A Codex skill for building small, source-linked working context without turning a summary, manifest, capsule, or event log into a second source of project truth.
It is intended for many-source, long-running, multi-agent, compaction-prone, or context-risky work. Canonical goals, tasks, decisions, research claims, proof, and policies stay with their existing owners. Context Manager selects exact slices, binds them to source and policy digests, generates a deterministic agent capsule, and rejects stale, mixed-revision, tampered, or over-privileged views.
Clone the repository into your personal Codex skills directory:
git clone git@github.com:oniivan/autonomous-context-manager.git \
~/.codex/skills/autonomous-context-managerRestart Codex or begin a fresh task so skill discovery reloads. Invoke it explicitly with:
Use $autonomous-context-manager to build and validate a source-linked context capsule
for this context-risky project without duplicating project truth.
- Name the canonical owner and exact locator for every selected artifact and policy.
- Create a pointer-only
context/manifest.json. - Define the task-specific selection in
context/views/<view-id>.json. - Build a deterministic capsule with
scripts/build_context_capsule.py. - Validate the manifest, view, capsule, source revisions, proof classifications, and trust/sensitivity/access policy before loading or transferring the capsule.
- Persist new domain facts with their canonical owners, then delete and rebuild derived context after relevant source or policy changes.
The complete operating contract is in SKILL.md. File schemas, ownership, integration boundaries, and evaluation oracles live under references/.
The runtime and test suite use Python's standard library.
PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tests -v
PYTHONDONTWRITEBYTECODE=1 python3 scripts/check_trigger_contract.py \
evals/trigger-cases.jsonOptional live compatibility checking for installed parent skills:
PYTHONDONTWRITEBYTECODE=1 python3 scripts/check_parent_adapter_contract.py \
evals/parent-adapter-cases.json \
--skills-root ~/.codex/skillsThe package's unit tests are standalone. The live adapter command intentionally checks the current installed versions of Project Onboarding, Research One-Shot, Source Distillation, Goal Compiler, and Chat Handoff. It verifies documented contracts and pointer payloads; it does not import, invoke, or modify those skills.
- This skill owns derived context routing and validation, not project status or domain truth.
- A real project must identify the canonical owner and exact locator for its trust/sensitivity/access policy. Missing or unresolvable policy fails closed.
events.jsonlis rejected in v0.1; event replay is not needed for safe rebuilds.- Project Steward support is control-only in v0.1. It tests projection isolation but does not render or refresh an owner-facing view.
- Parent-skill adapters are optional. The lifecycle core remains usable without those skills, GoalBuddy, HTML, media, or owner-view logic.
SKILL.md— trigger, workflow, ownership, and stop contract.references/— schemas, field ownership, adapters, and evaluation methodology.scripts/— deterministic build, validation, routing, and eval utilities.evals/— trigger, adapter, lifecycle, forward-test, and Steward control cases.tests/— package-local unit and adversarial tests.