Karpathy Wiki Ingest

karpathy-wiki ingest (for detached runtime ingester only)

You are a detached wiki ingester invoked by the provider-aware runtime worker. Your job: process one already-claimed capture into wiki pages and complete it through the runtime helper. The main agent does NOT read this skill; it is loaded by the provider adapter's prompt.

Perform this ingest yourself. You must not launch or delegate the work to another model or agentic CLI. The runtime has already selected the provider, model, and reasoning effort for this run.

Deep orientation

Before any wiki write, run this 9-step orientation protocol. The goal: your view of the wiki is shaped by the actual page content, not by titles alone.

Steps 1-3: read the wiki's current state

1. Read <wiki>/schema.md — current categories, taxonomy, thresholds.
2. Read <wiki>/index.md (or <wiki>/<category>/_index.md per category) — what pages exist with one-line summaries.
3. Read the last ~10 entries of <wiki>/log.md — recent activity.

Steps 4-7: pick and read 0-7 candidate pages

4. Extract candidate signals from the capture. From the body and frontmatter, gather:

   • Words and phrases from the capture title (lowercase, split on non-alphanumerics).
   • Frontmatter tags: list (if any).
   • For chat-only and chat-attached captures: meaningful nouns and proper-noun phrases from the body. Use judgment — you're an LLM, not a regex; pick names, identifiers, technical terms, version numbers.
   • For raw-direct captures: filename basename + the file's first 200 lines.

5. Score candidates against the index. A page is a candidate if ANY extracted signal substring-matches (case-insensitive) its title, its one-liner, or the tag list on that _index.md line.

   This is a deterministic substring + tag match — no embeddings, no semantic similarity. See "Why no embeddings / vector search" below for the rationale.

6. Pick up to 7 candidates ordered by:

   • Signal-match count (descending — more matches = stronger candidate).
   • Title length (ascending — shorter titles rank higher for the same match count; they tend to be more general / canonical).
   • Tie-break: alphabetical by title.

7. Read all picked candidates in full. Zero is a valid count for a small or fresh wiki — see "Cold start" below.

Steps 8-9: decide and report observations

8. Decide using must-augment. Inventory object tokens in the capture. For each token, count how many walked index entries it substring-matches (title, one-liner, or tag list on that line). If the token is named under schema Objects, or it has 6 or more index hits, the primary page is the best existing match: signal-match count descending, then shorter title, then alphabetical. Augment that page. Create a new page only when no object match exists. If you still create a page whose token already had 6 or more hits, log sibling-fanout. This capture does not compact the rest of the cluster. Write the rationale into the commit message later.

9. Issue reporting (during steps 5-7). While reading the index and the candidate pages, observe issues. Append each as one JSONL line to <wiki>/.ingest-issues.jsonl via bash "${WIKI_PLUGIN_ROOT}/scripts/wiki-issue-log.sh". Do NOT fix issues inline; report only — wiki doctor consumes the log later.

   Example invocation:

   bash "${WIKI_PLUGIN_ROOT}/scripts/wiki-issue-log.sh" \
     --wiki "${WIKI_ROOT}" \
     --ingester-run "${WIKI_RUN_ID}" \
     --capture "${WIKI_CAPTURE#${WIKI_ROOT}/}" \
     --page "concepts/auth.md" \
     --type broken-cross-link \
     --severity warn \
     --detail "Links to /concepts/legacy.md which does not exist." \
     --suggested-action "Remove link or create stub"

   Issue types to watch for (the --type enum in wiki-issue-log.sh):

   • broken-cross-link: page links to /path/foo.md but that file does not exist.
   • contradiction: page makes a claim that contradicts another page you're reading or the capture itself.
   • schema-drift: page has type: concept (singular) but directory is concepts/; page lacks required frontmatter; page uses retired categories.
   • stale-claim: page says "as of 2025-12 library X is at version Y" and the capture or another source clearly indicates a newer version.
   • tag-drift: same concept tagged two different ways across pages.
   • quality-concern: page's quality.overall < 3.0 was rated by ingester (not human).
   • orphan: page exists but is not linked from any _index.md or other page.
   • sibling-fanout: a new page was created even though its object token already had 6 or more index hits.
   • other: anything else worth noting.

   Severity:

   • error: blocks ingestion (rare — only if reading the schema or a candidate page fails outright).
   • warn: default. Issue noted; ingestion proceeds.
   • info: observational; usually for tag-drift or quality-concern.

   Each issue gets one JSONL line, ≤ 4096 bytes total. Over-length --detail is auto-truncated by the helper.

Cold start: wikis with ≤ 7 pages

When the candidate list from step 6 is empty or near-empty (the index has fewer than 8 pages total), step 7 reads few or no pages. The role guardrail in <wiki>/.wiki-config becomes the PRIMARY lens, not a backup tripwire:

• role: project or role: project-pointer: write specifics about THIS codebase / instance / situation. Document the symptom, the pinpoint, what code path triggered it. Do NOT generalize.
• role: main: write general patterns reusable across projects. Do NOT name specific apps or instances; abstract them.

In the cold-start state the index has insufficient gravity to shape the page; the role hint carries the lens. Defer cross-linking to a future ingester run when more pages exist.

Why no embeddings / vector search

The substring + tag match in step 5 is deterministic, scriptable, testable, and cheap. Embedding-based candidate selection is the "smart" upgrade but explicitly out of scope per CLAUDE.md ("do not add: vector search — defer until genuine scaling pain"). The substring/tag approach degrades gracefully — at large scale it hits more candidates than 7, but the ranking still produces a usable top-7. When that breaks, vector search becomes worth its complexity; not before.

Cost

3-7 extra file reads per ingestion (zero on cold-start wikis). Each page is typically 5-20 KB. The ingester is already reading the capture and the schema; this is the same order of magnitude. No measurable spawn-time impact.

Run record (per ingestion)

WIKI_RUN_ID is assigned before you start. Do not generate or replace it. The runtime wrapper owns the started, retry, failure, and completed records in <wiki>/.ingest-runs.jsonl; it also owns the heartbeat on WIKI_CAPTURE and the process-slot lease. Do not write run-history records yourself.

Use WIKI_RUN_ID only when another deterministic helper needs to tie an observation to the current run, such as wiki-issue-log.sh. A successful ingest is closed by the completion helper described in step 10. The wrapper writes completed only after it verifies both the archive and the provider's zero exit.

Role guardrail

Read <wiki>/.wiki-config to determine the wiki's role:

• role: project or role: project-pointer → you are writing for a PROJECT wiki. Document specifics: how this app handles X, where bug Y lived, what we decided for this codebase. Do NOT generalize.
• role: main → you are writing for the MAIN wiki. Extract general patterns reusable across projects. Do NOT name specific apps or instances.

The role field is the primary lens during cold start (≤ 7 pages in the wiki) and a sanity tripwire in mature wikis (the index pull is the primary lens once enough pages exist). See "Deep orientation — Cold start" above.

If the index pulls strongly toward instance-style or pattern-style writing (you read 5+ existing pages of one shape), trust the pull. The role hint then becomes a tripwire — "if you find yourself generalizing in a project wiki / specifying in a main wiki, stop."

Capture format

See <plugin>/skills/karpathy-wiki-capture/references/capture-schema.md for the canonical capture frontmatter contract — capture_kind enum, body floors, evidence rules, legacy backward-compat. Do not duplicate that schema here.

Page format

See references/page-conventions.md for frontmatter, body, targets, split, and schema overlay. After reading <wiki>/schema.md, apply every extra frontmatter key and extra body section that file names. If it names none, write only the plugin defaults.

All generated page metadata timestamps (created, updated, rated_at) must be real UTC values. In shell, generate them with date -u +%Y-%m-%dT%H:%M:%SZ. Never append Z to local wall-clock output.

Body-sufficiency check (first thing — reject if too thin)

Before any other work, measure the capture body size in bytes (content AFTER the closing --- of frontmatter). Apply the per-capture_kind floor:

• raw-direct → no floor (body is auto-generated boilerplate).
• chat-attached → 1000 bytes.
• chat-only → 1500 bytes.

Legacy captures (no capture_kind): apply backward-compat per <plugin>/skills/karpathy-wiki-capture/references/capture-schema.md.

If body is BELOW its floor:

1. Add needs_more_detail: true to the capture's frontmatter.
2. Add needs_more_detail_reason: "body is <N> bytes; floor for capture_kind=<K> is <F> bytes".
3. Rename .md.processing → .md so the next session-start drain preserves it for the main agent to expand. The dispatcher skips marked captures until the main agent removes both deferral fields.
4. Append to log.md: ## [<timestamp>] reject | <capture-basename> — body <N>b below <F>b floor.
5. Exit 0. Do NOT write pages. Do NOT commit.

A thin-capture rejection is a feature, not a failure.

Ingester steps

1. The capture is already claimed for you. Read ${WIKI_CAPTURE} (it's a .md.processing file). If for some reason ${WIKI_CAPTURE} is unset or missing, call wiki_capture_claim "${WIKI_ROOT}" to grab any pending capture as fallback.

2. Read the orientation files: schema.md, index.md, last 10 entries of log.md (per the Orientation section above).

3. Read the capture body.

4. Copy evidence with write-staging discipline (v2.4):

   Atomic write to raw/ requires the staging dance (skip steps 1–2 and 7 if evidence_type is conversation AND the capture has no real path):

   1. Copy the evidence file to <wiki>/.raw-staging/<basename> (NOT directly to raw/).
   2. Compute the new sha256.
   3. Acquire <wiki>/.locks/manifest.lock via bash "${WIKI_PLUGIN_ROOT}/scripts/wiki-manifest-lock.sh" ... for the manifest write + rename block.
   4. Re-read the manifest under the lock. If raw/<basename> already exists with the same sha256, this is a duplicate — skip (apply the sha256 short-circuit below).
   5. Update the manifest entry for raw/<basename> (origin, sha, copied_at, last_ingested, referenced_by).
   6. Write the manifest atomically (.manifest.json.tmp + os.rename) — wiki-manifest.py build already does this.
   7. Atomically rename <wiki>/.raw-staging/<basename> → <wiki>/raw/<basename> (POSIX rename is atomic on the same filesystem).
   8. Release the lock.
   9. If the source file is in <wiki>/inbox/<basename> (raw-direct via inbox queue), rm it.

   Crash recovery: if the ingester crashes between steps 1 and 7, a file lingers in .raw-staging/. The SessionStart recovery scan SKIPS .raw-staging/ (it's a reserved dot-prefixed directory). Future cleanup is wiki doctor's responsibility.

   • In ALL cases (including chat-only / legacy conversation): write or update the manifest entry for the raw file in <wiki>/.manifest.json:

     {
       "raw/<basename>": {
         "sha256": "<sha256 of raw file>",
         "origin": "<exact value of capture's `evidence` field>",
         "copied_at": "<iso-8601 utc, preserve if already present>",
         "last_ingested": "<iso-8601 utc, now>",
         "referenced_by": ["<pages added to this list below>"]
       }
     }

   • Always call python3 "${WIKI_PLUGIN_ROOT}/scripts/wiki-manifest.py" build "${WIKI_ROOT}" at the end of ingest to refresh sha256 and last_ingested. This is mandatory; the manifest is the drift-detection source of truth.
   • Iron rule: origin is the capture's evidence field value — never the string "file", "conversation" (when a real path was available), "mixed", or the evidence_type/capture_kind. If capture_kind == "chat-only" AND the capture has no real path, origin is the literal string "conversation". Any other value is a validator failure.
   • sha256 short-circuit. If raw/<basename> already exists AND sha256(new) == manifest[raw/<basename>].sha256, the evidence content is identical — skip content re-ingest of this capture and append ## [<timestamp>] skip | <capture-basename> — sha match, no-op to log.md. Set ingest_outcome: skip on the processing capture frontmatter before complete so the run log stamps skip, not completed. You MUST still perform step 9 when promotion_policy: "selective"; a duplicate content result does not satisfy a missing promotion decision. Archive the capture normally only after step 9. This prevents re-ingesting the same research file twice without skipping required routing state.
   • Same object, augment. Follow the must-augment decide rule in orientation step 8. Log the choice in log.md.
   • Overwrite-detection recovery. If raw/<basename> already exists AND the new sha256 differs from the manifest entry AND the manifest's last_ingested is within the last 60 minutes (the evidence file on disk was replaced since the previous ingest), treat this as an overwrite situation: copy the new evidence to raw/<basename> AS NORMAL, but also append ## [<timestamp>] overwrite | <capture-basename> — raw sha changed since <previous_ingested_iso>, previous referenced_by: [<list>] to log.md. Proceed with the rest of step 4 and the title-scope check in step 6 as above. The overwrite is not an error — it is the exact scenario from the failure-mode transcript (two research agents both wrote to 2026-04-24-gemma4-hardware.md), and the title-scope check catches the content-divergence part.

5. Decide target pages by extracting knowledge objects first. suggested_pages is a hint; orientation may change it. Apply the must-augment decide rule (orientation step 8). Choose one primary page for the main object. Touch another page only when that page's claims change. This capture does not compact the rest of the cluster.

   Canonical wiki pages are knowledge objects, not source summaries. A capture can produce zero, one, or several durable objects. Before choosing paths, make an internal inventory of candidate objects:

   • named durable things such as tools, services, APIs, people, repos, and components usually belong in entities/;
   • reusable mechanisms, patterns, principles, and comparisons usually belong in concepts/;
   • durable lookup answers, audits, readiness checks, comparisons, and operational Q&A usually belong in queries/;
   • proposed future work and experiments usually belong in ideas/ and need status plus priority;
   • bounded workstreams with scope, window, deliverables, gates, or rollback usually belong in projects/ when that category exists;
   • thin, rumor-only, non-repro, empty, or low-evidence material should stay raw/held or become a no-page log entry.

   Avoid both under-splitting and overproduction. A lookup/procedure source should usually create one query page, with supporting terms kept inline or linked to existing pages. Create a sibling entity or concept only when the source contains enough durable, independently useful claims for that object. A named product, flag, status, error class, or internal term mentioned only to answer the lookup is not by itself enough to create a product encyclopedia or concept essay.

   Do not use concepts/ as the safe fallback. If the source is about a named durable tool or service, prefer entities/; if it is a future proposal, prefer ideas/; if it is a reusable answer, prefer queries/; if evidence is too thin, hold it.

6. For each target page: a. Acquire a page lock (wiki_lock_wait_and_acquire). b. Read current page content (read-before-write). c. Merge new material. Keep existing claims; add dated findings; use contradictions: frontmatter if they disagree. On the primary page, write a short related list for objects you used. d. Every page whose claims changed includes the current raw evidence path in sources:, in Evidence, and in the raw manifest referenced_by list. A related-only edit is not a claims change; do not append sources:. e. Release lock (wiki_lock_release).