v0.4.0 — eval-harness: grade the fleet, publish the failures too

Grade your agent fleet against rules it must never break.

New: modules/eval-harness/ (v0.1.0)

Turns a multi-agent event log into reproducible traces and scores them against explicit safety invariants. Zero LLM tokens, pure stdlib, and the evaluator's exit code is the number of failed checks.

Four deterministic behavioural invariants:

• a human gate before a risky commit,
• an independent verify before commit,
• no duplicate-event storms,
• escalations get resolved.

Two benchmarks built from six weeks of our own production fleet: a readable showcase (public-live-v0) and the full 317-event corpus (consensus-safety-v0).

We publish our own failure alongside our own pass: human gate before a risky commit 100%, independent verify before commit 7.7%. The second number is bad, it is ours, and a harness that only ever reports the good invariant is a marketing asset rather than an eval.

Sanitisation ships as two deterministic paths — a structure-only whitelist, and allowlist curation that keeps real text — and the host role map is local config, never source.

Also in this release

• AGENTS.md — determinism and sanitize.py are the two load-bearing rules for anyone, human or agent, changing this repo.
• FOR-ROBOTS.md, changelog categories for these auto-generated notes, and the contributor deal inherited from one org-wide CONTRIBUTING.md instead of a local copy that silently shadowed it (we found six repos carrying the stale copy and deleted them).

About the timing

The module landed on 29 July and sat in Unreleased for six days while the README already described it as shipped. Under the release discipline this repo now follows, a shipped module is a release — cut when the work lands, not when someone remembers.

What's next

A Second Brain reference implementation (vault + RAG + reranker), the graph/associative recall layer with entity-vs-theme gating, the C(H+A)RM relationship layer, and the import pipelines. The release feed is how you tell what is real here from what is still a manifesto.

Full Changelog: v0.3.0...v0.4.0

v0.3.0 — RDR loop CLI

v0.3.0 — the RDR loop, as a CLI.

RDR is the spine of CharmOS: Recall → Deep Research → synthesis. Before acting on anything strategic, recall what you already know, research only the gaps, then decide — in that order. This module turns that loop into three deterministic commands (pure stdlib, zero LLM tokens), built directly on the TurnState ledger shipped in v0.2.0.

• rdr recall "<q>" — what do I already know? Whole-word search over your TurnState ledger (+ an optional notes dir).
• rdr research "<q>" — emits a Deep-Research prompt with your recall pre-loaded as CONTEXT, so research targets the gaps, not what you know.
• rdr memo "<q>" — scaffolds a Decision Memo (problem → what we know → research → options → risks → decision).

Quality: this module passed an independent adversarial review before publishing — defensive parsing of ledger cells, a schema-mismatch warning (a broken integration no longer masquerades as "no results"), non-UTF8 tolerance, and a whole-word scorer. See modules/rdr/README.md.

Full notes in the CHANGELOG.

v0.2.0 — TurnState: always-on memory

v0.2.0 — first runnable module.

CharmOS ships its first real code: modules/turnstate/, the always-on memory pillar from the manifesto.

TurnState gives an AI coding/agent session a durable, queryable working memory without spending a single LLM token to maintain it. After each turn it writes one deterministic row (ask · summary · files · tools · commands · decisions · evidence pointer) to a local SQLite ledger.

Two halves:

• turnstate_hook.py — real-time Stop-hook fast path. Pure stdlib, never raises into the session.
• turnstate_backfill.py — idempotent rebuild straight from transcripts + a --check freshness gate, so the ledger can never silently rot.

The raw transcript stays the source of truth; TurnState is the cheap derived index you recall against. Local-only, no data leaves your machine. See modules/turnstate/README.md.

Full notes in the CHANGELOG.

v0.1.0 — Manifest-first

CharmOS · {C(H+A)RM} — V1, manifest-first.

The first open-source framework that manages your human contacts AND your AI agents as first-class relationships, on top of a personal Second Brain, driven by one decision loop: RDR (Recall to Deep Research to synthesis).

This V1 ships the vision, architecture, and docs before the code:

• README.md, MANIFESTO.md
• /docs: architecture, the RDR loop, the C(H+A)RM category, privacy
• /examples: synthetic-only sample vault + contacts (a human and an AI agent)
• LICENSE (Apache-2.0), .env.example

Privacy: no real personal data is included — everything under /examples is synthetic.

Runnable modules land in later phases. We version with SemVer and keep a CHANGELOG. Feedback welcome.