feat(sessions): read OpenAI Codex CLI sessions

[lostmartian]

What

agentdiff now reads OpenAI Codex CLI sessions alongside Claude Code and runs the same findings on them: long context, cache rebuilds, stuck retries, repeated failures, re-reads.

agentdiff                     # Claude Code + Codex
agentdiff --agent codex       # Codex only (or: claude, all)
agentdiff --codex-home PATH   # default $CODEX_HOME or ~/.codex

Why

Roadmap P1.4. Covering more than one agent doubles the audience. It's also the defence against any one vendor building this into its own tool: no vendor will analyse a competitor's sessions.

How

The reader is built from the openai/codex source at rust-v0.163.0. sessions/codex.py documents the format notes; the main points:

• Where files are found: sessions/**/rollout-*.jsonl plus archived_sessions/, recursively.
  • Lines must have the wrapped format {timestamp, type, payload} (Codex 0.33+). Older unwrapped files carry no usage and are skipped.
  • .jsonl.zst files are read when compression.zstd (Python 3.14) or zstandard is available.
• Usage:
  • Read from token_usage_record, one per API response and de-duplicated by response_id.
  • Older logs fall back to event_msg/token_count last_token_usage, skipping repeats and rate-limit-only events.
  • input_tokens already includes cached tokens, and output_tokens already includes reasoning tokens. Both are handled.
• Model: taken from the latest turn_context, so it can change mid-session. thread_settings_applied.service_tier == "priority" marks the Fast tier.
• Pricing:
  • OpenAI list prices fetched 2026-10-19 (gpt-6.x, gpt-5.x, codex, o3/o4-mini).
  • New long_context_tokens field: above 272K input, input is billed at 2x and output at 1.5x, per OpenAI's pricing page.
  • Codex logs carry no cost, so the report labels Codex figures as API-equivalent estimates.
• Tools:
  • function_call / custom_tool_call / local_shell_call are paired with their outputs by call_id.
  • Exit codes are parsed from the output text (Process exited with code N / Exit code: N) or taken from paginated item_completed CommandExecution events.
  • Each tool call is attributed to the next usage record, because usage is logged when the response completes.
• Forks and subagents: fork de-duplication moved to sessions/dedupe.py and is shared by both agents. Legacy forks copy their parent's usage lines, and those are counted once. Threads with parent_thread_id nest under their parent.
• Analyzer: recognises Codex shell tools (exec_command, shell, local_shell), the cmd arguments they use, and bash -lc '…' wrapping.
• Report:
  • The header names the agents it read, e.g. "Claude Code + Codex".
  • An Agent column appears when there is more than one agent.
  • The footer notes Codex estimates and any requests on models with no known price.
• Tests: an autouse fixture points CODEX_HOME and CLAUDE_CONFIG_DIR at empty directories, so no test can read a developer's real logs.

Testing

make lint                  -> All checks passed
uv run python -m pytest    -> 475 passed (14 new in tests/test_codex.py)

What the tests cover:

• per-model pricing;
• a model change mid-session;
• the long-context and Fast tiers;
• the token_count fallback, including skipping repeats and null events;
• usage records taking priority over token_count;
• skipping legacy and broken files;
• finding nested and archived rollouts;
• de-duplicating a legacy fork;
• nesting a subagent under its parent;
• exit codes producing a stuck-retry finding;
• paginated CommandExecution events;
• loader labels and filtering;
• CLI text and JSON output;
• a full analysis producing a context-bloat finding.

By hand: my real Claude-only report is unchanged. A combined Claude + Codex demo renders the new header, the Agent column and the estimate note.

⚠️ Not yet tested against a real Codex session. Codex isn't installed on this machine, so every Codex test uses synthetic rollouts built from the source schema. Worth one real run before announcing Codex support:

codex      # do any small task
npx @kerrshift/agentdiff --agent codex --since 1d

Checklist

• Exactly one logical change in this PR
• make lint passes
• uv run python -m pytest is green
• Docs updated if user-facing
• CHANGELOG.md entry added under [Unreleased] (if feat/fix)
• Conventional title: type(scope): summary