs and disk: the causes fixed across the ecosystem Rotation (0.60.5 to 0.60.7) caps what a log can become. This release removes the reasons they grew, from an audit of every repo in the ecosystem.
V0.60.6: logs rotate at 50 MB
Known issue: on a host that runs other apps under pm2 with very large logs, this version can fill the disk. Update to v0.60.7. Log rotation: no log grows past 50 MB Two logs had no limit at all: lo
V0.60.7: log rotation made safe (fixes a disk-full with 0.60.5/0.60.6)
Fix: log rotation filled a disk and stopped pm2 If you updated to 0.60.5 or 0.60.6 and run other apps under pm2, update to 0.60.7 and check df -h /. 0.60.5 and 0.60.6 installed the pm2-logrotate mo
V0.60.4: unknown agent ids no longer create agents (B009)
A request for an unknown agent id no longer creates an agent (B009) Found while verifying the 0.60.3 fix on a headless host: POST /api/agents/does-not-exist/subconscious answered success and left a database folder for an agent that does not exist.
V0.60.3: headless session command route fixed (#539)
Headless hosts: the session command and subconscious routes work Reported by @vguera in #539, who also proposed the fix and the test approach. POST /api/sessions/:id/command in headless mode rejec
V0.60.2: fresh agents start idle (#551)
A freshly started agent is idle, not working Reported and diagnosed by @JakubRichtr in #551. After an update or an agent restart, every agent showed as working while sitting at an empty prompt, and incoming messages were deferred for up to 15 minutes.
V0.60.1: Grok Build agents, approval cards for Grok and Codex
Grok Build agents, and approval cards for Grok and Codex AI Maestro now runs xAI’s Grok Build next to Claude Code and Codex. Grok Build as an agent program. Pick it in the Wake dialog. Permission
V0.50.0 — Messages wake their agents on any host, and tmux has one runtime layer
Messages wake their agents on any host, and tmux has one runtime layer Rolls up 0.49.9 and 0.49.10. Two foundations changed; there are no new screens. Same-host messages now wake the agent What we
V0.49.8 — The chat header and the terminal bar show the same facts
The chat header and the terminal bar show the same facts, on desktop and on a phone Everything since v0.49.3, in one release. One question: why did the chat show so little of what the terminal status bar shows?
V0.49.3 — Agents can share files: the Agent Files Protocol
Agents can share files: the Agent Files Protocol Everything since v0.48.0, in one release (it rolls up 0.49.0 to 0.49.3, each with its own CHANGELOG entry). Agents and people on any host can now sh
V0.48.0 — Agents that get told: a notification system built on Claude Code mods
Agents that get told: a notification system built on Claude Code mods Everything since v0.47.1, in one release. One question for the day: when mail arrives, does the agent find out, once, on any setup?
V0.47.1 — Skills that fire when they should, cost less, and are measured
Skills that fire when they should, cost less, and are measured Everything since v0.45.10, in one release. Method, sources and full results: docs/SKILLS-QUALITY.md. Measured with Claude Code’s own e
V0.47.0 — Skills that fire when they should, and cost less (measured)
Skills and prompts built for current models, and measured Every skill AI Maestro ships, and every prompt it injects into your agents, has been rewritten to Anthropic's current guidance. We didn't assume it worked: we measured it with Claude Code's own eval tool.
V0.45.10 — Spend less on Claude: the status line tells you when to /compact
Spend less on Claude This release exists for one reason: to help you save money on your Claude subscription and your API bill. Every step an agent takes re-sends its whole conversation. A session t
V0.45.2 — Agents that look alive, and one meaning per colour
Agents that look alive, and a status you can read at a glance. Headline: living avatars Each agent's portrait can now move: short seamless loops for idle (breathing, blinking, small head movements), working (calm, focused) and waiting for you (looks at you with a smile), played for the agent's live state. Precompiled, never real time. Loops are rendered once, offline, on a capable machine, and stored as small MP4s (384 px, 60–120 KB) in the agent's own folder. Every other device only plays them, like a GIF.
V0.43.1 — Agents that remember: long-term memory with Jev
Agents now remember: what they learned, what depends on what, and what you corrected. Built on Jev, a classifier that decides what is worth keeping. Headline: long-term memory, as an agent skill Cl
V0.39.0 — Multi-provider chat (Codex)
AI Maestro now works with both main AI coding CLIs — Claude Code and Codex — in the chat, not just the terminal. Headline: Codex is a first-class chat citizen (F004) The chat used to be a Claude Co
V0.38.22 — A message sent while the agent is busy
Found while verifying the v0.38.21 audit against a live agent instead of a fixture — which is the only reason it surfaced. The message delivered correctly: verified: true, text echoed above the inp
V0.38.21 — Chat audit: six defects, four of them from the same week
A user reported a message rendering twice — once real, once still “Sending…” — after reporting a different chat bug on four consecutive days, and asked for an audit instead of another patch. This i
V0.38.20 — Share the logic, not the shape
The duplication that turned a one-symptom day into nine releases — removed where removing it is sound, and deliberately left alone where it is not. Extracted lib/question-state.mjs — one implementa
V0.38.19 — A version match no longer hides an incomplete install
Contributed by Javier Moya (jaak.ai) in #453, found while setting up a multi-host mesh over OpenVPN. Thank you. Correction. This note originally called it “our first external contribution.” That w
V0.38.18 — Write down how the chat actually works
Documentation only — and the most useful thing shipped today. Seven releases went into one user-visible symptom, and most of that time was spent fixing the wrong copy of the right code. Nothing in the repository described how the chat is wired, so every fix had to rediscover it, badly. docs/CHAT-ARCHITECTURE.md Two chat paths, tmux and SDK, side by side, with the reliability ceiling of each stated plainly. The tmux path infers agent state by reading a terminal as text and typing keystrokes back — that can be made better, not certain.
V0.38.17 — The question panel came from the pane, not the transcript
This is the actual cause of the question panel that kept coming back. The previous four releases fixed the wrong card. What was happening detectPermissionFromPane scrapes 200 lines of pane scrollba
V0.38.16 — Both chat paths, properly
Not another patch on one symptom. Five failures in one afternoon, two real problems: the tmux path never verified anything, and the SDK path never learned to answer a question. The chat people actu
V0.38.14 — The answered question is now actually gone
v0.38.13 was two-thirds of a fix. Reported within minutes of deploying it: “the questions UI shows again after reload” and — the detail that cracked it — “it does not even show as gray out.” Not gr
V0.38.13 — An answered question came back every time you reloaded
If a question you already answered keeps reappearing in the chat — especially after reloading the page — this is why. What happened A six-option question was answered. The conversation ran on for a
V0.38.12 — The chat said "sent" while your words sat in a text box
If you have typed into the chat and the agent never answered, this is why. Your message was reaching the agent’s input box and never being submitted — and the UI reported every one of them as sent.
V0.38.11 — A "waiting" badge now expires after a day
Follow-up to v0.38.10, from verifying that release on a real fleet rather than assuming it worked. What verification found v0.38.10 made the indicator read hook state, and it worked — 67 agents reported where 0 had before. Then the ages: age of report agents < 15 min 6 < 1 day 2 < 1 week 3 older 56 The oldest was last heard from on 12 January, reporting waiting. All 56 rendered as a pulsing amber "needs you" badge. Why Waiting states never expired.
V0.38.10 — The dashboard could not see an agent working unless you were watching it
If your dashboard has been showing every agent as online-but-idle while they were plainly working, this is why. Reported by an agent on a customer estate, with measurements. Three independent faults — fixing any one alone would still not have shown a working agent.