Autocontext

Overview

Autocontext is a control plane for evaluating agent behavior, preserving useful run artifacts, exporting training data, and distilling stable behavior into local runtimes. In Hermes, use this skill when the work calls for measurement, replay, datasets, local MLX/CUDA training, or read-only analysis of Hermes skill curation.

Hermes Curator owns Hermes skill mutation. Autocontext should inspect, evaluate, replay, export, and recommend. Do not use Autocontext as a replacement for Hermes Curator, and do not edit Hermes skills directly unless the user explicitly asks for that operation.

When to Use

• You need to run an Autocontext scenario from Hermes and inspect the result.
• You need machine-readable status for runs, solved knowledge, or training jobs.
• You need to inspect Hermes v0.12 Curator reports, skill usage counters, pinned state, or skill provenance.
• You need to export Autocontext knowledge into a reusable package or skill-like artifact.
• You need to prepare data for local MLX or CUDA training.
• You need to decide whether MCP is useful in a configured environment.

Do not use this skill for normal Hermes memory updates, direct skill consolidation, or user-local skill deletion. Those are Hermes Curator responsibilities.

Integration Surface Order

Use the CLI first. The autoctx CLI is the default surface because Hermes agents can run it with normal terminal tools, see stdout and stderr, preserve logs, and debug failures without special host configuration.

MCP is optional. Use MCP when the environment already has Autocontext MCP configured and the task benefits from typed schemas, constrained invocation, or tool discovery. Do not require MCP just to wrap a command that the CLI already exposes cleanly.

Use a native Hermes runtime or OpenAI-compatible gateway when Autocontext is calling Hermes as an agent provider. Use a Hermes plugin emitter only when the user specifically needs high-fidelity live traces beyond read-only import of existing Hermes artifacts.

CLI Quick Start

From a checkout of Autocontext:

cd autocontext
uv run autoctx --help

Inspect Hermes skill and curator state without modifying Hermes:

uv run autoctx hermes inspect --json

For a custom profile or test fixture:

uv run autoctx hermes inspect --home "$HERMES_HOME" --json

Install or refresh this skill into a Hermes profile:

uv run autoctx hermes export-skill --output ~/.hermes/skills/autocontext/SKILL.md --json

If the file already exists and the user wants to replace it:

uv run autoctx hermes export-skill --output ~/.hermes/skills/autocontext/SKILL.md --force --json

Running Autocontext From Hermes

Use --json whenever Hermes needs to parse the result.

RUN_ID="hermes_$(date +%s)"
uv run autoctx run grid_ctf --iterations 3 --run-id "$RUN_ID" --json
uv run autoctx status "$RUN_ID" --json
uv run autoctx replay "$RUN_ID" --generation 1

For a plain-language task:

uv run autoctx solve "Improve the support-triage response policy." --iterations 3 --json

For one-shot judgment or improvement:

uv run autoctx judge --task-prompt "..." --output "..." --rubric "..." --json
uv run autoctx improve --task-prompt "..." --rubric "..." --rounds 3 --json

Hermes Runtime Configuration

When Autocontext should call a Hermes-served model through an OpenAI-compatible gateway:

export AUTOCONTEXT_AGENT_PROVIDER=openai-compatible
export AUTOCONTEXT_AGENT_BASE_URL=http://localhost:8080/v1
export AUTOCONTEXT_AGENT_API_KEY=no-key
export AUTOCONTEXT_AGENT_DEFAULT_MODEL=hermes-3-llama-3.1-8b
uv run autoctx solve "..." --iterations 3 --json

Keep provider configuration outside the skill when possible. The user or profile should own secrets, base URLs, and model names.

Working With Hermes Curator

Hermes v0.12 writes Curator reports under ~/.hermes/logs/curator/<timestamp>/run.json and REPORT.md. It tracks skill usage in ~/.hermes/skills/.usage.json, and protects bundled or hub-installed skills through .bundled_manifest and .hub/lock.json.

Use:

uv run autoctx hermes inspect --json

Read the output as an inventory:

• agent_created_skill_count means Curator-eligible user or agent skills.
• bundled_skill_count and hub_skill_count are upstream-owned skills and should not be pruned by Autocontext.
• pinned_skill_count identifies skills Curator and agents should not modify.
• curator.latest.counts summarizes the latest consolidation, pruning, and archive activity.

Autocontext can use these signals for reports, datasets, and recommendations. Hermes Curator remains the writer for Hermes skill lifecycle changes.

Privacy Before Session and Trajectory Ingest

Curator decision reports are decision metadata and safe to import without redaction. Session and trajectory imports are different: they contain raw model prompts and responses, which may include secrets, tokens, or content the operator did not intend for external storage.

Before recommending or running autoctx hermes ingest-sessions or autoctx hermes ingest-trajectories, explain the privacy tradeoff: the importer is read-only against ~/.hermes, but the output JSONL contains the same content unless redaction is applied. Default is --redact standard (Anthropic/OpenAI keys, bearer tokens, emails, IPs, env values, paths, high-risk file refs). --redact strict adds user-defined regexes. --redact off writes raw content and the importer surfaces an explicit opt-in marker. Sessions in particular live in a SQLite store: an unwarranted ingest creates a new copy of every prompt and response. Prefer --dry-run first when the operator is unsure of the blast radius.

Training Path

For Autocontext-owned runs, export training data and train locally:

uv run autoctx export-training-data --scenario grid_ctf --all-runs --output training/grid_ctf.jsonl
uv run autoctx train --scenario grid_ctf --data training/grid_ctf.jsonl --backend mlx --time-budget 300 --json
uv run autoctx train --scenario grid_ctf --data training/grid_ctf.jsonl --backend cuda --time-budget 300 --json

Use MLX on Apple Silicon hosts. Use CUDA on Linux GPU hosts with a CUDA-enabled PyTorch install. Do not run host-GPU training inside a sandbox unless the user has already provided a host bridge or direct GPU access.

For Hermes Curator artifacts, train a narrow read-only advisor from exported decision rows:

uv run autoctx hermes export-dataset --kind curator-decisions --home ~/.hermes --output training/hermes-curator-decisions.jsonl --json
uv run autoctx hermes train-advisor --data training/hermes-curator-decisions.jsonl --logistic --checkpoint training/hermes-advisor.json --json
uv run autoctx hermes recommend --home ~/.hermes --advisor training/hermes-advisor.json --output training/hermes-recommendations.jsonl --json

Use --baseline first for the majority-class floor. Use --mlx on Apple Silicon or --cuda on PyTorch/CUDA hosts when the optional extra is installed. Curator reports are decision traces; they are best suited for advisor/ranker/classifier training, not full autonomous skill mutation.

MCP Workflow When Configured

MCP is optional. If the user has already configured Autocontext MCP, prefer it for structured tool calls that are easier or safer than shell commands. Otherwise, stay with the CLI.

Check the local integration guide before inventing tool names:

uv run autoctx serve mcp --help

Use MCP only when it adds value beyond the CLI: stable schemas, lower parsing burden, managed tool discovery, or a host policy that disallows shell access.

Common Pitfalls

1. Treating Autocontext as the Hermes Curator. Autocontext should inspect and recommend; Hermes Curator owns skill mutation.
2. Starting with MCP in an unconfigured environment. Use the CLI first unless MCP is already present and helpful.
3. Mutating ~/.hermes/skills after inspection. autoctx hermes inspect is read-only; keep it that way during analysis.
4. Training on raw curator artifacts without a target. First decide whether the target is ranking, consolidation classification, pruning advice, or model-routing advice.
5. Forgetting --json when Hermes needs to parse command output.

Verification Checklist

• Use autoctx hermes inspect --json before making claims about local Hermes skill state.
• Confirm pinned skills are not modified.
• Confirm bundled and hub skills are treated as upstream-owned.
• Prefer CLI commands for first-run workflows.
• Use MCP only when configured and materially better for the task.
• Keep Hermes Curator as the system of record for Hermes skill lifecycle changes.

References

Progressive-disclosure docs available alongside this skill. Load only when relevant.

• references/hermes-curator.md — How Hermes Curator and autocontext cooperate; who owns what; the read-only-first rule.
• references/cli-workflows.md — Exact autoctx commands for inventory, curator ingest, dataset export, judging, replay.
• references/mcp-workflows.md — MCP server setup, CLI-to-MCP tool name mapping, when to prefer MCP over CLI.
• references/local-training.md — How autocontext-exported datasets feed local MLX/CUDA advisor training; what the advisor predicts; expected scope.

Operators can write all references next to this skill via autoctx hermes export-skill --with-references --output <dir>/SKILL.md.