Skip to content

CHANGELOG.md

Latest commit

 

History

History
385 lines (324 loc) · 18.4 KB

CHANGELOG.md

File metadata and controls

385 lines (324 loc) · 18.4 KB

Changelog

All notable changes to AKC are recorded here. AKC follows semantic versioning; breaking changes to positioning or public interfaces bump the major version.

v2.0.0 — 2026-04-19

Major positioning release. AKC declares the cycle genre-neutral about the knowledge body that flows through it — behavioral patterns, domain expertise, or constitutional values all run through the same six phases. The security triplet that had sat in AKC through v1.x (Security by Absence, Single External Adapter, Untrusted Content Boundary) is extracted as genre-specific contemplative-agent content and no longer ships with AKC.

Why v2.0.0 and not v1.9.0

AKC's identity contract changes. Prior: "a knowledge cycle for AI agents, with Security by Absence as a first-class design principle." Current: "a knowledge cycle for AI agents, mechanism-only, genre-neutral about what flows through it." Academic citations resolving to v1.x will see a different Design Principles list and a different ADR set. ADR-0009 already established the precedent for positioning-breaking changes warranting a version bump; the security triplet extraction completes what ADR-0009 started, and the identity-contract shift is the direct reason this is a major bump rather than a minor one.

Removed

  • docs/adr/0001-security-by-absence.md — genre-specific to the contemplative-agent lineage. Extracted to local archive at _archive/akc-security-triplet-2026-04/ for future re-expression.
  • docs/adr/0006-single-external-adapter.md — genre-specific. Extracted to the same archive.
  • docs/adr/0007-untrusted-content-boundary.md — genre-specific. Extracted to the same archive.
  • docs/skills/llm-agent-security-principles.md — design-pattern skill that bundled the three security ADRs' concrete defense patterns. Extracted to the same archive.
  • Design Principle #7 (Security by Absence) from the README — removed from the design principles list. Former #8 (Code-LLM Layering) renumbers to #7; former #9 (Human cognitive resource) renumbers to #8.

Added

  • docs/adr/0011-cycle-applies-to-any-knowledge-body.md — new ADR declaring genre neutrality as a first-class positioning statement, with three falsifiable commitments: no future AKC ADR encodes domain-specific content, docs/skills/ hosts only cycle-mechanic skills, and external projects may cite AKC as their cycle substrate without AKC making claims about their content.
  • docs/adr/0009-akc-is-a-cycle-not-a-harness.md 2026-04-19 addendum — records the concrete security-triplet extraction that completes ADR-0009's demotion. Documents the consequences: ADR numbering gaps at 0001, 0006, 0007; design principles 9 → 8; supersession of the original "ADRs 0001–0008 unchanged" claim.
  • examples/constitution_amend/ — new examples entry that points at contemplative-agent's Constitution-amend workflow as a concrete instance of the AKC cycle applied to constitutional values. Descriptive only; the directory ships no implementation. Paired with ADR-0011's "Concrete instance in the wild" section, this makes genre neutrality observable: minimal_harness/ shows the mechanism on behavioral patterns, constitution_amend/ shows the same mechanism on values.

Changed

  • ADR numbering: gaps at 0001, 0006, 0007 are intentional and permanent. ADR numbers are identifiers, not positions in a sequence.
  • README.md / README.ja.md — repo tree, counts (8 ADRs, 8 design principles, 3 design-pattern skills), "Related Work" section (no longer mentions security-by-absence as prior-art inheritance).
  • docs/adr/README.md — index updated; note explaining the numbering gaps added.
  • docs/skills/README.md — design-pattern skill table shrunk from 4 to 3 entries.
  • docs/inspiration.md — "What AKC does not carry over" extended with the security triplet; prior-art table narrowed to ADR-0002 through ADR-0005 and the minimal_harness reference.
  • examples/minimal_harness/README.md — "What is not implemented" section collapsed around the one remaining stance (Layer 3 gated on human review). Footnote explains that the forbidden-substring validation pattern in knowledge_store.py is inherited from upstream genre work even though the ADR pointers no longer live in this repo.
  • examples/minimal_harness/knowledge_store.py — comment and docstring references to ADR-0007 replaced with a neutral description of the defensive pattern. Behavior is unchanged.
  • llms.txt / llms-full.txt — all counts, ADR lists, and skill lists updated to the v2.0.0 state. A new Q&A section in llms-full.txt ("What knowledge bodies can run through the AKC cycle?") replaces the former "What does AKC deliberately not implement?" section, reflecting the genre-neutrality shift.
  • CITATION.cff — version 2.0.0; abstract rewritten; the security-by-absence keyword removed and genre-neutrality added.

Migration notes for v1.x citations and downstream repositories

  • External links pointing at docs/adr/0001-security-by-absence.md, docs/adr/0006-single-external-adapter.md, or docs/adr/0007-untrusted-content-boundary.md at main HEAD will 404. Commit-pinned links (via GitHub permalinks) continue to resolve. The content is preserved locally at _archive/akc-security-triplet-2026-04/ on the author's machine; future re-expression is expected in the contemplative-agent lineage.
  • Academic citations pinned to v1.x continue to describe the state at that version. No action is required for prior citations.
  • Downstream projects that adopted AKC's security-by-absence design principle may continue to do so — the principle itself is sound — but should no longer cite AKC as its source. Contemplative-agent is the upstream.

v1.8.1 — 2026-04-18

Patch release. Completes the three-layer structure introduced by v1.8.0 (principle → pattern → implementation) by adding the design-pattern skill paired with ADR-0010, records the self-referential origin loop in ADR-0010's Notes, and extends inspiration.md with the daily-research upstream that preceded Contemplative Agent.

Added

  • docs/skills/signal-first-research.md — design-pattern skill paired 1:1 with ADR-0010. Covers the define-signal-before-search rule, the search-wide / intake-narrow asymmetry, three abstract worked examples (daily digest, session-scoped filter, exploration mode), and a diagnostic checklist. Written as a generic pattern guide — personal project references live in docs/inspiration.md, not in the skill body.
  • docs/inspiration.md new "Upstream: a daily signal filter" section — documents that AKC's origin chain extends one hop further back than Contemplative Agent, to the daily-research pipeline that surfaced Laukkonen et al. (2025) in the first place and later generated the report that triggered ADR-0010.

Changed

  • docs/adr/0010-...md Notes — extended with an "Origin loop" observation: the pipeline that named the signal-first principle had been practicing it for months. The principle was lived before it was stated.
  • README.md / README.ja.md — design-pattern skill count 3 → 4 in the repo tree and opening summary.
  • docs/skills/README.md — index table extended with signal-first-research ↔ ADR-0010 pairing.
  • CITATION.cff bumped to v1.8.1.
  • llms.txt — ADR-0010 entry cross-references the new pattern skill; design-pattern skills list extended.

v1.8.0 — 2026-04-18

Architectural depth release. AKC surfaces the implicit design philosophy that has shaped every phase since inception: human cognitive resources are the central constraint. The Research phase is redefined as signal-first, a new design principle is added, and "Why a cycle?" is extended with the cognitive-economy paragraph. No code changes; this is a documentation release.

See ADR-0010 for the full rationale.

Added

  • ADR-0010 Human Cognitive Resource as Central Constraint — articulates the implicit philosophy driving every AKC phase: as agent capability grows, the scarce resource shifts from compute to human attention and judgment. Names three concrete mechanisms: signal-first Research, cognitive economy as Design Principle #9, and pre-implementation dialogue as investment.
  • Design Principle #9 "Human cognitive resource is the bottleneck" added to the README. Nine principles total now.
  • "Whose cognitive budget is the cycle protecting?" subsection in the README "Why a cycle?" section — explicit statement that the cycle's target resource is human attention, with one line per phase tying the design back to that target.
  • v1.8.0 geo-writer snapshot — baseline and post-change scores on README are recorded in ADR-0010 as a self-application of the Measure phase to AKC's own documentation.

Changed

  • docs/akc-cycle.md Research phase — retitled from "Search before building" to "Search broadly, filter by signal". A new step 0 asks the agent to define the signal (what information would change the next action) before searching. An intake paragraph clarifies that signal-first is the default stance; exploration phases are legitimate exceptions when named explicitly.
  • README Research summary rows — both the cycle table and the rules table restate Research as "Search broadly, filter by signal".
  • CITATION.cff — bumped to v1.8.0. Abstract extended with the cognitive-economy framing. Added keywords: signal-first, cognitive-economy.
  • llms.txt — opening summary extended with the cognitive-resource axis. ADR-0010 added to the index.
  • docs/adr/README.md — ADR-0010 added to the index table.

Not changed

  • The six phases. Research → Extract → Curate → Promote → Measure → Maintain remain unchanged in number, order, and composition.
  • ADR-0001 through ADR-0009. All prior decisions remain in effect.
  • Reference implementation (examples/minimal_harness/). No code changes.
  • Tagline. "A knowledge cycle for AI agents — one that grows with the people who shape it" is retained. ADR-0010 adds depth beneath the ADR-0009 framing rather than replacing it.

v1.7.0 — 2026-04-11

Conceptual repositioning. AKC is no longer described as a "self-improving harness." It is now a knowledge cycle for AI agents — one that grows with the people who shape it. No code changes; this is a positioning release with significant documentation updates.

See ADR-0009 for the full rationale.

Added

  • ADR-0009 AKC is a Cycle, Not a Harness — articulates why the "self-improving harness" framing was wrong (layer confusion with ECC, two-pillar drift, missing bidirectional growth) and how AKC is now positioned. Catalogues the three concrete changes: cycle as the sole defining characteristic, Security by Absence demoted to a design principle, bidirectional growth made explicit.
  • README "Acknowledgments" section — explicit recognition that AKC grew out of daily use of Everything Claude Code (ECC). Without ECC as the practical baseline, AKC would not exist.
  • README "References" section — post-hoc theoretical resonances noticed after the fact: Thompson's Mind in Life (structural coupling) and Laukkonen, Friston, & Chandaria's A Beautiful Loop (recursive self-modeling). AKC was built from practice, not theory; these references are listed for readers who might find the resonance interesting.
  • README bidirectional growth language — the cycle is not a one-directional optimization loop. The human side changes too: judgment about what to keep and discard, intuition about when to adopt versus build, sensitivity to what makes a rule good versus vague.

Changed

  • README tagline — "A memory-centric, self-improving harness for AI agents" → "A knowledge cycle for AI agents — one that grows with the people who shape it."
  • README "What is AKC?" — removed Security by Absence as a co-equal pillar. The cycle is now the sole defining characteristic. Security by Absence remains in Design Principles #7 and ADR-0001.
  • docs/akc-cycle.md — removed the Security by Absence section. Security remains as a design principle and as ADR-0001; it is no longer presented as a phase-equivalent rule in the rules file.
  • CITATION.cff — bumped to v1.7.0. Abstract rewritten to match the new positioning. Removed harness keyword; added bidirectional-growth and human-ai-collaboration keywords.
  • llms.txt — opening summary rewritten. ADR-0009 added to the index.
  • README BibTeX note — updated to match new tagline.
  • GitHub repository description — updated to match new tagline.

Not changed

  • The six phases (Research → Extract → Curate → Promote → Measure → Maintain). The cycle skills (search-first, learn-eval, skill-stocktake, rules-distill, skill-comply, context-sync).
  • ADRs 0001–0008. Schemas. Reference implementation (examples/minimal_harness/). Design-pattern skills.
  • Security by Absence as a design principle and as a load-bearing guarantee. Only its rhetorical position changed, not its content.
  • The repository name, the abbreviation AKC, and the project identifier.

Why a minor bump

Pure semver would call a docs-only change a patch. v1.7.0 is a minor bump because the README is part of the project's identity contract: what AKC is changed in this release, even though no code did. The distinction matters for academic citation — work referencing AKC v1.6 encounters the harness framing, while v1.7 onward encounters the cycle framing.

v1.6.0 — 2026-04-09

Additive release. The eighth ADR and a new category of skills (design-pattern skills) — long-form "how" guides paired 1:1 with ADRs, shipped inside the repository alongside the ADRs they correspond to.

Added

  • ADR-0008 Code and LLM Collaboration — names the principle behind ADR-0004 (two-stage distill), ADR-0005 (approval gate), and ADR-0007 (untrusted boundary): code owns determinism and control flow, LLMs own meaning. Catalogues four layering patterns (LLM→Code guard, Code filter→LLM, LLM judge + Code enforce, Code orchestrator + LLM worker).
  • docs/skills/ — new directory for design-pattern skills:
    • when-code-when-llm.md — per-task decision framework (structural vs semantic) with worked examples in both directions
    • code-and-llm-collaboration.md — per-pipeline layering patterns with composition rules and a diagnostic checklist
    • llm-agent-security-principles.md — concrete defense patterns for ADR-0001 / ADR-0006 / ADR-0007 including HTTP, credential, and LLM-host hardening
  • README Design Principle #8 — Code-LLM Layering.
  • README "two kinds of skills" — explicit distinction between cycle skills (six external repositories, one per phase) and design-pattern skills (in-repo, paired with ADRs).

Changed

  • ADR-0001 Security by Absence — added an Audit test section with a ready-to-run git grep compliance check and a pointer to the llm-agent-security-principles skill for the full "how" reference.
  • docs/adr/README.md — ADR-0008 added to the index.
  • llms.txt — ADR-0008 entry and a design-pattern-skills section added.
  • CITATION.cff — bumped to v1.6.0, date updated.

Relationship to contemplative-agent

The three design-pattern skills are adapted from skills accumulated during contemplative-agent research work. The contemplative-agent repository itself is unchanged and remains an independent research project.

v1.5.0 — 2026-04-09

Additive release. The six-phase cycle, the six skills, and the rules install path are unchanged. v1.5 adds a second foundational principle (Security by Absence), seven ADRs, JSON schemas, and a dependency-free Python reference implementation adapted from the contemplative-agent research repository.

Added

  • docs/adr/ — seven new architecture decision records:
    • ADR-0001 Security by Absence as a First-Class Principle
    • ADR-0002 Immutable Episode Log as Source of Truth
    • ADR-0003 Three-Layer Distillation (Raw → Knowledge → Identity/Rules)
    • ADR-0004 Two-Stage Distill Pipeline (Free-form → Format)
    • ADR-0005 Human Approval Gate for Behavior-Modifying Changes
    • ADR-0006 Single External Adapter per Agent Process
    • ADR-0007 Untrusted Content Boundary
  • schemas/episode-log.schema.json and schemas/knowledge.schema.json — JSON Schema definitions for the two persistent layers, usable independently of any implementation.
  • examples/minimal_harness/ — a dependency-free Python reference implementation (~300 lines) demonstrating the three-layer memory architecture and the two-stage distill pipeline. Runs with a deterministic fake LLM; no network or API keys required.
  • docs/inspiration.md — prior-art credit to the contemplative-agent research repository.
  • docs/akc-cycle.md — new 7th principle: Security by Absence. Loaded as a rule, so the principle applies to every session that installs the rules file.

Changed

  • README.md — rewritten intro. AKC is now described as a "memory-centric, self-improving harness" built on two complementary principles (the cycle, and Security by Absence). The six-skill table, rules install path, harness-engineering comparison, and scaffold dissolution links are preserved. Design Principles gain a 7th entry.
  • CITATION.cff — bumped to v2.0.0, abstract updated to match the new positioning, date-released updated.

Archived

  • docs/history/README-v1.4.md — the v1.4.0 README is preserved verbatim for citation integrity.

Not changed

  • The six-phase cycle (Research → Extract → Curate → Promote → Measure → Maintain) and the six external skill repositories.
  • The Design Principles 1–6 (composable, observable, non-destructive, tool-agnostic, evaluation-scales, scaffold-dissolution).
  • docs/akc-cycle.md phases — only an additional principle is appended.

Relationship to contemplative-agent

AKC v2.0 absorbs specifications and reference code from contemplative-agent, which remains an independent research repository. AKC does not depend on contemplative-agent; contemplative-agent does not depend on AKC. See docs/inspiration.md.

v1.4.0 — 2026-02-22

Initial Zenodo release. DOI: 10.5281/zenodo.19200727.