v5.8.1: No More Frozen Agents — aios-shell Stall Fix and LLM-Judged Requirements Clarification¶
Two long-standing pain points get fixed in v5.8.1: coding agents (opencode, codex) freezing mid-command with no way to recover, and the workflow layer failing to notice when a request is actually vague.
The stall: long commands froze the whole agent¶
If you ran aios_shell with a command that took a while — a build, a test run, a multi-minute script — the agent went dead. No ping response, no cancel, no progress. You had to press Esc and send "continue" to get it moving again.
The root cause was architectural: the MCP server and the stdio proxy both processed JSON-RPC serially. One long-running command blocked every request behind it, including the notifications/cancelled message that Esc sends. The client literally could not cancel the command, and no response ever came back.
What changed¶
- The shell server main loop is now concurrent: ping, cancel, and other requests stay responsive while a command runs.
- The stdio proxy forwards concurrently too, so the proxy layer no longer blocks on an upstream command.
notifications/cancelledterminates the in-flight command by requestId immediately, instead of waiting for the timeout.- On Windows,
taskkill /T /Freaps the entire process tree — no more orphaned node/npm/git children aftercmd.exeexits. - Closing stdin cleans up all pending commands, so nothing is left hanging.
The aios-shell proxy chain is preserved: aios-mcp-proxy.mjs still attaches _meta.aios observation metadata and local refs. RTK and Caveman remain the client-side output compression (the proxy never actually compressed output — it forwards it unchanged, and SHELL_TOOL.description no longer claims otherwise).
As a safety net on top of the concurrency fix, generated MCP server configs now carry startup timeouts (startup_timeout_sec 60/30/30 for browser/auth/shell on Codex; experimental.mcp_timeout: 90000 injected for OpenCode).
The other fix: knowing when a request is vague¶
The workflow layer previously decided "this request is vague" with regex. Patterns like VAGUE_BEHAVIOR_PATTERN matched explicit wording — "优化一下" / "tweak the login logic" — but missed the far more common case: a request that names a concrete feature yet has no acceptance criteria, no scope, and no success definition. The regex fired only when the user happened to phrase things the pattern expected, so requirements clarification often never triggered and the agent built the wrong thing.
Grilling is now LLM-judged and in-execution¶
derive-facts.mjsno longer derives ambiguity from wording with regex.- The requirements capability activates only on a
grill/specintent (the LLM's semantic judgment) or a domain-vocabulary observation. rex-requirementswas rewritten around embedded in-execution grilling: ask one decision question at a time with a recommended answer, converge within three rounds, and only when you hit a real decision point. Grilling is not a front-loaded interrogation gate — the agent works first, looks up facts itself, and stops to ask only when a decision genuinely belongs to the user.- Because the workflow runtime re-selects the next capability at every stage boundary, clarification can be inserted mid-delivery and the original capability resumes afterwards.
The skill description is dual-trigger: the LLM may self-trigger on a vague, underspecified, or multi-interpretable request — even one that names a specific feature — and rex-harness may activate it as before.
Upgrade¶
aios update
No config migration is needed. Restart opencode/codex after updating so the new shell server and proxy take effect.
Also in this release¶
rex-code-reviewgained a scenario-based subagent acceptance mode: isolated, context-free acceptance runs covering normal, boundary, and abnormal scenarios, with evidence collected per finding.- All MCP config generators (Codex TOML, OpenCode JSON) now emit startup timeouts by default.