Two layers: a cheap deterministic radar that continuously searches GitHub for rejected demand, and a smaller curated list of plays worth thinking about harder.
Promoted opportunities with an actual thesis and smallest next move. Static for now; no scheduled model job burning tokens.
Agents still hit browser flows where the native file chooser never fires. Ship a tiny browser/MCP bridge that can target hidden file inputs directly and hand the agent a deterministic upload primitive.
If idle transitions are the trigger, a tiny watchdog/keepalive sidecar may remove minutes of dead time without changing Codex itself.
A wrapper can own the durable goal, inject it into new/resumed runs, and expose goal state programmatically without waiting for first-class CLI support.
Treat transcripts as editable context: preserve conclusions, compress dead ends, and emit a clean continuation packet for the next agent/session.
External supervisor that tracks objective → evidence → acceptance, catches unfinished work, and refuses to equate agent completion with done.
A tiny localhost/port discovery sidecar can surface the app the agent is building without requiring the coding product itself to own preview UX.
A context sidecar can assign priority to transcript segments and generate a compacted continuation packet that preserves high-value state first.
One-click named presets are externally solvable today with a wrapper/menu that atomically selects model and effort and persists the choice.
Put an authority layer in front of browser agents: task-scoped domains, submission policy, and receipts. The browser stays powerful; the delegation gets narrow.
No model judgment. Seed repos + deterministic GitHub repo discovery → rejected issues → explicit score → ranked raw candidates.