Conversation editor tasks

Open this repository folder in VS Code and choose Terminal → Run Task. The six Conversation: tasks are thin process adapters to AgentKit #495–#497. They require Node 22.18 or later and an existing AgentKit installation containing conversation.ts, conversation-memory.ts and conversation-host.ts. Set AGENTKIT_HOME before starting VS Code when that installation is outside ~/.agent-kit. No extension, graph database or new memory service is required.

Each task prompts for an existing private archive and an existing private request JSON file, both absolute paths. Prepare these with AgentKit's existing conversation schema. The archive needs its normal INDEX.md with exactly one <!-- notes --> marker. Keep personal requests, source files, archives and derived material in private Notes. Do not put them in this public Workspace checkout.

Task	Existing AgentKit entry point	Request action
Conversation: Fork	tools/conversation.ts	fork
Conversation: Attach	tools/conversation.ts	attach
Conversation: Return	tools/conversation.ts	return
Conversation: Review memory	tools/conversation-memory.ts	propose
Conversation: Store reviewed candidates	tools/conversation-memory.ts	store
Conversation: Host adapter request	tools/conversation-host.ts	Existing prepare, dispatch, record or reconcile action

The wrapper passes paths as literal process arguments without a shell. VS Code's task command echo is disabled. The terminal reports state, revision and a private Result file URI; open that file for the full packet, parent link, review or host receipt. Full results stay under <archive>/.workspace-conversation-results/. The directory inherits the archive's access controls; Unix result files request mode 0600. It is private transport output, not a new source of truth.

Raw child stderr, source bodies, proposal text and packets never print to the task terminal. A failure returns the AgentKit exit status and a private result when structured output exists. A process interruption or unreadable output gives a fixed recovery message: inspect the archive before retrying, because the operation may already have persisted. Existing operation IDs, revisions and replay rules remain AgentKit's responsibility. Reusing a request preserves its operation ID.

Walk through a branch

1. Prepare a fork request with a stable branch/operation ID, parent URI, exact source coverage, byte references and attributed handoff statements. Run Fork and open its private result. Use the portable packet in the intended chat.
2. Prepare an attach request for the same branch with a new operation ID and the returned revision. Select only the later context to attach. Run Attach; existing sources and attribution remain in the archive.
3. Prepare a return request at the current revision with the selected outcome. Run Return and open its result for the parent link and attributed packet.
4. Prepare a propose request with a new review ID and selected statement indices. Run Review memory and inspect its private review before choosing proposals.
5. Prepare a store request naming that review and only the selected proposal IDs. Run Store reviewed candidates. With memory off it returns NotConfigured (exit 2) and preserves the review. No store is created or configured by tasks.
6. With an explicitly configured existing MemoryKit adapter, rerun that store request. Items remain unverified candidates. Acceptance is a separate explicit action in AgentKit's existing memorykit.ts workflow; tasks never accept them.

See the installed AgentKit portable workflow, and its tools' schemas/tests for complete request contracts. The local conversation-task.test.mjs contains a synthetic fork/attach/return/review/store example using no personal content. Use the capability report before selecting optional adapters and existing MemoryKit setup when memory is wanted. Graphify projection remains the separate reviewed selected-context operation.

Native host boundary

An ordinary VS Code task has no native chat-tool transport. Fork, Attach and Return always provide the portable workflow. They do not create a chat, clone history, send a message or navigate a host implicitly.

When running in a host that actually exposes native tools, use Host adapter request with #497's existing capability evidence and explicit authorizations. Its preparation selects the existing native or portable adapter, and its outbox produces the call and receipt. The host workflow executes the named native tool, then records/reconciles the receipt through the same task. This task itself never invokes an invented VS Code command or MCP bridge. Recheck live tools before dispatch; a registration or manifest declaration is insufficient evidence.

Verification

Run node --test setup-llm/scripts/conversation-task.test.mjs from the repository root. CI checks literal arguments, private output and failure redaction on Windows, Linux and macOS. Set AGENTKIT_TEST_HOME to the existing AgentKit checkout to include its real portable entry points. Set MEMORYKIT_CLI to its existing built CLI to include isolated candidate storage. Tests create disposable synthetic data; they do not read the personal store or register/install any adapter.

Existing provisioning and MCP tasks retain their labels and arguments. Conversation tasks do not alter provisioning, host registration, extensions or global settings.

Windows editor smoke — 2026-10-28

VS Code 1.143.1 (x64), Node 26.12.1 and a worktree based on 8ab22f3 were tested with an isolated VS Code profile, extensions disabled, and explicit AgentKit/configuration environment variables. All input lived in a disposable synthetic archive outside this checkout. Using Terminal → Run Task and the actual path prompts produced:

Step	Observed result
Fork synthetic exact source	Forked, revision 1
Attach a distinct later source	Attached, revision 2; both source hashes retained
Return selected outcome	Returned, revision 2; private packet contains parent URI and source coverage
Review selected statement	Proposed; review persisted privately
Store while configuration absent	NotConfigured, exit 2; review retained
Rerun with isolated existing MemoryKit adapter	Stored; only selected item persisted as candidate, unverified

Task terminals displayed state and private result URIs without source bodies. Private results were inspected to confirm parent navigation data, both archived sources and candidate status. No native chat API or memory extension was needed. The unchanged provisioning preview entry point (setup.ps1 -Client Both -WhatIf) also completed with exit 0; all 13 original task definitions remain intact. The real AgentKit/MemoryKit wrapper test passed 2 tests with 0 skipped.

Desktop MCP discovery is a separate boundary: the desktop restarted, but this chat still exposed no MemoryKit tools. That unresolved host check is tracked in Workspace #64; the editor result above uses the existing CLI adapter and makes no desktop routing claim.