ai-provider

name	ai-provider
description	Reference for where Claude and Codex provider data is stored on disk — session logs, raw output, thinking/reasoning availability, and format differences between the two providers. Use whenever investigating provider output, debugging missing thinking blocks, reading session history, or asking "where does X get stored for Claude vs Codex?"

AI Provider Storage Reference

Claude

Session logs

~/.claude/projects/<path-based-id>/<session-id>.jsonl

The path-based ID is derived from the working directory (e.g., -Users-bill-Developer-personal-AIDevTools). Each line is a structured JSONL event covering tool calls, text output, thinking, and metrics.

App-level logs

~/Library/Logs/AIDevTools/aidevtools.log

Structured JSON log used by the Mac app and CLI. Filter with jq by label field.

Eval artifacts (raw provider output)

~/Desktop/ai-dev-tools/services/evals/AIDevTools/artifacts/raw/claude/<suite>.<case-id>.stdout

Only written during eval runs, not during chat sessions.

Thinking blocks

Claude thinking blocks are present in the exec stream as plain text events and are fully parseable. The ClaudeStreamFormatter.formatStructured() emits .thinking(text) events which the Mac app renders as purple collapsible blocks.

---

Codex

Native session logs

~/.codex/sessions/YYYY/MM/DD/rollout-<timestamp>-<uuid>.jsonl

Codex writes a session file for every invocation — both from the native TUI (originator: "codex-tui") and from codex exec (originator: "codex_exec"). Each line is a rich event object.

Key event types in the native format:

• session_meta — session ID, model, cwd, CLI version
• turn_context — model, effort, sandbox policy
• event_msg / agent_message — final text responses
• response_item / reasoning — encrypted thinking content (see below)
• event_msg / token_count — usage stats

Eval artifacts (raw exec stream output)

~/Desktop/ai-dev-tools/services/evals/AIDevTools/artifacts/raw/codex/<suite>.<case-id>.stdout

Only written during eval runs. Uses the codex exec --json stream format (different from the native session format).

Chat sessions from the Mac app and CLI

CLIClient always passes stdin: Data() (a pipe with immediate EOF). Codex detects the non-TTY stdin and skips writing both the session rollout file and the session_index.jsonl entry.

To work around this, CodexProvider+AIClient.run() manually appends an entry to session_index.jsonl after each successful run, using the thread_id parsed from the thread.started exec stream event. This makes sessions appear in the history picker.

What works: history listing, session resuming (thread_id is returned as AIClientResult.sessionId).
What doesn't: loading full message content from a past session — the rollout file in ~/.codex/sessions/ is never written, so loadMessages returns empty.

Thinking/Reasoning blocks

Codex reasoning is encrypted and not available in the codex exec --json stream:

• In native session files: appears as {"type":"response_item","payload":{"type":"reasoning","encrypted_content":"gAAAAA..."}} — the content is opaque ciphertext, cannot be decrypted or displayed.
• In codex exec --json stream output: reasoning events are not emitted at all — only item.completed and turn.completed appear.

Conclusion: Codex thinking blocks cannot be displayed. Do not attempt to infer thinking from agent_message content.

---

Format comparison: exec stream vs native session

The codex exec --json stream (what CodexStreamFormatter parses) uses a different format than the native session JSONL:

Event	Exec stream format	Native session format
Text response	{"type":"item.completed","item":{"type":"agent_message","text":"..."}}	{"type":"event_msg","payload":{"type":"agent_message","message":"..."}}
Tool execution	{"type":"item.completed","item":{"type":"command_execution",...}}	Similar but in response_item
Thinking	Not present	{"type":"response_item","payload":{"type":"reasoning","encrypted_content":"..."}}
Usage	{"type":"turn.completed","usage":{"input_tokens":N,"output_tokens":N}}	{"type":"event_msg","payload":{"type":"token_count","info":{...}}}

agent_message text format

For plain chat, text is plain string: "Hello."

For structured/schema-driven output (evals), text is JSON-wrapped: {"result":"AI DEV TOOLS JOKE: ..."}

CodexStreamFormatter must handle both. Extract result from the JSON wrapper when present; fall back to raw text otherwise.

---

Quick reference: finding a session

# List today's Codex sessions
ls ~/.codex/sessions/$(date +%Y/%m/%d)/

# Read a session's user messages
grep -o '"message":"[^"]*"' ~/.codex/sessions/YYYY/MM/DD/<file>.jsonl | head -10

# Check if a session came from the Mac app vs native TUI
grep '"originator"' ~/.codex/sessions/YYYY/MM/DD/<file>.jsonl | head -1
# "codex_exec" = Mac app / CLI invocation
# "codex-tui"  = native Codex terminal UI