Docs Overview
August 20, 2026 · View on GitHub
This directory is the maintainer-facing landing page for repository docs. Use it to find the right guide quickly and keep public documentation aligned when the repo changes.
Start Here
Using The Packages
Public examples use the canonical nested CLI paths: autoctx scenario create
for scenario authoring and autoctx serve mcp for MCP clients.
Integrating External Agents
- External agent integration guide
- Running the loop on your own hardware
- Hermes Curator + autocontext positioning
- Python and TypeScript extension hooks
- Sandbox and executor notes
- Capability-scoped research workspaces
- Provider-neutral remote execution sessions
- Persistent host worker
- Correctness-first external kernel evolution
- MLX host training notes
- Case study: recursive loop closed on local MLX
Contributing And Support
Architecture And Parity
- Cross-runtime CLI contract guide, machine-readable contract, wire-output schemas, and shared fixtures
- Interactive WebSocket protocol contract
The contract pins the TypeScript-only agent_progress_notes_v1 extension,
including its exact Autowork-compatible fixture, safe-copy and size limits,
earlier same-run evidence rules, exact durable replay, and finite retention.
Python intentionally does not advertise it until equivalent durable transcript
metadata is available.
- Core/control package split
- Runtime component lifecycle
- Runtime effect policy
- Runtime component graph
- Transactional runtime activation
- Runtime composition confluence harness
- Strategy package import side-effect contract
- Generic edge runtime compatibility spike
- Fetch adapter API reference
- Fetch host capability manifest examples
- Generated Fetch packaging guide
- Fetch conformance guide
- Fetch adapter troubleshooting guide
- Flue-inspired runtime decisions
- Scenario parity matrix — Python & TypeScript
- Scenario environment contract
- Role routing contract
- Typed rubric contract
- Run progress report
- Run utilization report
- Negative result ledger
- Campaign mode report
- Optional campaign scheduler
- Read-only campaign auditor
- Goal run report
- Playbook approval gate
- Immutable context bundles and outcome-gated promotion
- Campaign false-promotion calibration
- Derived lesson curation
- Soft structural hints
- Span-level credit attribution
- Ablation-backed context attribution
- Trainer-local statistical confirmation
- OPD/GKD + GRPO mixture experiment protocol
- Exploration collapse guard
- Browser exploration contract
- OpenTelemetry bridge
- Background session domain and parity contract
- Background execution trust boundaries and credential model
Execution Surfaces
simulate— modeled-world exploration with sweeps, replay, compare, exportinvestigate— evidence-driven diagnosis in synthetic harness or iterative LLM modesanalyze— interpret and compare outputs from all surfacescontext-selection— inspect persisted prompt context telemetry for run budget/cache tuningmission— real-world goal execution with adaptive planning and campaignsagent— TypeScript local runner/dev server and self-hosted Node build target for experimental.autoctx/agentshandlerstui— Node 22.19+ pi-tui operator client for local runs or remote TypeScript server attachment, with durable transcript replay and cockpit readstrain— distill curated datasets into scenario-local modelshermes— read-only Hermes v0.12 skill/Curator inspection plus Hermes skill export
Trace Pipeline
- Public trace schema v1.0.0 for cross-harness interchange
- Privacy-aware export with sensitive-data redaction (21 patterns)
- Publishing to local JSONL, GitHub Gist, Hugging Face (ShareGPT format)
- Dataset curation with gate filtering, top-quartile selection, held-out splits
- Model selection strategy (from-scratch / LoRA / full fine-tune)
- Training backends (MLX / CUDA) with promotion lifecycle
Maintainer Docs
Keep These In Sync
If a change affects commands, package names, published versions, environment variables, agent integration flows, or support expectations, review these docs in the same PR:
README.mdautocontext/README.mdts/README.mdexamples/README.mdautocontext/docs/agent-integration.mdCHANGELOG.mdSUPPORT.md