jetxu-llm/docmason/knowledge-construction

Knowledge Construction

Use this skill when docmason sync has prepared staged evidence and the agent must write knowledge.json, summary.md, or additive semantic_overlay/<unit-id>.json sidecars.

This is an internal follow-on workflow behind knowledge-base-sync. In the normal path, the main agent hands this workflow a bounded governed follow-up packet from the latest sync result before any source-level authoring starts. Ordinary users should not need to invoke it by name.

Required Capabilities

• local file access
• shell or command execution
• ability to inspect rendered images and extracted artifacts

If the agent cannot inspect rendered images, stop and explain that the environment is not capable enough for multimodal knowledge construction.

Procedure

1. If the current sync result or workflow handoff provides a governed follow-up packet path such as lane_b_follow_up.work_path, open that packet first.

• prefer the bounded packet path handed off by knowledge-base-sync
• treat that packet as the authoritative source and unit selection scope for this pass
• do not scan shared-job directories to invent a job id when the bounded packet path is already available

1. Read knowledge_base/staging/pending_work.json.
2. If knowledge_base/staging/hybrid_work.json exists, treat that file as the authoritative hard-artifact overlay queue inside the current bounded scope.
3. Work only on the staged items selected by the governed packet when one exists; otherwise work only on the staged items listed in pending_work.json. Treat each pending source or interaction memory as an independent bounded write scope.

• when both a governed packet and hybrid_work.json are present, use the packet's selected sources and units as the outer boundary and the queued hybrid targets as the inner overlay queue

1. For each assigned pending item, open:

• work_item.json
• source_manifest.json
• evidence_manifest.json
• knowledge_base/staging/hybrid_work.json when the current sync result reported candidate-prepared
• artifact_index.json when present
• pdf_document.json when present
• spreadsheet_workbook.json when present
• spreadsheet_sheet/*.json when present
• visual_layout/*.json when present
• derived_affordances.json when it already exists
• extracted text and structure files
• rendered assets referenced by the evidence manifest
• when present, the staged interaction-specific context file such as interaction_context.json

1. Build the source semantics from the richest published evidence available instead of defaulting to flattened text:

• for spreadsheets, prefer workbook, sheet, table, chart, metric, dimension, time-axis, hidden-sheet, and formula summaries over raw cell dumps
• for PDF and PPTX, prefer section paths, captions, continuation links, procedure spans, region roles, charts, tables, pictures, connectors, groups, and major regions over page-level text alone
• when the real support is artifact-level, include artifact_id in the citation instead of citing only the parent unit_id

1. Write knowledge.json with the required bilingual fields and only cite real evidence-unit IDs from the matching source.
2. Write summary.md with:

• # <title>
• a line that mentions the source ID
• ## English Summary
• ## Source-Language Summary

1. When the staged source includes high-value hybrid candidates and the environment can inspect renders, write additive semantic_overlay/<unit-id>.json sidecars only for units that are inside the current bounded scope and queued in hybrid_work.json.

• prefer overlay work where deterministic structure is already rich but cross-region or multimodal semantics are still missing
• keep the hard-artifact boundary intact:
• use the queued target_artifact_ids
• use the queued target_focus_render_assets first
• use the queued target_render_assets and target_render_page_span
• for image-only or scanned PDF pages, treat the published page-image artifact as the first-class target instead of pretending the text layer is enough
• if the baseline focus render is still not legible enough, use the targeted hi-res focus-render helper for that artifact instead of rerendering the whole source
• overlays must remain additive
• do not rewrite deterministic sidecars such as artifact_index.json, visual_layout/*.json, spreadsheet_*, or pdf_document.json
• bind overlay claims to consumed inputs, artifact ids when available, explicit uncertainty notes, and the current freshness contract:
• origin
• source_fingerprint
• unit_evidence_fingerprint
• covered_slots
• blocked_slots

1. Avoid placeholders, speculative citations, and unsupported related-source links.
2. Treat derived_affordances.json as a published sidecar rather than scratch output.

• the baseline affordance sidecar is generated deterministically by the repo
• if you enrich it, keep descriptors compact, evidence-backed, grouped by channel, and explicitly derived rather than source-authored fact

1. When evidence is weak, say so explicitly in known_gaps, ambiguities, confidence notes, or overlay uncertainty notes instead of inventing certainty.
2. After all assigned staged sources are complete, return control to the main agent so it can rerun docmason sync --json or docmason validate-kb --json.

Escalation Rules

• If a staged source or interaction memory requires render inspection and the environment cannot inspect renders, stop that item and report the blocker directly.
• If a cross-source relation is uncertain, omit it rather than guessing.
• If a chart, table, diagram, or region claim cannot be supported by the published artifacts, write the uncertainty explicitly rather than laundering it through a source-level summary.
• If hybrid_work.json says a source is candidate-prepared, do not declare the governed follow-up complete merely because deterministic publication already succeeded.
• Do not publish, validate, or sign off the final sync result from inside this workflow. That judgment belongs to the main agent.

Completion Signal

• The workflow is complete when every assigned staged source in the current bounded scope has updated knowledge.json and summary.md, or when a concrete capability blocker has been surfaced to the main agent.

Notes

• summary_source may equal summary_en when the source itself is English.
• related_sources should stay light in Phase 3. Only add a relation when the source evidence clearly supports it.
• semantic_overlay is for cross-region or multimodal semantics that add value beyond the deterministic substrate. It is not a second primary truth surface.
• If extraction is weak, record that weakness explicitly in known_gaps or ambiguities. Do not invent certainty.
• Interaction memories should remain explicit about lower trust tier, interaction-derived provenance, and the distinction from authored source documents.