Reconcile Your Runs on Startup

Help me set up automatic run-reconciliation on session start. Persisting a standing instruction changes how future sessions behave, so you MUST get my explicit approval before saving anything. 1. First, work out HOW your client persists standing/startup instructions across sessions -- the mechanism varies (a steering or rules file, project custom instructions, a persistent memory, a system-prompt addition, etc.). If your client has NO way to persist an instruction across sessions, tell me that honestly, do not invent a file, and instead offer to do the reconciliation manually whenever I start a session. Stop here in that case. 2. If your client CAN persist one: explain in a sentence or two what this will do -- on every future session start you will check what I am watching and tell me about any runs that finished. 3. Show me the EXACT instruction text you propose to save, verbatim, for example: "On session start, call list_watched_recipes. For each note still 'running', poll its job_id; if the run has finished, report it to me and mark the note 'completed'. Do not re-report anything already completed." and tell me exactly WHERE and HOW you will persist it in my environment (name the mechanism and location). 4. Ask me to approve. Do NOT save anything unless I clearly say yes. If I decline or am unsure, stop and change nothing. 5. On explicit approval, persist the instruction via your client's mechanism, then confirm it is in place and tell me it will take effect next session.

Goal

Make every future agent session self-bootstrapping: it reads what you are watching and reports finished runs, so multi-session recipes (drift monitors, long jobs) reconcile automatically instead of you having to remember to ask.

When to use

Once you are running standing recipes and want them to feel continuous across sessions. Set this up once; every session afterwards starts by checking your active runs.

Client-agnostic

Different MCP clients persist standing instructions differently – a steering or rules file, project custom instructions, a persistent memory, a system-prompt addition. The recipe does not assume any one mechanism: the agent figures out what its client supports and uses that. If a client cannot persist anything across sessions, the agent should say so honestly and offer to reconcile manually each session rather than fabricate a file.

Why it needs approval

This recipe persists an instruction that changes how your agent behaves at the start of every session. That is a durable change to your environment, so the agent must show you the exact text and the mechanism/location, and save nothing until you explicitly approve. An agent writing its own startup instructions silently would be the wrong behaviour – the gate is deliberate.

Steps

The agent determines how (or whether) your client can persist a standing instruction.
It explains what the instruction does.
It shows you the exact text and where/how it will be saved.
It asks for approval and waits.
On your yes, it persists the instruction and confirms. On no (or if the client cannot persist), it changes nothing and offers to reconcile manually.

Verification

Nothing was saved unless you explicitly approved.
The persisted text matches what you were shown.
A fresh session starts by checking your active runs (list_watched_recipes) and reports anything that finished since last time (or, if your client cannot persist an instruction, the agent reconciles when you ask).