config/agent.env.example

# Optional Pi agent backend credentials (copy to config/agent.env — do not commit secrets).
#
# Loaded by tools/config.py when APP_TYPE=agent (or legacy "pi"); set automatically by
# agent-redact/shared/gradio_app.py, shared/start.sh, and docker-compose pi-agent services. Override with
# APP_CONFIG_PATH if needed.
# APP_TYPE=agent also skips eager spaCy/Presidio load at startup (the agent orchestrates redaction remotely).
#
# Local Windows/macOS dev: leave unset to use {repo}/workspace (created automatically).
# Docker: compose sets AGENT_WORKSPACE_DIR=/home/user/app/workspace
# AGENT_WORKSPACE_DIR=
#
# Per-user session subfolders: {AGENT_WORKSPACE_DIR}/{gradio_session_hash}/ (default on).
# Set false only if you want one shared workspace tree for all browser sessions.
# AGENT_SESSION_WORKSPACE=true   # per-browser subfolder under workspace/ (default when unset)
# AGENT_SESSION_WORKSPACE=false  # single shared workspace root (legacy)
#
# Skills: on startup, repo skills/ are copied to {AGENT_WORKSPACE_DIR}/.pi/skills/ (read-only).
# Pi RPC uses --no-skills and loads only that folder; cwd is the session subfolder.
# AGENT_SKILLS_RESYNC=true   # force refresh from repo on next start
# AGENT_SKILLS_WRITABLE=true # skip read-only chmod (local debugging)
#
# Gradio upload temp (must match GRADIO_TEMP_DIR; local default: workspace/.gradio_uploads)
# AGENT_UPLOAD_ROOT=
# GRADIO_TEMP_DIR=
#
# GuardDuty Malware Protection for S3 — stage Gradio uploads before Start redaction.
# Requires RUN_AWS_FUNCTIONS=True and a non-empty MALWARE_SCAN_S3_BUCKET (same as main app).
# SCAN_UPLOADS_FOR_MALWARE=False
# MALWARE_SCAN_S3_BUCKET=
# MALWARE_SCAN_S3_PREFIX=upload-staging/
# MALWARE_SCAN_POLL_INTERVAL_SEC=2.0
# MALWARE_SCAN_TIMEOUT_SEC=120.0
# MALWARE_SCAN_SHOW_CHECKING_INFO=True  # also shows success toast when scan is clean
#
# Pi Gradio listen port (default 7862 local-docker, 7860 hf-space). Not GRADIO_SERVER_PORT (main app).
# AGENT_UI_PORT=7862
#
# Monorepo root (skills/, config/). Usually auto-detected on Windows/macOS; Docker sets /workspace/doc_redaction.
# AGENT_WORKDIR=
#
# Host-only: path to pi RPC CLI if not on PATH (Windows: ...\AppData\Roaming\npm\pi.cmd)
# AGENT_EXECUTABLE=
#
# doc_redaction Gradio app for Pass 1 /doc_redact and review APIs (not the Pi UI above).
# Local dev: main app on 7860 → DOC_REDACTION_GRADIO_URL=http://127.0.0.1:7860
# Docker compose: http://redaction-app-llama:7860 (see docker-compose_llama_agentic.yml)
# DOC_REDACTION_GRADIO_URL=http://127.0.0.1:7860
#
# When the main app keeps COGNITO_AUTH=True, set a dedicated Cognito user for gradio_client
# (not the Pi UI user's session). Optional on Express+Pi: main backend auth is off by default.
# DOC_REDACTION_GRADIO_AUTH_USER=
# DOC_REDACTION_GRADIO_AUTH_PASSWORD=
#
# --- Intro / branding (Pi Gradio UI) ---
# AGENT_GRADIO_TITLE=Agentic Document Redaction
# AGENT_INTRO_TEXT=intros/pi_intro.txt
# AGENT_INTRO_TEXT supports {USER_GUIDE_URL} and {AGENT_GRADIO_TITLE} when loading from a .txt file.
#
# Used by the pi-agent service in docker-compose_llama_agentic.yml.
# UI overrides in the Gradio app apply for the current container session only.

# Deployment profile: local-docker (default) | hf-space | aws-ecs (split container from doc_redaction)
# AGENT_DEPLOYMENT_PROFILE=local-docker
# On AWS ECS, Pi and doc_redaction are separate services — outputs are downloaded via
# gradio_client HTTP, not by grepping /workspace on the Pi container (see doc-redaction-app §2f).
# ECS/CDK typically sets AGENT_DEPLOYMENT_PROFILE=aws-ecs, AGENT_DEFAULT_PROVIDER=amazon-bedrock,
# and AGENT_DEFAULT_MODEL=anthropic.claude-sonnet-4-6 via task env (overrides this file).

# --- Agent orchestration backends (pi default | langgraph | agentcore) ---
# AGENT_ORCHESTRATOR=pi
# AGENT_ORCHESTRATOR=langgraph
# AGENT_ORCHESTRATOR=agentcore
# AGENTCORE_RUNTIME_URL=https://your-agentcore-runtime.example
# See agent-redact/agentcore/README.md for deploy steps and how to obtain the URL.
# AGENTCORE_API_KEY=
# LANGGRAPH_REQUIRE_REVIEW_APPROVAL=true
# LangGraph ReAct step budget (default 150). Raise if you see
# "Sorry, need more steps to process this request."
# LANGGRAPH_RECURSION_LIMIT=150
# LangGraph context compaction (trim older tool history before each LLM call)
# LANGGRAPH_COMPACTION_ENABLED=true
# LANGGRAPH_COMPACTION_RESERVE_TOKENS=28672   # falls back to AGENT_COMPACTION_RESERVE_TOKENS
# LANGGRAPH_MAX_OUTPUT_TOKENS=8192
# Also passed as ChatOpenAI max_tokens for llama-cpp. Large write_workspace_text
# bodies can still break tool-call JSON on local models; the runtime retries once
# with a compact-script nudge when the server returns a tool-arg parse error.

# Arize AX / Phoenix tracing (LangGraph + Pi).
# - LangGraph: auto LLM/tool spans via openinference-instrumentation-langchain.
# - Pi: custom AGENT/TOOL spans from the RPC event stream (LLM HTTP stays in Node
#   pi — no per-message LLM spans unless you instrument the model server).
# Requires arize-otel (AX) or arize-phoenix-otel (Phoenix); LangGraph also needs
# openinference-instrumentation-langchain — see requirements_agent.txt.
# Smoke: set vars below, send one Gradio chat turn, confirm spans in AX or local
# Phoenix; send a follow-up in the same browser session and check Sessions for a
# shared session.id (= Gradio session_hash).
# With ARIZE_TRACING_ENABLED unset, no collector traffic.
#
# Process KPIs to watch in Phoenix/AX (see agent-redact/eval/README.md):
#   - tool error rate / identical-error streaks
#   - turn latency p50/p95 (and per-tool: doc_redact vs review_apply)
#   - trajectory completeness (doc_redact … review_apply; workflow_incomplete)
#   - optional AgentCore Evaluations when AGENT_ORCHESTRATOR=agentcore
# ARIZE_TRACING_ENABLED=true
# ARIZE_BACKEND=ax
# ARIZE_BACKEND=phoenix
# PHOENIX_COLLECTOR_ENDPOINT=http://localhost:6006
# # Docker agent → Phoenix on host: http://host.docker.internal:6006
# PHOENIX_API_KEY=
# ARIZE_SPACE_ID=
# ARIZE_API_KEY=
# ARIZE_PROJECT_NAME=doc-redaction-langgraph
# ARIZE_ENDPOINT=europe
# ARIZE_ENDPOINT=us

# Default Pi orchestration backend: llama-cpp | google-gemini | amazon-bedrock
AGENT_DEFAULT_PROVIDER=llama-cpp
# AGENT_DEFAULT_PROVIDER=amazon-bedrock
# AGENT_DEFAULT_MODEL=anthropic.claude-sonnet-4-6
# AGENT_DEFAULT_MODEL=unsloth/Qwen3.6-27B-MTP-GGUF

# Default redaction task settings (Gradio UI + agent prompt; see agent-redact/shared/redaction_prompt.py)
# AGENT_DEFAULT_OCR_METHOD=hybrid-paddle-inference-server
# AGENT_DEFAULT_PII_METHOD=Local
# pi-bedrock compose profile overrides: AWS Textract service - all PDF types / AWS Comprehend
# pi-gemini compose profile overrides: tesseract / Local

# --- HF Space profile (AGENT_DEPLOYMENT_PROFILE=hf-space) ---
# AGENT_DEFAULT_PROVIDER=google-gemini
# AGENT_DEFAULT_MODEL=gemini-flash-latest
# HF Space: point DOC_REDACTION_GRADIO_URL at your doc_redaction Space (see top of file).
# HF_TOKEN=                          # auth to private redaction Space (Space secret on Pi Space)
# DOC_REDACTION_HF_TOKEN=            # alias mirrored to HF_TOKEN

# Local llama-cpp (match AGENT_LLAMA_MODEL_ID to the active compose profile / llama.cpp -hf repo)
# AGENT_LLAMA_BASE_URL=http://llama-inference:8080/v1
# AGENT_LLAMA_MODEL_ID=unsloth/Qwen3.6-27B-MTP-GGUF          # --profile 27b_36 (pi-agent)
# AGENT_LLAMA_MODEL_ID=unsloth/Qwen3.6-35B-A3B-GGUF           # --profile 35b_36 (pi-agent-35b)
# AGENT_LLAMA_MODEL_ID=unsloth/gemma-4-31B-it-GGUF           # --profile gemma4-31b (pi-agent-gemma-31b)
# AGENT_LLAMA_MODEL_ID=unsloth/gemma-4-26B-A4B-it-GGUF       # --profile gemma4-26b (pi-agent-gemma-26b)

# Google Gemini (Pi resolves GEMINI_API_KEY; GOOGLE_API_KEY is mirrored at startup)
# GEMINI_API_KEY=
# GOOGLE_API_KEY=
#
# Quota / rate-limit retries (Gradio UI loop and Pi auto-retry in settings.json; default 5)
# AGENT_MAX_RETRIES=5
# AGENT_QUOTA_RETRY_DELAY_S=60
# AGENT_GEMINI_RETRY_BASE_DELAY_MS=60000
# AGENT_GEMINI_RETRY_MAX_DELAY_MS=90000

# Pi session auto-compaction (written to ~/.pi/agent/settings.json on startup)
# AGENT_COMPACTION_ENABLED=true
# AGENT_COMPACTION_RESERVE_TOKENS=32768
# AGENT_COMPACTION_KEEP_RECENT_TOKENS=20000

# AWS Bedrock (uses AWS SDK credential chain — not a single API key)
# AWS_REGION=eu-west-2
# AWS_ACCESS_KEY_ID=
# AWS_SECRET_ACCESS_KEY=
# AWS_SESSION_TOKEN=
# AWS_PROFILE=