arc-atlas 0.1.1

Teacher/Student orchestration toolkit for Bring-Your-Own-Agent workflows.

Atlas SDK

Atlas SDK lets you wrap any Bring-Your-Own-Agent (BYOA) into a guided Teacher → Student → Reward loop. The toolkit focuses on sequential, high-trust workflows: you supply an HTTP endpoint, a Python function, or an OpenAI-compatible agent; Atlas handles planning, orchestration, evaluation, and persistence.

---

Key Features

• Bring-Your-Own-Agent (BYOA) Adapters – Drop in HTTP, Python, or OpenAI agents without rewriting core logic.
• Teacher / Student Loop – Plans and executes tasks sequentially with review, validation, and retry guidance.
• Reward System (RIM) – Runs configurable judges (process, helpfulness, custom) to score every step.
• Trajectory Capture – Emits intermediate steps that can be streamed, logged, or audited later.
• PostgreSQL Persistence – Ships with an async persistence layer and schema for sessions, attempts, guidance, and events.

---

Quick Start

python3.13 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -e .[dev]

Run an example configuration:

from atlas import core

result = core.run(
    task="Summarise the latest financial news",
    config_path="configs/examples/openai_agent.yaml",
)

print(result.final_answer)

Atlas returns an atlas.types.Result containing the final answer, the reviewed plan, and per-step evaluations.

---

Exporting Runtime Sessions

Atlas persists full execution traces whenever PostgreSQL storage is configured. Convert those sessions into training-ready JSONL with the bundled exporter:

# 1. Run tasks that log to Postgres (configure storage.database_url in your AtlasConfig)
atlas.core.run(...)

# 2. Export the captured sessions to JSONL
atlas.export --database-url postgresql://localhost:5432/atlas --output traces.jsonl --limit 25

# 3. Load the dataset inside the Atlas core repo
from trainers.runtime_dataset import load_runtime_traces
sessions = load_runtime_traces("traces.jsonl")

The CLI accepts repeatable filters such as --session-id, --status, and --trajectory-event-limit. Pass a standard PostgreSQL URL (including credentials) via --database-url. The exporter prints friendly counts of the sessions and steps written and emits newline-delimited JSON—one session per line.

Each session record follows the shared runtime schema consumed by the training stack:

• task, final_answer, plan – orchestration metadata for the run.
• session_metadata – persisted metadata plus status/timestamps.
• steps – executor traces with descriptions, outputs, reward breakdowns (score, per-judge details, tier samples), validation results, retry guidance, and executor metadata (including captured reasoning blocks under metadata.reasoning).
• trajectory_events – optional array of intermediate telemetry events for richer replay and debugging.

Once exported you can feed the file directly into load_runtime_traces or flatten it for RL pipelines with helpers in trainers/runtime_dataset.py from the core repository.

---

Configuration Guide

Configuration files live in configs/examples/. Each YAML document is validated against atlas.config.models.AtlasConfig.

Section	Purpose
agent	Adapter settings (endpoint, Python import path, OpenAI model) and tool schemas
student	Planner / executor / synthesizer prompts and token limits
teacher	LLM parameters for plan review, validation, and retry guidance
orchestration	Retry policy, per-step timeout, and trajectory emission flags
rim	Judge definitions, weights, aggregation strategy, thresholds
storage	Optional PostgreSQL connection info for persistence
prompt_rewrite	LLM used to derive planner / executor / teacher personas from the user prompt

During startup Atlas calls the rewrite LLM once to transform the BYOA system prompt into three personas:

1. Planner Student – drafts a dependency-aware plan
2. Executor Student – runs each step and returns a trace
3. Teacher – reviews plans, validates execution, and issues retries/guidance

By default the rewrite call reuses the same API credentials as your agent. Provide an explicit prompt_rewrite block if you need a dedicated model or different limits.

Example: HTTP Adapter (excerpt)

agent:
  type: http_api
  name: example-http-agent
  system_prompt: |
    You are an HTTP-based agent that can call external services.
  tools:
    - name: web_search
      description: Search the web for relevant documents.
      parameters:
        type: object
        properties:
          query:
            type: string
            description: Query string to search for.
        required: [query]
  transport:
    base_url: http://localhost:8080/agent
    timeout_seconds: 60

---

Architecture

1. core.run()                 # load config, adapter, context
2. Student.create_plan()      # ATLAS-derived planning graph via BYOA bridge
3. Teacher.review_plan()      # validates dependencies and tools
4. Orchestrator.arun()        # sequential execution, retries, telemetry
5. Evaluator.ajudge()         # process/helpfulness judges aggregate scores
6. Database.log_*()           # optional persistence of plans, attempts, trajectory events

Trajectory events stream through ExecutionContext.event_stream, enabling live console streaming and durable storage via atlas/storage/database.py and atlas/storage/schema.sql.

RIM Model Guidance

• Tier-1 judges (process/helpfulness): Gemini 2.5 Flash or Grok-4 Fast provide fast, low-cost scores.
• Tier-2 arbiter: Gemini 2.5 Pro reconciles disagreements with high fidelity.
• Supplied examples show how to point rim.judges[].llm and rim.arbiter at different providers if desired.

---

Terminal Telemetry

Atlas streams orchestration events directly to the terminal when core.run executes in an interactive shell. The default console renderer highlights the accepted plan, step attempts, tool invocations, reward scores, and the final synthesis without extra setup.

Example session:

=== Atlas task started: Summarize the Atlas SDK (2025-02-12 10:15:03) ===
Plan ready with steps:
  1. gather dataset A
  2. synthesise findings
[step 1] attempt 1 started: gather dataset A
[tool] web_search call -> {"query": "Atlas SDK release"}
[tool] web_search result <- {"result": "..."}
[step 1] completed: gather dataset A
  reward score: 0.91
[step 2] retry 2 started: synthesise findings
  guidance: cite the repository README
=== Atlas task completed in 12.4s ===
Final answer:
  Atlas SDK ships a teacher-student loop...
- gather dataset A | attempts: 1 | score: 0.91
- synthesise findings | attempts: 2 | score: 0.88
RIM scores | max: 0.91 | avg: 0.89

Disable streaming with core.run(..., stream_progress=False) when piping output or running in CI. Pass stream_progress=True to force streaming even when stdout is not a TTY. The renderer also works with core.arun and runs alongside PostgreSQL persistence, so stored sessions retain full telemetry.

See docs/examples/terminal_telemetry.md for a step-by-step walkthrough.

For a deeper look at how these events map onto the Atlas training stack—and why the SDK keeps telemetry lightweight—see docs/telemetry_overview.md.

---

Exporting Runtime Sessions

Use the atlas.export CLI to convert persisted PostgreSQL sessions into JSONL traces that match the core runtime schema.

atlas.export \
  --database-url postgresql://atlas:atlas@localhost:5432/atlas \
  --output traces.jsonl

Key flags:

• --session-id (repeatable) restricts the export to explicit sessions.
• --limit/--offset and --batch-size page through large archives.
• --trajectory-limit controls how many intermediate events are embedded per session.

Each line in the output is an AtlasSessionTrace record:

{
  "task": "...",
  "final_answer": "...",
  "plan": {"steps": [...]},
  "steps": [
    {
      "step_id": 1,
      "description": "...",
      "tool": "summariser",
      "reward": {"score": 0.92, "judges": [...]},
      "validation": {"valid": true, "rationale": "..."},
      "guidance": ["..."],
      "context": {"prior_results": {"1": "..."}}
    }
  ],
  "session_metadata": {
    "session_id": 42,
    "status": "succeeded",
    "trajectory_events": [...]
  }
}

The structure aligns with AtlasSessionTrace, AtlasStepTrace, and AtlasRewardBreakdown used by trainers/runtime_dataset.py, so you can immediately consume the file inside the core repo:

1. Run atlas.core.run(...) with PostgreSQL persistence enabled.
2. Execute atlas.export --database-url ... --output traces.jsonl.
3. Call load_runtime_traces("traces.jsonl") (from the core repo) to build training datasets.

Additional usage notes live in docs/examples/export_runtime_traces.md.

---

Testing

PYTHONPATH=. pytest tests --disable-warnings

The suite covers dependency parsing, prompt rewriting, student/teacher orchestration, RIM aggregation, adapter bridges, and database logging. Most tests rely on locally mocked adapters, so no external network calls occur.

---

Requirements & Notes

• Python 3.10+ (project is developed and validated with 3.13).
• Development extras (pip install -e .[dev]) install pytest tooling for local validation; core telemetry streams rely solely on the standard library.
• Vendored NeMo components live under atlas/roles/ and atlas/utils/reactive/; SPDX headers are retained and must remain intact.
• Aim for descriptive naming and concise docstrings so the intent is evident without extra commentary.

---

Contributing

1. Fork and clone the repository.
2. Use the provided pyproject.toml extras to install development dependencies.
3. Review existing modules before coding and keep commits focused and incremental to match the current style.
4. Add or update unit tests alongside feature changes.

Pull requests should include updated documentation or examples when behaviour changes.

---

License

Atlas SDK is released under the Apache 2.0 license. See LICENSE for full details. Vendored NeMo components retain their original licensing notices.