| name | x-ipe-assistant-knowledge-librarian-DAO |
| description | Central orchestrator for the knowledge pipeline. Discovers available knowledge skills at runtime, classifies requests, routes to constructors/extractors/ontology skills, and drives the full 格物致知 workflow. Triggers on requests like "build a user manual", "extract knowledge", "run knowledge pipeline", "discover ontology graphs". |
Knowledge Librarian — Assistant Skill
Purpose
AI Agents follow this skill to orchestrate the knowledge pipeline by:
- Discovering available
x-ipe-knowledge-* skills at runtime and classifying incoming requests
- Driving the 格物致知 workflow: plan (framework → overview → rubric → requests) then execute (extract → fill → critique → ontology → store → present)
- Tracking pipeline state step-by-step with graceful degradation on individual failures
Important Notes
BLOCKING: This skill coordinates knowledge skills — it does NOT perform extraction, construction, or ontology operations itself. Each step delegates to the appropriate knowledge skill.
CRITICAL: Knowledge skills CAN be invoked directly without this orchestrator as a degraded-mode fallback (AC-059E-08a/b).
CRITICAL: The 格物致知 backbone is the INTERNAL orchestration methodology — callers see only pipeline_summary output, never phase names.
CRITICAL: Best-Model Requirement. When this skill is delegated to a sub-agent, it MUST use the most capable (premium) LLM model available. The orchestration logic requires nuanced semantic matching and critique evaluation.
About
This skill serves as the central coordinator for the entire knowledge pipeline (Layers 1–3). Given a knowledge request (e.g., "build a user manual for app X"), it discovers which skills are available, classifies the request type, selects the best-matching constructor, and orchestrates the full plan-execute cycle.
CORE Backbone — 格物致知 (Investigate to Reach Understanding): The Librarian's orchestration follows a two-phase framework: 格物 (plan the knowledge acquisition) gathers framework, overview, rubric, and knowledge requests; 致知 (execute the plan) dispatches extractors, fills drafts, critiques quality, processes ontology, stores results, and presents output.
Key Concepts:
- Skill Discovery — Runtime glob scan of
.github/skills/x-ipe-knowledge-*/SKILL.md builds a categorized registry
- Semantic Routing — Request text matched against constructor descriptions for best-fit selection
- Critique Loop — Sub-agent evaluates draft against rubric; max 3 iterations before accepting partial quality
- Pipeline State — Each step tracked independently; failures are logged and pipeline continues where possible
When to Use
triggers:
- "Build a user manual for an application"
- "Extract knowledge from a website or memory"
- "Construct notes from source material"
- "Discover and link ontology graphs across domains"
- "Run the full knowledge pipeline"
- "Store knowledge to persistent memory"
not_for:
- "Direct knowledge skill invocation (skills work standalone as fallback)"
- "Task board management (use x-ipe-tool-task-board-manager)"
- "Feature workflow orchestration (use x-ipe-workflow-task-execution)"
- "Non-knowledge operations (code implementation, testing, etc.)"
Input Parameters
input:
request: "string"
request_type_override: "string | null"
target_constructor: "string | null"
max_iterations: 3
output_format: "markdown"
session_context:
task_id: "{TASK-XXX | N/A}"
feature_id: "{FEATURE-XXX | N/A}"
workflow_name: "{name | N/A}"
Input Initialization
<input_init>
<field name="request" source="Caller provides the knowledge task description">
<validation>MUST be non-empty string. FAIL FAST if missing.</validation>
</field>
<field name="request_type_override" source="Caller may override classification">
<steps>
1. IF caller provides a value in {construction, extraction, ontology_only, presentation, storage, behavior_learning} → use it
2. ELSE → null (perform classification in Phase 1)
</steps>
</field>
<field name="target_constructor" source="Caller may specify a constructor directly">
<steps>
1. IF caller provides a constructor name → use it, set request_type_override = "construction"
2. ELSE → null (perform semantic routing in Phase 1)
</steps>
</field>
<field name="max_iterations" source="Caller or default">
<validation>Integer >= 1. Default: 3.</validation>
</field>
<field name="output_format" source="Caller or default">
<validation>One of: structured, markdown, graph. Default: markdown.</validation>
</field>
</input_init>
Definition of Ready
<definition_of_ready>
<checkpoint required="true">
<name>Request provided</name>
<verification>input.request is non-empty string</verification>
</checkpoint>
<checkpoint required="true">
<name>Knowledge skills exist</name>
<verification>At least one .github/skills/x-ipe-knowledge-*/SKILL.md exists on disk</verification>
</checkpoint>
</definition_of_ready>
Execution Flow
| Phase | Steps | Action | Gate |
|---|
| 0 | 0.1 | 礼 — Greet | Announce identity |
| 1 | 1.1–1.6 | 格物 — Plan | Discover skills, classify request, framework, overview, rubric, plan |
| 2 | 2.1–2.6 | 致知 — Execute | Extract, fill, critique loop, ontology, store, present |
| 3 | 3.1 | 录 — Record | Write pipeline execution log |
| 4 | 4.1 | 示 — Present | Return pipeline_summary to caller |
BLOCKING: All phases execute in order. No phase may be skipped.
BLOCKING: Phase 1 (格物) adapts based on request_type — non-construction types skip framework/rubric steps.
Execution Procedure
<procedure name="knowledge-librarian">
<execute_dor_checks_before_starting/>
<phase_0 name="礼 — Greet">
<step_0_1>
<name>Announce Identity</name>
<action>
1. Print: "道 · Knowledge Librarian — ready."
2. Initialize pipeline_state with session_id = "librarian-{ISO-8601-timestamp}"
3. Create session working directory: x-ipe-docs/memory/.working/{session_id}/
</action>
<output>Greeting printed, pipeline_state initialized</output>
</step_0_1>
</phase_0>
<phase_1 name="格物 — Plan">
<step_1_1>
<name>Discover Skills</name>
<action>
1. GLOB: .github/skills/x-ipe-knowledge-*/SKILL.md
2. FOR EACH SKILL.md found:
a. READ YAML frontmatter → extract name, description
b. CATEGORIZE by prefix:
- x-ipe-knowledge-extractor-* → extractors[]
- x-ipe-knowledge-constructor-* → constructors[]
- x-ipe-knowledge-keeper-* → keepers[]
- x-ipe-knowledge-present-* → presenters[]
- x-ipe-knowledge-ontology-* → ontology[]
- x-ipe-knowledge-mimic-* → mimics[]
3. BUILD discovered_skills registry
4. IF discovered_skills is empty → pipeline_status = "failed", STOP
5. Mark pipeline_state.discovery = "done"
</action>
<constraints>
- No hardcoded skill names — registry built fresh each invocation
- Only x-ipe-knowledge-* namespace scanned (not x-ipe-tool-*, x-ipe-task-based-*, etc.)
</constraints>
<output>discovered_skills registry with categorized skills</output>
</step_1_1>
<step_1_2>
<name>Classify Request</name>
<action>
1. IF request_type_override provided → use it, skip classification
2. IF target_constructor provided → type = "construction", skip routing
3. ELSE analyze request text:
a. Construction verbs (build, create, construct, write) + domain → "construction"
b. Extraction verbs (extract, get, fetch, find) → "extraction"
c. Ontology verbs (discover, link, synthesize, normalize) → "ontology_only"
d. Presentation verbs (render, summarize, show, present) → "presentation"
e. Storage verbs (store, save, promote, persist) → "storage"
f. Behavior-learning verbs (learn behavior, track behavior, observe user flow, mimic website behavior, record user actions) → "behavior_learning"
g. No match → "classification_failed"
4. IF type = "behavior_learning":
a. SELECT best matching skill from discovered_skills.mimics[] whose description or operations include behavior tracking
b. Preferred operation: x-ipe-knowledge-mimic-web-behavior-tracker.start_active_tracking
c. IF no mimic skill exists → "classification_failed"
5. IF type = "construction" AND target_constructor is null:
a. FOR EACH constructor in discovered_skills.constructors[]:
Compare request against constructor description (semantic similarity)
b. Select highest-scoring constructor
c. On tie → alphabetical first, log tie-breaking
d. On zero matches → "classification_failed"
6. Mark pipeline_state.classification = "done"
</action>
<output>request_type + selected_constructor (if applicable)</output>
</step_1_2>
<step_1_3>
<name>Provide Framework</name>
<action>
1. IF request_type != "construction" → SKIP (mark "skipped")
2. INVOKE selected_constructor's `provide_framework` operation with request context
3. RECEIVE framework_document
4. SAVE to .working/{session_id}/framework/
5. Mark pipeline_state.格物_1_framework = "done"
</action>
<output>framework_document (or skipped)</output>
</step_1_3>
<step_1_4>
<name>Gather Overview</name>
<action>
1. IF request_type != "construction" → SKIP
2. SELECT appropriate extractor (extractor-web for URLs, extractor-memory for existing knowledge)
3. INVOKE extractor to gather high-level overview of the target
4. SAVE to .working/{session_id}/overview/
5. Mark pipeline_state.格物_2_overview = "done"
</action>
<output>overview_content (or skipped)</output>
</step_1_4>
<step_1_5>
<name>Design Rubric</name>
<action>
1. IF request_type != "construction" → SKIP
2. INVOKE selected_constructor's `design_rubric` with framework + overview
3. RECEIVE rubric_metrics[]
4. SAVE to .working/{session_id}/rubric/
5. Mark pipeline_state.格物_3_rubric = "done"
</action>
<output>rubric_metrics[] (or skipped)</output>
</step_1_5>
<step_1_6>
<name>Plan Knowledge Requests</name>
<action>
1. IF request_type != "construction" → SKIP
2. INVOKE selected_constructor's `request_knowledge` with framework + rubric
3. RECEIVE knowledge_requests[] specifying what, for which section, which extractor
4. SAVE to .working/{session_id}/plan/
5. Mark pipeline_state.格物_4_plan = "done"
</action>
<output>knowledge_requests[] (or skipped)</output>
</step_1_6>
</phase_1>
<phase_2 name="致知 — Execute">
<step_2_1>
<name>Execute Extraction</name>
<action>
1. IF request_type = "construction":
FOR EACH request in knowledge_requests[]:
INVOKE appropriate extractor (extractor-web for URLs, extractor-memory for existing)
COLLECT gathered_knowledge[]
2. IF request_type = "extraction":
INVOKE appropriate extractor directly with request content
COLLECT gathered_knowledge[]
3. IF request_type = "behavior_learning":
a. INVOKE selected mimic skill operation `start_active_tracking` with request context:
- target_app from request/session_context
- session_config.purpose from request
- active_config defaults: polling_interval_s=5, auto_reinject=true, auto_screenshot=true
b. WAIT for `start_active_tracking` to own polling until toolbar Analysis is clicked
c. RECEIVE operation_output.result.observation_payload from the mimic skill
d. APPEND observation_payload to gathered_knowledge[] for downstream processing
4. IF request_type in {ontology_only, presentation, storage}:
SKIP extraction (mark "skipped")
5. Mark pipeline_state.致知_1_execute = "done"
</action>
<output>gathered_knowledge[] (or skipped)</output>
</step_2_1>
<step_2_2>
<name>Fill Structure</name>
<action>
1. IF request_type != "construction" → SKIP
2. INVOKE selected_constructor's `fill_structure` with framework + gathered_knowledge
3. RECEIVE completed_draft (incomplete sections marked [INCOMPLETE: reason])
4. SAVE to .working/{session_id}/draft/
5. Mark pipeline_state.致知_2_fill = "done"
</action>
<output>completed_draft (or skipped)</output>
</step_2_2>
<step_2_3>
<name>Critique Against Rubric</name>
<action>
1. IF request_type != "construction" → SKIP
2. SET iteration_count = 0
3. LOOP:
a. INVOKE critique sub-agent (premium model) with (draft, rubric_metrics)
b. RECEIVE critique_result: { verdict: pass|fail, gaps[], scores{} }
c. IF verdict == "pass" → EXIT loop
d. IF iteration_count >= max_iterations → mark "partial_quality", log gaps, EXIT loop
e. iteration_count++
f. INVOKE constructor's `request_knowledge` for gap sections only (replan)
g. INVOKE extractors for new knowledge requests
h. INVOKE constructor's `fill_structure` for gap sections
i. CONTINUE loop
4. Mark pipeline_state.致知_3_critique = "done"
</action>
<constraints>
- Critique sub-agent MUST use premium model
- max_iterations prevents infinite loops (default: 3)
</constraints>
<output>critique_result + iteration_count</output>
</step_2_3>
<step_2_4>
<name>Process Ontology</name>
<action>
1. IF request_type = "ontology_only":
INVOKE ontology-synthesizer pipeline: discover_related → wash_terms → link_nodes
2. ELIF request_type = "construction" OR request_type = "extraction":
a. INVOKE ontology-builder's `build_ontology` on the draft/extracted content
b. INVOKE ontology-synthesizer pipeline on new entities
3. ELSE → SKIP
4. COLLECT ontology_result
5. Mark pipeline_state.致知_4_ontology = "done"
</action>
<output>ontology_result (or skipped)</output>
</step_2_4>
<step_2_5>
<name>Store Results</name>
<action>
1. IF request_type = "storage":
INVOKE keeper-memory's `store` with request content directly
2. ELIF request_type = "construction":
INVOKE keeper-memory's `store` with completed_draft
Determine memory_type: procedural for manuals/patterns, semantic for facts, episodic for session notes
3. ELIF request_type = "extraction":
INVOKE keeper-memory's `store` with gathered_knowledge
4. ELSE → SKIP
5. COLLECT stored_path
6. Mark pipeline_state.致知_5_store = "done"
</action>
<output>stored_path (or skipped)</output>
</step_2_5>
<step_2_6>
<name>Present Output</name>
<action>
1. IF request_type = "presentation" OR output_format != null:
INVOKE present-to-user's `render` with appropriate content + output_format
2. IF output_format = "graph" OR caller requests visualization:
INVOKE present-to-knowledge-graph's `connect` to push ontology to frontend
3. COLLECT presented_output
4. Mark pipeline_state.致知_6_present = "done"
</action>
<output>presented_output (or skipped)</output>
</step_2_6>
</phase_2>
<phase_3 name="录 — Record">
<step_3_1>
<name>Write Pipeline Log</name>
<action>
1. COMPUTE pipeline_status from pipeline_state:
- ALL steps "done" or "skipped" → "success"
- ANY step "failed" → "partial"
- Phase 1 failed → "failed"
2. LOG pipeline execution summary to session working directory
3. DO NOT write to x-ipe-docs/dao/ — this is an orchestrator log, not a DAO decision log
</action>
<output>Pipeline log written</output>
</step_3_1>
</phase_3>
<phase_4 name="示 — Present">
<step_4_1>
<name>Return Pipeline Summary</name>
<action>
1. ASSEMBLE pipeline_output:
- pipeline_status: success | partial | failed
- pipeline_summary: human-readable description of what was done
- completed_steps[]: list of step names and their status
- stored_path: path to stored artifact (if applicable)
- ontology_result: ontology processing results (if applicable)
- presented_output: rendered content (if applicable)
2. RETURN pipeline_output to caller
</action>
<output>pipeline_output returned</output>
</step_4_1>
</phase_4>
</procedure>
Output Result
pipeline_output:
pipeline_status: "success | partial | failed"
pipeline_summary: "Human-readable pipeline execution summary"
completed_steps:
- { name: "discovery", status: "done" }
- { name: "classification", status: "done" }
- { name: "格物_1_framework", status: "done | skipped" }
- { name: "格物_2_overview", status: "done | skipped" }
- { name: "格物_3_rubric", status: "done | skipped" }
- { name: "格物_4_plan", status: "done | skipped" }
- { name: "致知_1_execute", status: "done | skipped | failed" }
- { name: "致知_2_fill", status: "done | skipped | failed" }
- { name: "致知_3_critique", status: "done | skipped | failed" }
- { name: "致知_4_ontology", status: "done | skipped | failed" }
- { name: "致知_5_store", status: "done | skipped | failed" }
- { name: "致知_6_present", status: "done | skipped | failed" }
stored_path: "x-ipe-docs/memory/{type}/{slug}.md"
ontology_result: { entities_created: 0, relations_created: 0, terms_normalized: 0 }
presented_output: { format: "markdown", content: "..." }
Definition of Done
<definition_of_done>
<checkpoint required="true">
<name>Skills Discovered</name>
<verification>discovered_skills registry is non-empty with categorized skills</verification>
</checkpoint>
<checkpoint required="true">
<name>Request Classified</name>
<verification>request_type resolved to one of: construction, extraction, ontology_only, presentation, storage, behavior_learning (or classification_failed)</verification>
</checkpoint>
<checkpoint required="true">
<name>Pipeline Executed</name>
<verification>All applicable steps executed (non-applicable steps marked "skipped")</verification>
</checkpoint>
<checkpoint required="true">
<name>Pipeline Summary Returned</name>
<verification>pipeline_output returned with pipeline_status, pipeline_summary, and completed_steps</verification>
</checkpoint>
</definition_of_done>
Error Handling
| Error | Cause | Resolution |
|---|
DISCOVERY_EMPTY | No x-ipe-knowledge-* skills found | Return pipeline_status: failed with error message listing expected skill location |
CLASSIFICATION_FAILED | Request doesn't match any known type | Return available capabilities list; caller can retry with request_type_override |
CONSTRUCTOR_UNAVAILABLE | Selected constructor skill missing or errored | Mark step as "failed", return partial result |
EXTRACTOR_TIMEOUT | Extractor skill timed out | Log timeout, pass partial gathered_knowledge to next step |
CRITIQUE_UNAVAILABLE | Critique sub-agent failed to respond | Skip critique loop, proceed with draft as-is, mark "partial_quality" |
STORE_FAILED | keeper-memory store operation failed | Log error, continue to present step; caller still gets presented_output |
References
Examples
See references/examples.md for detailed execution examples.
