Aster Config

Configuring Aster (aster.yaml)

aster init writes aster.yaml at the repo root (or ~/.config/aster/aster.yaml with -g). Every field is optional. Precedence: CLI flags > shell env > aster.yaml > built-in defaults. API keys are NEVER read from this file; they come from ASTER_API_KEY or the key stored by aster init / aster login, so aster.yaml is safe to commit.

Review block

review:
  model: openai/gpt-4o-mini                      # fallback for stages with no override
  base_url: https://openrouter.ai/api/v1         # any OpenAI-compatible provider
  hypothesis_model: deepseek/deepseek-v4-flash   # cheap, high-recall pass
  verify_model: anthropic/claude-sonnet-5        # independent adversarial verify
  min_confidence: 0.6                            # drop findings below this (0.0-1.0)
  max_diff_bytes: 200000                         # cap the diff sent to the model
  analyzers: []                                  # [semgrep], [ast-grep], or both
  focus_areas: [correctness, security]           # bias the hypothesis pass
  include: []                                    # empty = everything except exclude
  exclude: ["target/**", "node_modules/**", "**/*.lock", "**/*.min.js"]

• Split models by stage: a cheap hypothesis_model for recall, a strong verify_model for precision. model fills any stage without an override. ASTER_MODEL env overrides all of it for a run.
• analyzers adds static-analysis backends whose findings also flow through verification: semgrep needs semgrep/opengrep on PATH; ast-grep needs ast-grep/sg on PATH plus ASTER_ASTGREP_RULES pointing at a rule file.

Permissions block

Gates what the agent edits, reads, and runs:

permissions:
  mode: edit              # plan | manual | auto | edit | yolo (prompts need the TUI; headless denies)
  allow: []               # e.g. ["Bash(cargo test:*)", "Edit(src/**)"]
  ask: []                 # e.g. ["Edit(migrations/**)"]
  deny: []                # e.g. ["Bash(npm publish:*)", "Edit(infra/**)"]
  use_default_rules: true # ask on .git/**, workflows; ask on sudo/rm/curl in `auto`; refuse secret reads

One rule language for all three tools: Edit(<glob>), Read(<glob>), Bash(<command>:*) for a prefix or Bash(<command>) for an exact line. A bare Edit, Read, or Bash covers everything that tool does. Precedence is deny, ask, allow, built-ins, then mode, so one allow entry overrides a single built-in without dropping the rest.

A Bash rule matches inside bash -lc "…", so chaining does not slip a command past it.

Keep use_default_rules: true unless you have a specific reason; it is what makes the agent confirm before touching CI workflows and git internals, and what stops it reading env and key files.

Environment variables

Env beats aster.yaml, and CLI flags beat env. Everything is optional except the API key.

Provider and models:

• ASTER_API_KEY - the provider API key (required; never read from aster.yaml)
• ASTER_BASE_URL - OpenAI-compatible endpoint
• ASTER_MODEL - fallback model for every stage
• ASTER_HYPOTHESIS_MODEL / ASTER_VERIFY_MODEL - per-stage overrides

Request tuning:

• ASTER_TIMEOUT_SECS / ASTER_MAX_RETRIES / ASTER_DEADLINE_SECS - HTTP timeout, retry count, overall deadline
• ASTER_MAX_TOKENS / ASTER_SEED / ASTER_REASONING_EFFORT - completion caps, deterministic seed, reasoning effort
• ASTER_LANGUAGE - language every reply is written in; unset follows the user
• ASTER_PRICE_PROMPT_PER_M / ASTER_PRICE_COMPLETION_PER_M - $ per million tokens, for cost reporting

Review:

• ASTER_ANALYZERS - comma-separated analyzer list, e.g. semgrep,ast-grep (empty = LLM only)
• ASTER_ASTGREP_RULES - path to the ast-grep rule file
• ASTER_VERIFY_CONCURRENCY - parallel verification workers
• ASTER_REPO - repo name used in the summary (defaults to local)

Conventions

• aster init -y writes the defaults with no wizard; --force overwrites an existing file.
• Deny-first: deny beats ask, which beats allow, which beats the built-ins.
• A headless run cannot answer a prompt, so anything reaching ask is refused there. Automation wants mode: edit with tight allow rules.