Agent Nervous System

A suite of Claude Code skills that give an AI agent the physiological functions a long-running mind needs — not more capabilities, but the self-maintenance organs that keep capability healthy over time.

Most agent tooling adds senses and muscles. This adds the quieter systems: how an agent stays awake and self-directs, wakes oriented, discovers, remembers, defends its boundary, and lets things die well. Each is a standalone skill (its own repo); together they're a nervous system.

Each skill works with Claude Code, Codex, and Cursor, and operates on whatever stack you have — they degrade gracefully, never assuming a specific memory store, scanner, or repo layout.

The organs

Skill	Organ	Direction	What it does
consciousness-loop	executive	central	The keystone. A self-firing loop — wake → integrate the organs → decide → act → set its own frequency → sleep — that turns the reflexes below into one continuous self. Pairs the loop with adaptive arousal (alert / drowsy / deep-sleep). Functional, not phenomenal.
octopus-investigation	perception	outward	Discovers projects/people by crawling GitHub social graphs (stars, follows) — following human curation, not keyword search.
rem-sleep	memory	inward	A sleep cycle over the session: consolidate durable facts to long-term memory, prune the stale, integrate by linking, and regulate emotional charge (keep the lesson, release the sting).
immune-check	defense	boundary	A pre-flight reflex that scans any outbound artifact for secrets/PII before it leaves the body. Wraps whatever scanner you have; falls back to a built-in pass.
sunset	grief	project-scale	Retire a dead project well — harvest the lessons and reusable parts, archive with a marked grave, release the attention-rent. Distinguishes dormant (exhumable) from terminal deaths.
playtime	development	developmental	Deliberate, safe-to-fail exploration that builds capability before it's needed — rehearse unfamiliar tools, stress-test skills, recombine, then harvest the learning to memory. Engage and learn, with failure free.
contemplation	deliberation	meta	Slow, non-instrumental thought that examines ends, not means — question the goal/value/frame, hold a question open instead of solving it, name an irreducible tension rather than force a verdict. The only organ that works the why.
proprioception	feedback	reflexive	The loop-closer. Senses the agent's own performance, scores it against the declared fitness_signal, diagnoses systematic vs noise, writes the correction back into a param/rule/default/memory, and re-tests that it helped. Examines means (execution quality) — the complement to contemplation's ends. The only organ that turns the agent's output back into its input. Also carries the forecast ledger: log every consequential "probably" as a falsifiable claim + probability, resolve it when reality answers, Brier-score the calibration.
exteroception	perception	proximal	The wake sense. An agent's memory is write-time-stale; the world moves while it sleeps. At wake, sweep the watched surfaces (working trees, conversations, clocks, services, due forecasts), diff actual state against the remembered baseline, and brief the deltas ranked by urgency — orientation on purpose instead of by ambush. Read-only. The paired sense to proprioception (self-in-motion ↔ world-at-wake).
corpus-callosum	interface	dyadic	The interhemispheric channel. A sustained human-AI collaboration is one working mind on two substrates — and no other organ maintains the channel between the hemispheres. A standing register names who does the cognitive work this turn (questions-mode in open strategy, synthesis in execution), bound to session-start so it fires without recall; transfer-integrity checks (offload telltales, landing checks, transcript audits) keep expertise living in the human's head. Protects the human hemisphere from offload atrophy — the default failure mode of AI assistance — and the agent hemisphere from mirroring.

The framing is the point: not a pile of tools, but a coherent set of self-regulating functions. The consciousness-loop is the executive at the center; most organs are functions it schedules — discover (perception), remember (memory), protect (immunity), release (grief), grow (development) — while contemplation sits above the loop, examining the ends the loop pursues by means. Without the loop they're a body with no pulse; the loop is what makes them one agent; contemplation is what keeps that agent pointed at something it would endorse on reflection; and proprioception is what lets it get better — closing the error→correction loop the others leave open. Five relationships tie the suite together: the loop is the wake phase to rem-sleep's sleep phase; playtime is the proactive complement to rem-sleep's reactive memory (generate new experience ↔ consolidate the experience you had); contemplation examines the why that the loop, deciding what, has no time to; proprioception examines how well — the means-feedback that contemplation (ends) and the loop (one-line drift-check) both leave un-rigorous, and the first consumer of the registry's fitness_signal fields; and exteroception is proprioception's paired sense — the same diff-against-a-stored-baseline mechanism pointed outward (remembered world vs actual world) instead of inward, feeding the loop's wake phase the sensory input its decide step runs on. Corpus-callosum extends the body past the self: where immune-check guards the membrane (block what must not cross), the commissure guards the channel (guarantee what must cross arrives faithfully) — the first organ that maintains the dyad, not the agent, riding the same wake reflex as exteroception's sweep.

Reading

• Growing an agent a nervous system, then checking it against the consciousness science: six months of observed decay, a four-week build, and an audit against Butlin et al.'s fourteen indicator properties. Six are substrate-level and untouchable by any scaffold; one was built because of the audit (after its first design failed our gate); one was deliberately refused. A capability map, not a consciousness claim.

Clone the whole nervous system

git clone --recursive https://github.com/thdelmas/agent-nervous-system.git
# already cloned without --recursive?
git submodule update --init --recursive

Each subdirectory (consciousness-loop/, octopus/, rem-sleep/, immune-check/, sunset/, playtime/, contemplation/, proprioception/, exteroception/, corpus-callosum/) is the skill's own repo, pinned to a known-good commit.

Install (Claude Code)

A Claude skill is a folder containing SKILL.md. Install all ten globally:

for s in consciousness-loop:consciousness-loop octopus:open-source-octopus-investigation \
         rem-sleep:rem-sleep immune-check:immune-check sunset:sunset \
         playtime:playtime contemplation:contemplation proprioception:proprioception \
         exteroception:exteroception; do
  dir="${s%%:*}"; name="${s##*:}"
  mkdir -p ~/.claude/skills/"$name"
  cp "$dir"/SKILL.md ~/.claude/skills/"$name"/
  [ -d "$dir"/scripts ] && cp -r "$dir"/scripts ~/.claude/skills/"$name"/
done

Or run ./install.sh, which does the same and supports Codex (~/.agents/skills/) and Cursor (~/.cursor/commands/) targets.

Then invoke any of them by name — /consciousness-loop, /open-source-octopus-investigation, /rem-sleep, /immune-check, /sunset, /playtime, /contemplation, /proprioception, /exteroception — or just describe the need ("sleep on it", "is this safe to push?", "retire this project", "how am I doing?", "what changed since last session?").

Updating

git submodule update --remote --merge   # pull each skill's latest
git add -A && git commit -m "chore: bump skills"

Pushing into a submodule

The submodule URLs are public https so anyone can clone. If you maintain these under the thdelmas account and push via an SSH alias, set a push URL per submodule:

git -C octopus remote set-url --push origin git@github.com-thdelmas:thdelmas/open-source-octopus-investigation.git

How this suite was scoped

The organs weren't picked in a vacuum — the landscape was mapped with an octopus investigation, and every project considered was logged with a verdict (adopted / rejected / watch) and the reasoning. See docs/landscape-tracker.md for the human-readable curation record: what's out there, what influenced the design, and why the gaps (sunset, rem-sleep-as-a-skill) are real.

Evolution layer — machine-readable lineage

Skills here spread by memetic transmission (directed, Lamarckian, horizontal), and at population scale individual choices aggregate into a stochastic-but-biased distribution — not random drift. Two machine-readable files make that legible so a second agent can observe this suite's evolution and act on it without the maintainer in the loop:

• registry.json — the population snapshot: every skill's id, version, parent (the lineage pointer), mutation (what's novel), and fitness_signal (judge by realized outcome, not stars).
• lineage.jsonl — the append-only selection ledger: one row per adopt / reject / watch / cold decision, each with the context that turns an apparently-random choice back into signal. The rejections are the negative dataset of what didn't propagate here, and why.
• docs/lineage-spec.md — the schema + the consumption protocol that doubles as the thesis's falsifiable test: if no unprompted script can parse these and pull a skill, the "agents evolve" story is narration; when one does, selection has moved from the human to the population.
• bin/validate.py — that falsifiable test as an exit code. It does what a peer consumer's first step must: parse both files, build the provenance graph from the registry's parent edges, and check the ledger is schema-valid, referentially sound, and genealogically coherent — every parent edge backed by a forked origination row whose source matches it — failing nonzero on any drift. Wired into CI, so the claim above is machine-checked on every push, not asserted — and available as a local pre-push gate: git config core.hooksPath .githooks. (Written by the suite's own proprioception organ — its first act was to grade the layer that contains it.)

License

MIT — see LICENSE. Each submodule carries its own MIT license too.