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).

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 — a bare session ID, "pick up where we left off" — do not guess it. Identify it first: try the Claude Code exact-session lookup in read-claude-code-history, then the Codex rollout locator in read-codex-history. Only a lookup that returns a verified identity decides which continuation skill runs; a plausible-looking ID prefix does not.

Provider scope — the job only this entry point routes

Each executor defaults to its own provider, so a request that spans providers never widens by itself. Naming the scope is this skill's whole job. It has two axes, and they use different flags — conflating them is the failure this section exists to prevent.

Both readers ship the same inventory command and its --source already defaults to all — but each reader's own task table pins it to that reader's provider (--source claude, --source codex), so the default never fires on its own. Search is the mirror image: it is Claude-only unless the other two stores are added explicitly.

Order the last two rows rather than picking one. A cross-provider content search is the expensive shape: it reads every event of every store, so the cost scales with the whole corpus rather than with the question. When an index exists and covers the providers in scope, recall answers in about a second and returns leads — sessions, dates, projects — that turn the exhaustive scan into a bounded one. Run it first, then scan only what the index does not hold. Skip straight to search when the request needs an exhaustive guarantee, because ranked recall returns top-K candidates and can never support an absence claim.

The index is optional. On a machine that never built one, recall exits non-zero saying the index does not exist — that is a routing signal, not a failure to report: fall back to the search row and say the sweep ran unindexed.

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 every flag beyond provider scope: --all-projects, --recursive, date bounds, --include-archived, --include-subagents, --include-automated, output format, and every detail of how each store is parsed. This skill names which providers are in scope and nothing else.

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

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.

Invariants that survive routing

• Completeness. A Claude inventory's source set is indivisible: the auto-discovered active homes (~/.claude, profile homes, the current CLAUDE_CONFIG_DIR) plus every archive registered in ~/.claude/history-sources.json. Never call a conversation absent unless the output shows the registered archives were covered. An unavailable required archive is a configuration error, not permission to return a partial answer. --claude-home is a diagnostic override and can never back a completeness 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 and a bounded search both return nothing for wording that exists under different words. Widen, or say what was searched — do not convert an empty result into "it never happened".
• 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.