feat(wiki): align with full Karpathy LLM Wiki pattern

The previous implementation was a truncated version of the pattern:
one-source-one-wiki-page with no entity pages, concept pages, index
catalog, or log. That's glorified RAG, not a compounding knowledge
base. User pointed at Karpathy's idea file and asked for the real thing.

Changes

  AGENTS.md — rewritten end-to-end
    * Three layers now named explicitly: raw / wiki / schema
    * Four page categories:
        research/wiki/sources/     one per raw file
        research/wiki/entities/    one per person/method/dataset/concept
                                   (accretes across sources)
        research/wiki/concepts/    higher-level themes
        research/wiki/queries/     filed-back substantive Q&A
    * research/wiki/index.md — content-oriented catalog, not chronological
    * research/wiki/log.md    — append-only chronological record with a
                                grep-friendly `## [YYYY-MM-DD] <op> | ...`
                                prefix convention
    * Citation format unchanged: `[src:<raw_filename>]`
    * Page template extended with frontmatter + See also + Conflicts
    * 10 hard rules now include: one ingest touches many pages, use
      edit_file for in-place updates, reciprocal cross-refs, log every
      substantive op

  Skills
    * wiki-curator/SKILL.md — rewritten. Canonical multi-page ingest
      flow: source summary → entity pages → concept pages → index.md →
      log.md → manifest flip. Explicit warning: if you only touched
      1-2 pages, you missed the point — redo it.
    * wiki-linter/SKILL.md — NEW. Wiki health check: unresolved
      contradictions, stale claims, orphan pages, missing cross-refs,
      missing entity/concept pages, data gaps, 2-5 suggested follow-up
      questions. Appends to log.md.
    * thesis-writer/SKILL.md — updated. Navigation order now
      index.md → concept → entity → source → raw (drill only when
      needed). Adds a file-back step: substantive synthesis from a
      drafting session gets filed to research/wiki/queries/ so the
      reasoning compounds instead of disappearing into chat.

  Agent main system prompt
    * Names the wiki subdirectories explicitly.
    * Instructs file-back on substantial chat syntheses.
    * Reinforces edit_file vs write_file (the latter refuses to
      overwrite by design).

  CLI
    * _ensure_workspace: pre-creates sources/ entities/ concepts/
      queries/ and seeds research/wiki/index.md + log.md with the
      expected section headers. First ingest no longer has to mkdir
      anything or guess at structure.
    * `thesis curate`: prompt updated — spells out the multi-page
      invariant and the log prefix format.
    * `thesis lint`:
        - no args  → wiki-linter (new default)
        - a file   → citation-linter on that chapter
        - --citations  → citation-linter on all chapters (old default)

Tests (15 new, 312 total passing)
  * Schema tests for:
      - All 6 skills present (was 5)
      - AGENTS.md declares the 4 page categories
      - AGENTS.md declares index.md + log.md + grep-friendly prefix
      - AGENTS.md states the "one ingest, many pages" rule
  * CLI tests split into three lint paths:
      - test_lint_defaults_to_wiki_linter
      - test_lint_specific_file_uses_citation_linter
      - test_lint_citations_flag_forces_citation_linter
  * End-to-end scripted-LLM tests:
      - test_multi_page_ingest_produces_all_categories: the canonical
        ingest produces pages in all 4 categories + updates index.md +
        appends log.md + flips manifest. Verifies the file structure
        at disk level, not just that the agent responded.
      - test_init_creates_full_wiki_layout: workspace scaffold seeds
        every wiki subdirectory and the two special files with the
        right section headers.

Ruff clean. Symlink sandbox test still skipped on Windows.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: NikitasT2003

AGENTS.md

@@ -1,11 +1,14 @@
 ---
 name: thesis-writer
 description: >
-  Use this skill when the user asks you to draft, write, or expand a thesis
-  section, chapter, paragraph, or the whole thesis. Enforces the grounding
-  rule (every factual claim cites an indexed source), loads the user's writing
-  style from `style/STYLE.md`, navigates the wiki for supporting material, and
-  writes to `thesis/chapters/<NN>.md` matching `thesis/outline.md` numbering.
+  Use this skill to draft, write, or expand a thesis section, chapter,
+  paragraph, or the whole thesis. Enforces the grounding rule (every
+  factual claim cites an indexed source), loads the user's writing style
+  from `style/STYLE.md`, navigates the LLM-Wiki-pattern knowledge base
+  (index.md → concept pages → entity pages → source pages → raw), and
+  writes to `thesis/chapters/<NN>.md` matching `thesis/outline.md`
+  numbering. When the drafting surfaced substantive synthesis worth
+  keeping, also files a query/synthesis page back into the wiki.
   Triggers "draft", "write section", "expand outline", "write chapter",
   "flesh out 2.1", "continue writing".
 ---
@@ -14,29 +17,66 @@ description: >
 
 ## Procedure (non-negotiable order)
 
-1. **Style check.** Read `style/STYLE.md`. If it does not exist, STOP and tell the user: "Run `thesis style` first to compile your writing-style guide from `style/samples/`." Do not draft without a style guide.
+1. **Style check.** Read `style/STYLE.md`. If missing, STOP and tell the
+   user: "Run `thesis style` first to compile your writing-style guide from
+   `style/samples/`." Do not draft without a style guide.
 
-2. **Outline check.** Read `thesis/outline.md`. Find the section the user asked for. If missing, ask the user to add it to the outline first.
+2. **Outline check.** Read `thesis/outline.md`. Locate the requested section.
+   If absent, ask the user to add it to the outline before drafting.
 
-3. **Wiki navigation.** Read `research/wiki/index.md`. Identify pages relevant to the section (match on tags, entity names, section keywords). Read each relevant wiki page.
+3. **Wiki navigation** (do this in order; do NOT jump straight to raw files):
+   a. Read `research/wiki/index.md` and identify the concept pages relevant
+      to the section.
+   b. Read those concept pages. Follow their `See also` links to
+      entity pages.
+   c. Read the entity pages. Follow their `See also` links to source-summary
+      pages where you need the source's exact wording.
+   d. Drill into `research/raw/<file>.md` ONLY when a source page lacks a
+      specific detail. Prefer wiki-first.
 
-4. **Drill-down (only if needed).** If a wiki page's summary/key-claims do not cover a point you need, read the corresponding `research/raw/<file>.md` for verbatim context. Prefer wiki-first.
-
-5. **Delegate to drafter.** Hand off to the `drafter` subagent with:
+4. **Delegate to the `drafter` subagent** with:
    - Section identifier + outline excerpt.
-   - The list of wiki pages that will ground the draft.
-   - A reminder to follow `style/STYLE.md`.
+   - The list of wiki pages (concept → entity → source) you consulted.
+   - A reminder to follow `style/STYLE.md` exactly.
+
+5. **Drafter writes** to `thesis/chapters/<NN>.md` (where `NN` matches the
+   outline numbering). If the chapter file already exists, extend it —
+   use `edit_file`, not `write_file` (which refuses to overwrite).
 
-6. **Drafter writes** `thesis/chapters/<NN>.md` where `NN` matches the outline's numbering (e.g. section 2.1 → chapter 02). If the chapter file already exists, extend it — do not overwrite silently.
+6. **Self-lint before returning.** Every paragraph making a factual
+   statement has at least one `[src:<raw_filename>]`. Transitions and
+   author commentary without facts are fine.
 
-7. **Self-lint before returning.** Every paragraph with a factual statement has at least one `[src:<filename>]`. Grab-bag transitions without factual content are fine.
+7. **File-back check.** If the drafting required non-trivial synthesis —
+   a cross-source comparison, a new connection, a causal argument the
+   sources didn't state outright — file that reasoning back into the wiki
+   as a query/synthesis page:
+   `research/wiki/queries/<YYYY-MM-DD>-<slug>.md`. Why: the reasoning
+   compounds if it's in the wiki, but disappears if it's only in a chapter
+   (where future questions can't reach it easily). Update `index.md` under
+   Queries and append to `log.md`:
+   ```
+   ## [YYYY-MM-DD] query | wrote section <N.M>
+   - Filed: [[queries/YYYY-MM-DD-<slug>]]
+   - Pages read: [[…]], [[…]]
+   ```
+   For routine drafting that just restates what sources already say, skip
+   this step.
 
 ## Hard rules
-- **Never write a factual claim without `[src:<raw_filename>]`.** If wiki + raw don't support it, output "no grounding in indexed sources — add material or remove claim" and stop.
-- **Never use pretraining knowledge** for facts, statistics, quotes, author attributions, or dates. The indexed sources are the only truth.
-- **Never fabricate** a source filename. Every `[src:X]` must correspond to an actual file under `research/raw/`.
-- **Match the style guide**: sentence length, hedging, jargon density, POV, transitions. The drafter should re-read `STYLE.md` before every handoff.
-- **Respect write scope**: drafter writes only under `thesis/**`.
+- **Never a factual claim without `[src:<raw_filename>]`.** Refuse with
+  "no grounding in indexed sources — add material or remove claim."
+- **No pretraining facts, no web.** The wiki is the truth.
+- **Never fabricate** a source filename. Every `[src:X]` corresponds to
+  an actual file in `research/raw/`.
+- **Match `STYLE.md`** — sentence length, hedging, POV, citation placement,
+  transitions. Re-read it at the start of each drafting turn.
+- **Respect scope**: drafter writes only under `thesis/**`; file-back
+  queries go under `research/wiki/queries/**`.
+- **Use `edit_file` to extend** existing chapters. `write_file` refuses
+  to overwrite.
 
 ## Output
-After drafting, report to the user: file written, approximate word count, and a 2-line summary of what you drafted. Do not paste the full chapter back into chat unless asked.
+Report back: the chapter file written, approximate word count, number of
+citations, and (if filed) the query page path. Do not paste the full chapter
+into chat unless asked.

skills/thesis-writer/SKILL.md

@@ -1,40 +1,112 @@
 ---
 name: wiki-curator
 description: >
-  Use this skill to build or update the research wiki from newly-ingested raw
-  sources following the Karpathy LLM Wiki pattern, one page per source file.
-  Read the template from AGENTS.md and produce `research/wiki/<source_stem>.md`
-  for every entry in `research/raw/_index.json` with status "pending", update
-  `research/wiki/index.md`, flag contradictions (flag only, no merging), then
-  flip status to "curated". Triggers include "curate", "build wiki",
-  "compile wiki", "update the wiki", "after ingest".
+  Use this skill to curate newly-ingested raw sources into the LLM Wiki
+  following AGENTS.md. A single ingest must touch multiple pages —
+  create/update the source-summary page, every relevant entity page,
+  every relevant concept page, update `index.md`, append to `log.md`,
+  flag contradictions, and flip the manifest status. One source typically
+  touches 5-15 pages. Triggers include "curate", "build wiki",
+  "compile wiki", "ingest into wiki", "update the wiki", "after ingest".
 ---
 
 # wiki-curator
 
 ## Goal
-Turn normalised raw sources into a compound, navigable knowledge base. The wiki is the retrieval layer for thesis writing — quality here determines drafting quality downstream.
+Turn normalised raw sources into a persistent, compounding knowledge base
+per the Karpathy LLM Wiki pattern. **The knowledge must accumulate across
+entity and concept pages — not sit in isolated one-per-source pages.**
 
-## Procedure (follow exactly)
+## Procedure (follow for every pending source)
 
-1. Read `research/raw/_index.json`. Collect every entry with `status == "pending"`.
-2. For each pending entry, in order:
-   a. Read the raw normalised markdown at `research/raw/<entry.filename>`.
-   b. Compose a wiki page at `research/wiki/<entry.filename>.md` (same stem + `.md`), following the template in AGENTS.md verbatim. Required sections: frontmatter (source, title, authors, date, tags), Summary, Key claims, Methods/datasets (if empirical), Key terms, See also, Conflicts.
-   c. Every claim on the page MUST carry `[src:<entry.orig>]`.
-   d. Update `research/wiki/index.md`: add a line under the appropriate topic tag cluster pointing to the new page (`- [[<stem>]] — <one-line teaser>`). Create the cluster if it doesn't exist yet.
-   e. Contradiction check: for each key claim, glob existing `research/wiki/*.md` and check if any previously-curated page asserts a conflicting claim on the same topic. If so, append `- ⚠ conflicts with [[<other_stem>]] on <topic>` to the `Conflicts` section of BOTH pages. Flag only — do not quote sides, do not merge.
-   f. Update `research/raw/_index.json`: set `entry.status = "curated"` and append `<entry.filename>.md` to `entry.curated_pages`.
-3. After all pending entries are processed, write a one-line summary to the user: `curated N sources (M conflicts flagged, K cross-refs added)`.
+1. Read `research/raw/_index.json`. Collect entries where `status == "pending"`.
 
-## Tagging guidance
-Tags should be short (1-2 words), lowercase, domain-specific. Reuse existing tags from `index.md` when possible; add new ones sparingly. Aim for 3–8 tags per page.
+2. For each pending entry, read the raw normalised markdown at
+   `research/raw/<entry.filename>`.
 
-## Cross-references (`See also`)
-Add a cross-reference when two sources discuss the same entity, method, or claim. Cross-references must be reciprocal: edit BOTH pages.
+3. **Write the source-summary page** at
+   `research/wiki/sources/<entry.filename>.md` following the AGENTS.md
+   page template. Sections: frontmatter, Summary, Key claims, Methods/datasets,
+   Key terms, See also, Conflicts. Every claim cites `[src:<entry.orig>]`.
+
+4. **Identify entities** in the source: named people, specific methods,
+   datasets, well-defined concepts, tools, systems. For each:
+   - Use `read_file` on `research/wiki/entities/<slug>.md` (via `ls` first to
+     check existence).
+   - If the page exists, **extend it**: use `edit_file` to append new claims
+     from this source, each carrying `[src:<entry.orig>]`. Add a reciprocal
+     `See also: [[sources/<entry.filename>]]` link.
+   - If it doesn't exist, **create it** with the new claims as seed content.
+     Include a `See also: [[sources/<entry.filename>]]` link back.
+
+5. **Identify major themes/concepts** the source touches (usually 2-5).
+   For each, same pattern as entities but under
+   `research/wiki/concepts/<slug>.md`.
+
+6. **Update `research/wiki/index.md`**:
+   - Add the new source under its topic cluster.
+   - List any new entity pages under `## Entities`.
+   - List any new concept pages under `## Concepts`.
+   - Preserve existing entries. Use `edit_file`.
+
+7. **Check for contradictions**: for each new claim, glob existing
+   entity/concept pages and `grep` for conflicting claims on the same topic.
+   On match, add `⚠ conflicts with [[<other>]] on <topic>` to the `Conflicts`
+   section of **both** pages. **Flag only** — do not quote both sides, do not
+   merge.
+
+8. **Append to `research/wiki/log.md`** using the exact grep-friendly prefix:
+   ```
+   ## [YYYY-MM-DD] ingest | <Source Title>
+   - New pages: [[sources/…]], [[entities/…]], [[concepts/…]]
+   - Updated pages: [[entities/…]] (+N claims), …
+   - Conflicts flagged: N
+   ```
+   Use `edit_file` to append (read current log, add new entry at the end).
+
+9. **Flip manifest status**: `edit_file` on `research/raw/_index.json` to
+   change `"status": "pending"` to `"status": "curated"` for the entry, and
+   populate `curated_pages` with the list of wiki pages you touched for that
+   source.
+
+10. After all pending entries processed, report:
+    `curated N sources, touched M pages, flagged K conflicts, suggested J questions`.
+
+## What a good ingest looks like
+
+For a single paper on, say, "attention mechanisms in transformers":
+- 1 source page: `sources/attention-is-all-you-need.pdf.md`
+- 3-5 entity pages (created or updated): `entities/transformer-architecture.md`,
+  `entities/self-attention.md`, `entities/multi-head-attention.md`,
+  `entities/vaswani.md` (author), `entities/machine-translation.md`
+- 1-2 concept pages: `concepts/attention-mechanisms.md`,
+  `concepts/sequence-to-sequence-models.md`
+- `index.md` (updated)
+- `log.md` (appended)
+
+**Total: 7-10 pages touched.** If you only wrote 1-2 pages, you're not
+building a wiki — you're making isolated summaries.
+
+## Slug convention (AGENTS.md §Slug)
+
+`<lowercase-kebab-case>`, under 60 chars, strip punctuation. Examples:
+- "Self-Supervised Learning" → `self-supervised-learning`
+- "GPT-4" → `gpt-4`
 
 ## What NOT to do
-- Do NOT write to `research/raw/**`. Those files are immutable.
-- Do NOT merge conflicting pages. Flag and move on; the user resolves.
+
+- Do NOT write to `research/raw/**` (sources are immutable). The exception is
+  updating `research/raw/_index.json` via `edit_file`.
+- Do NOT use `write_file` to overwrite an existing wiki page. It will refuse.
+  Use `edit_file` for in-place updates.
+- Do NOT merge conflicting claims. Flag on both pages and move on.
 - Do NOT invent authors, dates, or tags not supported by the source.
-- Do NOT skip the status flip — the manifest is how the user knows curation is done.
+- Do NOT stop after the source page. The value of this pattern is that
+  entity and concept pages compound — that's what the wiki is for.
+- Do NOT skip the `log.md` append. Every ingest must leave a chronological
+  trace.
+
+## Reciprocal cross-refs
+
+When you add `See also: [[X]]` on page A, also add `See also: [[A]]` on page
+X. Use `edit_file` on both. This is not optional — the graph breaks otherwise.

skills/wiki-curator/SKILL.md

@@ -0,0 +1,128 @@
+---
+name: wiki-linter
+description: >
+  Use this skill to health-check the LLM Wiki under `research/wiki/`.
+  Scans every page and reports: unresolved contradictions, stale claims
+  newer sources contradict, orphan pages with no inbound links, pages
+  that reference an entity without linking its page, entities/concepts
+  mentioned in source summaries but lacking their own page, thinly-covered
+  topics that would benefit from more sources, and 2-5 suggested follow-up
+  questions. Appends a `lint` entry to `log.md`. Triggers include
+  "lint wiki", "check wiki", "wiki health", "audit the wiki",
+  "review the knowledge base", "find orphans", "find gaps".
+---
+
+# wiki-linter
+
+## Goal
+Keep the wiki healthy as it grows. Humans give up on wikis because the
+maintenance burden outpaces the value. Your job is to keep that burden near
+zero by catching the things that silently break a compound knowledge base:
+orphan pages, stale claims, broken cross-refs, data gaps.
+
+## Procedure
+
+1. Read `research/wiki/index.md` and `research/wiki/log.md` to orient.
+
+2. Enumerate pages: `glob("research/wiki/**/*.md")`.
+
+3. For each category below, scan + build a list of issues. Numbered, so the
+   user can act on them one at a time.
+
+### Contradictions (unresolved)
+Glob for `⚠ conflicts with` markers. Report file + line + the conflicting
+pages. **Do not resolve** — surface them for the user to decide.
+
+### Stale claims
+An entity page may assert claim X from source A, and a later source B may
+contradict it. If neither page flagged the conflict, the old claim is stale.
+Strategy: read each entity page's `Key claims` section; for each claim, grep
+other pages that cite the same topic; if wording clearly disagrees, flag
+`⚠ possibly stale` on the older claim (and record both pages in the report).
+
+### Orphan pages
+A page is an orphan if no other page links to it via `[[...]]`. Expected
+orphans: `index.md`, `log.md`. Everything else should have at least one
+inbound link. Report orphan slugs + suggest where they should be linked from
+(usually `index.md` or the concept page closest to their topic).
+
+### Missing cross-references
+If page A mentions entity E (by name) but doesn't link `[[entities/<e-slug>]]`,
+flag it. Heuristic: for each entity page, grep wiki for its title (and common
+aliases listed in frontmatter); any hit without a `[[...]]` link is a missing
+cross-ref.
+
+### Missing entity / concept pages
+Scan source-summary pages for capitalized named entities and key terms listed
+in their `Key terms` section. For each, check if an entity page exists. If
+not, flag `missing entity page: <name>` and suggest the slug.
+
+### Data gaps (thinly covered topics)
+Count how many sources each entity / concept page cites. Flag any
+entity/concept page with fewer than 2 citing sources as a potential data
+gap — the user may want to find more reading on it. Report the slug and the
+current citation count.
+
+### Suggested follow-up questions
+End the report with 2-5 questions the user could ask next, generated from
+the wiki's current shape. Good questions:
+  - Point at thinly-covered topics: "What does the literature say about X?"
+  - Point at unresolved contradictions: "Which view of Y seems better
+    supported — the one in [[entities/a]] or [[entities/b]]?"
+  - Point at synthesis gaps: "How do [[concepts/p]] and [[concepts/q]]
+    relate, given the sources cover them separately?"
+Bad questions: vague prompts, anything generic that ignores the wiki's
+content.
+
+4. Output a grouped report like:
+
+```
+Wiki health report — N pages scanned
+
+== Contradictions (unresolved) ==
+1. [[entities/transformer]] ⚠ conflicts with [[entities/rnn]] on vanishing
+   gradients (line 42 + line 37). Consider: add a short comparison section
+   to [[concepts/sequence-modelling]] and note the trade-off.
+
+== Orphan pages ==
+2. [[entities/gpt-2]] — no inbound links. Suggested: add to index.md under
+   "Entities > Models" and reference from [[concepts/pretraining]].
+
+== Missing cross-references ==
+3. [[sources/attention-is-all-you-need]] mentions "Transformer" without a
+   link to [[entities/transformer]] (line 18).
+
+== Missing entity pages ==
+4. "Adam optimizer" appears in 3 source summaries with no entity page.
+   Suggested slug: entities/adam-optimizer.
+
+== Data gaps ==
+5. [[entities/layer-normalization]] cites only 1 source. Consider reading
+   more on this.
+
+== Follow-up questions you might ask ==
+- How does [[entities/adam-optimizer]] compare to SGD across the sources
+  that cover both?
+- Why does [[concepts/attention-mechanisms]] cite 8 sources but
+  [[concepts/positional-encoding]] only 2 — is that a real gap in the
+  literature or a gap in your reading?
+```
+
+5. **Append a log entry** to `research/wiki/log.md`:
+   ```
+   ## [YYYY-MM-DD] lint | pass <n>
+   - Contradictions: <count>
+   - Orphans: <count>
+   - Missing xrefs: <count>
+   - Missing pages: <count>
+   - Data gaps: <count>
+   - Questions suggested: <count>
+   ```
+
+## What NOT to do
+- Do NOT auto-fix. Offer; the user decides.
+- Do NOT run web searches — you have no web tool. Suggest where one could
+  help; the user can do it manually.
+- Do NOT write to `research/raw/**` or `data/**`.
+- Do NOT mark EVERY claim `⚠ possibly stale` — only those you're confident
+  a newer source contradicts. False positives train the user to ignore you.