ADAMANT Agent Engineering

ADAMANT · portable · agent-ready · evidence-first

Build at the level
the repository requires.

A working control system for coding agents: ten task skills, one dependency-free verification CLI, hash-chained evidence, isolated worktrees and real security gates.

Read the playbook Explore the skills Download the kit
sovware · main verification

$ ./.agents/bin/eng verify --profile full

check ok 2.35s exit=0

gitleaks ok 0.06s (0)

semgrep ok 1.53s (0)

osv-scanner ok 0.50s (0)

actionlint ok 0.03s (0)

PASS — evidence written

.agents/evidence/20261029T202043Z-791696ef.json

sha256:5273768155a6a471…

$

The gate is only green when the command actually ran.

01 — The loop

Understand. Plan. Prove.

Six phases, one rule: a skipped or blocked check is never a pass. The loop is deliberately small enough to run on every change.

02 — Operating principles

The rules behind the rules.

Seven cross-cutting commitments. They apply whether the task is a one-line fix, a kernel change or a release.

03 — Skills

Load the task, not the whole manual.

Ten entrypoints with activation conditions, permissions, workflow, verification and completion criteria. Open one to read the full source.

04 — Build paths

Start with the outcome.

Choose the work. Get the ordered steps, the skill to load, the commands to run and the evidence each step must leave.

05 — Toolbox

Composable tools, explicit limits.

What is adopted, what is conditional, and what is still a candidate. No tool is added because it is fashionable; each has a role and a failure mode.

06 — Evidence

A record, not a vibe.

Every verification writes one hash-chained record: exact commands, exit codes, durations, output hashes, Git state, config hash and scanner versions.

verdict: passfull · 5.44s
eng-evidence/v1 
{
  "commands": [
    { "name": "check", "status": "pass", "exit_code": 0 }
  ],
  "security": [
    { "tool": "gitleaks", "status": "pass", "findings": 0 },
    { "tool": "semgrep", "status": "pass", "findings": 0 }
  ],
  "git": { "head": "791696e", "dirty": false },
  "verdict": {
    "status": "pass",
    "required_failures": [],
    "unverified": []
  },
  "sha256": "5273768155a6a471…"
}
./.agents/bin/eng evidence verify chain intact
pass
Ran and succeeded

Every required command exited 0; required scanners completed with no findings.

fail
Ran and failed

A required command, timeout or security finding. The failing log stays with the record.

blocked
Gate unknown

A required scanner or config is missing. Blocked is not a pass and never becomes one.

skipped
Not applicable or absent

Reported as unverified. Optional scanner absence is visible, not silent.

error
Untrustworthy result

The tool ran but did not produce a report. Treated as unknown, not green.

Read the operating manual

07 — Repositories

Different trees, one contract.

Current rollout status. A blocked or review state is information, not a failure to hide.

08 — For agents

Give the next agent the shortest correct path.

The portal is also a machine-readable handoff. Point an agent at agents.json or llms.txt; download the same kit a human installs.

{ }

agents.json

Schema, start commands, skills, paths, tools, guardrails, repository states and limits.

 ≡

llms.txt

A compact brief with the commands, all ten skills, the adopted tools, the rules and the limits.

 ◇

site.json

The full portal payload: phases, principles, skill bodies, tools, paths and repo status.

 ↓

eng-kit.zip

The installable kit: CLI, skills, references, configs, hooks, CI template and manifest.
Agent handoff
Read AGENTS.md first.
Then read .agents/README.md and .agents/references/change-lifecycle.md.
Run ./.agents/bin/eng detect and ./.agents/bin/eng verify --profile fast.
Before editing shared code:
  ./.agents/bin/eng impact --base origin/main
Never report a skipped or blocked check as passing.

Known limits