otx decompose

Decompose a git repository or local folder into a language-agnostic specification.

Synopsis

otx decompose <source> [options]

Arguments

Argument	Description
source	Git URL or local folder path (required)

Options

Option	Default	Description
--project, -p	auto-detected	Project name used in output paths and prompts
--start-phase	0	Phase to start from. 0 = fresh run. Higher values resume from a checkpoint.
--end-phase	run all	Last phase to run, inclusive
--orchestrator	from settings	AI engine: ClaudeCode | OpenAI | Ollama
--api-key	OPENAI_API_KEY env	OpenAI API key (never saved to disk)
--endpoint	from settings	Custom OpenAI-compatible base URL (Azure, LM Studio, etc.)
--thick-model	from settings	Model for heavy phases (Phase 3, 4)
--regular-model	from settings	Model for normal phases
--thin-model	from settings	Model for light phases
--max-turns	from settings	Max agent turns per phase before forcing output
--max-tokens	from settings	Max output tokens per call. 0 = auto per phase weight
--keep-clone	false	Keep the cloned repository directory after completion
--hints	(none)	Free-text domain knowledge injected into every phase prompt

Examples

# Decompose a remote repository with the default backend
otx decompose https://github.com/org/repo

# Decompose a local folder, explicit project name
otx decompose ./my-project --project my-project

# Run only the first three phases (index, structure, init flow)
otx decompose https://github.com/org/repo --end-phase 2

# Resume a failed run from phase 3
otx decompose https://github.com/org/repo --project my-project --start-phase 3

# Use OpenAI with specific models
otx decompose https://github.com/org/repo \
  --orchestrator OpenAI \
  --api-key sk-... \
  --thick-model gpt-4o \
  --regular-model gpt-4o-mini \
  --thin-model gpt-4o-mini

# Use Ollama
otx decompose ./my-project \
  --orchestrator Ollama \
  --thick-model llama3.1:70b \
  --regular-model llama3.1:8b \
  --thin-model llama3.2:3b

# Inject domain hints
otx decompose ./trading-engine \
  --hints "Uses event sourcing with CQRS. Idempotency keys prevent double-execution."

Output

Files are written to Output/Decomposition/<project>/ in the current working directory. See Output Files for the full list.

Phase 6 automatically imports the composition inventory into the SQLite database (DB/opentransmute.db).

Resume on Failure

Phase state is persisted to Output/Decomposition/<project>/job.json after every phase completes. To resume after a failure, pass --start-phase N where N is the first phase that did not complete successfully. Outputs from phases 0 through N−1 are reloaded from disk automatically.