Scored across 1 tool
With a single tool there is no tool-to-tool confusion, but the eight embedded modes (recall, context, blame, fix, how, orient, remember, handoff) have adjacent boundaries — e.g. recall vs. context vs. orient, and how vs. fix — that an agent must resolve from prose alone. The descriptions do a good job of distinguishing them with concrete trigger conditions, so collisions are unlikely but not impossible.
A single tool named 'deja' cannot be internally inconsistent, and the mode names inside it are uniform lowercase verbs (recall, blame, fix, how, orient, remember, handoff). It loses the top mark only because the surface exposes no verb_noun tool naming pattern to be consistent with — the operation name is a string parameter rather than part of the tool name.
Eight genuinely distinct operations (search sessions, reconstruct a session, attribute file history, error lookup, command lookup, project orientation, write a memory, resume another session) are multiplexed behind one tool and a free-form 'mode' argument. The capability set is right-sized, but the surface collapses it into a single entry point, so mode selection becomes an unstructured parameter choice rather than a discoverable tool.
The lifecycle of session memory is well covered: read (recall, context, orient, how, blame, fix), write (remember), and resume (handoff). The main gap is memory management — 'remember' stores a settled decision but there is no way to list, revise, or forget stored decisions, nor to enumerate sessions without a query term.
