[New1Direction]

Summary

Makes Korg's core promise real for the data users actually generate: every Claude Code session is captured passively, cross-platform, into a tamper-evident korg-ledger@v1 chain that any third party can independently verify — turning the project from "a thing you set up" into "a thing that's just on."

This implements Phase 1 of the "verifiable flight-recorder for AI agents" direction (design + plans under docs/superpowers/). It unifies on the one verifiable ledger format, ships a zero-config capture hook, backfills history, and reserves the Phase-2 trust slots — all without breaking the frozen cross-language conformance vectors.

What's included

• adapters/korg-ledger-py/ (new) — a stdlib-only, third independent conformant producer of korg-ledger@v1 (alongside the Rust crate and JS verifier). LedgerWriter hash-chains JournalEvent JSONL with HLC ordering, a strictly-earlier causality gate, HMAC mode, resume-from-tip, and a concurrency-safe append (re-reads the chain tip under an exclusive lock so overlapping writers can't fork the chain).
• Zero-config capture (adapters/claude-code/) — make_canonical_emit + a korg-hook driver registered on PostToolUse/Stop/SubagentStop. It reuses the existing session parser, captures full richness (llm_inference, tokens, cache, tool results, the §2a causal spine) into a per-session verifiable ledger ~/.korg/sessions/<id>.jsonl, and never raises / always exits 0.
• One-command setup (adapters/korg-setup/) — claude_settings.py idempotently registers the hook in ~/.claude/settings.json; the macOS-only launchd daemon is demoted to opt-in (--daemon).
• Retroactive history — korg-backfill re-derives canonical per-session ledgers from every historical transcript (idempotent), plus a guarded migrator for the legacy flat ~/.korg/claude-events.jsonl.
• Format reservation (lockstep) — event_sig is excluded from the hash preimage across all four implementations (Rust, two Python, JS); the Rust JournalEvent gains a wire-compatible event_sig field; an additive LedgerRewind variant and the external-anchor path are reserved for Phase 2.

Key design decisions

• Per-session ledger files (not one interleaved file): one verifiable chain per causal session — dodges cross-session races, matches the per-session root, and is the shareable "verify in a browser" receipt Phase 3 will publish.
• Full-reparse + emitted_count capture: PostToolUse fires after a tool's result lands, so re-parsing the transcript each firing captures results the buffered parser fills in by mutation (which incremental tail mode drops).
• Backfill preferred over flat migration: source transcripts persist and are richer, so backfill produces output identical to live capture; migration is a guarded fallback that never duplicates backfill.

Test plan

All verified locally before this PR:

• 145 Python tests pass across the three packages (korg-ledger-py / claude-code / korg-setup), incl. conformance against the frozen vectors, cross-producer verification by the spec's own oracle, idempotency, and never-raises robustness.
• Rust: cargo test -p korg-ledger -p korg-registry green, incl. conformance_vectors_reproduce_the_frozen_oracle (frozen vectors byte-identical with the new field) and the LedgerRewind serialization test.
• JS: node spec/korg-ledger-v1/js/conformance.mjs PASS (incl. the new event_sig excluded assertion).
• Python spec oracle: python3 spec/korg-ledger-v1/conformance.py PASS.
• End-to-end: the real korg-hook stdin entry point and the korg-backfill CLI both produce per-session ledgers verified clean by the independent spec oracle.

Out of scope / follow-ups

• Plan 2b — JSONL-on-disk storage refactor for the server/runtime CapabilityJournal (internal; off the user path; capture ledgers are already JSONL).
• Plan 6 — point recall-mcp at the per-session ledgers, add a cross-producer conformance CI gate, and the honest-docs pass.
• Phase 2 — populate event_sig (per-event signing), external time-anchoring / transparency-log notarization, wire rewind() to append LedgerRewind.

🤖 Generated with Claude Code

[New1Direction]

Update: Plan 6 landed on this branch — Phase 1 is now complete in this PR.

The "Plan 6" item listed under Out of scope / follow-ups has been implemented and pushed:

• Readers unified — recall-mcp now reads the canonical per-session ledgers (_normalize_record handles both the canonical JournalEvent shape and the legacy flat format), and its default source is ~/.korg/sessions/.
• Cross-language conformance CI gate — a new conformance job runs the Python and JS spec oracles plus the producer/capture/setup/reader pytest suites on every push (Rust frozen-vector conformance is already gated by the existing cargo test). Verified locally: 164 green.
• Honest docs — the claude-code adapter README now describes the verifiable per-session capture path and marks the flat claude-events.jsonl as legacy.

Still genuinely out of scope (follow-up PRs): Plan 2b (JSONL-on-disk for the server-side CapabilityJournal), and Phase 2 (populate event_sig signing + external anchoring + wire rewind() → LedgerRewind).