Shared memory and context tools for agentic work.
Code Rooms
# Graph change tracking
Published in signed 1.6.5 at `363bc97f2336c73e90802916804dccd2109c5746`:
five existing `lock_*` handlers are discoverable again. The historical 1.6.3
release does not include this menu restoration; select a matching 1.6.5 wrapper
and native runtime using [the installation guide](INSTALL-1.6.5.md).
The default MCP menu remains 15 tools. Set `M1ND_TOOL_TIER=investigate` in the
**server's launch environment**, then restart/reconnect the MCP host. Its
`tools/list` advertises the core plus these five tools (20 total):
`lock_create`, `lock_watch`, `lock_diff`, `lock_rebase`, `lock_release`.
`M1ND_TOOL_TIER=full` advertises the full registry instead. Unknown values retain
core. Neither setting changes authority, proof, or host-binding requirements.
## First investigation
1. Use `north` to orient and `seek` or `search` to find the target. Copy a live
canonical `node_id` from the result; do not invent one from a filename.
2. Ask `help` with `tool_name: "lock_create"` for the live input schema. Create a
baseline using your `agent_id`, `scope: "node"`, and `root_nodes: [node_id]`.
3. Save the returned `lock_id`. Every later call uses that exact ID and the same
`agent_id`, in the same running session.
4. Call `lock_diff` after a graph update that is independently authorized. Read
changes before deciding whether to call `lock_rebase` to accept the new baseline.
5. Call `lock_release` when finished.
These are **graph snapshots**, not filesystem locks, reservations, mutual
exclusion, Git diffs, or source-content checks. Editing a file alone does not
update the graph. An empty graph diff does not establish unchanged source text.
Snapshots and watches are session-local and lost on restart; they are not a
resume/learning feature or durable investigation journal.
In published 1.6.5, graph-linked `search` results from a single-file
ingest root retain the graph's basename ID, including when the filesystem uses a
canonical Windows path. Check `graph_linked`: disk-only fallback rows are not
proof of graph membership. If distinct ingest roots make one relative graph ID
ambiguous, search omits the ambiguous file-content mapping rather than choosing
one root. This does not authorize the file's parent or siblings; existing
workspace/ingest read-root checks still apply.
`help(intent: "track changes in the graph")` offers this sequence. Help exposes
the full registry, but help text does not register tools in a host. If a host
cannot call an unadvertised tool, use `investigate` and confirm its actual
`tools/list` before calling it. Public stdio discovery/calls are testable here;
this is not a claim that every host UI has been tested.
## Advanced recipe
With the investigate menu selected, use the following arguments in consecutive
`tools/call` requests. Replace `LIVE_NODE_ID` with a returned canonical ID and
`RETURNED_LOCK_ID` with the create response; keep the process and agent unchanged.
```json
{"name":"lock_create","arguments":{"agent_id":"investigator","scope":"subgraph","root_nodes":["LIVE_NODE_ID"],"radius":2}}
{"name":"lock_watch","arguments":{"agent_id":"investigator","lock_id":"RETURNED_LOCK_ID","strategy":"on_ingest"}}
{"name":"lock_diff","arguments":{"agent_id":"investigator","lock_id":"RETURNED_LOCK_ID"}}
{"name":"lock_rebase","arguments":{"agent_id":"investigator","lock_id":"RETURNED_LOCK_ID"}}
{"name":"lock_release","arguments":{"agent_id":"investigator","lock_id":"RETURNED_LOCK_ID"}}
```
Between watch and diff, use only a graph update allowed in your current context.
For example, existing `ingest(mode: "refresh")` requires the exact declared root
and retains all its guards; the menu does not grant refresh permission.
`on_learn` observes admitted learning notifications and does not unlock `learn`.
- `node`: the explicit root set. `subgraph`: outgoing-edge BFS, radius 1–4,
default 2. `path`: roots plus optional `path_nodes`, an explicit node set;
it does not calculate or validate a connected path.
- Roots must be nonempty. Exact canonical IDs are preferred; an exact unique
label is also accepted. Unknown or ambiguous identifiers refuse. `query` and
`query_neighborhood` refuse rather than silently producing an empty baseline.
- Baselines capture scoped nodes and all incident edges, including incoming and
outgoing boundary edges. `baseline_edges` counts directed CSR arcs: a
bidirectional logical edge contributes two. Parallel arcs use deterministic weight-sorted multiset pairing, not stable
per-edge-instance identity. Edge keys are opaque collision-safe JSON tuples;
do not parse them as delimited filenames.
Diff reports canonical node additions/removals, internal edge changes,
boundary edge changes, weight changes and generation staleness.
Weight changes are reported only when the absolute delta exceeds 0.001;
`no_changes` means no detected change above that tolerance. `diff_truncated`
flags capped added-node, added-internal-edge and added-boundary-edge previews
only. `no_changes` is computed before those caps, so truncation cannot hide a
detected change behind an empty verdict.
- `manual` buffers no notifications. `on_ingest` and `on_learn` buffer coalesced
graph-trigger notifications for the session; diff drains them but keeps the
baseline.
The drained count is notification batches, not the number of mutations.
There is no background scheduler or filesystem watcher; `periodic` refuses.
- Rebase resamples the scope and preserves watch configuration. Review first:
rebasing discards the old comparison point. If the entire observed scope has
vanished, rebase refuses and preserves the old baseline so you can inspect
removals or release it. Release discards the snapshot and
watcher. Another agent ID cannot operate the snapshot; this ownership check
is not authentication against a caller that can choose its agent ID.
## Verification boundary
The regression suite covers registry/menu/help consistency and the runtime
snapshot lifecycle. An isolated public stdio probe must establish
`tools/list` → `tools/call`, graph change → diff, rebase and release. Exact
commands/results belong to the exact tested revision's report. Publication does
not establish that every host UI has been tested. This guide does not claim
durable trails, a learning authority change, or a new cross-platform integration run.