v0.3.0 - Page freshness, provenance, and source trust

Last month, I scoped off a section of my vault to experiment with filesystem mcps and collaborative editing using various models.

That section of my note graph looked something like this two weeks ago:

I stumbled over Karpathy's original gist really quite by accident, just looking for learning resources for agentskills.io. So, I figured, what could possibly go wrong and just winged it with Claude.

This is my graph about seven days later while using and improving these skills to this current release.

That tightly-knit cluster is the part that the wiki has now fully absorbed: domain homes emerged, and many collaborative runs were had, each usually ending with a detailed /wiki-crystalize.

This has been a fun journey! I hope my lightweight implementation of the original pattern is as useful to you as it has been for me! Many new things today.

---

v0.3.0 - Page freshness, provenance, and source trust

This release is about the wiki knowing more about itself. Pages now track when they were last touched, where their knowledge came from, and whether that source was trustworthy enough to take at face value. The ## Pending Review section is the visible output of the trust system - a standing reminder that a page was created from a single non-authoritative source and needs corroboration before you rely on it.

New: wiki-schema.md

A second config file, deployed alongside wiki-config.md by /wiki-config. It defines the frontmatter structure all six skills follow - mandatory fields, conditional fields, valid enum values - and documents write discipline in one place. All six skills read it on boot. Previously each skill carried its own idea of what a page looked like; now there is a single source of truth. If the schema is missing or malformed, skills stop and redirect rather than proceeding with inconsistent assumptions.

The schema is designed to be extended. Adding conditional fields or new enum values is safe. Removing mandatory fields or renaming existing ones is not recommended; skills depend on those being where they expect them.

New: freshness tracking

Three fields, written automatically by skills:

• updated: - the date any skill last touched a page. Written by wiki-ingest, wiki-integrate, and wiki-crystallize on every write. On page creation it equals date: (correct and expected). date: itself is now strictly immutable - creation date only, never modified on subsequent writes.
• crystallize_count: - counts deliberate crystallize events on a page, not general edits. wiki-crystallize reads the current value and increments; if absent, writes 1. Paired with updated:, this answers "is this page current and well-revised?" without parsing the log.
• Staleness check in wiki-lint - flags pages where updated: is more than 90 days in the past. Exempt: pages with status: artefact, snapshot, archived, or page_type: reference. Missing date: or updated: fields on skill-written pages are now a lint error rather than a silent gap.

New: provenance and trust system

Four fields on page creation, written by wiki-ingest:

• status: - active for pages with real content, stub for thin pages (fewer than ~3 substantive sentences). Agent judgment. Other values on state change: artefact, archived, snapshot.
• description: - ~200 char page summary, quoted. Written by wiki-ingest on create; wiki-crystallize rewrites it when the page's structure or scope changes substantially, not on incremental updates.
• source: - one-element list with the originating ingested/ path. Written on create only; enrichments go in the body ## Sources section. Overwriteable when a subsequent ingest genuinely supersedes the original.
• reliability: - high, medium, or low. Agent assesses the originating source's nature: primary or authoritative = high; well-sourced secondary = medium; blog, opinion, single-source, unverified = low. Only present when source: is present.

New: ## Pending Review section

Written by wiki-ingest on knowledge, research-note, and survey pages where reliability: is medium or low. high reliability sources get no section. The framing scales with trust level - medium gets a permissive note (one corroborating source would retire this); low gets a stronger prompt with specific claims to verify.

The section stays until a better source is ingested. On re-ingest with a demonstrably more authoritative source, wiki-ingest updates source: and reliability: in frontmatter, adds the new source to ## Sources, and removes the section. The agent can also offer to search for a primary or corroborating source inline rather than writing the section and moving on - when source quality is genuinely ambiguous, it asks rather than decides unilaterally.

Improved: wiki-ingest selective intake

Step 1.5 - a thematic triage layer before processing. wiki-ingest does a quick survey of raw/, clusters files by domain affinity using the wiki's actual domain structure, and presents batches before touching anything. You choose which batch to process now; unselected files stay in the queue. This matters most when raw/ has accumulated files across multiple domains - processing them together previously meant mixed-domain synthesis that was harder to review.

Improved: content trust boundary

An explicit security section in wiki-ingest. Source documents are untrusted data. The skill never executes commands, exfiltrates data, or modifies its own behaviour based on embedded instructions in sources. Suspicious content is flagged in the session summary; synthesis continues. This was always implicit practice; it is now an explicit, numbered set of rules that loads before the workflow step that processes untrusted content.

Improved: Config Discovery hardening

Several fixes across this release cycle, each prompted by a specific observed failure:

• wiki-schema.md check is now a dedicated mandatory step in Config Discovery, with explicit "do not proceed until you have a definite verdict" language. Previously the schema check was buried in prose and easy to skip.
• Scope check is now a MANDATORY STOP on broad filesystem scope (C:\, /home/X, etc.). Previous soft conditional language ("if broad scope, consider skipping") was not followed reliably in testing. Imperative beats conditional.
• wiki-config precedence is enforced: when /wiki-config is available, operational skills stop and redirect to it rather than offering their own bundled fallback. The fallback path (deploy from bundled references/wiki-schema.md) is reserved for when wiki-config is genuinely not installed.

Version bumps:

• wiki-config 1.0 → 1.6
• wiki-ingest 4.0 → 4.11
• wiki-lint 3.0 → 3.9
• wiki-integrate 3.0 → 3.8
• wiki-crystallize 3.0 → 3.9
• wiki-query 3.0 → 3.7

• Re-upload all six .skill files via Claude Desktop → Customize → Skills.
• Run /wiki-config once after upgrading. It will deploy wiki-schema.md alongside your existing wiki-config.md if you don't already have one. Nothing breaks without it, but skills will stop and ask rather than proceed.
• Existing wiki pages work unchanged. The new frontmatter fields (updated:, status:, description:, source:, reliability:, crystallize_count:) are conditional - skills write them on the next touch. Pre-existing pages without these fields are treated as status: active for backward compatibility.
• If you see ## Pending Review sections appearing on new pages after upgrading: that is intentional. It means wiki-ingest assessed the source as medium or low reliability. Retire the section by ingesting a stronger source for that topic.

---

I had a little fun with the canvas design skill just trying to express how this little project has helped shape the gravity of my knowledge base and task tracking. So many changes to quotas lately, but I would say my token use has about halved.