01 / signal
Codex is getting louder
The interesting part is not only the feature. It is the social reinforcement: more AI builders appear to be shifting serious workflows toward Codex.
2026-05-17 / AGENT CONTROL PANEL
Steer While It Runs
The workflow is shifting from prompt-and-wait to live direction: start clean, redirect mid-run, queue the next job, background side work, preserve context, and turn repetition into Chronicle-supported routines.
session-control.sh
/new
01 / fresh session
/steer
02 / redirect without stopping
/queue
03 / stack the next task
/goal
04 / unlock long loops
/background
05 / run side tasks
/compress
06 / save tokens
user notices wrong target → /steer “Actually make it for Hermes not OpenClaw” → task continues
“Old way was to stop, retype the whole thing, and lose context. /steer way is to nudge it mid-run: ‘Actually make it for Hermes not OpenClaw’ → task continues, corrected. It’s like editing a chef’s recipe while he’s still cooking.”
CHRONICLE
Repetition Wants a Memory Layer
Codex Chronicle landed as a practical signal: repeated tasks should not keep rebuilding their scaffolding from scratch. Chronicle is the durable layer; slash commands are the live control surface.
02 / scaffold
Loops deserve context
Repetitive personal workflows — coding, YIL drafting, betting-board analysis, QA — are candidates for durable context instead of copy-pasted setup.
03 / question
What goes first?
The page should keep the audit open: which recurring Dan workflow should become Chronicle-backed before the rest?
OLD MODEL
Stop. Retype. Hope.
- Notice halfway through that the prompt was wrong.
- Kill the run and lose the useful partial context.
- Rewrite the whole request with the correction baked in.
/STEER MODEL
Nudge. Continue. Correct.
- The run keeps its working memory.
- The human stays in the loop without restarting the loop.
- Correction becomes part of directing, not evidence that prompting failed.
SESSION STACK
A Vocabulary for Long-Running Agent Work
/new
fresh session
/steer
redirect without stopping
/queue
stack the next task
/goal
unlock long loops
/background
run side tasks
/compress
save tokens
WORKFLOW
From One-Shot Prompts to Directional Systems
The day's synthesis is less about any one command and more about the shape they reveal together: live direction, queued intent, background execution, compressed context, and reusable memory.
The useful move is to preserve the open questions instead of pretending the workflow is already solved. Where does `/steer` change behavior first: coding tasks, writing and YIL drafting, betting-board analysis, or QA workflows?
Follow-up to keep alive
What should the “full stack” pattern for Hermes sessions be when `/new`, `/steer`, `/queue`, `/goal`, `/background`, and `/compress` are used together?
Design Notes
YIL 2026-05-17 — Steering While the Agent Is Still Cooking
A day about the move from one-shot agent prompts to durable, interactive agent work: Codex Chronicle for repetitive tasks and Hermes slash commands for steering, queueing, backgrounding, and compressing live sessions.
accent-1 / Chronicle
accent-2 / Steering
accent-3 / Async work
accent-4 / Compression
This draft uses an agent control-panel surface because the captured learning is about a user directing a live worker rather than submitting a perfect one-shot prompt. The command rail is the visual language; `/steer` is centered because it reframes mid-run correction as normal collaboration. Chronicle sits above the command stack as the durable memory layer for repeated workflows, while the follow-up questions remain unresolved by design.