cell-runner: walking-skeleton spike — runnable CC→PC→WC agentic loop (proof-of-life)

[usurobor]

CDD Trace

Step	Artifact	Skills loaded	Decision
0 Observe	the three-cell loop (CC→PC→WC) exists only as a hand-cranked bootstrap; no code turns it	—	proceed
1 Select	milestone: "cnos moves to three cells running"	—	walking-skeleton first
4 Gap	see below	—
5 Mode	see below	—
6 Artifacts	src/packages/cnos.cell-runner/**	—	built
8 Review	PR #this	—	Pending (external β)
9 Gate	—	—	Pending
10 Release	—	—	Pending

Gap (step 4)

What: the three-cell agentic loop (CC→PC→WC) exists today only as a hand-cranked bootstrap (κ + external β + subagents). There is no runnable machinery that turns the loop.

Why it matters: the milestone — cnos running its own development through the three cells, and tsc adopting the same structure — needs code that runs the loop, not only a specification of it (the doctrine plan is PR #672; the implementation wave is #627). Designing forever without running hides the real integration seams.

What fails if skipped: we keep perfecting a plan and never discover, by running, where CM ⇄ CC ⇄ PC ⇄ WC actually breaks.

Mode + Active Skills (step 5)

• Mode: MCA
• Work shape: runtime (exploratory spike)
• Level: L6
• Active skills: reuse-house-idiom (declarative FSM table + generic evaluator, per cnos.issues/issues-fsm); fail-closed; honest-provenance (spike ≠ canonical doctrine)
• Dominant risk: a spike quietly presenting itself as authored doctrine, or a validator that only looks mechanical — both explicitly guarded here.

Changes

Added

A runnable Go+CUE proof-of-life for the three-cell loop under src/packages/cnos.cell-runner/ (new go.work module cnos.dev/cnos/cell-runner; the only edit outside the package is adding it to the root go.work).

One loop turn, mechanically, git-native, no human in the machinery:

CC.measure → CM object + judgment{request_planning}   (incoherence detected)
  → PC.plan → wave{ one WC contract }                 (relation graph)
    → WC.execute → α produce · β review · γ close · V validate · δ decide · receipt
      → CC.re-measure → judgment{coherent}            (incoherence cleared)
loop terminates: incoherence 1 → 0, receipt chain closed.

• The three cells are distinguished ONLY by output telos (WC→artifact, PC→relation_graph, CC→judgment) and run the same kernel.
• Kernel is a declarative transition table (kernel/transitions.json) driven by a generic evaluator — mirroring the cnos.issues/issues-fsm idiom; never switches on a state name.
• Judgment seams sit behind an Actor interface; v0 ships deterministic stub actors — agent-backed actors (the real α/β/CC judgment) drop into the same interface next.
• The CM object (schema/cm.cue) is the runner/provider-produced measurement — the exact seam a tsc-provided TSC measurement plugs into (WC-2 of cnos#671: cell-runtime doctrine wave (CM-grounded) — PC plan under review #672); it is NOT the CC's judgment (the CC consumes it).
• Receipts are computed at γ from transition evidence (not free-text), shaped after schemas/cdd/receipt.cue, and cue-vet green.

Built against the R12 cell-runtime doctrine design (PR #672).

Not in scope

Agent-backed actors; multi-WC / parallel waves; the full wave FSM (WC-3b) beyond one node; real CM scoring beyond defect-count; operator-authorization gate wiring; self-hosting on a real cnos change; tsc adoption. Each is a named next increment in DESIGN.md, not a silent gap. This spike is not canonical runtime — it is reconciled to the authored doctrine (#672 WCs / #627 S2–S8) when that lands, or discarded.

Acceptance Criteria

• go run ./cmd/cell-runner --workspace <tmp> --task testdata/toy → exit 0, transcript shows incoherence 1 → 0 across CC→PC→WC→CC
• every emitted CM / wave / receipt / judgment cue-vets against schema/*.cue
• fault injection (--fail-injection=v-fail|beta-reject) drives the kernel to a fail-closed held state with a stop receipt and non-zero exit
• go build ./..., go vet ./..., go test ./... green; gofmt -l empty
• no Python (find … -name '*.py' empty)
• CI green
• external β review

Known Debt

v0 actors are deterministic stubs — the loop's judgment is not yet real (that is v1: agent-backed actors behind the same Actor interface). The runner receipt schema is a minimal spike shape reusing the cdd transmissibility gate, not the full cnos.cdd.receipt.v1. Emitted objects are JSON (stdlib-only; JSON is a YAML/CUE subset). All recorded in DESIGN.md §Build notes.

Pre-Review Gate (alpha/SKILL.md §2.6)

• Rebased on current main (branched from main tip)
• PR body carries CDD Trace through step 7
• Every AC mapped to evidence (run transcript + cue-vet output)
• Peer enumeration — n/a (single new isolated package)
• Schema / shape audit — all emitted objects cue-vet against schema/*.cue
• Harness audit — new module added to go.work; no existing surface changed
• Known debt explicit
• CI green on head commit (PR remains draft)

---

🤖 Generated with Claude Code

---

Generated by Claude Code