o3co /dpd-fill
Announce: "Using dpd-fill skill."
When to invoke
- User wants gap analysis on the current graph
- User asks "何か抜けてる?" / "穴はある?" / "assumptions をチェックして"
- After
/dpd-import, to detect gaps in the imported doc - Before finalizing a subgraph (pre-
mark_reachedsanity check)
Tool calls
1. Get session baseline
get_session_state(session_id=<session_id>)
Returns session mode, active roots, focus node.
2. Walk the graph
For each active root (or focus root if scoped):
walk_subtree(session_id=<session_id>, root_id=<root_id>)
Collect all nodes. Note: state, type, provenance, text for each.
3. Inspect Pool
pool_list(active_only=True, scope=<sub-scope>)
Pool items are observations not yet attached to the graph. Check for items that imply graph gaps.
Inference pass
With the full graph + Pool in context, run the following prompts internally:
- Missing decompositions: "Are there nodes that should have children but don't? Which open questions remain undecomposed?"
- Unstated assumptions: "What assumptions are implicit in the existing decisions/answers that have no
assumptionnode recorded?" - Unexamined hypotheses: "Are there hypotheses that were never explored or closed? Are there alternative hypotheses not surfaced?"
- Gap candidates: "What relevant considerations appear absent from this graph entirely?"
- Pool signals: "Do any Pool items imply a node or edge missing from the graph?"
Collect inferences. For each:
- Draft node:
type,text, targetparent_id, rationale for why it's missing - Classify stakes: high (structural gap, affects decisions) / low (supplementary detail)
User opt-in flow
Present all inferred additions as a numbered list before calling any tool:
/dpd-fill found N candidate inferred nodes:
1. [high] assumption under <parent>: "<text>"
Why: <rationale>
2. [low] hypothesis under <parent>: "<text>"
Why: <rationale>
...
Apply all? (Y/N/select numbers)
Wait for user response. Do NOT call add_node before user confirms.
On confirmation (full or partial), for each approved node:
add_node(
session_id=<session_id>,
parent_id=<parent_id>,
type=<type>,
text=<text>,
provenance='inferred'
)
After each: one-line added <node_id> [inferred]: <text>
Per-turn self-check verification pass [v0.3.1]
Before presenting the candidate list to the user, run the following checks on each proposed inferred node. This mirrors the ambient-mode per-turn self-checks (SKILL.md §4.8).
| Check | What to verify for each proposed inferred node |
|---|---|
| #1 End modification | Does this inferred node implicitly modify or expand the End's scope? If yes, flag it — it requires user confirmation, not silent inference. |
| #2 End scope | Does this node extend the subgraph beyond the End's original achievement criteria? If yes, downgrade or mark: "[out-of-scope — propose as separate subgraph?]" |
| #3 Factual / vendor-spec claim | Does the node text assert a vendor fact, API availability, or external compatibility? If yes, mark as "unverified" and recommend WebSearch before applying. |
#4 decision node without source |
Is this a decision-type inferred node without a source evidence node in context? If yes, add a note: "requires derived_from source to be identified before applying." |
| #5 Flat overcrowding | Does this inferred node bundle N≥3 distinct concerns? If yes, propose splitting into sub-tree before adding. |
#6 contributes_to fanout |
Does adding this node trigger a cascade of contributes_to edges to the End? If yes, apply §4.2.2 norm. |
For each flagged node, annotate inline in the candidate list:
2. [high] decision under <parent>: "<text>"
Why: <rationale>
⚠ Check #3: vendor-spec claim — verify before applying
/fcot result: <...>
Unflagged nodes proceed normally to user opt-in.
/fcot orchestration
For each high-stakes inferred node (stakes = "high"), invoke /fcot after proposing but before applying:
/fcot "<inferred node text>"
/fcot will attempt to falsify the inference. If /fcot finds the inference unsound, downgrade or drop that candidate. Report /fcot verdict inline with the proposal.
Pattern:
1. [high] assumption: "<text>"
Why: <rationale>
/fcot result: <Confirmed sound | Falsified — <reason> | Quick check only>
For low-stakes nodes, /fcot is optional (user may request it explicitly).
Notes
- All inferred nodes get
provenance='inferred'— this distinguishes them from conversation-grounded (grounded) and manual edits (manual) in the graph audit trail. /dpd-fillis safe to run multiple times. Re-running after new conversation will surface new gaps.- If run after
/dpd-import, the imported graph (provenance='imported', state='archived') is included in the walk. This enables systematic gap analysis of external docs (§7.1 pipeline). - Advanced
/dpd-fill(goal-driven auto-decomposition) is deferred to v0.3.2+. Current implementation is manual inference pass + user opt-in.
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.