Local Conversation History

List project-scoped local histories without reconstructing ad hoc rg, stat, jq, or SQLite pipelines. The bundled script performs provider and archive discovery, schema introspection, filtering, de-duplication, title extraction, internal-time sorting, positive-only Codex writer-lock observation, and rendering in one process.

Decide the job before calling a tool

Treat the user's intent as the routing key; the word “history” alone does not mean inventory.

A topic clue wins over inventory wording. For example, “find our historical conversation records about DINO” is full-content search even though it says “conversation records.”

When routing a content search, preserve the unknown parts of the user's scope:

• If the provider is unknown or the user says “our history,” require the finder to cover Claude plus Codex and Kimi CLI; a Claude-only result cannot support an absence claim.
• If the project is unknown, require the finder to search all projects instead of guessing the current workspace.
• Exclude the current session before treating a fresh hit as historical evidence; the user's query and the agent's search command are recorded in the current transcript and otherwise self-match.

Let the finder own its exact commands, query widening, source diagnostics, and result interpretation. This skill owns only the inventory-vs-search decision.

Completeness invariant

For a normal Claude Code inventory, the source set is indivisible:

1. auto-discovered active homes (~/.claude, profile homes, and the current CLAUDE_CONFIG_DIR), and
2. every archive registered in ~/.claude/history-sources.json.

Do not claim that a Claude conversation is absent unless the output shows that the registered archives were searched. A required archive that is unavailable is a hard configuration error, not permission to return an incomplete result. Explicit --claude-home is a diagnostic scope override and intentionally bypasses the registry; never use it for a completeness claim.

Route the request

• List/recent/show/browse local conversations: run the bundled script once.
• Triage Codex sessions that may still be in use: use the automatic writer-lock file held marker. The command probes every admitted Codex row and appends any positive hit outside the recent-row limit. A hit proves only that a process held Codex's canonical advisory lock during the snapshot; it does not identify that process or prove the session is running. An unmarked row does not prove that a session stopped.
• Understand what a marked Codex session is doing: pass its exact session ID to daymade-claude-code:continue-codex-work. A title, recent timestamp, or held writer-lock marker identifies a thread lock, not its holder, current task, or progress.
• Restrict to one provider: pass --source claude, --source codex, or --source kimi (Kimi CLI, a.k.a. kimi-code).
• Point at a non-default Kimi CLI home: pass --kimi-home <dir>; resolution order is --kimi-home > KIMI_HOME > ~/.kimi-code.
• Include child workspaces under a directory: pass --recursive.
• List every workspace: pass --all-projects; omit --cwd.
• Include archived Codex threads or archived Kimi CLI sessions: pass --include-archived.
• Restrict by conversation date: pass --from-date and/or --to-date.
• Include internal agents or obvious smoke prompts only when explicitly asked: pass --include-subagents or --include-automated.
• Search inside full transcripts, recover deleted files, or analyze tool calls: use the daymade-claude-code:claude-code-history-files-finder skill instead.
• Reconstruct and continue a Claude Code session with daymade-claude-code:continue-claude-work; use daymade-claude-code:continue-codex-work for a Codex thread.

Run exactly one inventory command

Resolve scripts/list_local_history.py relative to this SKILL.md. Do not search the machine for the script and do not recreate its logic inline.

On macOS or Linux, execute the script directly when its executable bit is available; otherwise use Python 3. On Windows, use py or python:

<skill-dir>/scripts/list_local_history.py --cwd <workspace> --limit 10 --language en
py <skill-dir>/scripts/list_local_history.py --cwd <workspace> --limit 10 --language en

Choose --language zh when the user is speaking Chinese. If the user supplied no path, pass the shell's current working directory explicitly. Use forward slashes in Windows command examples, while allowing the actual --cwd value to use the platform's native path form.

Expected output is already presentation-ready Markdown:

# Local conversation history
Scope: `<workspace>`

## Codex — 3 conversations
Runtime: `writer-lock file held` proves lock contention, not holder identity; an unmarked row is not evidence that a session stopped.
| Updated | Title | Session ID | Flags |
|---|---|---|---|
| 2026-01-15 10:30 +00:00 | Review authentication flow | `019...` | writer-lock file held |

Return that output directly, with at most one short observation. Do not run follow-up find, rg, stat, or database calls merely to restate the result. When the user asks what a marked Codex thread is doing, the sanctioned follow-up is daymade-claude-code:continue-codex-work for that exact ID, not a process-name or cwd guess.

Preserve the evidence boundary

Treat the command as an inventory, not a transcript export:

• Keep the script read-only. It never resumes, renames, archives, deletes, or repairs a conversation.
• Report titles only; do not paste raw JSONL or full prompts unless the user asks for a specific session afterward.
• Keep every displayed timestamp's explicit timezone offset.
• For Claude Code, treat the minimum and maximum valid top-level timestamp values across the JSONL as the session range. Never substitute file mtime: copying or migrating an archive changes mtime without changing conversation time.
• For Codex, prefer the state database's internal created/updated fields. If the database is unavailable, compute the rollout range from internal top-level event timestamps plus session_meta.payload.timestamp; never use rollout mtime or database-file mtime as chronology.
• Treat writer-lock file held as evidence only that some process owned the canonical per-thread advisory lock during the snapshot. It does not identify that process or prove an open UI, an executing agent, ongoing tool use, business progress, repository permission, or a project lease. Every in-scope Codex row is probed; positive hits outside --limit are appended. Never invert an absent marker into “inactive”.
• For Kimi CLI, prefer state.json's createdAt/updatedAt (epoch milliseconds). If state.json is missing or lacks them, compute the range from internal wire time fields (plus the metadata record's created_at, also ms); never use file mtime as chronology.
• A date-only filter means the whole local calendar day. A datetime filter must include Z or an explicit UTC offset. Sessions without internal timestamps are excluded with a visible warning while a date filter is active.
• Preserve provider labels and session IDs exactly as printed.
• State warnings from the script instead of silently hiding a missing, unreadable, or unsupported store.
• Do not claim Claude Desktop native chats are included. The Claude source here is Claude Code history; Codex covers local Codex CLI/Desktop thread stores; Kimi CLI covers the local kimi-code session store, not the Kimi web product.

Handle source configuration and failures

The script honors CLAUDE_CONFIG_DIR, CODEX_HOME, and KIMI_HOME. Register durable Claude archives once in ~/.claude/history-sources.json; the default command then searches them on every run. Use --history-sources <file> to test another registry. Use --claude-home <dir>, --codex-home <dir>, or --kimi-home <dir> only when the user explicitly requests an exact single-store diagnostic scope.

If no conversations appear, use the diagnostics already printed by the same command. Read references/storage_and_portability.md when the format, path, or writer-lock observation needs diagnosis; it documents the source registry, inspected stores, internal-time policy, lock semantics, Windows path normalization, and known boundaries.

Maintainer verification

In the source repository, daymade-claude-code/_conversation_core/ is the code SSOT shared by this skill, claude-code-history-files-finder, continue-claude-work, and continue-codex-work. The four skills remain self-contained at install time because sync_core.py copies that package into each scripts/_core/. Never edit a bundled _core copy directly.

After changing shared code, synchronize and verify all four bundles, then run this skill's standard-library regression suite:

uv run python ../sync_core.py sync
uv run python ../sync_core.py check
python -m unittest discover -s tests -p "test_*.py"

The test suite builds isolated Claude and Codex fixtures, including SQLite and raw-JSONL paths, so it never depends on the maintainer's personal conversation content. Development trigger cases live in evals/evals.json.