local-conversation-history

Local Conversation History — router

This skill decides which skill runs. It does not parse history itself, does not own commands for a single provider, and never re-implements what an executor already does. If you find yourself explaining flags for one provider, you are in the wrong skill — hand off and stop.

Route by platform × action

Establish two things before routing: which platform the conversation lived on, and whether the user wants evidence (what was said/done) or resumption (take the work forward).

Platform	Read evidence	Continue the work
Claude Code	daymade-claude-code:read-claude-code-history	daymade-claude-code:continue-claude-code-work
OpenAI Codex	daymade-claude-code:read-codex-history	daymade-claude-code:continue-codex-work
Kimi CLI	read-claude-code-history, with the Kimi scope named — see Provider scope	no continuation skill exists

Resumption always follows a read. The continuation skills require a verified read receipt; routing straight to them without one is a defect, not a shortcut.

When the platform is not stated, separate an exact identity from a memory of content. With a complete session ID, first use the Codex exact-ID locator, then the Claude Code exact-session lookup if Codex does not verify it. The verified original record decides the provider and continuation skill; an ID prefix does not. With only a title, date, quote, or remembered topic, use the inventory or content-search route below and verify the original messages before assigning a provider.

Provider scope — the job only this entry point routes

Name the requested providers before handing off. Inventory and content recall use different executors; do not assume a provider-specific default covers a cross-provider request.

Cross-provider need	Route to	Name this scope
Inventory — "what have I been working on", "list my recent chats", session titles/dates/IDs	Indexed metadata where available, then exact-session readers	Do not use list_local_history.py --source all: its date and output limits are applied after Claude session bodies are read. Report a gap if no indexed inventory exists.
Content search — "did we ever discuss X", find a quote, file, or tool result	read-claude-code-history, its indexed recall, then exact-session reader	Select the relevant indexed providers; inspect index coverage and freshness. Open only candidate sessions.
Ranked recall — wording may have drifted	read-claude-code-history, its hybrid recall index	The index states which providers it holds; do not infer coverage for an unindexed provider.

Indexed recall accepts explicit Claude, Codex, and Kimi provider filters.

Use supplied clues before broad discovery. For a pasted quote with a known project, date, or title, use indexed metadata to narrow candidates, then verify their original messages. For Codex, follow Locate a quoted exchange in read-codex-history. Inventory alone never establishes a content match; if its candidates miss or its scope is incomplete, report the gap and refine the indexed query. Keep the current Session excluded. A request for only an ID stops at verified message evidence; it does not require reconstructing every unrelated conversation.

Search history through the index or an exact known session. A raw sweep of every conversation is prohibited, including when an output limit or scan timeout is set. The current raw-search date flags filter after files are read, so they do not make that command bounded. If the index is missing, stale, or lacks a provider, report that coverage gap; do not fall back to a corpus scan. Ranked top-K results cannot support an absence claim.

Kimi CLI has no other entry anywhere — no dedicated skill exists for either axis, so both routes above land in read-claude-code-history, which documents its own Kimi home resolution.

Let the executor own exact-session lookup, indexed provider filters, coverage, and output format. This skill names which providers are in scope.

Intent decides the route — the word "history" does not

The user's requested result	Route
A list of conversations: titles, dates, session IDs	The indexed inventory row under Provider scope; if unavailable, state the gap
The conversation where a topic, quote, file, or tool result appeared — "find that old chat", "did we ever discuss X"	The indexed content search row under Provider scope; use supplied clues to narrow candidates first. Listing titles alone is not evidence the content exists
Their own raw inputs in chronological order, verbatim	The matching reader's verbatim-input path. Preserve duplicates and session boundaries; duplicates are part of the ledger, not noise
Recent requests that may have stalled — "anything we recently forgot to finish?" / "最近提过却没推进的事"	The indexed inventory/content-search routes under Provider scope, then the matching exact-session readers. Verify candidates against current project records; a project registry alone does not cover unregistered requests
Picking work back up from an identified session	The matching continuation skill, after a read

The requested output wins over the background motivation. If someone explains a problem and then asks for a window of their own raw inputs, return that window — the explanation's topic clues do not convert the request into a content search.

For a recent-unfinished-work request, keep the retrieval read-only. Establish the requested time window and relevant providers, inspect index freshness, and open only selected sessions through their owning readers. Pair each candidate's original user request with the latest relevant handoff or completion evidence; check the project's current decision/status source before calling it unfinished. Keep completed work, user-paused work, external blockers, and work awaiting acceptance distinct. Treat reminders, assistant plans, and uncompleted reference bookmarks as leads, not proof of an outstanding commitment. Report the candidate, next action, evidence source/date, and any unresolved coverage gap. Stop after the requested inventory; do not resume old work or migrate tasks without that authorization. The calling agent performs this reconciliation; the readers own retrieval, and no mechanical check establishes real-world completion for them.

Invariants that survive routing

For remembered facts, identify the exact proposition before searching: account count, account users, and an event's execution identity are different questions. Check three evidence lanes in selected sessions: direct user words, assistant retellings as locators, and original external records previously read by tools. An assistant's quote locates the source; it does not replace that source. Use the provider reader's complete tool-evidence route when the target is a prior read.

When the user says “search again”, change a concrete search dimension: source lane, indexed provider/date/project scope, literal versus semantic matching, or the selected session's original records. Do not end the retry with the same ranked prose query and different synonyms alone. State which coverage changed. Keep a found subclaim separate from an unfound remainder; current explicit user facts remain authoritative even when their historical wording is not located. Preserve original evidence values; do not silently redact, replace, or summarize the stored text while exporting a record.

• Coverage. Read the index's provider and source scope plus its freshness frontier. An archive or provider outside that scope is unknown. A missing required source cannot support an absence claim.
• Self-match. The current session records the user's question and this agent's own commands, so it matches almost any query about itself. Exclude the current session ID before treating a hit as historical evidence.
• Zero results are not absence. Ranked recall can miss different wording and unindexed records. State coverage; do not convert an empty result into "it never happened".
• Ask whether a word is the topic's name or the way you talk about it. A search for 技术选型 over a corpus where the user actually said 用哪个 and 不要闭门造车 returns almost nothing — measured: 10 hits against 163 real occurrences, because a Chinese phrase only matches on a token boundary. When searching one person's corpus, search the way they speak: imperatives, negations, and concrete scene sentences, not the label a design document would give the topic. A zero hit on a topic name cannot support an absence claim; widen to the phrasing that person would have used first, and say which forms you tried.
• A zero has three causes and only one of them is "no history". The other two are a home that was never found and a scope that excluded everything, and none of the three looks different in an empty table. This bites Kimi hardest: its documented default home is not where every install puts it — a desktop client can bundle the CLI inside its own runtime and keep sessions there — and its sessions belong to their own workspaces, so a default run inside some other repository returns nothing on a store full of conversations. Rule out both before reporting absence: name the home that was actually read, and say what project scope was in effect. The inventory prints a diagnostic line when a home is missing, so quote it when it appears — but a located home that yields zero prints no diagnostic at all, and the search path prints none in either case, so never treat a silent empty result as the reader confirming absence. read-claude-code-history owns how to locate a home and which scope flags the inventory needs.

Do not

• Do not run provider-specific parsing, SQLite, rg, jq, or JSONL pipelines here. Every one of those belongs to an executor that already handles its store's schema, archives, and failure modes.
• Do not copy an executor's flags into this file beyond the three that name provider scope (--source, --codex, --kimi) — those are this skill's own subject. Every other flag changes on the executor's schedule; copying one here makes this file drift silently and then teach the wrong command.
• Do not route to a continuation skill to answer a question about the past. Reading is evidence; continuing changes the world.