CLI in Action

Real CLI output from a production development session. No staging, no mockups, exactly what you see when DevArch is installed. Follow a complete workflow from project init through commit.

Setup: install and initialize

Run the installer you get with your purchase: one self-contained file, no git, no dependencies. Then one command per project.

install · init

~ % bash devarch-install.sh
[ok] ~/.devarch populated
[ok] Generated ~/.devarch/DEVARCH.md
[ok] Install complete! 'devarch' is on PATH.

~/my-project % devarch init
       AI-Assisted Development, Disciplined.  ·  v5.3.0
[ok] Created CLAUDE.md (imports the DevArch ruleset)
[ok] Written .devarch/descriptor.json
[ok] Project initialized!

Always start with bash devarch-install.sh. Downloading from the web strips the file's run permission, so ./devarch-install.sh fails with permission denied, and sudoisn't needed (it installs to your home folder, not the system). Running it through bash sidesteps both. On Windows, run it from your WSL shell.

Session start: recover context

Every session opens with a recap of what happened last time. No digging through git logs.

/devarch:recap

Searched for 1 pattern, read 1 file

Last session recap (2026-03-20): Massive architecture session - completed the embedding spike, wrote ADR-135 (bridge protocol, accepted), decomposed the engine into 368 implementation blocks, defined V1 scope (276 blocks, 75%), and restructured the repo into a clean Rust workspace.

Status: COMPLETE. Next step is Phase 20: define the pattern catalog schema and TOML format, starting with PUZ-001 as a working reference entry.

Audit: surface blockers before you start

The pre-session audit reads prior work, detects blockers, and forms an action plan, automatically.

pre-session-audit

● Previous session recap (2026-03-21, ~1 hour ago):

Audited the agents and found a gap: no agent enforces test-writing or doc comments. Ran session-planner to decompose the catalog into Phases 20–24. Phase 20 is marked CURRENT but not yet started.

Blocker: CLAUDE.md needs testing and documentation requirements codified before Phase 20 begins.

● Good, I have the full picture. Two things to do:
  1. Unblock: Update CLAUDE.md with Rust-specific documentation requirements
  2. Implement: Phase 20: types, TOML schema, PUZ-001, tests

Planning: decompose work into budgeted phases

Non-trivial goals get a plan before any code is written. Each phase has a budget, entry state, deliverables, and exit criteria.

session-planner · docs/context/plan.md

## Phases

### Phase 1: Template Restructuring
- Budget: 100 tool calls
- Entry state: Three monolithic templates exist.
- Deliverable: new base template + two capability add-ons
- Exit state: Three new files. Old three deleted.
- Status:  DONE

... 5 more phases, each with entry state, deliverables, and exit state

Implementation: build, then verify

After a phase completes, the audit confirms all deliverables are in place and the codebase is healthy.

phase completion + audit

● Phase 20 is complete. Here's what was done:

CLAUDE.md blocker resolved:
- Added Rust /// doc comments to the documentation standards list
- Added rule: #[deny(missing_docs)] in every lib.rs, doc comments on public items

Phase 20: Catalog Schema and TOML File Format:
- lib.rs: all catalog types; catalog.toml: PUZ-001 reference entry
- tests.rs: 10 tests covering load, fields, detection signals, codegen recipe

All 10 tests pass, full workspace compiles. Phase 20 exit criteria met.

● Pre-session audit check - both items now resolved. Clean bill of health otherwise.

Check in: standup in one command

Quick status check: what got done, what's next, any blockers. Works mid-session or at the start of a new one.

/standup

Standup
────────────────────────────
Done:
  - Patch 1.3.3: scripts-only commit agents, /fin removed, permissions tested.
Next:
  - No active plan. Pick up next task.
Blockers:
  - None

Lifecycle
Gate:     clear
Session:  session-20260323-2054-main.md
Plan:     no active plan

Ship it: summarize, commit, push

One command writes a work summary, commits everything, and pushes. The next session picks up right where you left off.

/fin

● work-summary-writer (Write session work summary)
  Done (8 tool uses · 36.6k tokens · 58s)

● commit-remote (Commit and push session work)
  Done (8 tool uses · 14.4k tokens · 41s)

● Committed and pushed as 78237f0.

Every example here is from a real DevArch session, the same discipline the methodology enforces, captured as it ran.