Everything the model is told before you type a word. Assembled live, so a change to your coach notes shows up here on reload.
System prompt — click any section to read it
Recent turns — what it actually did
Each turn with its cost, latency and tool sequence. Expand one to see the tools in order, how long each took and how much came back.
Tool definitions — sent in full every turn
Ordered by cost. These are JSON schemas, not prose: they are what the model dispatches from. Writes are never executed directly — they become proposals you confirm.
Sources — what feeds the prompt, and where to change it
Every store below is injected into the system prompt above. They were
scattered and several were invisible: saved notes could only be written by
the coach and never read back, and day-patterns drove every
plan while being hidden from the UI. A store you cannot see is one you
cannot correct. Edits here take effect on the next turn — no deploy.
Notes — what these components are
An aide-memoire, stored as a workfile rather than in code, so it can be edited
without a deploy —
PUT /api/coach/harness-notes, or ask the coach to append to
harness-notes.