Local testing: auto-save per-session chat logs under .claude/tmb/logs/

[ZaxShen]

Pain

During live dogfood, all diagnostic signal lives in the interactive chat scrollback. When a session does something unexpected (see #47 for an example: 21.7k tokens, 0 MCP tool calls, 18s latency), the only way to debug is to screenshot or copy-paste the output into an issue. No structured capture, no grep-able history, nothing to attach to a bug report.

Proposal

Per-session log files, auto-saved under .claude/tmb/logs/<YYYY-MM-DD-HHMMSS>-<session-id>.log, capturing:

• Timestamped user prompts.
• Agent spawn names (tmb:bro, tmb:architect, etc.).
• Per-spawn token counts + latency + tool use count (the (0 tool uses · 21.7k tokens · 18s) line).
• Every MCP tool call and its JSON response.
• Error responses / permission denials / forbidden-role rejections.

Not in scope

• Full chat transcript capture — Claude Code already owns the transcript; we shouldn't duplicate.
• User-facing notifications about what's logged — the file is a developer tool.
• Upload/export — everything stays local under .claude/tmb/, gitignored.

Options

A. MCP-only capture via audit table (cheapest)

Every MCP call already passes through the server. Have the server write a per-session JSONL line to <db-dir>/logs/<timestamp>.log in addition to the audit table. ledger_log calls are already there. Just add a file writer.

• Pro: zero CC-level integration; server-owned file; gitignored path already.
• Con: only captures MCP calls, not agent spawns or tool-use counts from CC.

B. CC SessionStart / SessionEnd hooks + tee

A SessionStart hook opens the log file and a SessionEnd hook flushes. Agent spawns and tool counts visible from CC's event stream. Requires figuring out whether CC exposes per-spawn stats to hooks.

• Pro: full picture — agent spawns, token counts, tool counts all logged.
• Con: depends on CC hook event surface; may need CC-side feature.

C. Wrapper script (tmb-test.sh)

Users run ./tmb-test.sh instead of claude --plugin-dir .... The script tees stderr/stdout to a file. No plugin change.

• Pro: trivial; works today.
• Con: opt-in per session; easy to forget.

Recommendation

Ship A + C together:

• A as a plugin-native feature: MCP server writes a structured JSONL log per session. Cheap, always on, gitignored.
• C as a documented helper in docs/local-testing.md for capturing chat scrollback alongside the MCP log.
• B deferred until CC exposes richer hook events — file a sub-issue referencing this when the hook API is clearer.

Acceptance criteria

• MCP server writes JSONL to <db-dir>/logs/<YYYY-MM-DD>-<session-id>.log, one line per tool call with: timestamp, tool name, agent arg, args (sans sensitive keys), response status, response body (truncated at 4KB).
• Log rotation: keep last 50 session files; delete older.
• docs/local-testing.md adds a "Debug logs" section showing where to find them and an optional claude ... 2>&1 | tee pattern.
• .gitignore excludes .claude/tmb/logs/ (already covered by .claude/ blanket, but document explicitly).
• Integration test: a session that calls identity_set produces exactly one matching log entry.

Related

• Surfaced during PR 📝 dogfood setup: SCENARIOS.md + bun workspaces + doc fixes (closes #40, #42) #41 dogfood.
• Paired with bro first-response latency — 18s with 21.7k token context, 0 MCP tool calls #47 — we're only debugging the 0-tool-uses bug by staring at chat output. A log file would have shown the tool-call attempts vs. rejections in seconds.

[ZaxShen]

Issue tracking moved to Linear. This issue continues as TRU-8: https://linear.app/trustmybot/issue/TRU-8

GitHub issues are no longer the active tracker for this project — see the Linear workspace for current state. Closed-as-migrated, not closed-as-resolved.

[ZaxShen]

Disregard the prior 'closed-as-migrated' comment. GitHub remains the public issue tracker for this project — file new issues, PRs, and discussion here as normal. TMB team mirrors this issue internally to Linear (TRU-8) for our own workflow, but that mirror is auto-synced from GitHub — contributors don't need a Linear account.

[ZaxShen]

Closing — focusing on v0.8.x. Planning lives on GitLab milestones (v0.8.0 / v0.8.1). Will re-mirror active items to GH once tri-source schema sync (#48) lands.