Session replay: capture mode (auto/plan/YOLO) + approvals (count/decision/edit-diff) + wire transcript viewer

[vivekchand]

Context

Today the session UI shows nothing about how a session ran:

• Auto vs interactive vs plan vs --dangerously-skip-permissions — Claude Code emits {"type":"permission-mode","permissionMode":"default"} on every session (verified in ../clawmetry-pro/tests/fixtures/runtimes/claude_code/…jsonl:1), but _parse_v3_event at clawmetry/sync.py:3186 never consumes it. No sessions column stores it. grep permissionMode|acceptEdits|bypassPermissions|autoAccept across clawmetry/, routes/, app.js returns zero hits.
• Approvals asked count — the approvals DuckDB table (clawmetry/local_store.py:610) IS populated by the sync daemon (sync.py:7346). /api/session-governance/<sid> (routes/sessions.py:2887) returns {approvals[], decision_count, denied_count}. /api/session-insight/<sid> folds it into governance = {decision_count, denied_count}. The transcript viewer never fetches either endpoint.
• User replies to approvals — approvals.decision + decision_reason + resolver are stored, rendered only in the NemoClaw tab (app.js:8177).
• "Edit" approvals — where the user modifies the tool args before approving (a real Claude Code flow) — has no column and no ingest path.

Deliverables

1. Ingest mode + policy signals

• clawmetry/sync.py _parse_v3_event — consume permission-mode, mode, permissionMode (Claude Code); turn_context.approval_policy + sandbox_policy + permission_profile (Codex); session.auto_mode_resolved (Copilot); exec_approvals_config (OpenClaw); steps.permissions per step (Antigravity).
• New sessions columns: mode_permission, mode_sandbox, mode_collaboration, approval_policy_default, resolved_at (latest resolved timestamp — modes can change per turn in Codex + Claude Code Plan Mode).
• Per-turn mode changes stored as kind=mode.changed events in the new replay_events table (see OSS-01).

2. Extend approvals capture

• Add edit_diff JSON column to approvals table — {before: tool_args, after: tool_args} captured whenever the user modifies the tool_use payload before approving. Claude Code's toolDenialKind maps to decision=denied; a subsequent tool_use with different args after the same sourceToolUseID maps to decision=edited with edit_diff.
• Capture apiRefusalCategory / apiRefusalExplanation for model-side refusals as resolver=model.
• Capture hookInfos / hookErrors / hookAdditionalContext for hook-driven auto-decisions as resolver=hook.

3. Wire the transcript viewer

• /api/replay-tree/<id> (from OSS-01) includes approvals[] per turn.
• New UI in clawmetry/static/js/app.js:
  • Session-level mode chip in the transcript header — auto / plan / interactive / YOLO — with a tooltip listing sandbox + collaboration.
  • Per-turn mode marker — small badge on turn chapters where mode changed (Codex per-turn, Claude Code Plan Mode toggles).
  • Approvals rail running down the right side of the transcript, one row per approval, click to jump to the tool_use it gated. Colors: green approved, yellow edited, red denied, gray timeout.
  • Per-turn approvals count on the TURNS TOC (next to the existing 🔧 N / ✕ N).
  • Inline approval bubble rendered above the gated tool_use chip, showing prompt (if captured — see hook-lifecycle issue), decision, decision_reason, resolver, edit_diff.

Acceptance

• Every ingested session has mode_permission populated (or explicitly unknown for runtimes without a signal — see per-runtime issues)
• /api/replay-tree/<id> returns approvals[] matching what /api/session-governance/<id> returns today
• Opening a session with ≥1 approval renders the approvals rail; clicking scrolls to the gated tool
• Fixture session with an edited approval shows the diff (before/after tool args)
• Sessions with bypassPermissions/--dangerously-skip-permissions show a red YOLO chip

Depends on

• OSS-01 (schema)

Blocks

• Every per-runtime mapper issue (they need the columns + endpoint contract to write into)

Gotchas

• Approval outcome ≠ approval prompt. Almost every runtime persists only the decision, not the prompt text the user saw. Rebuilding the prompt for replay needs a runtime hook (see hook-lifecycle issue).
• Auto mode is per-turn (Codex, Claude Code), per-session (Copilot, OpenClaw), or per-project (opencode permission table). A single session-level "auto" chip is misleading — badge at the turn level where the runtime supports it.
• Do not conflate mode.changed events across runtimes with different vocabularies — keep the raw runtime value in payload, only normalize the top-level enum.

References

• clawmetry/local_store.py:610 (approvals schema)
• clawmetry/sync.py:7346 (approvals ingest today)
• clawmetry/sync.py:3186 (_parse_v3_event — the seam to extend)
• routes/sessions.py:2887 (/api/session-governance/<sid> — already returns most of what we need)
• routes/sessions.py:2905 (/api/session-insight/<sid> — folds governance into waste_flags)
• clawmetry/static/js/app.js:8177 (NemoClaw approvals renderer — reference implementation for the bubble)

[vivekchand]

Hi @vivekchand! Took a careful look at the codebase before commenting. Here's the plan:

The problem
Sessions show no information about how they ran — no mode chip (auto / plan / YOLO), no approval count, no inline approval outcomes — even though the Claude Code JSONL emits permission-mode events on every session and the approvals DuckDB table is already populated.

Current codebase state (as of today)
The foundation from #4813 is already in the tree:

• clawmetry/replay_schema.py — canonical schema with PermissionMode, ModeChip, ApprovalInfo, ReplayEvent types ✅
• replay_events DuckDB table with indices ✅
• /api/replay-tree/<session_id> endpoint in routes/sessions.py:2984 ✅
• window._cmReplayTree JS module in app.js:26152 with renderTree, fetchReplayTree, registerKindRenderer ✅

The one remaining piece of wiring in #4813 is noted in the code: // Wire-up into openTranscriptModal happens in #4814 (mode + approvals) (app.js:26148).

Approach

The full issue has three deliverables. Here is how they break down:

1. Ingest mode + policy signals — clawmetry/sync.py _parse_v3_event and a new Claude Code JSONL walker need to emit kind=mode.changed rows into replay_events when they encounter permission-mode events. New sessions columns (mode_permission, mode_sandbox, …) would also land here. This overlaps significantly with the Claude Code mapper in Session replay: Claude Code mapper — Task/Agent/Workflow fanout, inline sidechains, nested delegation #4815 and involves schema changes — best done as part of that PR.

2. Extend approvals capture (edit_diff column on approvals, apiRefusalCategory, hookInfos) — also a schema change; leave for a focused follow-up.

3. Wire the transcript viewer — call window._cmReplayTree.fetchReplayTree + renderTree from viewTranscript (app.js:17729). This is independently safe: when row_count === 0 (always true today, until adapters land) the existing flat renderer is untouched. When adapters populate replay_events the mode chip and turn tree appear automatically. This piece follows the exact same fire-and-forget pattern as _loadAuthorityPanel at app.js:17825.

Implementing now: deliverable 3 only (wire-up)

1. clawmetry/static/js/app.js:17825 — add _loadReplayTree(sessionId) call alongside the existing _loadAuthorityPanel(sessionId) call (both are fire-and-forget async).
2. clawmetry/static/js/app.js:17928 — define async function _loadReplayTree(sessionId) that uses window._cmReplayTree.fetchReplayTree + renderTree, falls back silently on any error or empty tree.

Test plan

• node --check clawmetry/static/js/app.js — syntax clean
• Existing tests/test_appjs_units.js (220 tests) must stay green
• Manual: opening any transcript still renders the flat viewer unchanged (row_count is 0 until adapters land)
• Future: once Session replay: Claude Code mapper — Task/Agent/Workflow fanout, inline sidechains, nested delegation #4815 populates replay_events, the mode chip and delegation tree appear without further wiring

Estimated change
~28 lines across 1 file (clawmetry/static/js/app.js). Deliverables 1 and 2 (schema + ingest) are left for the Claude Code mapper PR (#4815) and a focused approvals follow-up.

Drafting a PR now for the wire-up piece — will link here when ready. If you'd rather drive this yourself, drop a comment and I'll back off.

---

Generated by Claude Code

[vivekchand]

Draft PR opened: #4857. Please review when you have a moment.

---

Generated by Claude Code