config.example.yaml

# Teich configuration
#
# Quick start:
# 1. Choose agent.provider: codex, pi, claude-code, hermes, or chat
# 2. Set model.model to the model you want to run
# 3. Keep prompts in prompts.jsonl, or add inline prompts below
# 4. Prefer TEICH_API_KEY / OPENROUTER_API_KEY / OPENAI_API_KEY in your environment for secrets
# 5. Run: uvx teich generate -c config.yaml

agent:
  # codex = Codex CLI in Docker
  # pi = Pi coding agent in Docker
  # claude-code = Claude Code CLI in Docker
  # hermes = Hermes Agent CLI in Docker
  # chat = direct text-only dataset generation via an OpenAI-compatible API
  provider: pi

  # Codex-only: use your ChatGPT subscription instead of an API key. Teich seeds
  # one auth.json snapshot from your host login ($CODEX_HOME/auth.json or
  # ~/.codex/auth.json) into auth_dir, then runs a host-side broker that owns the
  # rotating refresh token. Each container gets its own seeded auth.json (refresh
  # token swapped for a per-run secret) and refreshes through the broker, which
  # single-flights rotation, so any max_concurrency is safe.
  # NOTE: the first rotation invalidates your host `codex` login; run `codex
  # login` again afterward to restore it. auth_dir holds live credentials, so
  # Teich keeps it out of output/sandbox/failures and gitignores it.
  # codex:
  #   use_host_auth: true
  #   host_auth_file: null            # default: $CODEX_HOME/auth.json or ~/.codex/auth.json
  #   auth_dir: ./.teich/codex-auth

  # Claude Code-only settings.
  # Subscription auth (Pro/Max): create a long-lived token with `claude
  # setup-token` on the host and export CLAUDE_CODE_OAUTH_TOKEN (or set
  # oauth_token below). When a token is available and api.base_url is unset,
  # Teich passes it into each container and withholds ANTHROPIC_API_KEY (an API
  # key silently wins over subscription auth inside Claude Code and would bill
  # the API). Usage counts against your plan's rate limits, not API credits,
  # and the token is safe at any max_concurrency.
  # fallback_model forwards as --fallback-model (a model or list, tried in
  # order on overload); always_thinking writes alwaysThinkingEnabled into the
  # container's ~/.claude/settings.json; max_thinking_tokens sets
  # MAX_THINKING_TOKENS in the container (0 disables thinking where allowed).
  # claude:
  #   oauth_token: null               # prefer CLAUDE_CODE_OAUTH_TOKEN in the env
  #   fallback_model: [sonnet, haiku]
  #   always_thinking: true
  #   max_thinking_tokens: null

  # Trace each agent session to Langfuse (https://langfuse.com). Works for Codex
  # and Claude Code -- each uses its own native integration (Codex plugin, Claude
  # Code Stop hook) and Teich passes the credentials into the container. Side-
  # channel only: it reads transcripts, fails open, and doesn't change agent
  # behavior or Teich's output files. All three credentials are required when
  # enabled. base_url may be Langfuse Cloud or a self-hosted URL; for a host-local
  # instance use http://host.docker.internal:<port>.
  # langfuse:
  #   enabled: true
  #   public_key: pk-lf-...
  #   secret_key: sk-lf-...
  #   base_url: https://cloud.langfuse.com

model:
  # Model id passed to the selected agent/provider.
  model: deepseek/deepseek-v4-flash

  # Codex approval behavior. Common values: never, on-request.
  approval_policy: never

  # Codex sandbox mode.
  sandbox: danger-full-access

  # Optional reasoning / thinking level.
  # - Codex: forwarded as model_reasoning_effort
  # - Pi: normalized to low / medium / high when supported
  # - Claude Code: forwarded as --effort (low | medium | high | xhigh | max on
  #   supported models)
  # - Hermes: model/provider specific
  # Hermes also enables built-in toolsets:
  # safe,terminal,file,skills,memory,session_search,delegation
  reasoning_effort: medium

  # Optional Codex reasoning-summary detail, forwarded as model_reasoning_summary.
  # Values: auto | concise | detailed | none. Set to "detailed" to capture rich
  # reasoning summaries in traces (Codex's default can emit empty summaries).
  # Leave null to use Codex's default.
  reasoning_summary: null

  # Optional Codex service tier. Set to "fast" to enable Codex fast mode: it
  # runs ~1.5x faster at a higher credit rate and requires ChatGPT subscription
  # auth (agent.codex.use_host_auth) plus a supported model such as gpt-5.5 or
  # gpt-5.4. Leave null for the standard tier.
  service_tier: null

  # Optional context length override for providers that support it.
  # Useful for Hermes custom endpoints when /v1/models reports a served
  # context below Hermes' minimum but the endpoint supports a larger window.
  context_length: null

  # Optional Pi-specific provider overrides.
  # Teich already defaults maxTokens to 131072 even if this block is omitted.
  # Uncomment to customize the Pi/OpenRouter model entry further.
  # pi_model_overrides:
  #   maxTokens: 131072
  #   compat:
  #     maxTokensField: max_tokens

# Optional API/provider configuration.
# Leave this section commented out to use the runtime defaults.
#
# Common setups:
# - OpenAI: provider=openai
# - OpenRouter: provider=openrouter, base_url=https://openrouter.ai/api/v1
# - Local OpenAI-compatible server: provider=openai, base_url=http://localhost:1234/v1
#
# You can also put secrets in env vars instead of committing them here:
#   TEICH_API_KEY=...
#   OPENROUTER_API_KEY=...
#   OPENAI_API_KEY=...
#   TEICH_BASE_URL=...
#   TEICH_PROVIDER=...
#   TEICH_MODEL=...
api:
  provider: openrouter
  base_url: https://openrouter.ai/api/v1
  api_key: null
  # Pi uses OpenRouter through chat/completions even if this is set to responses;
  # the Responses adapter can stall before the first native session event.
  wire_api: responses

# Optional MCP servers exposed inside the agent runtime.
# mcp_servers:
#   - name: filesystem
#     command: npx
#     args: ["-y", "@modelcontextprotocol/server-filesystem", "/workspace"]

# Prompts can come from a file, inline below, or both.
# Relative paths are resolved from the location of this config file.
# JSONL is recommended because it safely supports multiline prompts, prompt-level system messages, and follow_up_prompts.
# CSV is still supported, but prompt files with commas/newlines are easier to maintain as JSONL.
prompts_file: prompts.jsonl

# Optional inline prompts. Use objects when you need metadata or chat follow-up turns.
# prompts:
#   - prompt: "Draft a compact project plan"
#     system: "Answer as a concise project manager."
#     follow_up_prompts:
#       - "Revise it for a solo developer"
#       - "Add a risk checklist"
prompts: []

output:
  # Where generated .jsonl files are written.
  # - codex / pi: normalized copies of native agent session traces
  # - claude-code: native Claude Code transcript JSONL from .claude/projects/...
  # - hermes: native Hermes session export JSONL from generated runs, including delegated subagents
  # - chat: text-only training rows with messages/prompt/response/thinking fields
  traces_dir: ./output

  # Where Teich stores workspace snapshots for each run.
  sandbox_dir: ./sandbox

  # Where Teich stores failed or interrupted traces for debugging.
  # These files are excluded from resume detection, conversion, README generation, and uploads.
  failures_dir: ./failures

  # Used in the generated trace README.
  pretty_name: "My Agent Traces"

# Optional direct upload to a Hugging Face dataset repo.
# Tags are auto-generated from the provider and model.
publish:
  repo_id: null
  hf_token: null
  private: false

# Number of prompts to run in parallel.
max_concurrency: 1

# Per-session timeout in seconds.
timeout_seconds: 600

# Legacy global API key field.
# Prefer env vars or api.api_key above for new configs.
openai_api_key: null

# Optional developer/system instructions injected into every agent run.
# Applied across codex (developer_instructions), claude-code and pi
# (--append-system-prompt), and hermes (AGENTS.md) — additive, so the agent's
# built-in base prompt is preserved. Handy for nudging chain-of-thought into the
# trace, e.g.:
#   developer_instructions: |
#     Think out loud so your problem-solving process is visible. Before each tool
#     call or edit, briefly explain what you're doing and why; after a command or
#     test runs, state what you concluded before the next step.
developer_instructions: null