Journal — 2026-03-08

2026-03-08 — Federation Phase 1 Shipped

mark_index and mark_resolve MCP tools merged. Agents can now crawl servers to build hash indexes and resolve content by hash across the federation.

Key features:

• mark_index: crawls source server, publishes hash index to hub, manifest checks (hard block + force override), dry_run, merge with existing aggregated indexes, truncation warnings at 1000 docs
• mark_resolve: looks up hash in hub index, tries each server, verifies content-hash on response
• client/internal/index package: Parse, Build, Merge with server URL canonicalization

PR review hardened the implementation:

• Sentinel error for truncated indexes (visible warning instead of silent)
• Server URL canonicalization (trailing slash, default port, case)
• Negative expected_version validation
• Source path used as crawl root (subtree indexing)
• Hash validation on crawled content-hash values
• Fail on existing index fetch failure (don't silently drop other servers' entries)
• Aggregated index uses target URL in Source header
• Extracted checkManifests to fix cyclomatic complexity
• Tool parameter descriptions use urlDesc() for bare path support

"The Agent as the Librarian" — core architectural principle documented. Server is a bookshelf, agent is the librarian.

2026-03-08 — Persistent Graph Store + Backlinks

Shipped the persistent graph store and backlinks — Phase 4's first functional increment after the Hub pattern.

What Shipped

• client/internal/graphstore package — new persistent graph store at ~/.mark/graph.json. StoredNode tracks URL, title, status, link count, etag, and crawl timestamp. StoredEdge tracks directed links. Atomic writes (.tmp + os.Rename), schema versioning (version: 1), incremental merge with deduplication.
• EtagFetcher — shared adapter implementing graph.Fetcher that collects etags during crawl. Replaces duplicated etag collection in MCP and TUI.
• CrawlAndPersist — unified method on *Store that runs graph.Crawl, merges results + etags, and saves. Nil-safe (nil store = crawl without persist). All three clients (CLI, TUI, MCP) now share this single code path.
• mark_backlinks MCP tool — reverse edge lookup from the persistent graph. "What links here?" Returns sorted markdown list with document titles.
• TUI graph seeding — graph view loads instantly from stored graph while background crawl runs. No more blank screen on graph toggle.
• CLI persistence — demarkus graph now also persists to the store (was missing before).
• 10 unit tests — Load, Save, Merge, Backlinks, ToGraph, CrawlAndPersist, nil-safety, atomic writes.

The Sharing Evolution

This increment went through three rounds of consolidation at Fritz's push for cohesive client design:

1. First pass: Duplicated etag collection + merge/save in both MCP and TUI
2. Second pass: Extracted EtagFetcher into graphstore, shared by MCP and TUI
3. Third pass: Extracted CrawlAndPersist into graphstore, shared by all three clients

The principle: CLI, TUI, and MCP are all clients. Core fundamentals must be shared. Differences are fine when unavoidable, but the default is elegant reuse.

Also This Session

• Dropped native Windows builds — removed windows from goos in both .goreleaser.yml files. WSL users run Linux binaries; native Windows builds were untested overhead.
• Updated roadmap — Phase 3 marked COMPLETE (federation was already shipped but roadmap hadn't been updated). Phase 4 now shows persistent graph, backlinks, CrawlAndPersist, and graph seeding as Done.

What's Next

• Graph as content — export the crawled graph as a markdown document with mark:// links, publish it to a server, others fetch and import it
• Graph-aware navigation — TUI shows explored topology, graph proximity for related docs
• Agent discovery — agents crawl the graph to build knowledge maps

Note to Future Sessions

CrawlAndPersist is the entry point for all graph crawling. Don't bypass it — it handles etag collection, merge, and atomic save. If the store is nil, it still works (just doesn't persist). The EtagFetcher is also in graphstore, not in the graph package — it's a persistence concern, not a graph concern.

2026-03-08 — Persistent Graph: Copilot Review Hardening

Post-merge review pass on the persistent graph store. Copilot raised ~15 comments; about half were valid fixes, the rest were over-engineering or duplicates.

Fixes Applied

1. Data race in CrawlAndPersist — nodeCount for MaxNodes cap was a plain int incremented from concurrent crawler goroutines. Fixed with atomic.Int32. Race detector confirmed clean. Test's OnNode callback had the same race — fixed too.
2. Error wrapping in Load — read/parse errors returned raw without file path context. Wrapped with fmt.Errorf("read graph store %q: %w", ...) and "parse graph store %q: %w", matching the bookmarks store pattern.
3. Schema version validation — schemaVersion constant existed but was never checked on load. Added doc.Version != schemaVersion guard with clear error message.
4. MaxDepth semantics mismatch — graphstore.CrawlOptions documented 0 = default 2 but graph.CrawlOptions uses 0 = start node only. Initially added a 0→-1 translation, then Copilot correctly pointed out this breaks -depth 0 from CLI. Reverted to pass-through with corrected comment: 0 = start node only, -1 = default 2.
5. Graph store load errors surfaced — CLI now warns to stderr, TUI shows in status bar (preserves both bookmark + graph errors with | separator), MCP logs with log.Printf.
6. MCP handler holds graphStore field — instead of loading from disk on every mark_graph/mark_backlinks call, store is loaded once at startup and shared. Enables testability.
7. Happy-path backlinks test — added TestHandlerMarkBacklinks_HappyPath with pre-populated temp store, asserting returned titles and URLs. Also TestHandlerMarkBacklinks_NilStore.
8. Tool description accuracy — mark_graph description now says "When a local graph store is available" instead of unconditionally claiming persistence.
9. Displaced comment — assertIsToolError doc comment had drifted above a test function after insertion. Moved back.

Copilot Comments Rejected

• Loop variable aliasing (×2) — Copilot claimed n := doc.Nodes[i]; &n aliases all entries. Wrong: := creates a copy, and Go 1.22+ scopes loop vars per iteration.
• Windows os.Rename doesn't overwrite — not applicable, we dropped Windows native builds this session.
• URL validation in backlinks — "no backlinks found" is the correct answer for any URL not in the graph, regardless of format.
• Flat tree from store — valid observation (stored nodes have depth 0), but intentional trade-off. Flat list is still useful as instant preview; crawl provides depth.

What I Learned

The data race was the most important catch. The nodeCount++ inside OnNode was called from multiple crawler goroutines — a real -race failure, not theoretical. atomic.Int32 is the right fix: no mutex needed, the counter is independent of other state.

The MaxDepth saga was instructive: Copilot's first comment (translate 0→-1) was right about the documentation mismatch but wrong about the fix. Its second comment (revert the translation) was right about the fix. The lesson: when two layers have different zero-value semantics, the wrapper should either document and pass through, or fully own the translation. Half-translating creates a third set of semantics that nobody understands.

Note to Future Sessions

Graph store is now hardened: race-free, error-wrapped, version-checked, properly surfaced in all three clients. The MCP handler holds the store as a field — tests can inject pre-populated stores directly.

2026-03-08 — Graph as Content & Documentation Updates

Graph as Content (Phase 4)

Implemented graph export as publishable markdown — the graph becomes a document you can publish to a demarkus server and others can crawl to discover the topology.

New files:

• client/internal/graphstore/export.go — Export() renders the store as markdown with mark:// links in tables; ParseExport() parses it back
• client/internal/graphstore/export_test.go — 4 tests: export, empty, parse, round-trip (with escaped pipes and backslashes)

CLI: demarkus graph export [-o file.md] — exports stored graph to stdout or file MCP: mark_graph_export — returns graph as markdown for agents to publish

Design insight: No special import function needed. Crawling the exported doc discovers all the mark:// links naturally — the same link extraction the crawler uses parses the graph document. This matches the DESIGN.md vision.

Copilot review fixes:

• Cell escaping for titles containing | and \ (escape backslashes first, unescape in reverse order)
• Regex captures link destination (url) not display text [label] for canonical URL
• Reject extra positional args in graph export subcommand
• Rejected: strconv.Atoi overflow (regex constrains to digits), URL special chars (mark:// doesn't use them)

Documentation Updates

Updated 5 in-repo docs on main to document persistent graph store and backlinks:

• README.md — tool descriptions mention persistent graph and backlink queries
• docs/DESIGN.md — Phase 4 implementation status block
• docs/site/client/index.md — graph persistence, MCP tools list
• docs/site/architecture/index.md — expanded Document Graph section
• docs/site/philosophy/agent-cookbook.md — new backlinks discovery recipe

2026-03-08 — Phase 4: Graph-Aware Navigation Complete

Shipped the last Phase 4 feature: graph-aware navigation in the TUI.

What shipped:

• Three sub-views in graph mode: Links (d), Backlinks (r), Topology (t)
• Link density indicators [N←] showing backlink counts per node
• Topology view: all explored nodes sorted by importance (backlink count descending)
• graphstore.BacklinksEnriched() — shared enrichment logic used by both TUI and MCP
• graph.InDegrees() — O(E) single-pass in-degree computation, replaces per-node O(N*E) lookups
• Rune-safe truncation (truncateRunes()) for multi-byte UTF-8 status icons
• Null-safe sub-view switching during active crawls

Design decisions:

• Removed Store.InDegrees() to avoid duplicating graph.Graph.InDegrees() — callers use store.ToGraph().InDegrees() instead (topology view is user-triggered, not a hot path)
• graph.InDegrees() guards against phantom nodes (edges to URLs never added as nodes, e.g. from cancelled crawls)
• AllNodes() exists on both Graph and Store but returns different types (*Node vs StoredNode) — not duplication

Phase 4 marked COMPLETE on roadmap. Remaining item: Agent discovery (not started).