fix(mcp): anchor REMEMBER to the turn, log tool names, add a live handshake probe

[harrymove-ctrl]

Rebased onto dev after #706 landed. This PR is now only what #706 did not cover.

Dropped, because #706 solved it better

This branch originally carried two more fixes. Both are superseded, and I removed them rather than resolve the conflict in my favour:

1. Signed-out vs signed-in descriptions. fix(mcp): make Claude Code auto-remember/recall actually fire #706 added buildToolDefinitions(proactive) with separate SIGNED_OUT_* / SIGNED_IN_* constants. My version pushed proactive wording into the one shared list, which would have told signed-out clients to save proactively using tools that cannot work without credentials. fix(mcp): make Claude Code auto-remember/recall actually fire #706's design is correct; mine was worse.
2. memwal_recall annotations. fix(mcp): make Claude Code auto-remember/recall actually fire #706 set readOnlyHint: true, destructiveHint: false, going further than the readOnlyHint: false I had, with the same reasoning. memwal_analyze keeps dev's destructiveHint: true — it writes memories, so that value is defensible.

1. REMEMBER had no turn anchor

With the #706-equivalent fixes live on dev, I ran T1-T3 from WALM-368, scored from the MCP log rather than the tool cards:

Test	Result
T2 recall	PASS — memwal_recall fired unprompted and answered from it
T1 remember	FAIL — no tools/call at all
T3 control	PASS

Recall firing is new. Both earlier dogfood passes recorded zero tools/call for an entire session. Since RECALL and REMEMBER ship in the same instructions block, to the same client, in the same session, this is a controlled comparison rather than a guess — and the difference is in our own text:

	Trigger	Anchored to
RECALL	"before answering anything that touches..."	a turn event
REMEMBER	"...call memwal_remember without waiting to be asked"	nothing

REMEMBER asked the model to classify a statement and then self-start, with nothing tying the call to a moment in the turn. In T1 it replied "Got it — <fact>. What do you need done with it?", treating a durable fact as task setup and moving on.

REMEMBER now anchors with "in that same turn, before you finish replying", adds configuration values (hostname, port, region, id) since the failing case was a hostname and port, and closes the observed behaviour: do not ask whether to save, and acknowledging a fact in the reply does not store it.

#706's scoping is preserved verbatim ("skip one-off tasks, the current file or bug, and small talk"). Applied to all three copies that carry this contract; the two that must be byte-identical were verified so.

2. Tool calls were not scorable from the log

The host logs method="tools/call" id=N with no tool name, so "remember never fired" and "remember fired and failed" are indistinguishable — the exact ambiguity under investigation, and why the earlier dogfood runs could not be reconciled.

The bridge now logs {"event":"bridge.tool_call","tool":"memwal_recall","id":7} — name only, never arguments, since memory text is the user's private data.

grep -o '"event":"bridge.tool_call","tool":"[^"]*"' ~/Library/Logs/Claude/mcp-server-memwal.log | tail

3. packages/mcp/test/handshake-contract.mjs

node test/handshake-contract.mjs --dev

Drives the real built server against a chosen env and asserts what a client is actually handed: instructions present with both halves anchored, memwal_recall advertised read-only, and cold start agreeing with upstream field by field.

tool-definitions.test.mjs from #706 covers the static list. This covers the deployed path, which is where the earlier runs went wrong:

1. Tool metadata lives in services/server/scripts, so a published npm package alone changes nothing for a signed-in user. Testing before the relayer redeploys measures the old server.
2. Cold start and upstream can disagree, so the answer depended on whether you looked before or after the re-list.

Deliberately not named *.test.mjs: it needs a live relayer and credentials, and the npm test glob must not collect it. On failure it names the broken precondition; on success it prints the T1-T3 runbook, since those are model-behaviour cases a script cannot automate.

4. Finishing the cold-start sync #706 started

The probe immediately earned its keep, catching drift on four tools #706 did not touch:

Tool	Cold start	Upstream
memwal_remember_bulk	"Remember Several Facts"	"Remember Multiple Facts"
memwal_health	"Check Relayer Health"	"Check Walrus Memory Health"
memwal_analyze	150 chars	340 chars
memwal_restore	165 chars	427 chars

Less severe than the remember/recall case since these are thinner rather than contradictory — memwal_restore upstream explains when to reach for it ("recall returns nothing even though facts were saved before"), while the cold-start copy described only the mechanics. A client that caches its first tools/list keeps the thin version all session.

Values were copied programmatically from a live upstream tools/list rather than retyped, so they cannot drift again through transcription.

Testing

packages/mcp             28 pass, 0 fail
services/server/scripts  221 pass, 0 fail
handshake-contract --dev 13/13   (dev relayer deployed from this branch)

login-handoff.test.mjs pinned the two stale titles. Its own comment requires pre-login discovery to expose "the same safety metadata clients will receive after the bridge hands off," so the expectations were wrong, not the change. Correction, not relaxation.

Honest scope

Items 2, 3 and 4 are tooling and correctness, and stand on their own. Item 1 is a hypothesis with a controlled comparison behind it, not a proven fix. It needs another T1 run on a real client to confirm. Nothing here should be read as "proactive memory is fixed" until that run happens — the probe exists precisely so that run is attributable.

Found in passing, NOT fixed here

The dev relayer crash-loops on redeploy, which produced the 502s during this work:

thread 'main' panicked at src/main.rs:758:14:
Failed to initialize legacy security-delete database:
  Internal("legacy migration failed: ... canceling statement due to lock timeout")

LEGACY_DB_URL and DATABASE_URL point at the same Neon database, so the legacy security-delete migration contends with the main migration runner; during a redeploy the draining container holds the lock and the new one panics. Two separate problems: a transient lock timeout should not be fatal at main.rs:758, and the legacy path should not compete with the main runner on every boot. Both are in the Rust server, out of scope here, and worth their own ticket.

Refs WALM-368, WALM-324

---

Measured against dev's shipping artifacts

Added MEMWAL_MCP_SERVER so the probe can run a published package instead of this checkout's build. That lets it answer the question that matters before merging: what does dev ship today?

Published @mysten-incubation/memwal-mcp@dev (0.0.10-dev.4) against the dev relayer at e4bdcc97 (#706), signed in:

11/13 passed
FAIL  REMEMBER is anchored to a turn event
FAIL  cold-start and upstream lists agree
      (drift: remember_bulk, analyze, restore, health)

Both defects this PR fixes reproduce with none of my code in the path — dev's own package, dev's own relayer. The same probe against this branch reports 13/13.

Reproduce:

npm pack @mysten-incubation/memwal-mcp@dev
# extract, npm i --omit=dev, then:
MEMWAL_MCP_SERVER=<extracted>/dist/bin/memwal-mcp.js \
  node packages/mcp/test/handshake-contract.mjs --dev