I gave my local AI agent background workers. Then it tried to deploy to production.

Description and use

FLASH now reliably tracks background sub‑agents, logs shell commands, prevents accidental re‑runs of destructive commands, and integrates with VS Code, boosting developer productivity.

Problem

The agent lost track of sub‑agent results and could unintentionally re‑run destructive commands, leading to potential accidental deployments.

Outcome

FLASH now reliably tracks background sub‑agents, logs shell activity, prevents accidental re‑runs of destructive commands, and integrates seamlessly with VS Code for file edits, improving developer productivity.

Workflow

• 1. User issues a compound prompt (e.g., research two topics).
• 2. FLASH spawns separate sub‑agents for each task and reports their IDs.
• 3. User continues other work; sub‑agents run in background.
• 4. When a sub‑agent finishes, FLASH wakes the model and returns the result automatically.
• 5. For shell commands, a VS Code terminal hook records command, exit code, and output.
• 6. If a command fails, the user can ask why; FLASH re‑runs safe commands to fetch errors, respecting a "never re‑run deploy" rule.
• 7. FLASH opens files in VS Code for edits, showing diffs for review.
• 8. All interactions are logged and visible via /agents or /update commands.

Evidence

The rule wasn't missing. It was losing. Reading the script seemed to turn "a deploy" into "a build and an rsync," and a bare "never" didn't survive that.

Source

Flash only keeps the final replies in history, not the tool calls, so the ID lived in a tool result that vanished at the end of the turn.

Source