DEV Community

Hector Flores
Hector Flores

Posted on Edited on Originally published at htek.dev

Copilot CLI Weekly: Workflows Get Guardrails, Admins Get Regex

A workflow that can launch copies of itself without limit isn't a feature. It's a fork bomb with a billing address. Copilot CLI just closed that hole, and it's the most important line in a week that shipped four stable releases: v1.0.109 through v1.0.112, with v1.0.113 prereleases already queued behind them.

I read the release notes so you don't have to. Most entries are terminal polish, but three themes matter: agent safety, enterprise policy, and the CLI becoming a place you live in all day.

What Shipped

Release Date Headline
v1.0.109 Nov 10 Stable cut of the prior week's work
v1.0.110 Nov 12 Experimental masked password prompts, /compact credit limits
v1.0.111 Nov 13 Regex permission rules, workflow self-spawn fix
v1.0.112 Nov 13 /subagents fixes, sandbox docs, /copy formatting

Guardrails for Agents That Run Agents

In v1.0.111, with experimental dynamic workflows enabled, a workflow's own agent can no longer start or resume workflow runs. The release note says it plainly: a workflow can no longer start copies of itself without limit when all tools are allowed.

That's the failure mode I warn about whenever autonomous loops come up. If you hand an agent every tool, including the one that launches agents, runaway recursion is a matter of time. If you're building multi-agent setups, my guide to GitHub Agentic Workflows covers why a bounded blast radius is the whole game.

Two related changes landed in v1.0.110:

  • The workflow run tool no longer tells the model that a user reviews every run's limits when all tool permissions are pre-approved. The model was being told something false. That's a prompt-honesty bug, and it matters because agents act on what they're told.
  • /compact now respects the session AI credit limit, so compaction can't push you past your cap.

Every autonomous capability needs a hard ceiling that the agent itself cannot raise. Credit limits and self-spawn blocks are exactly that.

Enterprise Policy Gets Real Matching

v1.0.111 adds regular expression support to enterprise-managed Shell, Bash, and PowerShell permission rules. Anyone who has tried to allowlist git subcommands with prefix matching knows the pain: you either allow too much or maintain a sprawling list.

Regex lets an admin write one rule that permits git status and git diff but not git push --force. The sharp edge, in my opinion: regex rules are easy to write wrong. An unanchored pattern will match more than you intend, so test policies against adversarial command strings before rolling them out.

Also in v1.0.111:

  • --host-allow-session-creation lets remote clients create sessions on a CLI host for one launch, without changing saved settings. That's the right shape: opt-in per launch, never persisted.
  • Hook approval dialogs show control characters in tool names and arguments as visible text instead of hiding them.

And in v1.0.110, the / picker no longer lists or runs a skill for a word that the skill's description hides inside a terminal escape sequence.

That's a theme worth noticing. Terminal escape sequences are an injection surface: text can be hidden from the human while the model still reads it. v1.0.112 continues this: opening a sub-agent's view no longer writes escape sequences or control characters from its task prompt to your terminal. Three releases in a row hardening the same class of problem tells me the team takes it seriously. It lines up with what I've argued in my piece on context engineering: everything in the context window is attack surface.

Passwords, Sandboxes, and Living in the Terminal

Masked password prompts

v1.0.110 adds an experimental feature that answers git, ssh, and sudo password prompts from agent commands in a masked dialog. Previously a command that wanted a password could simply stall. Enable it with:

copilot
> /experimental on
Enter fullscreen mode Exit fullscreen mode

I'd still prefer SSH agents and credential helpers over typing secrets at all, but stalling forever was worse.

Sandbox clarity

copilot help sandbox now documents that on macOS, sandbox host rules block direct TCP connections to local ports. A build client reaching its daemon is the example. Localhost stays reachable only through the local proxy, as HTTP or a CONNECT tunnel from a proxy-aware client. If a build daemon mysteriously fails inside the sandbox, that's why.

Workflow sign-in options

copilot workflow run --help now lists a Microsoft 365 account among the sign-ins that can run workflows, alongside usage-based billing and a custom model provider (BYOK) while signed out of GitHub. The CLI keeps loosening its coupling to a GitHub login.

Sessions and polish

The remaining volume is the Sessions sidebar, /usage, /plugin, and /mcp getting the unglamorous work that makes a TUI trustworthy:

  • /rewind with Conversation + files no longer reports a file as restored when later turns had already put it back or deleted it again.
  • Follow-ups typed while the agent is working now run in the order you typed them.
  • The /usage activity graph counts only prompts you typed, not sub-agent or autopilot prompts.
  • MCP OAuth no longer fires duplicate reconnects, and canceled authorization leaves servers in needs-auth.
  • /session prune skips sessions open in other processes and reports how many it skipped.
  • /context shows loaded custom instructions as their own token count rather than burying them in System Prompt.

That /context split is quietly useful. If you've been tuning instruction files, as in my custom agents mistakes list, you can now see what they actually cost you.

What It Means

Look at the shape of this week. Very little of it is new capability. It's all about control: who can run what, how many times, at what cost, with what visible to the human. That's a tool shifting from "impressive demo" to "something an enterprise can actually permit."

If you've followed the extensibility story in my extensions complete guide, the plugin fixes this week fit the same pattern of honest status reporting: /plugin no longer shows a plugin that failed to load as enabled, and a failed copilot plugin enable now names config.json instead of pretending it worked.

My opinion: a stable build nearly every working day is a double-edged sword. It's fast for fixes, but pin your enterprise rollout to a version and read release notes before upgrading. Compare that to the feature-burst pace I covered in the biggest week yet; this is the consolidation phase.

Bottom Line

Update to v1.0.112, check your managed permission rules against the new regex support, and try masked password prompts in a throwaway session first. The v1.0.113 prerelease already lets administrators disable built-in Browser Use through the features.browserUse managed setting, including in running sessions. That signals where this is heading: every capability the CLI gains will ship with an admin kill switch attached.

Top comments (0)