From idea to production, with receipts.
The ten-minute tour. Every output below is the real CLI; nothing is mocked.
step 01Install into the repo you are already in
cd /path/to/your/repo npx arsdlc init # skills + instructions + commands, detected for your harness # skills only, through the open skills CLI: npx skills add AnshRoshan/arsdlc
This writes the skills into your harness's skill directories, a docs/ process tree and an arsdlc.json marker. Claude Code can take the plugin instead: /plugin marketplace add AnshRoshan/arsdlc, then /plugin install arsdlc@arsdlc, which also brings the guard mod and the /arsdlc-next, /arsdlc-status and /arsdlc-report commands. Every arsdlc command below is that same binary, whichever way you got it.
step 02Open a feature with a lane
arsdlc new "Payment retries" --kind backend --lane standardPick a lane: spike (a question), quick (bounded change), standard (new capability), regulated (money, PII, auth). The lane decides which gates apply, and it only moves up. arsdlc kinds prints what each kind has to verify.
step 03Talk to your agent
use arsdlc-brainstorm, I want to build payment retries
The brainstorm skill classifies the work, interviews you with a recommended answer beside every question, and routes to arsdlc-discover then arsdlc-spec. The spec is EARS format and you sign §5. Nobody signs it for you.
step 04Let the gates referee
arsdlc next payment-retries --json
state: BUILD · owner: human · blocking: G1 approval (0/1 on current plan hash)Every stage ends with this. It reads the files, derives the state, names the blocking items and the one skill to load. The agent obeys the disk, not the chat, so a fresh session costs nothing to orient.
step 05Approve as a human
arsdlc approve G1 payment-retries --by "Ansh" ✗ segregation of duties: "Ansh" is the author of payment-retries
Approvals are bound to the hash of the artifact. Edit the plan afterwards and the approval goes stale and the state moves backwards. Agents cannot approve at all, and the author cannot approve their own work.
step 06Verify, do not trust
arsdlc evidence payment-retries --label full -- npm test … 41 passed (41), exit 0, 8.2s evidence/2026-10-04T09-12-03-full.json recorded
arsdlc evidence runs the command itself and records the exit code and output hash. A test run the agent watched instead of recorded is not evidence, which is exactly what the guard mod holds in Claude Code.
step 07Get a second model over the diff
arsdlc gate G4 payment-retries ✗ review ran on a different model than wrote the code both on opus
G4 reads Author model and Reviewer model out of acceptance.md and fails a same-family pair. If only one family is configured, the honest line is Cross-model: degraded (why), and the gate prints that it is not the guarantee.
referenceThe command set
| arsdlc init | install skills, instructions, slash commands and read-only subagents, detected for your harness |
| arsdlc doctor | what is installed where, which companions, which test gate |
| arsdlc skills / kinds | list the bundled skills; print each kind's verify model and each lane's gates |
| arsdlc new | open a feature with a kind and a lane |
| arsdlc next / status | derive state from disk: blockers, next task, the skill to load; or every feature at once |
| arsdlc check | validate a spec (EARS) or a delta file |
| arsdlc trace | requirement to task matrix |
| arsdlc gate / gates | run one of G1-G6, or all six |
| arsdlc scan | secrets over tracked files |
| arsdlc report | what the evidence and approval logs say: runs, pass rate, staleness, where the feature waited |
| arsdlc evidence | run a command and record the run as evidence |
| arsdlc approve | human gate approval (G1, G4, G5); refuses agents and refuses the author |
| arsdlc lane | show or raise a feature's lane; lowering needs --force with a note |
| arsdlc audit | approval chain integrity plus the evidence log |
Next: all 20 skills, or read the README. Every claim on this page is checked in CI against the CLI's own command table, so a command that stops existing fails the build.