Shared memory and context tools for agentic work.
Code Rooms
# m1nd Agent Skills
This directory is the installable agent doctrine for `m1nd`.
It is intentionally host-neutral:
- `m1nd-first` is the short rule set agents should load before repository work.
- `m1nd-operator` is the deeper operating manual for routing, recovery, `L1GHT`,
document binding, multi-agent coordination, and runtime refresh.
- `m1nd-guardian` is the M1ND-10 continuation doctrine: one active front,
proof-state separation, same-candidate receipts, stop rules, authority boundaries,
and cold-agent succession.
- `m1nd-operator/references/full-spec-agent-os.md` is the full operating layer:
a route table for the whole m1nd/L1GHT tool system and high-value tool
combinations.
- `m1nd-universal-agent-pack.md` is the portable prompt-pack form for hosts that
do not have a native skill directory.
The general agent pack carries the trained-agent loop measured in internal bug-hunt rounds:
`north(task)` as the in-session front door (one round-trip for trust, task
context, prior memory, sufficiency, and one `next_move`; `needs_ingest` → follow
the `recovery_playbook`, which names `ingest` only where the mutation policy
allows it → re-north), then respect the evidence limits (`closure`,
`trust_envelope`, `insufficient_evidence`), prove final truth directly, and
`memorize` the durable finding at close. Under the current
[calibration containment](../docs/CALIBRATION-QUALIFICATION.md), all v1 rows are
unqualified: prediction abstains, the seek envelope stays at `reverify` or
`unprovable`, and running calibration does not enable `act`. Recovery before absence,
workspace-binding clarity, and evidence logging stay part of the loop.
Use the npm installer from the repo root to install the right shape:
```bash
npm install -g .
m1nd install-skills codex
m1nd install-skills generic --project /path/to/project
```
For a published package, the same flow becomes:
npm install -g @maxkle1nz/m1nd@beta
m1nd init --host codex