akc-cycle

# akc-cycle

> The install entry point (MIT) for the Agent Knowledge Cycle (AKC) — six phases (Research, Extract, Curate, Promote, Measure, Maintain) plus Scaffold Dissolution. Two install paths: (1) a single behavioral rules file — copy `rules/common/akc-cycle.md` into your harness rules directory and the cycle runs through natural conversation, loaded every session; (2) a Claude Code plugin carrying skills for each of the six phases (`/plugin marketplace add shimo4228/akc-cycle`). The rules file is the recommended minimal floor; the plugin adds the skill layer.

## Core documentation

- [llms-full.txt](llms-full.txt): self-contained Q&A reference for AI agents (Project Facts, question-form H2 sections)
- [README.md](README.md): human-facing entry (install, what the plugin adds and runs, per-phase table, citation); editions and sync model sit in its closing collapsed section
- [rules/common/akc-cycle.md](rules/common/akc-cycle.md): the rule file itself (self-contained edition, owned by this repo) — six phases with trigger conditions, Scaffold Dissolution with its evidence standard, and digests of the AKC worldview mechanisms (ADR-0024/0025/0026)
- [.claude-plugin/plugin.json](.claude-plugin/plugin.json): Claude Code plugin manifest for the skill layer (skills in `skills/`)

## Architecture

- Two-layer install: the rules file is loaded deterministically every session (the floor); the plugin's skills are probabilistically triggered (the depth), except those marked `disable-model-invocation: true`, which run only when called by name. Claude Code plugins have no slot for always-loaded rules (as of 2026-10-08), so the rules file stays a copy-install and is not part of the plugin payload.
- Plugin payload, by phase (matching the AKC phase table): `search-first` (Research); `learn-eval`, `skill-writer` (Extract; skill-writer writes a skill behind a fresh-context draft gate and the author's sign-off, ADR-0005); `harness-stocktake` (one report of every audit's verdicts, one approval, applied on a branch), `skill-stocktake` / `skill-health` / `rules-stocktake` / `agent-stocktake` / `hook-stocktake`, `generation-audit` (re-audit on a model-generation change, ADR-0023), `harness-boundary` (Scaffold Dissolution at design time) (Curate); `rules-distill`, `review-to-lint` (code-LLM layering, ADR-0008) (Promote); `skill-comply`, `llm-as-judge` / `jev-judgment-design` (the judge pattern, ADR-0008), `author-calibrated-eval` (intent alignment) (Measure); `context-sync` / `repo-asset-stocktake`, `decision-history` (dated, past-tense decision records), `verify-bootstrap` (machine gates as the enforcer of LLM-first readability, ADR-0025) (Maintain). The repo is its own marketplace (`.claude-plugin/marketplace.json`, source `./`).
- Six phases, each a principle + explicit trigger: Research (search broadly, filter by signal), Extract (capture reusable patterns), Curate (audit accumulated knowledge), Promote (elevate recurring patterns to rules), Measure (verify behavioral change quantitatively), Maintain (keep documentation roles clean)
- Scaffold Dissolution: the rules are scaffolding; success is the cycle running without explicit invocation, by absorption inward (conversation patterns) or downward (harness substrate). Evidence standard: completion evidence is held-out transfer (same-context ablation is only necessary evidence); negative-delta artifacts are actively deleted, audited at each model-generation turnover
- Two editions: this repo owns the self-contained edition (no skills assumed); the author's harness runs a pointer edition delegating each mechanism to installed skills ([claude-harness rules/common/akc-cycle.md](https://github.com/shimo4228/claude-harness/blob/main/rules/common/akc-cycle.md)) — the reference shape the rules file can shrink into once the plugin is installed
- Design lineage: the rules-file install target for the [Agent Knowledge Cycle](https://github.com/shimo4228/agent-knowledge-cycle) ([DOI 10.5281/zenodo.19200726](https://doi.org/10.5281/zenodo.19200726))
- Sync model: canonical copies of the plugin skills live in the author's Claude Code harness (`~/.claude/`); this repo is a one-way mirror for those via `scripts/sync-from-local.sh` (fixed allowlist, aborts on any missing component or origin marker). The rules file is NOT synced — this repo owns the self-contained edition (since 2026-09-01)

## Sibling research lines

- [agent-knowledge-cycle](https://github.com/shimo4228/agent-knowledge-cycle): the AKC research line this rule belongs to ([DOI 10.5281/zenodo.19200726](https://doi.org/10.5281/zenodo.19200726))
- [contemplative-agent](https://github.com/shimo4228/contemplative-agent): autonomous agents grounded in four contemplative axioms ([DOI 10.5281/zenodo.19212118](https://doi.org/10.5281/zenodo.19212118))
- [agent-attribution-practice](https://github.com/shimo4228/agent-attribution-practice): harness-neutral ADRs on accountability distribution ([DOI 10.5281/zenodo.19652013](https://doi.org/10.5281/zenodo.19652013))