… JSONL

Three endpoint flips at once, all reading from CC#3's
${KORA_HOME}/kora_audit_log.jsonl (KR-AUDIT-JSONL-SINK, PR #139).
Backend-only — no FE changes; TS interfaces already match the
projected shapes. Same stub-then-real pattern as PR #137's
slack-dm flip.

Shared reader (NEW):
  * kora_cli/audit/jsonl_reader.py with read_audit_entries(
    seam, limit, since) — single helper used by all 3 endpoints.
  * Per-call get_kora_home() lookup so monkeypatch in tests works
    without ContextVar plumbing (per the #137 fixture lesson).
  * Tolerates missing file, malformed JSON, non-dict lines,
    Pydantic ValidationError on a single row — log + skip.
  * Newest-first by emitted_at descending.

Flip 1 — /api/agent-activity/recent (mcp.tool_called):
  * Filters audit rows where seam == "mcp.tool_called", projects
    to AgentCall shape.
  * K-DG drift caught: spec said details has duration_ms /
    tool_status / result_summary. Actual writer at
    kora_cli/listeners/mcp_tools.py:714-724 only emits tool_name
    / tool_kind / caller_actor_kind / args_keys / result.
    Projection: duration_ms=0; status="ok" (audit only fires on
    success path today; KR-MCP-RUNTIME-SURFACE follow-on extends);
    result_summary=details.result (writer docstring confirms this
    is pre-filtered to a short string).

Flip 2 — /api/reasoning/recent (reasoning.tool_called):
  * Filters audit rows where seam == "reasoning.tool_called".
  * GROUPS by caller_session_id so a multi-tool reasoning
    iteration collapses into ONE ReasoningCall row with
    tools_used: [name1, name2, name3]. tools_used is a FE
    extension field — existing TS ReasoningCall ignores it;
    follow-on FE bucket renders.
  * Status derivation per spec: all-ok → ok; any not_allowed →
    halted+capability_denied; any execution_error → failed+
    handler_error.
  * Limitations (per spec §2 Flip 2): model_used / tokens /
    response_text_truncated_200 set to null — those fields live
    in slack_dm_log.jsonl outbound entries, not audit. The
    KR-REASONING-PANEL-MODEL-XREF follow-on bucket cross-
    references by timestamp + thread_ts. cost_rung_at_call set
    to "unknown" (lowercase Enum.value — preserves the K-DG pin
    from PR #132 against engine.py:47-49 CostLadderRungName
    literal). Keeping this PR small per spec §4.
  * Aggregate counts (total_recent_24h, by_status_24h) use
    INDIVIDUAL audit rows — not groups — so headline reflects
    total reasoning activity volume.

Flip 3 — /api/webhooks/events/recent (webhook.dead_letter):
  * Filters audit rows where seam == "webhook.dead_letter".
  * status pinned to "dead_letter" (this seam ONLY emits dead-
    letters; verified happy-path events stay in chain log; rate-
    limited from slowapi doesn't currently call emit_audit).
  * SECURITY: source_ip OCTET-MASKED at the projection edge.
    Audit writer at webhook_dead_letter.py:142 passes RAW
    peer_ip; endpoint enforces _mask_ipv4_last_two_octets
    ("54.203.99.142" → "54.203.x.x") per the KR-WEBHOOK-EVENTS
    #109 contract. IPv6 / unexpected shapes → "—" defensively
    (never leaks unmasked).
  * SECURITY: details sub-set to {reason, header_present} —
    never the full audit details dict (which carries
    body_bytes / request_id / headers the panel hasn't vetted).

4-layer SECURITY contract carry-forward (preserved across all 3
endpoints):
  * Walk-payload sweeps for Anthropic key shapes, Slack token
    shapes, HMAC-secret shapes, full IPv4 leaks.
  * Per-field caller_actor_kind label-shape pin (no hash/base64
    runs).
  * result_summary no-raw-JSON pin.
  * cost_rung lowercase literal pin.
  * details sub-set enforcement (webhook).

Fixture-isolation discipline applied per #137 lesson: all 3
endpoint test files monkeypatch get_kora_home in ALL THREE
module namespaces (kora_constants, kora_cli.config,
kora_cli.web_server). Reader tests also use the pattern. The
reader's own resolution uses a local kora_constants import so
the test patch takes effect on a fresh call.

Tests:
  * tests/kora_cli/audit/test_jsonl_reader.py — 14 reader tests:
    missing file, empty file, seam filter, since filter, naive
    datetime UTC assumption, limit cap, malformed JSON tolerance,
    non-dict line skip, Pydantic ValidationError per-line skip,
    blank lines, KORA_AUDIT_LOG_PATH env override.
  * test_web_server_agent_activity.py — 11 tests (rewrite):
    empty log, shape, mcp.tool_called projection, seam filtering,
    ?limit cap, newest-first, walk-payload SECURITY, label-shape
    caller_actor_kind, no-raw-JSON result_summary, by_caller_24h
    reconciliation, cron-regression sanity.
  * test_web_server_reasoning.py — 15 tests (rewrite): empty,
    shape, single-tool projection, multi-row session grouping,
    different sessions → separate rows, duration_ms sum, status
    derivation (ok / halted / failed), seam filtering,
    walk-payload SECURITY, lowercase cost_rung pin, ?limit on
    groups, by_status_24h counts individual rows not groups,
    cron-regression sanity.
  * test_web_server_webhook_events.py — 15 tests (rewrite):
    empty, shape, dead_letter projection, status pinned, source
    → endpoint mapping, IPv4 octet-mask enforcement, walk-payload
    no-full-IPv4 sweep, IPv6/dash → "—" fallback, details sub-
    set enforcement, seam filtering, newest-first, ?limit cap,
    cron-regression sanity.
  * Full admin-panel regression: 357/357 across 27 suites.

K-DG drift summary:
  * spec §2 Flip 1 said details has duration_ms / tool_status /
    result_summary — only `result` is present in the actual
    writer. Documented + handled in projection.
  * spec §2 Flip 2 said cost_rung could be null — FE
    ReasoningCostRung union requires a value; used "unknown"
    (which IS in the enum) instead. Preserves the lowercase
    Enum.value contract from PR #132.

Follow-on buckets cited:
  * KR-REASONING-PANEL-MODEL-XREF — cross-ref to
    slack_dm_log.jsonl for model_used / tokens / response_text
  * KR-MCP-RUNTIME-SURFACE follow-on — extends audit writer with
    duration_ms + failure-path emit (status taxonomy)

Refs:
  * rafe-walker/kora-docs 17_cc_bucket_prompts/KR-AUDIT-PANEL-ENDPOINTS_three_flips.md
  * PR #139 — KR-AUDIT-JSONL-SINK (writer + AuditEntry shape)
  * PR #137 — KR-SLACK-DM-PANEL-FLIP (fixture-isolation pattern)
  * PR #114 / #132 / #109 — original stub panels being flipped

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>