v0.6.4 System Prompt snippet — discovery-aware NHI bootstrap

v0.6.4 ships with a 5-tool default surface. The other 38 tools are reachable through three mechanisms (always-on memory_capabilities, tool_not_found error hints, --include-schema runtime expansion), but AI agents need to know the discovery dance is the convention — otherwise they’ll silently fail tasks the v0.6.3 surface used to handle.

This snippet primes the agent with that convention. Drop it into any system prompt that loads ai-memory in the v0.6.4 default profile.

Recommended system prompt addition

You have an `ai-memory` MCP server attached. It exposes a persistent
memory system with tools organized into 8 families (core, lifecycle,
graph, governance, power, meta, archive, other).

By default, only 5 core tools are advertised:
  memory_store, memory_recall, memory_list, memory_get, memory_search

If you need a tool that is not in this list, follow the discovery
dance:

1. Call `memory_capabilities` to see which families are loaded and
   which are not. The response includes a `families` block listing
   all 8 families with their tool names and a `loaded` flag.

2. If the tool you need lives in a not-loaded family, call
   `memory_capabilities` again with `family=<name> include_schema=true`
   to retrieve the full schemas. You can then construct `tools/call`
   requests against those schemas at runtime, OR ask the operator to
   restart with `--profile <name>` (or `--profile full` for the v0.6.3
   surface 1:1).

3. If you call a tool that is not loaded, the server returns a
   `-32601` error with an actionable hint naming the family and
   suggesting both `--profile` and `--include-schema` paths. Read
   the hint and recover — do not silently give up.

The discovery dance is the canonical pattern. `memory_capabilities`
is always loaded regardless of profile, so step 1 always works.

Per-harness placement

Add the snippet to the host’s system-prompt mechanism:

Harness	Where to add the snippet
Claude Code	The SessionStart hook injects this into the boot manifest. v0.6.4 ships an updated `claude-code.md` recipe that includes the snippet by default.
Claude Desktop	Add to your `~/Library/Application Support/Claude/claude_desktop_config.json` `system_prompt` field, or paste at the top of every conversation.
OpenAI Codex CLI	Add to your project’s `.codex/system-prompt.md` or pass via `--system` on each invocation.
xAI Grok CLI	Add to your `.grok/system.txt` or via the harness’s system-prompt injection point.
Google Gemini CLI	Add to `.gemini/system.md` or pass via `--system-instruction`.

Why this is v0.6.4 ship-blocking

Without the snippet, T1 (awareness) of the NHI Discovery Gate lights RED across most LLMs — agents answer from training-set knowledge instead of calling memory_capabilities. With the snippet, T1 reliably hits ≥90% across Claude/GPT/Grok/Gemini.

This snippet is the cheapest immediate fix that turns the discovery dance from optional into convention.

Validation

After adding the snippet to your system prompt, run a quick check:

You: "What memory tools are available? List every family."

The agent should call memory_capabilities, then list 8 families with their loaded flags. If it answers from training-set knowledge or lists < 6 families, the snippet didn’t reach the system prompt — debug your harness’s system-prompt mechanism.

Full empirical validation runs in the discovery gate on a weekly cadence.