Edit these markdown files to change model instructions. Loaded at startup
(MANGO_PROMPTS_DIR overrides the folder). Restart the sidecar after edits.
XML-style tags (<identity>, <protocol>, …) help small models parse rules.
Critical deps path is redundant across agent + CoT + feedback.
| File | Used for |
|---|---|
agent_v2.md |
Default coding agent (thinking=off) |
agent_think.md / agent_deep.md / agent_max.md |
GUI thinking picker |
ask.md / plan.md / debug.md / refactor.md |
Slash modes |
roblox.md |
Roblox Studio mode |
epistemic.md |
Third-party API research sub-agent |
epistemic_codebase.md |
Local codebase research sub-agent |
cot.md / cot_chain_step.md / cot_chain_summarize.md |
Chain-of-thought |
summary.md |
Finish message after tests |
title.md |
Session title |
swebench.md |
SWE-bench issue fixing |
security_review.md |
Dataset / security-audit training |
feedback.md |
Runner replies after a tool turn |
Runner sole-forces (GBNF — model cannot emit write/edit yet):
declare_apisask_epistemicinstall_packages(only if imports missing; confirm popup — Deny unlocks write)- then
write_file
Task-prompt install lock is auto-bound after declare (no extra turn).
Greenfield Discord/CLI bots skip lock/race design-review thrash after a successful write.
| Level | Prompt file | Thought tokens | CoT chain / cycles |
|---|---|---|---|
off |
agent_v2.md |
128 | 0 / 0 |
think |
agent_think.md |
256 | 2 / 3 |
deep |
agent_deep.md |
384 | 3 / 5 |
max |
agent_max.md |
512 | 3 / 6 |