Getting started

Install, configure, and run your first coder session. Names verified in src/Command/*Command.php::defaultName() and config/panifex.php.

Install

Requirements: PHP >=8.2, CakePHP ^5.4.

composer require crustum/panifex
bin/cake plugin load Crustum/Panifex

No vendor:publish step. Configuration lives in config/panifex.php (Panifex.*, every value has a PANIFEX_* env override).

Provider and model

Anthropic (Claude) by default:

ANTHROPIC_API_KEY=sk-ant-...
PANIFEX_PROVIDER=anthropic
PANIFEX_MODEL=claude-sonnet-4-6

Any provider in crustum/cakephp-ai works. Unknown models need explicit rates so the budget stays meaningful (local example, zero cost):

PANIFEX_PROVIDER=ollama
PANIFEX_MODEL=qwen3-coder
PANIFEX_PRICE_INPUT=0
PANIFEX_PRICE_OUTPUT=0

Extra models go under Panifex.pricing.models (config/panifex.php). Switch per session with coder code --model=... [--provider=...], or mid-session with /model.

Verify

bin/cake coder health
bin/cake coder health --probe-provider

Checks config, API key, git repo state, and integration wiring. --probe-provider makes one minimal live model request. Before changing environments, commit or stash work. The agent edits files and clean git is the undo button.

First session

bin/cake coder code

Type a task, follow up naturally, exit to end. Be specific ("add published_at to Articles with a finder and migration", not "add a feature"). Review with git diff after each task.

Read-only exploration: bin/cake coder code --shell=off (nothing executes). Worktree isolation: --worktree / --no-worktree (production defaults on; clean up with bin/cake coder prune). Plan-first: --plan or /plan <task> (read-only plan, then Execute / Revise / Cancel).

Headless (CI/cron, no terminal): bin/cake coder run "task" --output=json. Same agent and safety layer, one run, JSON + exit code (0 ok, 1 error, 2 budget, 3 denied, 4 steps).

Project instructions and memory

Every agent loads PANIFEX.md from the project root (falls back to AGENTS.md, then CLAUDE.md, ignoring tool names it doesn't have). Scaffold one with bin/cake coder init (--force to overwrite): conventions, boundaries, gotchas. Capped at 20,000 chars; never overrides the safety layer.

Panifex.memory (PANIFEX_MEMORY, file | none): file (default) resumes the transcript on the next run (--session=name keeps separate histories; --session on coder run joins one, otherwise runs stay anonymous). Transcripts are JSON; /clear forgets, /compact summarizes, /sessions lists.

Application map

The agent runs inside the booted app, so it asks the framework instead of guessing from files. Tier one is an index in every system prompt (one line per table + route count); detail comes on demand via DescribeTables / DescribeRoute / bin/cake coder map:

bin/cake coder map                  # index
bin/cake coder map Articles         # one table in full
bin/cake coder map --all            # everything
bin/cake coder map --route=articles.view  # middleware, bindings, rules, auth
bin/cake coder map --fresh          # rebuild cache

Real columns/types off the live connection, casts, finders, behaviors, routes with expanded middleware. Schema and metadata only, never rows. Panifex.app_map (enabled / index / cache) stays on in every environment.

Configuration essentials

Panifex.shell (per APP_ENV, production off): off | allowlist | approve (default) | yolo. shell_allowlist (composer, npm, bin/cake). RunCake also needs cake_allowlist (bake, migrations, seeds, cache, routes, test locally) and cake_destructive entries always confirm. Budgets: budget_usd (PANIFEX_BUDGET, 1.00), max_steps (PANIFEX_MAX_STEPS, 40, a ceiling under each agent's own attribute). Context: max_tool_result_chars (48000), max_context_chars (600000). Prompt cache (Anthropic, on by default) covers prefix and conversation tail.

Next steps

• First session — the interactive agent in depth
• Tools — the 35 built-in tools and extending them
• Reference — troubleshooting and limits