Changelog Adopt

Claude Changelog Adopt

The self-improvement loop that keeps OmegaOS current with Claude Code itself. Every day it reads the one authoritative list of what Claude Code shipped, works out which entries should change how OmegaOS or its agents behave, and hands you a decision report — the same move we made by hand when /loop shipped and became R-LOOP's native-loop doctrine, now standing and automatic.

It is a loop in the R-LOOP sense: a verifiable goal (absorb every new changelog entry), a memory (the last-absorbed version + a per-entry decision store), and a hard ceiling (it proposes; the operator, or an armed + gated oracle dispatch, decides — it never silently rewrites the OS).

Complementary to /ecosystem-watch: that is the broad, noisy X radar that also feeds the public @Agentik_os feed. This is the narrow, authoritative source — the official changelog — pointed straight at OmegaOS self-upgrade, with no marketing side-output.

Pipeline

cron (daily) → [1] FETCH  official CHANGELOG (GitHub raw: anthropics/claude-code)
             → [2] DIFF   vs state/last_version → NEW entries only (first run seeds forward)
             → [3] CLASSIFY (claude, opus)  per entry: relevance · category · target surface ·
                                            in_scope · concrete proposal (what/where/why) · 0-10
             → [4] GATE   (claude, opus)  adversarial: is the feature real per the entry text? is
                                          the patch correct, minimal, in-scope, non-duplicative?
                                          default-REJECT on doubt (anti-fabrication, R-VERIFY)
             → [5] REPORT self-contained HTML in ~/.omega/artifacts/  (R-HTML)
             → [6] ALERT  Telegram (Alerts topic) via omega-alert-send.sh
             → [7] ADOPT  (only if ARMED)  dispatch each vetted proposal to an OmegaOS oracle,
                                           bounded (R-LOOP), In-Review handoff — never auto-Done

Deterministic control (fetch, diff, version watermark, dedup, dispatch side-effects) lives in shell + small python. Judgment (classify, propose, refute) lives in the model. The gate is an INDEPENDENT adversarial pass whose only job is to reject — default-reject on doubt.

Auto-adoption scope (operator decision, non-negotiable in the runner)

The engine may only PROPOSE auto-patches to these surfaces:

• Doctrine — crates/omega-core/src/rules.rs (the Laws/Rules SSOT) + its markdown mirror.
• Agent identities — agents/*.md (oracle, worker, Atlas, Nova/companion, the AISB matrix).
• Skills — skills/<name>/ (new or amended skills, e.g. adopting a new Claude Code primitive).
• Installer — install.sh parity for any of the above (L0).

A changelog entry whose only sensible adoption touches the core Rust orchestration (crates/ beyond rules.rs — executor, loop_guard, dispatch, TUI) is classified surface: core-rust, in_scope: false, and surfaced as needs-human in the report. The engine never proposes an auto-patch to the compiled core; that stays an explicit operator call.

Adoption safety rails (non-negotiable)

Acting on OmegaOS itself (dispatching an adoption) requires ALL of:

• armed flag present (~/.omega/state/changelog-adopt/armed) — the kill-switch. Absent = proposals are report-only. rm the flag to stop all self-modification instantly.
• the adversarial gate returned keep=true for that proposal.
• in_scope: true (surface is doctrine/agents/skills/install.sh — never core-rust).
• the entry fingerprint is not already in the seen-store (never adopt the same entry twice).
• the per-run adoption cap (CA_DAILY_ADOPT_CAP, default 3) is not exhausted.

Even armed, an adoption is DISPATCHED to an oracle that runs the full quality gate and hands back In-Review — the operator still approves the merge. The engine proposes and prepares; it never self-marks Done and never force-pushes (L0, R-VERIFY, R-LOOP).

Run

bash scripts/adopt-run.sh                 # full run (fetch → classify → gate → report → alert)
CA_DRYRUN=1 bash scripts/adopt-run.sh     # never dispatches even if armed (report only)
CA_FORCE_VERSION=2.1.100 bash scripts/adopt-run.sh   # re-process everything newer than X (backfill)

Env knobs: CA_CHANGELOG_URL (GitHub raw default), CA_CLASSIFY_MODEL (claude-opus-5), CA_GATE_MODEL (claude-opus-5), CA_DAILY_ADOPT_CAP (3), CA_DISPATCH_PROJECT (OmegaOS), CA_SEED_VERSIONS (1 — how many latest versions the FIRST run treats as new).

Arm / disarm

touch ~/.omega/state/changelog-adopt/armed   # allow gated auto-dispatch of adoptions
rm    ~/.omega/state/changelog-adopt/armed    # kill-switch: proposals stay report-only

State (runtime, gitignored — never in the repo, R-ENV)

• ~/.omega/state/changelog-adopt/last_version — the forward watermark (last version absorbed).
• ~/.omega/state/changelog-adopt/seen.jsonl — per-entry decision fingerprints (dedup memory).
• ~/.omega/state/changelog-adopt/armed — the kill-switch flag (absent by default).
• ~/.omega/state/changelog-adopt/runs/<date>/ — per-run raw changelog, new entries, classify, gate, report.

Cron

# 08:15 daily — DISARMED by default until the operator arms it.
15 8 * * * PATH=$HOME/.local/bin:/usr/local/bin:/usr/bin:/bin OMEGA_DIR=$HOME/.omega \
  bash $HOME/.omega/skills/changelog-adopt/scripts/adopt-run.sh \
  >> $HOME/.omega/logs/changelog-adopt.cron.log 2>&1