[cyharyanto]

• Pull cognition parsing out of the kernel by introducing strong-typed adapter outputs (ActionHint, DecisionLine, JudgeOutput), a shared runtime::canonicalize utility for cross-vendor canonicalization (markdown fences, prose-around-JSON, balanced-brace extraction), and renaming CognitionResponse.text → content with role-specific typed fields populated by adapters.
• Give the Actor an agentic inner loop bounded by DEFAULT_INNER_STEP_BUDGET so pure-reasoning replies continue the turn while structured ActionHint responses commit, replacing the prior single-turn JSON-parse-or-fail surface.
• Replace the findings-specific Tool with a generic record substrate by removing record_finding, renaming Finding/*FindingsBackend to Record/*RecordStore, routing record appends through modify_artifact with kind=record-store, and letting the Charter define record content shape while the kernel injects _* provenance.
• Add the Skills primitive as Actor-side cognition instrumentation by adding core::skill::Skill (content-hashed, SKILL.md convention), folding skills into the Snapshot identity hash, loading them from <charter_dir>/skills/*.md, and composing them into the Actor system prompt under grounding.
• Drop filesystem IO from the kernel by splitting charter_loader into pure parsers (parse_charter_def, parse_role_context_def) inside core/ and deployment-side IO (load_charter_def, load_role_context_def, load_skills) inside runtime::charter_loader.
• Unify durable streams on one async append-only primitive by adding chartered_dispatch::JsonlSink serving Receipts, the cognition log, and the record store, making ReceiptStore::append async, and codifying the discipline in a new spec §Persistence.
• Make Snapshots persisted and addressable by adding chartered_dispatch::snapshot_store (one immutable file per <chartered_dir>/snapshots/<id>.json), persisting on Snapshot creation, and adding spec §Snapshot Lifecycle so every Receipt's snapshot_id resolves to a record on disk.
• Rework the v1 protocol by giving Receipt receipt_id/task_id/attempt_id/steward_id, replacing EnforcementLevel with structured GovernanceMode (grounding+evaluation toggles), promoting Verdict.frame_id to a Steward-scoped FrameRef, renaming ToolCall.tool_name → tool while dropping its timestamp, and removing Metrics.
• Add the Gemini cognition backend as a sibling adapter to OpenAI by introducing runtime::gemini_backend with lazy per-model outputTokenLimit/thinking caching, native role mapping (System → systemInstruction, Assistant → model), and an explicit non-forwarding policy for LLM-native tools (search/codeExec/urlContext are cognition, not governed Tools).
• Rename the OpenAI environment variables from LLM_* to OPEN_AI_* across .env, runtime, dashboard launcher, docs, and tests so per-vendor configuration is decoupled from the abstract LLM name now that Gemini exists alongside OpenAI.
• Make the dashboard read its Charter view from the kernel by adding chartered-runtime --print-charter (JSON emit of the parsed Charter), replacing the dashboard's parallel TOML+markdown parser in local-api.mjs with a subprocess call, so the Charter parser has one source of truth.
• Codify three-tier test discipline by formalizing unit (stateless, CI) / integration (vertical cut, tempdir-isolated, CI) / e2e (real LLM via production transport, local-only, #[ignore]d, no soft-skip) in AGENTS.md §Verification, adding docs/IMPLEMENTATION_CHECKLIST.md §Test Tier Discipline, renaming tests/e2e.rs → tests/binary_integration.rs, adding tests/persistence.rs, and gating llm_e2e.rs behind #[ignore].
• Expand the specification to cover the newly-incorporated subsystems by adding §Persistence, §Snapshot Lifecycle, §Skills, §Subagents, §User-Facing Integration Boundary, the "Cognition is commodity" framing for §Cognition Layer, and the "LLM-side native tools are cognition, not Tools" clarification under §Tools.
• Add Transcription Fidelity discipline to AGENTS.md requiring user-intent artifacts to contain only the user's stated words, prohibiting silent assistant extrapolation into mandates and briefs.