Repository navigation
feat(handoff): continue chats across agents and named accounts - #669
Open
Adam-Dalloul wants to merge 161 commits into
Open
Adam-Dalloul wants to merge 161 commits into
Adam-Dalloul wants to merge 161 commits into
Conversation
Adam-Dalloul
force-pushed
the
feat/agent-handoff
branch
2 times, most recently
from
September 6, 2026 22:07
622245a to
194e574
Compare
A conversation can now change agents without starting over. Two paths, decided per pair of agents: Native (Claude to Claude, including an isolated custom Claude slot): the session transcript is copied into the target's config dir and the target session/loads it under the same id, so nothing is summarized. The load is verified before the row moves; a target that opens a fresh session instead is disconnected and the copy removed. Summary (any other pair): the target starts a new session seeded with a briefing built from the conversation (a digest of earlier turns, the last turns in full, the user's note), bounded and reported as truncated when earlier turns had to be dropped. Either way the conversation row keeps its id: rebind_for_handoff moves it to the new agent and session, and a conversation_handoff record keeps the earlier segment reachable. The detail read splices each earlier segment in front of the current session behind a divider tagged _meta["codeg.handoff"], placed by user-turn count for a carried session.
One action from three places: the conversation header menu, the sidebar card's context menu, and the composer's agent control of a persisted conversation. The dialog reuses the agent selector, takes an optional note for the next agent, and states which path the backend will take before the user confirms (native load, briefing, or why it cannot run). On success the open tab is swapped onto the new agent so the same conversation reopens under it; a failure leaves the tab where it was. The seam renders like the compaction divider: a chrome-less rule naming both agents, whether the context moved natively or as a briefing, the note, and the briefing folded behind a toggle. The seeded briefing never shows as a user bubble, live or reloaded.
Adam-Dalloul
force-pushed
the
feat/agent-handoff
branch
from
September 7, 2026 15:29
194e574 to
400571d
Compare
Settings MCP still targets the built-in family row. Extra custom agents whose spec.env sets CLAUDE_CONFIG_DIR, CODEX_HOME, GROK_HOME, GEMINI_CONFIG_DIR, or OPENCODE_CONFIG_DIR now receive the same server on upsert and lose it on remove. Claude extra homes only get enabledPlugins; server defs stay in the shared ~/.claude.json. Auth files are never copied.
Custom agents whose spec.env sets a family isolator now have a Sign in card on the Agents page. That opens the official CLI login (claude/codex/grok/gemini/opencode) with the isolator env set. No tokens are copied. Account 1 stays on the built-in home. Headless and unit-test paths return the command instead of spawning a terminal.
Settings MCP still targets the built-in family row. Extra custom agents whose spec.env sets CLAUDE_CONFIG_DIR, CODEX_HOME, GROK_HOME, GEMINI_CONFIG_DIR, or OPENCODE_CONFIG_DIR now receive the same server on upsert and lose it on remove. Claude extra homes only get enabledPlugins; server defs stay in the shared ~/.claude.json. Auth files are never copied.
Custom agents whose spec.env sets a family isolator now have a Sign in card on the Agents page. That opens the official CLI login (claude/codex/grok/gemini/opencode) with the isolator env set. No tokens are copied. Account 1 stays on the built-in home. Headless and unit-test paths return the command instead of spawning a terminal.
…tches A connection whose task ends on its own (the agent failed to load or exited, or the task panicked) now parks its child in the manager's draining list, as every explicit teardown does, so a missing map entry is never taken as proof that the process stopped. The deferred cleanup is spawned on the connection's runtime handle, because the guard drops outside any runtime context. Account switches wait for every writer of the session before copying it and before rolling a target back, and a child parked after its task finished leaves the list as soon as it is reaped. A switch's staging copies are removed once the switch is committed, a rollback completes, or staging itself fails. They are kept only when a rollback fails.
Launching a named account re-applies the shared MCP servers to config files that the account's own CLIs rewrite too. The config writers now leave a file alone when its content would not change, and the config staleness check no longer syncs account folders at all. Pi accounts mirror pi's own server list instead of the empty ACP forwarding view, which removed their servers again on every launch. A missing account folder no longer fails the import scan or every MCP change; that account is skipped until its folder is restored.
…ount A conversation switched to another account keeps its session in that account's folder. Work-task retries, returns and merges, their fresh-session fallback, and chat-channel resumes now launch with the conversation's saved account instead of the Default one, which the prompt gate then refused.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Continue an existing chat with another agent or account while keeping its tab, unsent draft, title, pin, tags and visible history. The account picker, composer and conversation menus share one explicit handoff dialog. Account setup is reachable from the chat, including supported official sign-in and API-key methods.
Built-in Claude-to-Claude can use verified native resume and the shared account-transfer safeguards: source writers stop, session files and session-local tasks move, and the requested model is verified after initialization. Failed load/commit preserves the source. Live context resets are refused before copying or disconnecting the source, including when the database has already rebound to a private session ID.
Other supported local agents start a fresh native session with the complete saved Codeg conversation available as context, including tool records, exit codes and embedded images. Long histories use a read-only, conversation-scoped MCP reader; agents without wire MCP receive the local transcript. Any handoff involving a custom agent uses this saved-context path. A custom ACP session ID does not establish that a native file contains all its history. OpenClaw gateway handoff is unsupported. A different provider's private native state and context window do not transfer.
The account binding and immutable history segment commit together. Context survives restart and backup, and a failed first prompt retains it for retry. Historical @mentions do not become new routing frames; fresh mentions keep their selected native account. Older native-session copies remain marked as imported, while independent later native work remains importable. Retained custom-Claude source IDs and their continuations also stay covered by the upstream duplicate-import rule after switching agents; deleted chats and legacy entries without saved history do not retain those IDs. Per-segment usage/model information is retained. Existing frozen usage facts from earlier previews remain readable without adding overlapping transcript totals; this branch no longer writes speculative native-recovery snapshots.
Includes #919 and #472, retaining current upstream workspace tabs, activity rendering, ACP 3.4, context-MCP stdin/stdout and CLI ownership. The filter fix #1029 and recording fix #1048 are now merged upstream and included here with the maintainer's fixes for model labels, relayed-message usage and duplicate imports. The account import paths retain those protections for Default and named accounts. Native usage recovery remains a separate policy discussion in #1047; API-list estimates shipped with #503. All previous migration IDs and published branch history remain intact.
Validation after the usage merges: production server/MCP/CLI compilation and build, all-target Clippy, the production frontend export and 1,144 UI regressions passed. Focused backend tests cover account-scoped imports, Codex environment handling, handoff, token usage, reported counters and ACP transcript projection. Real HTTP/MCP/browser fixtures cover native task restoration, unavailable-model refusal, rollback, paged history and images, account/agent switching, drafts/titles, @ routing, retry, restart, import deduplication and reset refusal, including phone and RTL layouts. A custom-agent recording fixture also verifies concrete model labels, no duplicate import through a named account, reopening, and schema-triggered refresh. The old database and frozen usage fixtures remain intact. These use isolated synthetic data and CLI fixtures, not live sign-in/generation for every provider or OS. A final real HTTP regression reproduced and fixed a custom source becoming importable again after handoff. The fix was rebuilt and rechecked with imports, handoff tests and Clippy; the explicit duplicate import is now skipped and usage totals stay unchanged. The frontend is unchanged from the browser run. CI runs again on the refreshed commits.
Maintainer edits are enabled. See
docs/agent-accounts.mdfor the flow and support boundaries. The fixtures do not establish live provider sign-in/generation for every agent or OS.