BongSuCHOI Coding Agent Sessions
Use the bundled finder before answering session-history questions from memory. When a platform is named, read its storage reference only if location or linkage details are needed.
Platform references
| Platform | Reference |
|---|---|
| Codex | references/codex.md |
| Claude | references/claude.md |
| OpenCode | references/opencode.md |
| Senpi | references/senpi.md |
| Other or unknown | references/all-platforms.md |
Search and read
Unless the user supplied an exact path, begin with the finder:
uv run --isolated python "<skill-root>/scripts/find-agent-sessions.py" list --limit 20
uv run --isolated python "<skill-root>/scripts/find-agent-sessions.py" find --query "deploy" --query "token usage" --from 7d
uv run --isolated python "<skill-root>/scripts/find-agent-sessions.py" read <session-id>
Resolve <skill-root> to this skill's actual absolute path. For vague recall, use a few discriminative aliases, repository names, exact errors, IDs or Korean/English variants with repeated --query. Narrow with --platform, --cwd, --model, --from and --to; add --include-subagents when delegated work may contain the only evidence. Run the result's absolute-path detail_hint or read command before ad hoc filesystem digging.
For Codex comparisons, narrow with --reasoning-effort <level> when effort is
material. List and search hide internal host sessions such as
codex-auto-review by default and report excluded_internal_count; add
--include-internal only when approval, retry, or host-review traces are part
of the question. Exact-ID read remains available for internal sessions.
--model and --reasoning-effort match any recorded value in the session's
ordered history, not only its final setting.
list and find hide child sessions by default but report subagent_count; read shows the child tree. Codex results label root, subagent, and internal session kinds and expose observed model and reasoning-effort histories. These are recorded settings and stored reasoning summaries, not hidden token-by-token chain of thought. The finder normalizes formats, while each result's raw path remains the source of truth for exact claims.
The finder uses only the Python standard library. Run it with uv run --isolated. If the default uv cache fails specifically with a sandbox permission error, retry once with UV_CACHE_DIR=/tmp/codex-uv-cache; do not relabel package resolution, network, script, or target failures as cache failures. If uv is missing, propose the exact installation command and wait for approval. Treat ordinary finder failures as execution errors, not missing dependencies.