Conversation capabilities

Run this opt-in report before choosing integrations for the portable conversation workflow in AgentKit #495, reviewed extraction in #496, and host outbox in #497. It inventories existing capabilities. Graph projection belongs to Workspace #57.

Run the report

PowerShell 7.4 or later is required for this optional report. From the repository root:

./setup-llm/scripts/Report-ConversationCapabilities.ps1 -AsJson
./setup-llm/scripts/doctor-workstation.ps1 -Client Codex -ConversationCapabilities -AsJson

The standalone report exits 0 even when optional integrations are absent or unknown; invocation failure exits 2 with a fixed error code. The doctor still exits 1 for unhealthy required MCP profiles and 0 otherwise. Without the switch its JSON remains the existing six-row-per-client array and it runs no new probes. With the switch JSON is an object containing profiles and conversationCapabilities.

Additional standalone options are -GraphifyPackagePath, -MemoryConfigPath, -Namespace, -ExtensionsPath and -EvidencePath. The doctor accepts -GraphifyPackagePath and -CapabilityEvidencePath; use the standalone report to select other paths/scopes. Reports are printed, never automatically persisted.

Status	Meaning
`supported`	The stated probe or declaration succeeded; inspect its evidence and scope.
`unavailable`	A command/directory is absent, or an operator verified a host tool is absent.
`unconfigured`	A registration, configuration, namespace selection or extension declaration is absent.
`unknown`	Failed/unrecognized probe, missing evidence, or runtime access not verified.

supported does not imply a successful graph build, activated extension, or authorized host action. Each entry includes a fixed evidence code, safe details and a next step. Only graphify --version and codex mcp list --json execute, with ten-second timeouts per probe. Configuration commands, extension code and graph source are never executed. Raw probe output, configuration values, namespace names, conversations and store contents are excluded from the report. No setup, downloads, indexing, store opening or settings changes occur. Review any report before sharing it; tool IDs and versions are operational metadata. Personal sources and derived personal material remain in private Notes.

Graphify and the live context graph

Use uv tool dir to locate the existing graphifyy environment, then identify its graphify Python package (for example Lib/site-packages/graphify on Windows). Pass that directory explicitly as -GraphifyPackagePath. The report does not install Graphify or assume a package path from an installer. It records the CLI version and SHA-256 hashes of supplied detect.py and validate.py when they declare:

.md and .txt document inputs;
extraction edges with source, target, relation, confidence, source_file;
extra_excludes and .gitignore discovery mechanisms.

These are source declarations, not an end-to-end conversation adapter test. The caller must verify that the supplied directory belongs to the CLI on PATH. An unrecognized source/version remains unknown. In Graphify 0.9.91, tracked files bypass gitignore discovery filtering: .gitignore alone is insufficient to protect private material. Validate explicit exclusions and corpus boundaries before any future indexing operation.

No live context-graph tool is assumed from Graphify installation or historical AgentKit design state. Default context-graph.live is unknown. To record a live owner manually, add contextGraph to an evidence document with status, live: true, owner, and contractPath pointing to the current owner's contract. Supported evidence requires a readable hashed contract outside a design/ directory. This is operator evidence; the script cannot authenticate ownership or test that API. The report retains the hash and an owner-identified boolean, not the owner/path. Locate and verify this contract before #57; do not resurrect the retired AgentKit design graph or create a database.

MemoryKit, memory MCP and VS Code

MemoryKit configuration follows AgentKit's AGENTKIT_MEMORYKIT_CONFIG override, otherwise HOME / USERPROFILE plus .agent-kit-state/memorykit.json. The report checks only the configured CLI's presence, memory.db presence, capture/context flags and whether a namespace was selected. It never executes that configured CLI. Existing MemoryKit status and namespace get open the service, which can initialize history or SQLite sidecars; they are deliberately excluded from this read-only report. Thus even a selected namespace and present store remain unknown for actual access. Verify access separately through the existing reviewed AgentKit memory workflow before extraction. No store is enabled and no candidate is captured automatically.

Codex's memory MCP registration is a separate capability. An enabled server named memory yields registered-runtime-not-probed, still unknown for functional access. An empty list is unconfigured; CLI absence is unavailable; failure or unfamiliar JSON is unknown. Other names are not guessed. Claude Code registrations remain part of the baseline doctor; this report does not claim a Claude memory runtime probe.

VS Code inspection reads extension manifests in ~/.vscode/extensions or the supplied -ExtensionsPath. It reports exact publisher/name IDs, versions, hashes and contributed command IDs whose ID or title mentions memory, plus extensions named for memory. Other extension behavior may remain undiscoverable. Malformed manifests do not stop remaining inventory. Manifest declarations stay unknown for functional use: they do not prove activation, the active profile, remote extensions or command execution. For manual verification, open the intended VS Code profile, inspect installed extensions and each extension's contributed commands/documentation, then test an actual documented command within the intended scope. Never invent a command ID based on a feature name.

Manual host evidence

A shell cannot discover this chat's host tools. Inspect the active session contracts and provide an operator evidence file instead. It expires after seven days and rejects duplicate tools, unknown tool names, invalid states and future timestamps beyond five minutes. Example (replace the date/version with observed values):

{
  "schemaVersion": 1,
  "observedAt": "2026-10-27T14:50:00Z",
  "host": {
    "name": "codex",
    "version": "26.1021.3705.0",
    "tools": [
      { "name": "create_thread", "status": "supported", "verification": "contract" },
      { "name": "fork_thread", "status": "supported", "verification": "smoke" },
      { "name": "navigate_to_codex_page", "status": "supported", "verification": "smoke" },
      { "name": "read_thread", "status": "supported", "verification": "contract" }
    ]
  }
}

Allowed tool states are supported, unavailable, unknown; verification is contract or smoke. Omitted tools remain unknown. contract records exposure only; smoke records an authorized test. Evidence does not authorize creating or forking a chat. Missing access remains unknown rather than an assumed implementation.

The agentKitHost field projects supported tool names into #497's existing {host, tools, evidence} contract, with exact mcp__codex_app__ names. Recheck active session exposure before using it. It creates no host operation, memory write or graph.

Adapter selection

Operation	Existing adapter and evidence needed
Fork / attach / return	#495 portable package; available when every optional integration is absent.
Selected handoff	#497 `create_thread`, attributed selected prompt, explicit new-chat authorization.
Full-history fork	#497 `fork_thread`, explicit consent to the complete inherited history. No selective native clone.
Native return	#497 navigation to the saved parent. Portable attach retains the selected outcome; no automatic message or history rewrite.
Memory extraction	#496 reviewed selection using existing MemoryKit; verify store and namespace access separately.
Graph projection	Deferred to #57 after live owner, relationship mapping and private exclusion scope are verified.

Workstation verification, 2026-10-27

The recorded report was produced by the standalone entry point on Windows with the installed Graphify package directory and operator host evidence. The evidence cites AgentKit's authorized synthetic smoke test: create, full-history fork and parent navigation succeeded; test chats were archived. Current tool contracts were also exposed. read_thread records contract exposure only. Host version comes from installed Codex app metadata, not a tool API version.

Observed: Graphify 0.9.91 with the declarations above; no default MemoryKit configuration, no Codex memory MCP registration, and no memory-related declaration in the inspected local VS Code extension directory. No live context-graph owner/API was exposed to this session. Namespace access and functional editor memory use were not probed. This is a dated inventory, not a guarantee for another profile or future installation.

Pester fixtures cover absent tools, failures, partial hosts, supported declarations, invalid/stale evidence, subprocess failure/timeouts, secret redaction, store immutability and unchanged doctor exit behavior. CI runs them on Windows, Linux and macOS.

Run the report​

Graphify and the live context graph​

MemoryKit, memory MCP and VS Code​

Manual host evidence​

Adapter selection​

Workstation verification, 2026-10-27​