coding-agent-sessions skill

Coding Agent Sessions

Find local coding-agent sessions across agent products before answering from memory. Prefer the bundled finder for broad cross-platform search, then read the selected session or raw file when you need exact evidence.

PHASE 0 - PLATFORM ROUTER

1. IF the user names a platform, load its reference first.

1. Run the broad finder first unless the user gave an exact file path. For fuzzy recall, expand the query first.

When the user remembers a task vaguely ("that OpenCode bug", "the dashboard PR", "when did we fix X"), derive 3-6 short query lanes before searching: product/tool aliases, repo/package names, exact error text, issue/PR/session IDs, English/Korean phrasing, and likely verbs such as fix, review, plan, deploy, or merge. Run the lanes together with repeated --query so match_reasons shows which wording found the hit.

python3 scripts/find-agent-sessions.py list --limit 20
python3 scripts/find-agent-sessions.py find "commit" --from 7d --platform senpi --platform opencode
python3 scripts/find-agent-sessions.py find "proxy" --platform openclaw --platform droid --platform amp
python3 scripts/find-agent-sessions.py find "refactor" --platform oh-my-pi --platform gajae-code
python3 scripts/find-agent-sessions.py find --query "deploy" --query "token usage" --workers 64
python3 scripts/find-agent-sessions.py find --query "opencode bug" --query "fix opencode" --query "OpenCode parent session" --include-subagents --workers 64
python3 scripts/find-agent-sessions.py read <session-id>

Use python instead of python3 on systems where that is the available executable.

1. Use explorer-style parallel lanes when one query batch is not enough.

If scope is broad (multi-month, many repos/platforms, or a vague "what happened with X"), split independent searches by names/errors, repos/cwds, platforms/models, and time windows. Use available subagent/delegation tools for these lanes when they exist; otherwise run the finder calls in parallel. Merge candidates by id/path, then read the most likely sessions.

1. Read details from search results before ad hoc digging. Search results include detail_hint; run that read --platform command to see the first user prompt, last user prompt, events, and child sessions together.

1. Verify by opening raw transcripts for claims. The finder normalizes formats; the raw path remains the source of truth.

Output Contract

The finder prints JSON for stdout and jq. Every result includes:

Each matchreasons entry includes query, platform, field, and snippet, so you can tell which platform matched and what content caused the hit without opening every transcript. field is usermessage when the query matched a user prompt that the firstusermessage/lastusermessage previews do not cover, such as a prompt in the middle of the session.

Filters

list and search share these filters:

When --platform is omitted, the finder searches every registered platform in parallel. Each optional platform first probes fixed, known transcript roots and returns immediately when the product has no local store, so broad default searches stay cheap. Use repeatable flags such as --platform openclaw --platform droid only when narrowing. Comma-separated platform values are intentionally unsupported.

For OpenCode, the finder uses opencode db path plus direct SQLite queries first, then opencode session list --format json as a fallback. It avoids heavy messages/ or parts/ scans during normal list/search, and only falls back to file joins when the OpenCode DB/CLI is unavailable or explicit --root values request a nonstandard store.

Usage-only sources such as Copilot OTEL, Mux, Antigravity tokscale cache rows, Synthetic provider retagging, and Cursor IDE usage CSV are excluded from default transcript search because they do not reconstruct user prompts.

Subagent / Child Sessions

list and find/search return main sessions only by default, each annotated with subagentcount. read (alias: get) always returns a prompts object and a subagents array containing every child session (id, agent label, parentid, prompt preview, raw path), so opening a main session reveals its whole delegation tree. read works too and returns that child's own events.

When the user asks whether some specific work was ever done, search with --include-subagents — delegated work often lives only in child transcripts, not in the main session. Workflow journal.jsonl files are orchestration logs, not sessions.

Codex Notes

For Codex sessions, use the same broad finder. It reads state_*.sqlite, rollout JSONL, and archived rollout files:

python3 scripts/find-agent-sessions.py list --platform codex --from 7d
python3 scripts/find-agent-sessions.py find "deploy" --platform codex
python3 scripts/find-agent-sessions.py read <session-id> --platform codex

Use references/codex.md for Codex storage details.

Troubleshooting

Activation

Use this skill for any request to find, read, or inspect a local coding-agent session, regardless of product name — including memory-recall questions ("what did I work on a few days ago", "did we already migrate X", "when did I fix Y"). If the user only says "that session where we did X", do not rely on one literal query: expand to multiple discriminative terms first, add --include-subagents when the work may have been delegated, then narrow by --platform, cwd, model, and time (--from 7d for "a few days ago"). Prefer the detail_hint from the chosen search result for the next read step. If the first query batch is still ambiguous, use explorer-style lanes by keyword, repo/cwd, platform/model, and time window before summarizing.