Agent identity on every core MCP tool, and write verbs that point the right way

[samrusani]

Two defects found while testing whether an agent can install Alice and use it unattended, which is the distribution model these tools have to serve.

1. Five of eleven core tools rejected agent identity

alice_recall, alice_resume, alice_recent_decisions, alice_explain and alice_memory_correct had additionalProperties: false with no identity properties, while the other six accepted them. An agent that stamps agent_id on every call, which the commit tool's own schema invites, got a hard tool_request_failed on every one.

Note alice_memory_correct is a write verb, so this was not only a read-path problem.

Verified against the shipped binary:

before:  alice_recall {query, agent_id} -> isError=true
after:   alice_recall {query, agent_id} -> isError=false, count=1

2. The descriptions steered agents into the review queue

tool	said	problem
alice_capture	"use this whenever you learn something worth keeping"	but it is review-gated, so nothing becomes recallable
alice_memory_commit	"on the user's instruction ('remember this')"	tells the agent to wait to be asked

So the tool inviting ambient use was gated, and the tool that works unattended told the agent not to use it unprompted. An agent recording something it learned got status: "imported", isError: false, and content that alice_recall never returns.

alice_capture behaviour is unchanged. It is review-gated by design and no persona alters that, confirmed on fresh stores with and without ALICE_MEMORY_PERSONA. Only the wording changes: capture now states plainly that alice_recall will not return the text and names commit as the verb for immediate recall; commit now says to use it whenever you learn something worth keeping, including when the user has not asked.

On the frozen-definitions digest

It moves because the definitions moved. Before updating it, all eleven core tools were diffed between HEAD and this branch: exactly five changed, each gaining only the identity properties, six untouched.

The guard was then re-checked for vacuity by deleting one splice and confirming the count actually dropped 11 → 10 before running the test. An earlier attempt used a non-matching string, changed nothing, and passed for the wrong reason.

Not in scope

alice_explain errors on every call, with and without identity, verified identical on stashed HEAD. Pre-existing and unrelated; separate PR.

Verification

4544 unit passed, 2 skipped.

Upgrade Overview

Protected Areas

• continuity APIs
• memory schema
• evidence pipeline
• trust rules
• promotion logic

Compatibility Impact

Additive only on the request side. Five core MCP tools gain optional agent
identity properties they previously rejected; no property is removed, renamed or
made required, and no response shape changes. Callers that never sent identity
are unaffected. Callers that did send it moved from a hard tool_request_failed
to a successful call, so the only behavioural change is that previously failing
requests now succeed.

Two tool descriptions change text. No tool is added, removed or renamed, and the
core tool count stays 11. alice_capture behaviour is unchanged: it remains
review-gated and its content is still not returned by alice_recall, which the
new wording now states explicitly instead of implying the opposite.

Migration / Rollout

None required. No database migration, no schema change, no configuration change,
no data backfill. The change is contained in the MCP tool definitions and takes
effect when the process restarts.

Operator Action

None. Operators do not need to change environment variables, configuration files
or deployment settings.

Worth knowing for anyone integrating an agent: an agent that stamps agent_id
on every call previously failed on alice_recall, alice_resume,
alice_recent_decisions, alice_explain and alice_memory_correct, and now
succeeds. If any integration carried a workaround that strips identity fields
before read calls, that workaround is now unnecessary, though harmless.

Validation

4544 unit tests passed, 2 skipped, including the frozen-definitions guard
test_mcp_registry_order_definitions_and_alias_identity_are_frozen.

Verified against the shipped alice-memory mcp binary over stdio, on a fresh
data directory, with no agent key and no environment variables:

before:  alice_recall {query, agent_id} -> isError=true  tool_request_failed
after:   alice_recall {query, agent_id} -> isError=false count=1

Before updating the definitions digest, all 11 core tools were diffed between
HEAD and this branch: exactly 5 changed, each gaining only the agent identity
properties, 6 untouched. The guard was then re-checked for vacuity by deleting
one splice and confirming the count dropped from 11 to 10 before running the
test.

Rollback

Revert the commit. There is no persisted state, no migration and no artifact
tied to this change, so a revert fully restores the previous behaviour with a
process restart. Agents that had started sending identity on read calls would
return to receiving tool_request_failed, which is the pre-change behaviour.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.