Context-light cross-backend handoff: compact session map + recall tools on every backend

[saucam]

Problem

Switching or forking a session onto a warm backend (Claude, pi, Codex, Gemini CLI) seeds the incoming backend's first prompt with a rendered text transcript of the conversation (seedFromHistory → renderHistorySeed). Two costs:

1. Front-loads the context window. The seed is budgeted at ~70% of the target model's window (seedBudgetChars = window × 0.7 × 3.5, src/daemon/providers/context-windows.ts). For a long session on Codex (256k window) that is up to ~180k tokens consumed before the model does any work.
2. Lossy at the tail, with no recovery. When the session exceeds the seed budget, the oldest turns are dropped (newest-first truncation + a user notice). The dropped turns are still in the verbatim memory store, but the recall tools (recall / recall_file / timeline) are wired only into the Claude provider (src/daemon/providers/claude/index.ts ~L326). A switched-in Codex/pi/OpenAI/Gemini cannot pull them back.

Net: the cross-backend switch is the lowest-fidelity path in the meta-harness. The per-turn path (canonical history → each backend's native messages + tool calls, never summarized) is unaffected and stays lossless — this is only about the handoff.

Proposed design

Flip the handoff from "push the whole transcript" to "seed a compact map, pull detail on demand", uniformly across all backends.

1. Expose the recall tools on every backend, not just Claude.
recall(query), recall_file(path), timeline(limit), plus a new exact-retrieval get_turn(id) / get_turns(ids). Exposure per backend:

• Codex — native MCP via the app-server (it already surfaces mcpToolCall).
• pi — the extension/tool bridge (same path the approval bridge already uses).
• OpenAI / Gemini — as function/tool declarations added to each runTurn() request.
• Claude — already done (in-process MCP); keep as the reference implementation.

All read the same workspace-scoped episode store, so recall returns the real bytes regardless of which backend asks.

2. Seed a compact session map, not the full transcript.
On switch/fork, the incoming backend's first prompt carries a short anchor:

• the last user turn, verbatim;
• a one-line-per-turn index of the session (turn id, role, provider/model, a short label, files touched);
• the workspace memory index.

The model starts with a near-empty context and a directory of what happened, and calls recall / get_turn to pull any message in full when it needs it.

3. Retrieval by id.
The turn index gives each turn a stable id; get_turn(id) returns the exact canonical turn (content + structured tool calls + results). "Jump to any message in history" becomes a single tool call, and it behaves as if that turn were native to this model's session.

Result: cross-backend switching is context-light (near-empty start) and lossless on demand (any turn retrievable) — the ~180k-token front-load and the tail-truncation both disappear.

Prior art in the codebase (reuse, don't reinvent)

• MemoryEngine + the verbatim episode store already hold the full history; the Claude memory MCP already exposes recall / recall_file / timeline.
• The rotation seed strategy B: task-anchor (Layer D, /rotate + auto-rotation) already does exactly this shape within Claude: capture the last user turn, inject a compact anchor, let the model recall prior detail. This issue generalizes that from "Claude rotation" to "any cross-backend switch on any backend."
• The workspace memory index (Layer C) is already the compact per-turn map to reuse for the anchor.

Conceptually: a cross-backend switch should behave like a /rotate or a compaction — the new backend inherits a compact anchor plus the ability to recall everything, not a wall of transcript.

Scope / acceptance

• recall / recall_file / timeline + a new get_turn(id) available on Codex, pi, OpenAI, Gemini, Gemini CLI (Claude already has them).
• Turn index carries stable ids; get_turn returns the exact canonical turn (content + structured tool calls + results).
• Switch/fork seeds the compact map (last user turn + turn index + workspace index) instead of the full renderHistorySeed transcript.
• Switch a long (>seed-budget) session onto any backend: it starts context-light and can retrieve any prior turn, including turns that a transcript seed would have dropped.
• Keep the renderHistorySeed transcript path as the fallback when memory is disabled (CODEOID_MEMORY=0), since there is no episode store to recall from.

Why now

This is the highest-leverage follow-up to the v0.3.0 multi-provider release: it closes the one real fidelity gap in the cross-backend story and makes the "any harness, full history, never summarized" claim hold on every backend, not just Claude.