Dpd Find Similar

/dpd-find-similar — Retrieval-augmented proposal (v0.3.2)

Announce: "Using dpd-find-similar skill."

This skill is the user-pull entry point for DPD's retrieval-augmented proposal pipeline (v0.3.2 spec §6).

When to invoke

User invokes /dpd-find-similar [query?]. Claude may NOT auto-fire this skill (v0.3.2 §3.4: user-pull only).

Step 1 — Query resolution

If query argument is present, use it directly.

Otherwise, propose a query candidate in this priority and explicitly confirm before running the search:

1. focus_node text (call get_session_state to fetch).
2. Recent conversation topic (cluster the last few turns around the most prominent anchor).
3. Active Pool item cluster (call pool_list(active_only=True) and look for a shared keyword among items whose attach point is undetermined).

Wording: "query: <candidate> で検索する、OK?" — Wait for explicit confirmation. Do not silently proceed.

Step 2 — Run find_similar

find_similar(query="<confirmed>", top_k=5, include_open=False)
  → { "results": [SubgraphSummary, ...] }

SubgraphSummary fields: start_node_id, session_id, root_id, scope, start_text, end_text, achievement_conditions, state (closed / archived), score, matched_snippet, closed_at.

Step 3 — Present results

Render in conversation prose (not a graph diagram). Format:

過去類似 subgraph (state=closed/archived):

1. [score=8.21] scope=<scope> root=<root_id> start=<start_node_id> "<start_text>"
   End: "<end_text>"
   Conditions: "<achievement_conditions>"
   Snippet: "...<matched_snippet>..."

2. ...

Allowed prose: factual descriptions of what was retrieved (above format counts). Forbidden prose: lesson-style claims like "past X did Y, so we should Y" — those belong in Step 4 as graph operands.

If results is empty: offer "該当なし、include_open=True で再検索しますか?" and on confirmation re-run with include_open=True.

Step 4 — Distill into graph candidate (D3)

After the user selects one or more past subgraphs:

1. Read each selected subgraph's Start / End / decisions / rationales.
2. Hypothesize what to add to the current focus subgraph: which nodes, which edges, which neighboring modifications.
3. Format as a hierarchical list (the v0.3.1 §4.5 ambient-mode proposal format), with inline annotations:
   • (NEW) decision ← "..."
   • [→ add edge type=derived_from from=X to=Y]
   • (NEW) rationale ← "..."
4. Reuse v0.3.1 §4.6 reaction handling: Full OK applies all; Partial reject calls pool_reject on rejected items; Full reject calls pool_reject on all and stays in ambient.

Strict prose ban: do NOT write lesson-style commentary. If a justification belongs in the proposal, encode it as a rationale node, not as prose.

Tone

Custodial, not transactional. Past judgments are inputs the user gets to inspect — they are not authoritative.

Sub-skill boundaries

• This skill is invoked only by the user. Other skills (/dpd-fill, /dpd-import) may call find_similar internally when their own user-pull semantics already justify it.
• This skill never touches End nodes directly (v0.3.1 §5.3 End modification gate still applies to any End edits in the distilled candidate — confirm before applying).
• This skill never auto-consults retrieval during ambient mode (§3.4 user-pull discipline).

Feedback footer

After completing a meaningful response (not for trivial status output), print exactly one line at the very end:

💬 Hit a bug or have feedback on DPD? Run /dpd-feedback "<short description>" or open an issue at https://github.com/o3co/agent-dpd/issues/new

This surfaces the dogfood feedback path without interrupting the main interaction. Keep it to one line. Do not repeat across multiple turns within the same exchange — once per skill invocation is enough.