Harden parser + backend: structured status, turn attribution, Postgres CLI

[kkrlstrm]

Acts on a round of technical review. I verified every claim against the real Codex rollout data in ~/.codex/sessions before changing anything — a few review points turned out to be raw-GitHub-view rendering artifacts, and the fixes for the real ones are grounded in what the rollouts actually contain.

Real bugs fixed

1. Postgres CLI was broken. sessions / inspect / stats hard-coded the SQLite table names (FROM sessions, FROM tool_calls), but the Postgres backend stores codex_* tables and query() only translated placeholders — never the table names. So every read command failed against Postgres (ingest worked, which is why it wasn't obvious). Fixed with a store.table() indirection the CLI uses to resolve the physical name per backend.

I deliberately did not use the reviewer's "create unprefixed PG views" option — an unprefixed sessions view reintroduces the exact name collision the codex_ prefix exists to avoid when sharing a warehouse with cc-logger. The table() method is the lighter version of their "backend methods" option.

2. Tool-call status was shell-centric. Success/failure was scraped from the string Process exited with code N, which non-shell tools don't emit — so on real data apply_patch was 9/9 unknown. The rollouts already carry authoritative outcomes the parser ignored: exec_command_end.exit_code and patch_apply_end.success. Consuming those:

tool	before	after (real data)
apply_patch	9 unknown	8 success, 1 failure
exec_command	27 unknown	7 unknown (rest resolved by exit_code)

The string scrape is now a fallback that never downgrades a structured result.

3. No turn-level attribution. Added turn_id to tool_calls and messages (parser + both schemas + upserts). 289/289 tool calls on real data now carry their turn, enabling "which prompt caused the failed patch / which subagent turn burned the most tokens."

4. launchd was non-portable. The committed plist hard-coded /usr/bin/python3 and a stale WorkingDirectory (kai-gtm-agents/codex-logger, which no longer exists). Replaced with a generated install-launchd command — writes a plist wired to the real interpreter + repo path, plutil-valid — plus a placeholder template in launchd/.

Docs tightened to match the code

• Softened "captures everything"; the event table now states exactly what is and isn't resolved.
• Fixed a docstring implying event_msg agent_message/user_message were parsed — they're streamed duplicates of response_item/message (which already captures all 164 assistant + 76 user messages on real data), so re-ingesting them would double-count.
• Added a Privacy & security section (this stores prompts, tool output, and secrets that scroll through a terminal).
• Noted the CLI now works against Postgres; kept the honest "no live-DB CI yet" caveat.

Verified as non-issues (not changed)

The reviewer read the GitHub raw view and several claims were rendering artifacts, confirmed against source: the Python files are normally formatted (not "one long physical line"), pyproject.toml parses as valid TOML, and the plist was valid XML.

Tests

Added coverage for status-resolved-from-*_end-events, per-turn attribution, and turn_id round-trip. Full suite green; also validated pip install -e ., the codex-logger console script, and install-launchd output (plutil -lint: OK).

🤖 Generated with Claude Code