@deepseek-ai/dsh-evolution-learning-graph

Learning graph over skills and memory

Graph surface

/graph renders the learning graph. A skill node is its name; a memory node is memory: :, where is memory|user and `` is the position in that file's entries.

Memory node ids carry a trailing snapshot token (memory: :: — an 8-hex digest of the node label's first line). It exists so graph edit/graph delete can detect index drift: any memory write between the last render and the command shifts indices, and a stale id is rejected (requiring a re-run of /graph) instead of mutating a different entry.

Memory→skill edges are word-level, not substring: the entry is tokenized on non-letter/digit/hyphen runs and a skill name links only when it is a whole token. This prevents a skill named run from linking the words running/grunt.

graph edit/graph delete route through the evolution approval seam when it is mounted (soft-probed; the write executes directly when it is absent) — for BOTH skill nodes and memory nodes (P2-6, v15; memory staged args mirror the memory tool runner's replay shape). Before staging, the seam's hasRunner is checked (P2-7, v15): with approval enabled, a session policy that is not never, and the write actually staging (foreground stages per approval stageForeground; subagent origins always stage), a missing replay runner's row (tool-skill-manage / tool-memory) refuses the write; allow-direct combinations execute unchanged instead of creating a pending record no approver could replay. Each approved/executed edit bumps the skill's patch counter and a delete archives it, matching skill_manage — including the no-op gate: an edit whose content is byte-equivalent to the current file (noop) writes nothing and does not bump the patch counter, exactly as skill_manage treats an unchanged update/patch. The command invocation's session rides the approval request (v12 N1), so a never-policy session stages nothing instead of deriving every graph write as foreground.

Model Experience

Indirect model surface

@deepseek-ai/dsh-evolution-learning-graph registers no direct prompt or tool schema itself. Model-visible effects are owned by the packages that consume this service.

Token effect

Zero direct token effect from this package; consumers add any model-visible tokens.

KV Cache effect

Independent of request-prefix construction. This package does not alter the assembled prompt or tool list.

Known Limitations and Deferred Work

• The memory drift snapshot covers only the node label (first line, first 80 chars): a change confined to a later line of the same entry is not detected, because the rendered node label is the comparison anchor.
• A hand-typed bare memory: : id has no snapshot and skips the drift check (legacy path; the rendered ids always carry the snapshot).