submission.env.example

# ─────────────────────────────────────────────────────────────────────────────
# Submission runner config. Copy to `submission.env`, then choose exactly one route:
#
#   Model API in Docker:  make runner-pull && make preflight && make run
#   Coding-agent on host: make run-local LOCAL_BACKEND=codex \
#     LOCAL_REPO_DIR=... LOCAL_OUTPUT=... LOCAL_LOG_DIR=...
#
# Do not run Codex, Claude Code, or another coding-agent CLI through Docker. API runs use
# mini-swe/LiteLLM; CLI runs select their harness only with the Make variable LOCAL_BACKEND.
# (submission.env is gitignored — it may hold your API key; never commit it.)
# ─────────────────────────────────────────────────────────────────────────────

# ── REQUIRED ─────────────────────────────────────────────────────────────────
# Model identity recorded in the submission. For mini-swe it must be LiteLLM-routable:
#   anthropic/claude-...   openai/gpt-...   openrouter/...   gemini/...
#   openai/<model> together with API_BASE for an OpenAI-compatible endpoint
MODEL_NAME=openai/gpt-5.4
# API_KEY=...        # generic key for the model API route
# Prefer the provider's own env var? Use it INSTEAD of API_KEY, e.g.:
#   OPENAI_API_KEY=...   ANTHROPIC_API_KEY=...   OPENROUTER_API_KEY=...   GEMINI_API_KEY=...

# ── MODEL API ONLY (OpenRouter / gateway / local vLLM / Azure …) ─────────────
# Nothing is hardcoded to anthropic/openai. Reach any model with these:
# API_BASE=https://my-gateway.example/v1
# API_KEY=...          # generic key when no PROVIDER_API_KEY name fits
# MODEL_KWARGS={"custom_llm_provider":"openai","extra_headers":{"X-Org":"acme"}}
#   # ^ JSON of extra litellm.completion kwargs (api_version, temperature, …)
# MSWEA_COST_TRACKING=ignore_errors
#   # ^ required when your model isn't in litellm's price map (typical for custom
#   #   endpoints) — mini-swe otherwise aborts the session on the cost-calculation
#   #   error. Token usage is still recorded from API responses.

# ── LIMITS ────────────────────────────────────────────────────────────────────
# MAX_TOKENS=8192      # per-call output ceiling
# SUBMIT_LIMIT=100      # counterexample submissions across the complete run; accepted and
#                       # rejected calls both consume one attempt

# ── CODING-AGENT CLI AUTH (HOST `make run-local` ONLY) ───────────────────────
# Select the harness on the make command line: LOCAL_BACKEND=codex or
# LOCAL_BACKEND=claude-code. LOCAL_BACKEND is not an entry in this env file.
# For claude-code, auth comes from ONE of (the litellm key plumbing above doesn't apply):
# CLAUDE_CODE_OAUTH_TOKEN=...   # Claude subscription — generate with `claude setup-token`
# ANTHROPIC_API_KEY=...         # or plain API billing
# MODEL_NAME may be bare (claude-opus-4-8) or anthropic/-prefixed. API_BASE, MODEL_KWARGS
# and MAX_TOKENS are ignored. Claude Code self-terminates; no turn cap is passed.
# For codex, use an existing `codex login` or set OPENAI_API_KEY. MODEL_NAME may be bare
# (gpt-5.4) or openai/-prefixed. The runner uses an ephemeral workspace-write sandbox and
# ignores ambient Codex config/rules so local customizations do not change the benchmark.
# OPENAI_API_KEY=...
# CODEX_SANDBOX=workspace-write

# ── OUTPUT / LOGS ────────────────────────────────────────────────────────────
# Docker defaults follow. `make run-local` instead requires LOCAL_OUTPUT and LOCAL_LOG_DIR
# as separate explicit paths.
# OUTPUT=/out/submission.json  # authoritative submission output
# TRAJECTORY_DIR=/out          # where whole-repo persists its raw/final log;
#                              # default = OUTPUT's directory

# ── CUSTOM AGENT PROMPT (mount the files too: -v "$PWD/cfg:/cfg") ─────────────
# AGENT_STRATEGY_FILE=/cfg/strategy.md   # extra hints into the prompt's {{strategy}} slot
# AGENT_CONFIG=/cfg/config.yaml          # full prompt-config replacement

# ── VERSION PINS (baked into the image from PR_REF; override only to debug) ────
# EXPECTED_PRED_VERSION=    # empty string disables the version check
# EXPECTED_PRED_COMMIT=

# ── METADATA ─────────────────────────────────────────────────────────────────
# SUBMITTED_BY=your-hf-handle   # also fillable at submit time on the Space