context-graph: serve recall to Claude Code and Codex over MCP

[antejavor]

Closes #418. Part of #390; implements #394 points 1, 2, 4 and 9.

What

• Tool surface in agent-context-graph: components register tools under a new agent_context_graph.tools entry-point group (tools.py).
  • agent-context-graph mcp serves them over stdio MCP, behind a new mcp extra.
  • agent-context-graph recall "<question>" [--json] runs recall from a shell.
• recall(question) in sessions-graph (sessions_graph/tool.py): the user and the [recall] lane and width overrides come from config.toml, never from the call, so a model can only read its own user's memory.
• SessionStart: one line of context saying the tool exists, added only when sessions-graph is enabled. No retrieved content is pushed, and if a tool fails to load the hook still succeeds.
• Plugins: Claude Code and Codex bundle the server through .mcp.json, and both bootstrap.sh scripts install agent-context-graph[mcp].
• Config and doctor: config set and bootstrap now keep hand-edited [recall] keys when they rewrite the file, and doctor checks that the server can start and serves recall.
• Docs: a Recall section in the agent-context-graph README (with the overrides table), Codex specifics in its plugin README, a CONTEXT.md glossary entry, and AGENTS.md.

Change from #394 point 4: results are text only

#394 chose "text for the model, JSON beside it". Checked end to end, that choice backfires: when a result has both, Claude Code and Codex both show the model the structuredContent JSON instead of the text. Codex's source does this explicitly (as_function_call_output_payload prefers structured_content), and Claude Code was observed sending the model the JSON. Sending both meant the model never saw the benchmarked text. The MCP result now carries only the text; the JSON is printed by recall --json.

Checked and written down (#418 "done when")

	Claude Code 2.1.x	Codex 0.158
Text + structuredContent	Model sees the JSON only	Model sees the JSON only (text dropped)
Plugin bundles an MCP server	.mcp.json at the plugin root; tool appears as mcp__plugin_context-graph_context-graph__recall	"mcpServers": "./.mcp.json" in .codex-plugin/plugin.json
SessionStart additionalContext	Injected ("provided additionalContext (203 chars)")	Injected; the model called recall without being told about memory

Codex specifics, now in the plugin README:

• Plugin hooks run only after the user trusts them (startup review or /hooks); codex exec never prompts.
• Each recall call asks for approval unless default_tools_approval_mode = "approve" is set.
• MCP servers don't inherit the environment, so the Codex .mcp.json forwards CONTEXT_GRAPH_CONFIG via env_vars.

End-to-end runs

Both runs used a fresh MAGE Memgraph and the worktree plugins.

• Claude Code: session 1 told the model a fact and was recorded by the hooks; the session-end embed ran (embedding_status = completed). In session 2 the model called recall and answered correctly from the text result.
• Codex: with hooks and the tool allowed, the model called recall without being prompted and answered from the Claude Code session's memory.

Tests

• agent-context-graph: 122 passed. New test_tools.py covers the MCP server through an in-memory client, the CLI, the SessionStart hint, and [recall] surviving rewrites.
• sessions-graph: 93 passed, 1 skipped. New test_e2e_recall_tool.py runs against a real Memgraph: own user only, config widths, refusals, and registration.
• ruff and ty are clean.

Not in this PR

• actions-graph: record Claude Code assistant replies so recall can return them #425: the Claude Code adapter never records assistant replies (Stop's last_assistant_message is ignored), so in real sessions recall only sees user turns.
• Release: the plugin manifest versions and the >=0.1.9 pin in bootstrap.sh need bumping when agent-context-graph with mcp is published.