Repository navigation
Upstream sync — Claude Code v2.1.162 → v2.1.195
Captured + incorporated 4 changelog cycles via /forge watch + /forge update. Verbatim verification against GitHub anthropics/claude-code/CHANGELOG.md before incorporation (caught and rejected 6 hallucinated findings from the watch agent).
BREAKING (upstream)
domain/hook-architecture.md— Hook matchers with hyphens (e.g.code-reviewer,mcp__brave-search) now exact-match instead of substring-matching (v2.1.195). Audit dotforge agent matchers (code-reviewer,security-auditor,session-reviewer,test-runner) — they're exact-match candidates.
Security / governance
sandbox.credentialsblocks credential file + secret env var reads (v2.1.187)sandbox.allowAppleEventsopt-in for macOS (v2.1.181)requiredMinimumVersion/requiredMaximumVersionmanaged settings (v2.1.187)attribution.sessionUrlto omit claude.ai links in public repos (v2.1.187)
Model / agents
- Claude Fable 5 (Mythos-class) — new top tier above Opus (v2.1.170)
- Sub-agent 5-level nesting (v2.1.172)
- Agent Teams implicit team (v2.1.178)
claude_code.assistant_responseOTel event (v2.1.193, PII opt-in)claude agents --json waitingForfield (v2.1.162)/effortpersistence confirmation (v2.1.162)
Hooks / observability
post-sessionlifecycle hook for self-hosted runners (v2.1.169)autoMode.classifyAllShellroutes every shell call through classifier (v2.1.193)
CLI / session
--safe-mode/CLAUDE_CODE_SAFE_MODEfor triage (v2.1.169)/cdcache-preserving working-dir change (v2.1.169)/rewindafter/clear(v2.1.191)--toolsnative-build conditional (v2.1.162)CLAUDE_CODE_DISABLE_MOUSE_CLICKS(v2.1.195)disableBundledSkills(v2.1.169)- Bedrock reads AWS region from
~/.aws(v2.1.172) /plugin list(v2.1.163)
Non-breaking for dotforge
All edits are domain-rule content additions — no template, hook, or settings.json changes. Existing projects continue to work; running /forge sync propagates the updated domain knowledge.
Pipeline
- 4 captures:
inbox/→active/withincorporated_in: ["v4.1.0"] - 4 metrics entries (1 monitoring on BREAKING hook matcher, 3 not-applicable)
- Researcher hallucination rate: 6/25 findings invented (24%). Verification step (WebFetch CHANGELOG verbatim) caught all. Lesson reinforced: never trust agent-reported version numbers without source quote.