We use two layers to build a profile to guide AI agents on your behalf.
- Existential Layer: Asks questions to your introspective content like journals to uncover drivers and aspirations.
- Operational Layer: Processes your public content, work output, emails, social posts to see how you carryout actions.
These layers build on one another to create datasets to inform choices and scaffold actions with existential underpinnings.
Basically, we run systems that will explain why you choose things to AI.
This internal monologue annotates dataset with reasoning traces to introspect better and explain to AI models why/how it should do something.
Skills are unified under workspaces/skills and exposed in two downstream skill
shapes, alongside translation-layer and alignment outputs:
| Output | Purpose |
|---|---|
lib.artifacts.skills.files.<skill-name> |
Flat skill content keyed by skill name |
lib.artifacts.skills.names |
Available skill names |
lib.artifacts.skills.categorized |
Categorized skill tree shaped as <category>/<skill>/... |
lib.artifacts.alignment.spec |
Generated alignment spec |
lib.artifacts.alignment.translationLayer |
Generated translation-layer orchestrator soul (SOUL.md) |
lib.artifacts.alignment.interactionPosture |
Generated interaction posture (INTERACTION_POSTURE.md) |
lib.artifacts.operational.toolSpecs.{memory,tasks} |
Operational tool specs |
{ pkgs, inputs, ... }:
let
cognitive = inputs.cognitive-assistant;
system = pkgs.stdenv.hostPlatform.system;
artifacts = cognitive.lib.artifacts;
operational = artifacts.operational;
alignment = artifacts.alignment;
in
{
# Categorized Hermes-style skill tree: <category>/<skill>/SKILL.md.
environment.etc."hermes/skills".source = artifacts.skills.categorized;
# Operational profile tool specs.
programs.opencode.tools = {
memory = builtins.readFile operational.toolSpecs.memory;
tasks = builtins.readFile operational.toolSpecs.tasks;
};
# Generated alignment spec.
environment.sessionVariables.ALIGNMENT_SPEC = "${alignment.spec}";
}The repo runs as one unified pipeline parameterized by a layer profile
(existential or operational) for profile-specific build steps. Model and
provider setup is in docs/Getting-Started.md.
# Existential profile
python -m core --profile existential ask-questions
python -m core --profile existential build-prompts
python -m core --profile existential build-skills
# Operational profile
python -m core --profile operational ingest-corpus
python -m core --profile operational ask-questions
python -m core --profile operational build-prompts
python -m core --profile operational build-skills
python -m core --profile operational build-tool-specs
# Cross-profile / shared commands
python -m core enhance-skill
python -m core build-translation-layer
python -m core build-alignment-spec
python -m core review-queue [--dry-run] # refill the Langfuse review queue
# Everything after ask-questions, for every profile
python -m core updateQuestion answering retrieves from configured QMD collections. Existential uses
Journal and Personal; operational uses Base, Consulting, and Customers.
Raw source notes are not copied into this repository.
build-skills reads the profile's human_profile.md and writes only to
workspaces/skills/<profile>/<skill-name>/SKILL.md, removing skills in that
profile folder that the profile no longer declares. Cross-system consumers
should read skills only from workspaces/skills or lib.artifacts.skills.*,
not from profile artifact directories. Imported skills live under purpose
categories such as workspaces/skills/workflow/<skill-name>/SKILL.md.
build-translation-layer reads both profile artifacts plus the runtime
lineage and produces the interaction posture and the bootstrap SOUL:
workspaces/skills/<profile>/<skill-name>/SKILL.md
workspaces/alignment/artifacts/INTERACTION_POSTURE.md
workspaces/alignment/artifacts/SOUL.md
workspaces/alignment/artifacts/absorbed.json
The alignment command sits above both profiles. It reads unified skills from
workspaces/skills and writes a personalized production-readiness checklist.
# Build the spec (requires build-skills to have been run)
python -m core build-alignment-specRegenerate the spec whenever skills change. See
docs/Alignment-Artifacts.md and
profiles/alignment/README.md.
The repo builds the bootstrap SOUL. What Hermes runs is the living SOUL under
CA home ($CA_HOME, default $XDG_DATA_HOME/cognitive-assistant): the
bootstrap plus experience records learned from real sessions.
python -m core improve-run # analyze sessions, propose records (TTY)
python -m core improve-apply # review; accepted records rebuild the living SOUL
python -m core soul status # generations and verdictsSee docs/Improvement-Loop.md and docs/Living-SOUL.md. The full docs index is docs/README.md.
nix develop --command python -m unittest discover -s tests -v
nix develop --command bash -c 'cd improve && npm test'