Ingest raw source documents into the knowledge base. Select mode to control the ingestion method:

• add: Add a local file or content string (immutable, SHA-256 verified). Supports directory imports. Single images (<10MB) returned inline — you MUST immediately call wiki_write to describe them.

• fetch: Download a file from a URL into raw/ (arXiv abstract URLs auto-converted to PDF). Single images returned inline — you MUST immediately call wiki_write to describe them.

• import_confluence: Recursively import Confluence pages with attachments and hierarchy. Supports both Cloud (*.atlassian.net/wiki/...) and Server / Data Center ({host}/spaces/...). Defaults to reading the CONFLUENCE_API_TOKEN env var; pass auth_env to point at any other variable. Token format accepted: email:api-token (Cloud Basic), Bearer <pat> (explicit), or a bare PAT (Bearer prefix added automatically).

• import_jira: Import a Jira issue with comments, attachments, and linked issues. Supports both Cloud and Server / Data Center; auto-falls-back to REST API v2 on older Server / DC. Defaults to reading the JIRA_API_TOKEN env var; pass auth_env to point at any other variable. Token format accepted: email:api-token (Cloud Basic), Bearer <pat> (explicit), or a bare PAT (Bearer prefix added automatically).

List all versions of a raw file, sorted by version number, with the latest version marked. Given a base filename (e.g. 'report.xlsx'), returns all matching versions (report.xlsx as v1, report_v2.xlsx, report_v3.xlsx, etc.) with metadata and a 'latest' field pointing to the newest file.

List all raw source documents with metadata (path, source URL, download time, hash, size).

Report which raw/ files are not yet referenced by any wiki page. Answers 'what should I compile next?' — returns uncovered files sorted by recency/size, plus overall coverage ratio. Matches frontmatter 'sources' and inline 'raw/...' body references. Parsed artifacts (raw/parsed/) are excluded.

Read a raw source document's content and metadata. Raw files are immutable — this is read-only. Text/SVG files return content as string; document files (PDF, DOCX, XLSX, PPTX) have text extracted automatically; other binary files (images, etc.) return metadata only.

Pagination by format:

• PDF: use 'pages' for page ranges (e.g. '1-5')

• PPTX: use 'pages' for slide ranges (e.g. '1-10')

• XLSX: use 'sheet' to read a specific sheet; response always includes 'sheet_names'

• DOCX / text: use 'offset' + 'limit' for line-based pagination (default limit: 200)

For large documents, paginate rather than reading all at once.

Read one or more wiki pages. Single page: pass page. Multiple pages: pass pages array — reads all in one request, saving N-1 round trips. RECOMMENDED WORKFLOW: (1) call without 'section' to get the TOC for large pages, (2) use wiki_search 'section' field to jump directly to the relevant heading, (3) call with 'section' to read only that part. Large pages without 'section' are truncated at 200 lines — check 'truncated' and 'toc' in the response.

Create or update a wiki page. Content should include YAML frontmatter (title, type, tags, sources) and Markdown body. Timestamps (created/updated) are auto-managed. Auto-routes root-level pages to matching topic subdirectories (via frontmatter topic field or tag matching). Auto-links: scans body for mentions of existing page titles and injects [[slug|text]] links automatically (skips code blocks, existing links, URLs). Wiki pages are MUTABLE — they represent compiled knowledge that improves over time. Set return_content: true to include the final written content in the response — eliminates the follow-up wiki_read call in write-then-reference workflows.

Delete a wiki page. Cannot delete system pages (index.md, log.md, timeline.md).

List all wiki pages, optionally filtered by entity type or tag.

Full-text keyword search across all wiki pages. Returns paths, scores, and snippets sorted by relevance. Uses BM25 by default; switches to hybrid BM25+vector re-ranking when search.hybrid: true is set in .agent-wiki.yaml (requires wiki_admin action:rebuild to build the vector index first). Set include_content=true for simple inline content (with optional inline_budget cap). Set read_top_n to additionally read the top N unique matching pages (deduplicated) — enables combined search+read in one call; returns a pages array with full content and nextReads for unread matches. Use type or tags to narrow results without a separate wiki_list call. When no results are found, returns a knowledge_gap field with a suggested page slug, title, type, and tags — use it to decide what to create with wiki_write.

Wiki administration and maintenance. Select action to control behavior:

• init: Initialize a new knowledge base — creates wiki/, raw/, schemas/ directories and default templates.

• config: Show current workspace configuration: directories, lint settings, search settings, entity templates.

• format-check: Validate the portable OKF package manifest (agent-wiki.yaml), package inventory, and conformance findings against OKF v0.1.

• rebuild: Rebuild index.md, timeline.md, code knowledge graphs, and optionally the vector index (when search.hybrid is enabled). Set evidence_report=true to also regenerate the evidence report and persist it to wiki/evidence-report.md as part of the rebuild. Set okf_report=true to persist evidence/okf-report.json.

• lint: Run comprehensive health checks: contradictions, orphan pages, broken links, raw file integrity (SHA-256), synthesis page integrity. Set apply_fixes=true to auto-repair fixable issues.

• evidence-report: Aggregate evidence-first telemetry into a corpus-level Markdown report (source coverage, lineage diagnostics, 4-week write trend, Phase 2b readiness gates). Set write=true to also persist the report to wiki/evidence-report.md.

Execute multiple tool calls in a single request. Reduces tool-call count for LLM subscriptions that bill per-request (e.g. GitHub Copilot). Supports ANY combination of tools — e.g. read 5 wiki pages, write 3 pages, add 2 raw files, search, all in one call. Wiki index rebuild is automatically deduplicated (runs once at the end, not per-write). Each operation is independent — one failure does not abort the batch. Nested batch calls are not allowed.

Knowledge ingestion pipeline. Select mode to control behavior:

• batch: Batch import, extract, chunk, and pack source documents into digest packs. Scans a directory (or single file), imports to raw/, extracts text with structural provenance (per-page PDF, per-sheet XLSX, per-slide PPTX), chunks into fixed-line segments, then packs into markdown digest packs under raw/digest-packs/{topic}/. Files already in raw/ are skipped.

• digest_write: Write LLM-generated digest summaries to wiki with structured provenance. Creates or updates one or more wiki pages from digested content, linking back to source raw files and digest packs. Index is rebuilt once at the end. Each page gets auto-classified, auto-routed, and timestamped.

Parse a source file from raw/ into structured code knowledge (AST, normalized model, summary). Currently supports COBOL (.cbl, .cob, .cpy). Persists artifacts under raw/parsed/cobol/. Optionally traces all references to a variable.

Query parsed code knowledge. Select query_type to control behavior:

• trace_variable: Trace all references to a variable in a parsed source file — shows where it is read, written, or passed, grouped by section/paragraph.

• impact: Query the compiled knowledge graph for downstream impact — returns affected programs, copybooks, or datasets grouped by dependency depth, with evidence and uncertainty markers.

• procedure_flow: Query parsed procedure/section PERFORM flow for one source file — returns section-level and paragraph-level flow, optionally focused on one procedure.

• field_lineage: Query compiled field-lineage artifacts — returns deterministic and inferred matches for one field, optionally narrowed to a copybook or qualified name.

• dataflow_edges: Query field-level dataflow edges — MOVE/COMPUTE/ADD assignment, EXEC SQL host-variable, and CALL USING parameter edges. Filter by from/to, or set field+transitive: true to follow chains across the full graph.