log.md

Log

Chronological history of vault activity. Append-only. Each entry starts with ## [YYYY-MM-DD] <action> | <subject> so grep "^## \[" log.md parses cleanly.

[2026-05-21] bootstrap | Vault initialized

Created the LLM Wiki scaffold per the pattern in raw/articles/llm-wiki.md.

• Folders: raw/{articles,transcripts,assets}, wiki/{sources,entities,concepts,topics,syntheses}.
• Files: CLAUDE.md (schema), index.md (catalog), log.md (this file).
• Domain: AI/ML engineering. Workflow: autonomous batch ingests.

[2026-05-21] ingest | llm-wiki

• Source: raw/articles/llm-wiki.md — the LLM Wiki idea doc itself (Source #1, used to bootstrap the system).
• Created (10): sources/llm-wiki, concepts/incremental-wiki-pattern, concepts/rag-retrieval-augmented-generation, concepts/memex, concepts/wiki-maintenance-bookkeeping, entities/obsidian, entities/obsidian-web-clipper, entities/qmd, entities/marp, topics/personal-knowledge-management.
• Updated: index.md.
• Contradictions: none (single source).
• Follow-up questions: see sources/llm-wiki → "Open questions".

[2026-05-22] update | llm-wiki author attribution

User-provided fact: [[llm-wiki]] is authored by Andrej Karpathy (the doc text itself is unsigned).

• Created: entities/andrej-karpathy.
• Updated: sources/llm-wiki (author field + related links), index.md (entities now 5; source line cites author).

[2026-05-21] ingest | 5-agent-skills + grill-with-docs-launch

Batch ingest, 2 sources from [[matt-pocock]] / [[ai-hero]].

• Moved into raw subfolders: raw/5 Agent Skills I Use Every Day.md → raw/articles/, raw/I stopped using grill-me for coding…md → raw/transcripts/.
• Created sources (2): sources/5-agent-skills, sources/grill-with-docs-launch.
• Created entities (3): entities/matt-pocock, entities/ai-hero, entities/claude-code.
• Created concepts (9): concepts/agent-skills, concepts/grill-me, concepts/grill-with-docs, concepts/ubiquitous-language, concepts/bounded-context, concepts/domain-driven-design, concepts/architectural-decision-records, concepts/test-driven-development, concepts/design-tree.
• Created topics (1): topics/llm-engineering-workflow.
• Updated: index.md (sources 1→3, concepts 4→13, entities 5→8, topics 1→2).
• Contradictions / tensions: None contradictory; the two sources are temporally related — [[5-agent-skills]] presents [[grill-me]] as Pocock's favorite, [[grill-with-docs-launch]] introduces [[grill-with-docs]] as its successor for coding contexts. The rule of thumb Pocock gives: code → grill-with-docs, no code → grill-me. The earlier source already lists grill-with-docs in its "more where these came from" callout, so the two are consistent rather than contradicting.
• Wikilink stubs (not yet pages, intentionally deferred): [[to-prd]], [[to-issues]], [[tdd]], [[improve-codebase-architecture]], [[tracer-bullets]], [[vertical-slice]], [[deep-modules]]. Pages should be created when future sources cover them substantively.
• Follow-up questions: see topics/llm-engineering-workflow → "Open threads", and per-source "Open questions" sections.

[2026-05-21] ingest | mattpocock/skills repo (6 skills + repo meta)

Primary-source deep ingest of [[matt-pocock]]'s skill kit. Closes the wikilink stubs from the previous ingest and grounds the existing concept pages in primary text.

• Cloned to: /tmp/pocock-skills (commit --depth 1). Copied 20 files into raw/skills/mattpocock-skills/ preserving upstream directory structure (README, CLAUDE, CONTEXT at root; per-skill SKILL.md + supporting files for grill-me, grill-with-docs, to-prd, to-issues, tdd, improve-codebase-architecture).
• Schema evolution: updated CLAUDE.md to (1) recognize raw/skills/ as a new source bucket for primary skill/agent-config files; (2) add skill-md and repo-doc as valid source_kind values; (3) add book as a valid entity_kind.
• Created sources (7): sources/pocock-skills-repo, sources/pocock-grill-me, sources/pocock-grill-with-docs, sources/pocock-to-prd, sources/pocock-to-issues, sources/pocock-tdd, sources/pocock-improve-codebase-architecture.
• Created concepts (7): concepts/tdd, concepts/to-prd, concepts/to-issues, concepts/improve-codebase-architecture, concepts/deep-modules, concepts/vertical-slice, concepts/seam. (Closes 4 of the deferred wikilink stubs from the previous ingest; tracer-bullets and deep-modules consolidated into [[vertical-slice]] and [[deep-modules]] respectively.)
• Updated concepts (6) — full rewrite to incorporate primary-source data: concepts/grill-me (verbatim text, evolution note), concepts/grill-with-docs (full operational rules), concepts/test-driven-development (Pocock's "good tests" philosophy + horizontal-slicing anti-pattern), concepts/architectural-decision-records (three-criteria rule, format, qualifying enumeration), concepts/ubiquitous-language (CONTEXT.md format rules), concepts/agent-skills (4-problems framing, full kit catalog, evolution evidence).
• Created entities (5): entities/john-ousterhout, entities/kent-beck, entities/eric-evans, entities/pragmatic-programmer (book), entities/skills-repo.
• Updated entities (1): entities/matt-pocock (now reflects full picture: 9 of 10 vault sources are his; cites the canon he references).
• Updated topics (1): topics/llm-engineering-workflow (4-problems table, full feature loop, expanded "current understanding" and open threads).
• Updated: index.md (sources 3→10, concepts 13→20, entities 8→13, topics 2 unchanged).
• Contradictions / evolutions flagged:
  • /grill-me: article quoted as 3 sentences; current SKILL.md is 5 sentences ("For each question, provide your recommended answer" and "Ask the questions one at a time" are later refinements).
  • /to-prd: significant evolution — article described the skill as including grilling ("Interview the user relentlessly using the grill me skill"); current SKILL.md says "Do NOT interview the user — just synthesize what you already know." Grilling is now strictly the upstream job of /grill-with-docs. Flagged on both source pages.
  • /to-issues: current SKILL.md introduces the HITL/AFK labeling not present in the article.
  • /tdd: current SKILL.md contains a much richer philosophy + anti-pattern section than the article describes; horizontal-vs-vertical TDD framing is primary-source-only.
  • /improve-codebase-architecture: major evolution — article described conversational candidate presentation; current SKILL.md outputs an HTML report with Tailwind+Mermaid, recommendation-strength badges, and a top-recommendation section.
• Deliberately deferred (still wikilink stubs, no page yet): [[michael-feathers]] (mentioned on [[seam]] without entity page), plus skills not in scope this batch: /diagnose, /triage, /setup-matt-pocock-skills, /zoom-out, /prototype, /caveman, /handoff, /write-a-skill, and misc/personal/in-progress/deprecated buckets.
• Follow-up questions: see updated topics/llm-engineering-workflow → "Open threads"; per-source "Open questions" sections in the 7 new sources/pocock-* pages.

[2026-05-21] ingest | Claude Code official docs (10 pages)

First-party docs from [[anthropic]] at code.claude.com/docs. Closes the [[claude-code]] stub and lifts the wiki's coverage from "Pocock-only" to "Pocock + Anthropic protocol."

• Fetched via WebFetch. New raw/docs/claude-code/ subfolder.
• Schema evolution: added raw/docs/ to CLAUDE.md as a recognized source bucket.
• Raw files saved (9 of 10): overview.md, quickstart.md, how-claude-code-works.md, features-overview.md, memory.md, permission-modes.md, prompt-caching.md, common-workflows.md, agent-sdk-overview.md. The 2 interactive pages (claude-directory, context-window) returned JSX-heavy React component code; raw not saved this round (auto-mode classifier also blocked the workaround cp from internal cache). Their source pages note this; content is well covered via cross-references.
• Created sources (11): cc-overview, cc-quickstart, cc-how-it-works, cc-features-overview, cc-memory, cc-permission-modes, cc-prompt-caching, cc-common-workflows, cc-agent-sdk-overview, cc-claude-directory, cc-context-window.
• Created concepts (11): agentic-loop, claude-md, mcp, subagents, hooks, plan-mode, plugins, agent-sdk, auto-memory, context-window, prompt-caching.
• Updated concepts (1): agent-skills — major rewrite to bring Anthropic's protocol-level treatment alongside Pocock's content.
• Created entities (1): anthropic.
• Updated entities (1): claude-code — major rewrite from stub to full entity. Now reflects the 5 surfaces, agent SDK, extension layer, session storage, and runtime relationship with this wiki itself.
• Updated topics (1): llm-engineering-workflow — restructured to surface the two-perspectives framing (Anthropic protocol + Pocock content).
• Updated: index.md (sources 10→21, concepts 20→31, entities 13→14).
• Notable observations:
  • CLAUDE.md vs CONTEXT.md layering is genuine: CLAUDE.md (process rules), CONTEXT.md (vocabulary), ADRs (decisions), MEMORY.md (Claude-observed). All persistent, all loaded at session start. Worth a synthesis page on "the project memory stack."
  • Auto mode (research preview) uses a separate classifier model — boundaries stated in conversation aren't durable (lost via compaction). For hard guarantees: deny rules + hooks.
  • /to-prd evolution flagged in earlier ingest is mirrored in Anthropic's docs: skills are described as composable atoms, and Anthropic's heuristic ("always do X" → CLAUDE.md; "do X when I tell you" → skill) cleanly supports Pocock's grilling-out-of-/to-prd decision.
  • Agent SDK pricing transition on 2026-06-15 — separate monthly Agent SDK credit on subscription plans. Worth tracking for anyone embedding SDK in production.
  • Prompt caching is the engineering core — three-layer (system → project context → conversation) prefix model and exact-prefix-match rule explain why mid-session CLAUDE.md edits don't apply, why MCP server disconnects invalidate everything, and why opusplan toggles are expensive.
• Wikilink stubs (not yet pages, intentionally deferred): michael-feathers (still); Anthropic-side: Routines, Settings, Hooks-guide page (/en/hooks), Sub-agents page (/en/sub-agents), Skills page (/en/skills), MCP page (/en/mcp), Plugins page (/en/plugins), Plugin-marketplaces, Ultraplan, Agent teams, Background agents. Each likely deepens an existing concept page.
• Internal slug note: a couple of source pages reference [[context-window-concept]] — actual concept page is [[context-window]]. Minor wikilink rot to clean up in a future lint pass.
• Follow-up questions: see updated topics/llm-engineering-workflow → "Open threads"; per-source "Open questions" sections in the 11 new sources/cc-* pages.

[2026-05-21] ingest | Karpathy — From Vibe Coding to Agentic Engineering (Sequoia AI Ascent 2026)

First substantive technical content from [[andrej-karpathy]] in the vault. Closes the long-standing gap where the [[andrej-karpathy]] entity existed but had no actual Karpathy frameworks attached. Provides the paradigm vocabulary that bridges [[anthropic]]'s protocol and [[matt-pocock|Pocock]]'s practice.

• Source moved: raw/Andrej Karpathy From Vibe Coding to Agentic Engineering.md → raw/transcripts/.
• Created sources (1): karpathy-vibe-to-agentic.
• Created concepts (6): vibe-coding, agentic-engineering, software-3-0, llms-as-ghosts, jagged-intelligence, verifiability.
• Updated entities (1): andrej-karpathy — major rewrite from "author of LLM Wiki idea doc" stub to a full entity with biographical context (OpenAI / Tesla / Eureka Labs), the framework table he's pushed into this vault, and the closing thesis ("you can outsource your thinking but not your understanding") connection to [[incremental-wiki-pattern]].
• Updated concepts (1): incremental-wiki-pattern — Karpathy explicitly cites LLM knowledge bases by name in the talk as a [[software-3-0]] use case; added a "Karpathy's Software 3.0 framing" section with verbatim quotes.
• Updated topics (1): llm-engineering-workflow — restructured around three converging perspectives (Karpathy paradigm + Anthropic protocol + Pocock curriculum) instead of the previous two. Added a "Karpathy's vocabulary" subsection to the map.
• Updated: index.md (sources 21→22, concepts 31→37, entities/topics unchanged).
• Cross-cutting insight surfaced: Karpathy's "agentic engineering" terminology cleanly maps onto everything Pocock has been doing. Pocock's skills + canon + 4-problem framing is a concrete curriculum for agentic engineering. Karpathy supplies the name; Pocock supplies the practice.
• Direct primary-source endorsement of this vault's pattern: Karpathy at ~7:00 of the talk cites LLM knowledge-base wikis as a textbook example of new capabilities only possible in Software 3.0. Returns to the theme in the closing thesis on understanding.
• Notable observations:
  • December 2025 inflection point (per Karpathy): agentic tools crossed a quality threshold where chunks "just came out fine." Worth treating as a calibration anchor in this topic.
  • The Karpathy hint at an unnamed verifiable RL domain ("there is one domain that I think is very… I don't want to give away the answer") — worth watching.
  • Aesthetics not in the RL signal yet — explains why even top models produce "really gross" code. Karpathy is hopeful but it's open research.
  • "OpenClaw" appears multiple times in the transcript; likely a YouTube auto-caption error for an actual tool name (possibly OpenCode, Cline, or similar). Flagged for resolution next Karpathy ingest.
• Wikilink stubs deferred (still): Eureka Labs, Sequoia Capital, Stephanie Zhan, micro-GPT / nanoGPT, Gemini, Nano Banana, MenuGen — light mentions only. Worth elevating to entity pages if they recur.
• Follow-up questions: see updated topics/llm-engineering-workflow → "Open threads", and sources/karpathy-vibe-to-agentic → "Open questions" for talk-specific items.

[2026-05-21] ingest | Pocock — /handoff is my new favourite skill

Pocock video + canonical SKILL.md for /handoff — closes one of the deferred wikilink stubs from the original Pocock ingest. Also introduces the "dumb zone / smart zone" framing of context windows that didn't appear in Anthropic's docs.

• Source moved: raw/handoff is my new favourite skill.md → raw/transcripts/.
• Copied canonical SKILL.md from the (still-extant) /tmp/pocock-skills clone into raw/skills/mattpocock-skills/skills/productivity/handoff/SKILL.md.
• Created sources (2): pocock-handoff (SKILL.md primary), pocock-handoff-launch (video transcript).
• Created concepts (1): handoff — closes the previously-deferred wikilink. Frames as the "DIY subagent" alternative to runtime [[subagents]]; cross-runtime portable (markdown is the lingua franca).
• Updated concepts (1): context-window — added the "dumb zone / smart zone" framing (Pocock's ~120K-tokens-into-1M anchor). Cross-links to [[handoff]] as the recommended escape valve before hitting the dumb zone.
• Updated entities (2):
  • matt-pocock — added new sources; added cohort-course fact; skills repo at almost 100,000 stars as of May 2026.
  • skills-repo — same star milestone.
• Updated: index.md (sources 22→24, concepts 37→38).
• Notable observations:
  • argument-hint skill frontmatter field appears here for the first time in any ingested SKILL.md. Claude Code surface convention worth verifying against the Anthropic skills doc page (not yet ingested).
  • Pocock's canonical pattern: grill → handoff → prototype → handoff back → grill. Constructs a tree of focused sessions out of flat markdown handoff docs. User-orchestrated, runtime-agnostic.
  • The dumb-zone framing is empirical (Pocock's experience), not Anthropic-published. Worth a synthesis page when more practitioner data accumulates on context-degradation curves per model.
  • Cross-runtime adversarial review: passing a handoff doc from [[claude-code]] to Codex or Copilot CLI is a clean primitive Pocock highlights. Pattern worth tracking for future ingests on multi-agent workflows.
• Deferred (still wikilink stubs): Sand Castle (Pocock's project), TLDraw SDK, AI Coding For Real Engineers (course), /prototype skill (mentioned, not yet ingested). Each could become an entity/concept if it recurs.
• Follow-up questions: see sources/pocock-handoff and sources/pocock-handoff-launch → "Open questions" sections.

[2026-05-21] correction | promote dumb-zone framing to its own concept page

User noted (correctly) that the "dumb zone / smart zone" framing should be its own page, not a subsection of [[context-window]]. The framing is a named Pocock concept with its own empirical anchor (~120K tokens), its own workflow implications, and likely future recurrence — same shape as [[vertical-slice]], [[deep-modules]], [[seam]], [[verifiability]], [[jagged-intelligence]], which all got their own pages. Embedding it in [[context-window]] was an inconsistent call.

• Created concepts (1): dumb-zone — full treatment with verbatim framing, attention-quadratic mechanism, workflow implications, pairing with Anthropic's mechanisms, and convergence with [[jagged-intelligence]].
• Updated concepts (1): context-window — trimmed the inline "dumb zone framing" subsection to a brief pointer to the new [[dumb-zone]] page.
• Updated: index.md (concepts 38→39).

[2026-05-21] ingest | Cherny — Why Coding Is Solved, and What Comes Next (Sequoia AI Ascent 2026)

First Anthropic-insider primary source in the vault — [[boris-cherny|Boris Cherny]], creator of [[claude-code]]. Counterpoint to [[matt-pocock|Pocock]] (heavy user) and [[andrej-karpathy|Karpathy]] (paradigm namer): Cherny is the builder whose product bet captured the [[product-overhang|overhang]] Karpathy later described as the December 2025 inflection.

• Source moved: raw/Anthropic's Boris Cherny Why Coding Is Solved, and What Comes Next.md → raw/transcripts/.
• Created sources (1): cherny-coding-solved.
• Created concepts (2): product-overhang (Anthropic internal term for model-capability-not-yet-captured-by-products); agent-loops (Cherny's "loops are the future" — /loop + Routines + Desktop scheduled tasks as the persistent-agent primitive).
• Created entities (1): boris-cherny.
• Updated entities (2):
  • anthropic — added Mike Krieger as CPO + Anthropic Labs incubator history + the explicit "all our SQL is written by models" / "Claudes talking to Claudes over Slack" internal practice. Added boris-cherny to related.
  • claude-code — added origin story (Anthropic Labs 2024, Boris Cherny, TS+React stack, Opus 4 inflection in May 2025). Added related links to boris-cherny, product-overhang, agent-loops.
• Updated: index.md (sources 24→25, concepts 39→41, entities 14→15).
• Notable observations / tensions surfaced:
  • Boris vs Pocock: Boris bets the harness becomes less important as models improve — explicitly including "permission modes, human in the loop, all this kind of stuff." Pocock's entire framework bets on durable named workflows + persistent project docs. Strong synthesis candidate: do skills outlast the harness, or get absorbed?
  • Boris vs Karpathy: Both describe the same inflection. Karpathy frames it as December 2025 felt from the user side; Boris frames it as Opus 4 in May 2025 captured the product overhang he deliberately built for 6 months pre-PMF.
  • Concrete scale anchors: Cherny runs ~5–10 sessions, hundreds of agents during the day, few thousand at night, peak day 150 PRs, 0 lines hand-written in 2026. Most extreme single-practitioner data point in this vault.
  • Anthropic internal practice: "no more manually written code anywhere at the company," "all SQL is written by models," "Claudes talking to Claudes over Slack." Where Anthropic is ahead of external builders is org structure and process, not technology.
  • The TS+React stack choice for Claude Code was an "on-distribution-for-the-model" decision — fascinating data point for new agent-native projects.
  • The printing-press analogy for software democratization is Cherny's preferred framing — 1400s parallel, but "much faster than 50 years."
• Wikilink stubs deferred: Mike Krieger (CPO), Anthropic Labs (incubator), Claude Design (Anthropic product mentioned), Co-work (Anthropic knowledge-work product), Hamilton Helmer's 7 Powers (referenced framework), Lauren Reeder (interviewer). Light mentions only — promote if they recur.
• Follow-up questions: see sources/cherny-coding-solved → "Open questions" for talk-specific items.
• Lesson for future ingests: when a source introduces a named framing with its own implications, default to a concept page even if it feels sub-thematic to an existing concept. Cross-link aggressively rather than embed.

[2026-05-24] ingest | conductor.build docs (40 pages)

First-party deep ingest of [[conductor]]'s public docs at conductor.build/docs. Closes the gap left by [[gstack-readme]] and [[tan-ai-engineering-team]], which talk about Conductor without primary text. Substantially elevates Conductor from "third-party orchestrator gstack mentions" to a fully ingested first-party source.

• Discovery: site sitemap fetched + matched against site:conductor.build/docs search → 40 doc pages (concepts × 5, guides × 7, reference × 26, troubleshooting × 1, plus intro/install/first-workspace/faq).
• Raw files saved: all 40 pages to raw/docs/conductor/, kebab-case names with category prefixes (concept-*, guide-*, guide-repo-*, ref-*, ref-scripts-*, troubleshooting-*). Schema note: added raw/docs/conductor/ alongside existing raw/docs/claude-code/. No CLAUDE.md change required.
• Created sources (40): cond-introduction, cond-first-workspace, cond-installation, cond-faq, cond-workspaces-and-branches, cond-workflow, cond-parallel-agents, cond-agent-modes, cond-testing, cond-how-to-run, cond-issue-to-pr, cond-migrate-from-cursor, cond-providers, cond-review-and-merge, cond-linking-multiple-directories, cond-monorepos, cond-agent-behavior, cond-big-terminal-mode, cond-checkpoints, cond-checks, cond-cities, cond-conductor-json, cond-deep-links, cond-diff-viewer, cond-environment-variables, cond-files-to-copy, cond-keyboard-shortcuts, cond-mcp, cond-privacy, cond-security-and-permissions, cond-settings, cond-shells, cond-slash-commands, cond-todos, cond-scripts, cond-scripts-setup, cond-scripts-run, cond-scripts-share-with-teammates, cond-scripts-spotlight-testing, cond-troubleshooting-issues.
• Created concepts (2): [[workspace-isolation]] (Conductor's core primitive — five-level model: Project → Repository → Workspace → Branch → Working tree → Running environment; primary text is [[cond-workspaces-and-branches]]) + [[spotlight-testing]] (Conductor's escape-hatch pattern when workspace isolation is too expensive; primary text is [[cond-scripts-spotlight-testing]]).
• Major rewrite — [[conductor]] entity: rebuilt with first-party data. New sections: bundled-binaries paths, five env vars, three script slots, conductor.json fields, managed settings + JSON Schema URL, six repository-preference fields, six agent-mode controls, 7 agent presets in Big Terminal Mode, checkpoints in private git refs, Spotlight, 295-city naming pool, conductor:// deep links + four link shapes, provider routing, MCP integration, security posture. Cross-linked to all 40 source pages.
• Updated concepts (3): [[parallel-agentic-sessions]] (added Conductor's first-party share-vs-split decision table + workspace-as-delegation-unit framing), [[mcp]] (added three-way Cursor/Claude/Codex MCP file mapping + Codex's streamable HTTP MCP example), [[claude-md]] (added cross-vendor convention-file table; AGENTS.md ≡ Codex's CLAUDE.md).
• Updated topic (1): [[llm-engineering-workflow]] — Conductor docs noted as deep-ingest reference; new open thread on Conductor's position on the Cherny axis (orchestration-UI layer is about humans coordinating agent work, which Cherny's harness-shrinks thesis specifically doesn't claim model improvement replaces).
• Updated: [[index.md]] — sources 42→82 (added 40 conductor source pages); concepts 51→53 (added [[workspace-isolation]], [[spotlight-testing]]); last-updated stamp.
• Major findings from primary text:
  • Workspace isolation is a real named primitive — the five-level model in [[cond-workspaces-and-branches]] is more rigorous than the existing vault treatment. One workspace = one branch = one working tree = one running env. Now extracted to its own concept page.
  • 295-city naming pool ([[cond-cities]]) — probability table puts 99% likelihood of seeing all cities at ~26,964 spawns. At Tan-scale usage (10–50 PRs/day), exhausting the pool takes 200+ days. Useful upper-bound anchor for "how parallel can this realistically go."
  • Spotlight testing is a named pattern for the run-isolation-vs-shared-stack tradeoff. Mechanism: workspace edits → on-change checkpoint commit → check out at repo root. One-way sync. Tracked files only.
  • 7 agent presets in Big Terminal Mode ([[cond-big-terminal-mode]]): Claude, Codex, OpenCode, Amp, Pi, Copilot, Gemini. First vault confirmation Conductor is truly cross-vendor — not just Claude + Codex talk.
  • conductor:// URL scheme ([[cond-deep-links]]) — four link shapes, including async?repo=…&plan=<base64> which is a clean external-handoff primitive. Pairs naturally with [[handoff|Pocock's /handoff]] as the cross-tool delivery format.
  • Six repository-preference fields ([[cond-agent-behavior]]) injected into specific UI actions — Conductor's prompt-injection surface. Confirms [[cond-faq]]'s "Conductor injects system prompts that explain to the agent that it's running inside Conductor, what a workspace is, and other context."
  • MCP file format is identical between Cursor and Claude Code ([[cond-migrate-from-cursor]]) — copy-paste migration. Codex uses TOML. Codex supports streamable HTTP MCP — first such example documented in vault.
  • Codex's instruction-file convention is AGENTS.md ([[cond-migrate-from-cursor]] + [[cond-agent-behavior]]) — now documented on [[claude-md]] with the three-way Cursor / Claude / Codex table.
  • CONDUCTOR_PORT allocates 10 ports per workspace ([[cond-environment-variables]]) — the dev-server-parallelism primitive. Pair with nonconcurrent runScriptMode for projects that can't tolerate parallel servers.
  • Managed settings ([[cond-settings]]) — orgs can deploy ~/.conductor/settings.json with 4 supported keys. JSON Schema published at conductor.build/schemas/settings.json for editor autocomplete. First vault example of an agent-host with a published settings schema.
  • No-sandbox stance is doubled down: [[cond-security-and-permissions]] + [[cond-faq]] both explicit. "Agents run directly on your system without sandboxing." "Workspace isolation is development isolation, not a security boundary."
• Wikilink stubs deferred (not yet pages): OpenCode, Amp, Pi (3 of the 7 Big Terminal Mode presets — already mentioned in [[gstack]] but not entities yet), Fly Postgres, PostHog (Conductor's analytics stack — too tangential unless they recur), .worktreeinclude convention (could become a tiny concept page if it propagates), Linear (issue-tracker integration, recurring across vault). The 5 changelog-derived model anchors (Opus 4.7, Sonnet 4.6, Haiku 4.5) are already vault-known from [[claude-code]] and [[cherny-coding-solved]].
• Schema notes: No CLAUDE.md schema changes required. source_kind: repo-doc covers all 40 pages. The slug-prefix convention cond- mirrors the existing cc- convention for Anthropic docs.
• What's NOT in this ingest: 165 changelog entries from the sitemap. Each is a per-release set of bug fixes / model bumps; not load-bearing for the wiki. The substantive feature names already surface in the docs pages (e.g., checkpoints, plan mode, multiple chats, diff viewer). If a specific changelog entry becomes interesting (e.g., the gstack-Conductor integration release), ingest individually.
• Follow-up questions: see [[conductor]] entity → Open threads / Direct tensions; [[workspace-isolation]] + [[spotlight-testing]] → Open questions; [[llm-engineering-workflow]] → Open threads (Conductor's position on the Cherny axis).

[2026-05-23] ingest | gstack — Garry Tan's Claude Code skill kit

First third-party-ecosystem source on [[claude-code]] outside [[matt-pocock|Pocock]]'s output. [[gstack]] (by [[garry-tan|Garry Tan]], YC President & CEO) is the opposite philosophical pole from Pocock: 30+ role-organized skills enforcing a fixed [[sprint-pattern|sprint sequence]] (Think → Plan → Build → Review → Test → Ship → Reflect) with /autoplan orchestrator and cross-host distribution to 10 AI coding agents. This adds a fourth perspective to the [[llm-engineering-workflow]] topic alongside Karpathy / Anthropic / Pocock.

• Source moved: raw/garrytangstack Use Garry Tan's exact Claude Code setup….md → raw/articles/gstack-readme.md.
• Created sources (1): [[gstack-readme]].
• Created entities (6): [[gstack]], [[garry-tan]], [[y-combinator]], [[conductor]], [[openclaw]], [[gbrain]].
• Created concepts (5): [[sprint-pattern]], [[role-based-skill-kit]], [[parallel-agentic-sessions]], [[cross-model-review]], [[diataxis]].
• Updated entities (4): [[andrej-karpathy]] (cited at gstack opening; flagged unverified No Priors podcast + forrestchang/andrej-karpathy-skills repo refs), [[matt-pocock]] (now has heavyweight peer [[gstack]] for contrast), [[anthropic]] (third-party heavyweight kit added to ecosystem related), [[claude-code]] (gstack added as primary host + Conductor/OpenClaw/GBrain related).
• Updated concepts (2): [[agent-skills]] (added the two-poles framing; pocock-lightweight vs gstack-heavyweight; gstack added to Sources), [[auto-memory]] (added /learn and gstack memory sync + cross-link to [[gbrain]]).
• Updated topics (1): [[llm-engineering-workflow]] — restructured from three to four converging perspectives. Added the heavyweight pole map, the Pocock-vs-gstack contrast table, parallel-sessions as a fourth form of leverage, and several new open threads (Cherny-vs-gstack on harness thickness; memory layering; cross-model review as routine; OpenClaw credibility; gstack productivity claims).
• Updated: [[index.md]] (sources 25→26, concepts 41→46, entities 15→21).
• Notable / unresolved:
  • OpenClaw credibility still unresolved. This is the second source to mention "OpenClaw" — first was [[karpathy-vibe-to-agentic|Karpathy's transcript]] where the prior ingest flagged it as a likely caption error. [[gstack-readme]] confidently treats it as a real product (github.com/openclaw/openclaw, 247K stars, author Peter Steinberger). Neither verified in vault. [[openclaw]] page documents the unresolved status and lists what to check next.
  • Productivity claims are author-attributed, not vault-verified. 810×/240× framings and the methodology doc (docs/ON_THE_LOC_CONTROVERSY.md) flagged as claims rather than facts; same for the forrestchang/andrej-karpathy-skills 17K-star reference.
  • Strong tension surfaced — [[boris-cherny|Cherny]] vs gstack on harness thickness. Cherny argues the harness shrinks as the model improves; gstack bets on a thick role-organized harness with /careful / /freeze / /guard safety scaffolding. Documented on [[gstack]] entity, [[role-based-skill-kit]], and [[llm-engineering-workflow]] open threads. Strong synthesis candidate for future writing.
  • Pocock-vs-gstack philosophical contrast is now the topic's central comparative axis. Documented in [[role-based-skill-kit]] (the pattern), [[agent-skills]] (the umbrella), and the [[llm-engineering-workflow]] table. Both shipped, both popular; the split appears to map onto user type (experienced → atomic verbs, first-time → role labels).
  • First-vault appearance of: [[cross-model-review|cross-model review as a kit feature]] (/codex), [[diataxis|Diataxis as agent-kit organizing schema]], cross-vendor browser-coordination primitive (/pair-agent), productized prompt-injection defense (22MB classifier + Haiku transcript check + canary token + verdict combiner), iOS live-device QA over USB CoreDevice with opt-in tailnet exposure.
• Wikilink stubs deferred (not yet pages): Peter Steinberger, AquaVoice, Whisper, Greptile, Codex CLI, OpenCode, Cursor (as agent kit, not editor), Factory Droid, Slate, Kiro, Hermes, ClawHub, Conductor's Mike Krieger / Anthropic Labs (still), Daniele Procida (Diataxis author), ACP (Agent-to-agent Communication Protocol), forrestchang/andrej-karpathy-skills repo. Each could become an entity if it recurs.
• Follow-up questions: see updated topics/llm-engineering-workflow → "Open threads"; per-page "Open questions" sections in the 5 new concept pages and the new source page.
• Schema co-evolution note: no changes to CLAUDE.md needed for this ingest. The existing source-kind repo-doc covered the README; raw-organization conventions (raw/articles/) held.

[2026-05-23] ingest | gstack — repo deep ingest (50 SKILL.md files + meta docs)

Follow-up to the README-only ingest. User asked: "did u ingest the specific skills that are in tans repository?" — answer at the time was "no, only the README." This batch closes that gap: cloned github.com/garrytan/gstack (depth-1) and copied all 50 SKILL.md files plus repo meta-docs into raw/skills/gstack/ preserving upstream structure.

• Repo verified: MIT-licensed, Copyright (c) 2026 Garry Tan, version 1.43.3.0. LICENSE confirms authorship.
• Raw copy: 50 per-skill directories under raw/skills/gstack/<skill>/SKILL.md, plus README.md, ETHOS.md, AGENTS.md, ARCHITECTURE.md, CLAUDE.md, root SKILL.md, VERSION, LICENSE, and 4 docs (OPENCLAW.md, ON_THE_LOC_CONTROVERSY.md, skills.md, explanation-diataxis-in-gstack.md).
• Created sources (15): [[gstack-repo-deep]], [[gstack-ethos]], [[gstack-office-hours]], [[gstack-plan-ceo-review]], [[gstack-cso]], [[gstack-review]], [[gstack-qa]], [[gstack-codex]], [[gstack-investigate]], [[gstack-ship]], [[gstack-learn]], [[gstack-autoplan]], [[gstack-retro]], [[gstack-loc-controversy]], [[gstack-openclaw-doc]].
• Created concepts (4): [[boil-the-lake]], [[search-before-building]], [[user-sovereignty]], [[iron-law]].
• Created entities (2): [[simon-willison]] (cited in ETHOS), [[ben-vinegar]] (slop-scan, cited in LOC controversy).
• Major updates (significant rewrites):
  • [[openclaw]] — resolved partially. Now: real ACP-based EA orchestrator per [[gstack-openclaw-doc]]; Steinberger/247K-stars framing remains unverified (not in gstack's architectural docs).
  • [[gstack]] — actual skill count corrected to 50 (not 23/30+). Real usage stats added (75K stars, 14,965 installs, 305,309 invocations, top-skill table). Auto-generated template architecture documented. OPENCLAW_SESSION env var contract.
  • [[garry-tan]] — productivity claims now methodology-backed (Tan/Vinegar/script-driven). New frameworks-pushed-into-vault rows.
  • [[andrej-karpathy]] — added "Iron Man suit" framework row + updated mentions.
  • [[role-based-skill-kit]] — direct skill-size comparison table (5 sentences for Pocock's /grill-me vs 3,093 lines for gstack's /ship). The two-orders-of-magnitude evidence.
  • [[sprint-pattern]] — primary-text artifact path conventions (~/.gstack/projects/<slug>/*-design-*.md).
  • [[cross-model-review]] — Filesystem Boundary + DIFF_START/DIFF_END delimiters + known-bad version list + "Codex caught 15+ findings four Claude reviews missed" data point.
  • [[agent-skills]] — gstack frontmatter vocabulary (preamble-tier, gbrain: context queries, triggers, benefits-from).
  • [[auto-memory]] — /learn mechanics: append-only JSONL at ~/.gstack/projects/<slug>/learnings.jsonl, latest-wins, confidence scores, "Prior learning applied" surfacing.
  • [[gstack-readme]] — caveat updates reflecting partial resolution of OpenClaw + LOC controversy.
• Updated: [[index.md]] (sources 26→41, concepts 46→50, entities 21→23).
• Major findings from primary text:
  • 50 SKILL.md files (not 23). README's title number is a stylistic undercount; AGENTS.md is the accurate catalog.
  • Skill-size data is the central evidence for [[role-based-skill-kit]]: Pocock's /grill-me = 5 sentences; gstack's /ship = 3,093 lines. Two-to-three orders of magnitude.
  • Auto-generated from SKILL.md.tmpl: every skill = ~750-line common preamble + skill-specific body. The boilerplate auto-injects the three [[gstack-ethos|ETHOS principles]].
  • gbrain: frontmatter is structured, not advisory — declares context_queries (filesystem globs, list filters) the runtime pulls before the skill body. Real federated-memory protocol per-skill.
  • OPENCLAW_SESSION env var contract baked into every preamble. Confirms [[openclaw]] integration at the runtime level.
  • The Iron Law ([[gstack-investigate]] verbatim): "NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST." Three phases (root cause → pattern → hypothesis) with 3-strike rule. Auto-freezes scope to affected module.
  • /plan-ceo-review's 18 cognitive patterns are a real reading list operationalized: Bezos (one-way/two-way doors, Day 1, regret minimization), Munger (inversion), Grove (paranoia, wartime), Jobs (focus as subtraction), Horowitz (people-first, wartime), Hastings (talent density), Altman (willfulness, leverage), Chesky/Graham (founder mode), Rams (subtraction default).
  • /codex's Filesystem Boundary + DIFF_START/DIFF_END delimiters are real cross-AI hygiene primitives — defensive against prompt confusion and prompt injection.
  • gstack-loc-controversy documents the methodology behind 810× productivity claims with sensitivity analysis (at pathological 10× verbosity, still 81×). Includes [[ben-vinegar|Vinegar]]'s slop-scan 5.24 → cut 62% in one session. Real-usage stats: 75K stars in 5 weeks, 14,965 installations, 305,309 skill invocations.
• OpenClaw status update (was the biggest open question): architecturally real per docs/OPENCLAW.md (ACP runtime + sessions_spawn + ClawHub registry + OPENCLAW_SESSION env var). Steinberger/247K-stars framing NOT in gstack's architectural docs — appears only in the README's intro prose. Net update: treat OpenClaw as a real EA-shaped peer orchestrator; treat the celebrity-builder claim as rhetorical color.
• Wikilink stubs still deferred: Daniele Procida (Diataxis author), Peter Steinberger, Clawvisor, SOUL.md convention, slop-scan tool (could become entity if recurring), Bookface, plus the previous batch's deferred list (AquaVoice, Whisper, Greptile, etc.).
• What's NOT in the deep ingest: 35 SKILL.md files not given individual source pages (the iOS skills, design skills, browser skills, setup helpers, power tools beyond /freeze). They're in raw/skills/gstack/ for any future concept page to draw from.
• Lesson for future deep ingests: when a kit ships multiple times more SKILL.md content than the previous lightweight comparator, measure. The 5-sentences-vs-3,093-lines data point would have been invisible from the README alone. Primary text is where the real differences live.

[2026-05-23] synthesis | pocock-vs-gstack

User asked how [[garry-tan|Tan]]'s software practices relate to [[matt-pocock|Pocock]]'s — shared philosophy? similar agent workflows?

• Created: wiki/syntheses/pocock-vs-gstack.md — first synthesis page in the vault.
• Thesis: same diagnosis (agents need encoded named workflows; no blank prompts; root-cause discipline; HITL), opposite scaffolding. Workflows isomorphic at the shape level (interview → plan → build w/ tests → review → iterate) but 2–3 orders of magnitude apart in skill size and process rigidity. gstack extends through ship/reflect that Pocock leaves out-of-scope.
• New analytical primitive surfaced: the Cherny axis as the long-term-bet decider — Pocock's atomic skills implicitly bet with [[boris-cherny|Cherny]]'s harness-shrinks thesis ([[cherny-coding-solved]]); gstack's 3,093-line /ship bets against it. Mutually exclusive predictions; 2–3 model generations until verdict.
• Canon-divergence observation: Pocock cites Beck/Ousterhout/Evans/Feathers/Pragmatic Programmer (software-craft canon). gstack cites Willison + Karpathy (AI-pragmatism canon). Coherent stylistic tell.
• Updated: [[index.md]] (syntheses 0→1; last-updated stamp), [[garry-tan]] (related), [[matt-pocock]] (related).
• Contradictions: none new — synthesis builds on [[role-based-skill-kit]] without conflicting with existing pages.
• Open threads forwarded: hybrid kits (Pocock-atomic + gstack-role-layer); 100-skill catalog limit; the Cherny verdict.

[2026-05-24] ingest | tan-ai-engineering-team — Garry Tan's YC video on gstack

Tan's public introduction to [[gstack]] on the [[y-combinator|YC]] YouTube channel (2026-04-23). The first primary source where Tan himself (vs the README) names the design slogan, walks through /office-hours → adversarial review → /design-shotgun live, and articulates the user-facing cross-model mental model. Closes the prose-without-Tan's-voice gap from [[gstack-readme]] / [[gstack-repo-deep]].

• Source moved: raw/How to Make Claude Code Your AI Engineering Team.md → raw/transcripts/tan-ai-engineering-team.md.
• Created sources (1): [[tan-ai-engineering-team]].
• Created concepts (1): [[thin-harness-fat-skills]] — Tan's named architectural slogan. Partitions the Cherny axis precisely: agrees the runtime gets thinner, refuses that workflows (skills) get absorbed with it.
• Created entities (3): [[palantir]], [[posterous]], [[bookface]] — Tan's pre-YC biographical narrative (now anchors the productivity-claim credibility chain).
• Updated entities (4): [[garry-tan]] (full bio narrative, "thin harness fat skills" + "ADHD CEO/autistic CTO" frameworks, ~400 PRs + 10–50/day personal anchors, 70K-stars/3-weeks origin), [[gstack]] (28 vs 50 skill-count reconciliation, /qa + /browse origin story, /design-shotgun cross-model image-gen routing, adversarial-review numbers 6→8/16-issues-autofixed, "level 7 of 8 software factory"), [[conductor]] ("Gary mode" reasoning-trace surface + +-icon work-item creation flow + recommended-onboarding-host status), [[claude-code]] (Tan's "ADHD CEO / autistic CTO" mental model layered on the model menu).
• Updated concepts (5): [[agent-skills]] (thin-harness-fat-skills as gstack's design principle), [[sprint-pattern]] (live-demo walkthrough table + adversarial-review numbers), [[parallel-agentic-sessions]] (Tan's first-person 10–15 anchor + +-icon-replaces-TODO flow), [[cross-model-review]] (ADHD CEO/autistic CTO mental model + second use of cross-model routing for modality via /design-shotgun → Codex image-gen), [[role-based-skill-kit]] (Tan's verbal articulation of the role-named rationale: "a team with roles with process with review").
• Updated topic (1): [[llm-engineering-workflow]] — "thin harness, fat skills" added to the gstack lens description; the Cherny-axis open thread refined to its precise form (runtime vs workflow layer); two new open threads (the "level 8 software factory" scale; the 28-vs-50 skill-count reconciliation).
• Updated synthesis (1): [[pocock-vs-gstack]] — the Cherny-axis section now distinguishes runtime-axis from skill-axis. Tan's bet is now articulable as: with Cherny on the runtime, against Cherny on the workflow layer. Mutual exclusivity sharpened.
• Updated: [[index.md]] (sources 41→42, concepts 50→51, entities 23→26, last-updated stamp).
• Major findings from primary text:
  • The "thin harness, fat skills" slogan is Tan's own. Articulated verbatim in this transcript and no earlier vault source. It is the cleanest single-sentence articulation of gstack's architectural identity and the cleanest engagement with Cherny's harness-shrinks thesis: agreement on the runtime axis, refusal on the workflow axis. Worth tracking as a named primitive.
  • /qa origin story: built because Claude in Chrome MCP was "one of the worst pieces of software I've ever used." Wrapped Playwright at the CLI level instead. First vault example of MCP failure motivating CLI replacement — relevant to the broader [[mcp]]-vs-direct-tool tradeoff.
  • /design-shotgun farms image-gen to OpenAI Codex — first vault detail. Cross-model routing for modality, not just code review. Pattern generalizes beyond /codex.
  • Adversarial review concrete numbers: 6/10 → 8/10, 16 issues auto-fixed, 3 deferred. First numerical anchors for what /plan-ceo-review actually produces.
  • Posterous comparable-output anchor: "essentially built all of Posterus" — a project that originally took 2 years, $10M, 10 engineers — solo + agents in months. The most legible single before/after framing in the vault. Complements (doesn't replace) the LOC-based 810×/240× framings in [[gstack-loc-controversy]].
  • The "ADHD CEO / autistic CTO" framing is the user-facing mental model for [[cross-model-review]]. Tells the user when to invoke /codex, not just how.
  • Conductor's + icon replaces the inbox — Tan: "I don't have a to-do list anymore." First example in the vault of a parallel-session orchestrator with externalized-TODO ergonomics.
  • Skill-count discrepancy: README says ~23, Tan says 28, repo file count is 50. Most plausible split: user-facing slash commands (28) vs all SKILL.md files including setup/iOS/power tools (50). Documented on [[gstack]] entity for future resolution.
• Wikilink stubs deferred (not yet pages): level-8-software-factory (the implied scale Tan references; source unresolved); "Gary mode" (Conductor feature — fine inline on [[conductor]]); CPA marketplace / 1099 aggregator demo app (just a demo, no persistent entity); the YC partner office-hours method (referenced as the seed for /office-hours; could become a topic if YC-method ingest expands).
• Transcript artifact note: the YouTube auto-caption introduces several speech-to-text errors that future search might hit — "Boris Churnney" (= [[boris-cherny|Cherny]]), "Andre Karpathy" (= [[andrej-karpathy|Karpathy]]), "Palunteer" (= [[palantir|Palantir]]), "Posterus" (= [[posterous|Posterous]]), "GSAC" (= [[gstack|gstack]]), "OOTH" (= OAuth), "1099 ins" (= 1099-INTs). Flagged on the source page.
• Contradictions / tensions surfaced:
  • vs [[gstack-readme]]'s 23: video says 28. Repo file count is 50. Discrepancy reconciled as a user-facing-vs-total-files split (inference).
  • vs [[gstack-loc-controversy]]'s 75K stars: video says 70K. Both are snapshots; growth is rapid. Not contradictory.
  • vs Cherny's harness-shrinks thesis: tightened, not contradicted — Tan agrees the runtime shrinks but partitions skills out of "harness" and bets they stay fat. The disagreement is now precise rather than thick-vs-thin all-the-way-down.
• Schema co-evolution note: added a new informal source_kind: talk value for this YouTube screencast/talk hybrid. Already present elsewhere as "transcript" — both work; this transcript is also a recorded talk, so talk reads more accurately. No CLAUDE.md change needed; the source_kind list is already non-exhaustive.
• Follow-up questions: see [[tan-ai-engineering-team]] → "Open questions"; updated open threads on [[llm-engineering-workflow]].

[2026-05-27] ingest | sandcastle (repo) + pocock-afk-factory (video)

Ingested [[matt-pocock|Pocock]]'s [[sandcastle]] — a provider-agnostic TypeScript library for running coding agents in isolated [[agent-sandboxing|sandboxes]] — from both the launch video and a deep clone of github.com/mattpocock/sandcastle (v0.6.4). First execution + isolation-layer tool in the vault; first tool that ships real sandboxing (containers/microVMs) rather than [[workspace-isolation|worktree isolation]].

• Raw added: cloned repo → raw/repos/sandcastle/ (2.8M, .git stripped). New raw/ subfolder convention raw/repos/ for cloned code libraries (vs raw/skills/ for skill kits) — recorded in CLAUDE.md.
• Source moved: raw/I Open-Sourced My Own AFK Software Factory.md → raw/transcripts/pocock-afk-factory.md (matches source slug).
• Created sources (2): [[pocock-afk-factory]] (transcript), [[sandcastle-repo]] (repo-doc — README + CONTEXT.md + CLAUDE.md + 17 ADRs + templates + permissions research doc).
• Created entities (1): [[sandcastle]] (entity_kind: library).
• Created concepts (2): [[afk-agents]] (the autonomy axis — unattended agents; distinct from temporal [[agent-loops]] and spatial [[parallel-agentic-sessions]]), [[agent-sandboxing]] (isolation as a security boundary; the spectrum none → worktree → container → microVM; branch strategies).
• Updated entities (3): [[matt-pocock]] (work expands from skills into execution infra), [[ai-hero]] (@ai-hero npm scope publishes Sandcastle), [[claude-code]] (CC as the default agent provider invoked inside Sandcastle sandboxes).
• Updated concepts (4): [[parallel-agentic-sessions]] (added a third scaling pattern — "programmatic factory" via Promise.allSettled over sandboxes), [[workspace-isolation]] (resolved the long-standing "compare to remote-runner sandboxes" open thread — Sandcastle supplies the security boundary Conductor declines to), [[cross-model-review]] (the execution-layer expression — vendor as a swappable agent provider per run(); quotes Pocock on adversarial review + best-of-N), [[agent-loops]] (Sandcastle's code-driven iteration/backlog loop + its deliberate "RALPH" terminology ban).
• Updated topic (1): [[llm-engineering-workflow]] (Sandcastle added to platform map; orchestration now has two surfaces — GUI/Conductor vs code/Sandcastle; new [[agent-sandboxing]] + [[afk-agents]] entries under Permissions/safety; new "Sandcastle vs Conductor" synthesis-candidate open thread).
• Updated: [[index.md]] (sources 82→84, concepts 53→55, entities 26→27; new source + concept groups; last-updated stamp).
• Major findings from primary text:
  • The thesis in one line: AFK autonomy is bought with isolation. To run agents unattended you must bypass all permission prompts; doing that on the bare host is reckless ("delete your home directory… exfiltrate data"); a sandbox makes "never ask, never escape" safe. This is the single load-bearing idea behind Sandcastle.
  • Sandcastle is the direct counterpoint to [[conductor|Conductor]]: both solve parallel agents, but Conductor is explicit it is "development isolation, not a security boundary" and "agents run directly on your system without sandboxing." Sandcastle exists precisely to provide that boundary. (It uses worktrees under a container — the two compose.)
  • Unopinionated by design: ships no workflow/task model — "you write the prompt, and the engine executes it." Orchestration is a plain .sandcastle/main.ts you own. Contrast gstack's fat [[sprint-pattern|sprint]] kit.
  • Provider-agnostic is the enabling bet: agent is swappable (Claude Code / Codex / Cursor / Pi / OpenCode / Copilot) → cross-model adversarial review and best-of-N implementations are one-token changes. Sandbox is swappable too (Docker / Podman / Vercel microVM / Daytona / no-sandbox).
  • CONTEXT.md is a textbook [[ubiquitous-language]] instance and pointedly bans "RALPH" for the iteration loop, "container" (too specific), and "workspace" (reserved) — strong vocabulary hygiene worth citing.
  • docs/research/permissions-systemic-fix.md is a candid diagnosis of why container sandboxing is hard (host-UID≠image-UID, single-file bind-mount parent ownership, missing SELinux :z labels) — concrete evidence for why Pocock found off-the-shelf Docker sandboxes painful AFK.
  • !`command` shell-expansion prompt syntax is explicitly "copied from Claude skills" — a cross-pollination data point between the skills ecosystem and this tool.
• Contradictions / tensions surfaced: none against existing pages. Sandcastle resolves a tension previously only flagged as an open question on [[workspace-isolation]] (worktree isolation ≠ security boundary).
• Schema co-evolution: added raw/repos/ to the CLAUDE.md raw-subfolder list (cloned code libraries, vs skill kits in raw/skills/).
• Wikilink stubs deferred (not yet pages): [[codex]] (now 3+ mentions across vault — gstack-codex, cross-model-review, sandcastle — a genuine missing-page candidate per lint rules), [[host]] (Sandcastle CONTEXT.md term; fine inline for now), Docker / Podman / Vercel / Daytona / Beads / Effect (provider/dep entities — left as unresolved links, low compounding value as stubs).
• Suggested follow-ups / synthesis offer: a synthesis page "Sandcastle vs Conductor — code+containers vs GUI+worktrees" is a strong candidate (parallels [[pocock-vs-gstack]] one layer down, at the execution tier). Offered to user, not auto-created.

[2026-05-27] lint | full health-check + fix pass

Ran the lint workflow across all 171 wiki pages and fixed every actionable finding ("fix all").

• Broken links repaired (5): [[context-window-concept]]→[[context-window]] (×2: cc-context-window, cc-how-it-works); [[ousterhout]]→[[john-ousterhout]] (agentic-engineering); [[syntheses|synthesis]] de-linked to plain text (×2: pocock-skills-repo, personal-knowledge-management — syntheses is a folder, never a page).
• Stale claim fixed: [[matt-pocock]] "9 of the 10 ingested sources are his" (true at 10 sources; vault now has 84) → "most frequently recurring author… 13 source pages."
• Missing pages created (2): [[michael-feathers]] (originator of [[seam]]; completes Pocock's engineering canon) and [[codex]] (OpenAI's coding agent; the recurring non-Anthropic agent across [[cross-model-review]], [[sandcastle]], [[conductor]]). Both wired bidirectionally; index entities 27→29.
• Orphans fixed (4): linked the 4 previously-orphaned Conductor doc sources ([[cond-introduction]], [[cond-issue-to-pr]], [[cond-keyboard-shortcuts]], [[cond-slash-commands]]) from the [[conductor]] entity's Mentions.
• Frontmatter standardized: added self-referential sources: [<slug>] to the 25 older source pages that lacked it (cc-, pocock-, cherny, karpathy, llm-wiki, etc.), matching the newer cond-/gstack- convention. Every wiki page now carries sources:.
• Bidirectional links (171 gaps → 0): added a consolidated "Also linked from:" reciprocal-navigation bullet to the Related/Variants section of 53 target pages. Chose one consolidated bullet per page (vs 171 scattered bespoke bullets) to honor the bidirectional rule without cluttering the curated Related sections — reads as an explicit "what points here" aid.
• Verification: re-ran the full lint — reciprocity gaps 0, orphans 0, frontmatter gaps 0, all prior broken links resolved.
• Left intentionally: deferred 1-ref wikilink stubs [[host]], [[continuous-checkpoint-mode]], [[gstack-plan-eng-review]] (below the 3+ missing-page threshold); [[wikilinks]] literal-in-backticks prose (false positive); the 4 single-source peripheral entity stubs ([[marp]], [[obsidian-web-clipper]], [[palantir]], [[qmd]]) — flagged but fine as-is.

[2026-05-27] prune | Conductor docs collapse

User flagged the Conductor docs ingest as content that shouldn't be in the vault — 40 of 84 source pages (48%) were one third-party product's documentation, much of it product-reference trivia (cities, keyboard shortcuts, diff viewer, deep links, privacy/PostHog, shells, settings) that doesn't compound and goes stale on the next Conductor release. Approved a collapse.

• Deleted (33 source pages): cond-introduction, cond-first-workspace, cond-installation, cond-faq, cond-testing, cond-how-to-run, cond-issue-to-pr, cond-migrate-from-cursor, cond-review-and-merge, cond-linking-multiple-directories, cond-monorepos, cond-agent-behavior, cond-big-terminal-mode, cond-checkpoints, cond-checks, cond-cities, cond-deep-links, cond-diff-viewer, cond-environment-variables, cond-files-to-copy, cond-keyboard-shortcuts, cond-mcp, cond-privacy, cond-security-and-permissions, cond-settings, cond-shells, cond-slash-commands, cond-todos, cond-scripts, cond-scripts-setup, cond-scripts-run, cond-scripts-share-with-teammates, cond-troubleshooting-issues.
• Kept (7 concept-feeding source pages): [[cond-workspaces-and-branches]] (primary text for [[workspace-isolation]]), [[cond-scripts-spotlight-testing]] (primary for [[spotlight-testing]]), [[cond-parallel-agents]], [[cond-workflow]], [[cond-agent-modes]], [[cond-providers]], [[cond-conductor-json]].
• Nothing lost: all 40 raw docs remain untouched in raw/docs/conductor/ (storage layer). The salient facts of the deleted pages were already absorbed into the [[conductor]] entity — env vars, scripts, settings, preference fields, Big Terminal presets, checkpoints, cities, deep links, MCP, security posture. Added a consolidation note to the entity and converted ~20 dangling _(per [[cond-X]])_ citations to plain "per Conductor docs."
• Inbound links repaired (survivors): [[spotlight-testing]], [[mcp]], [[claude-md]], [[agent-sandboxing]], [[workspace-isolation]], [[parallel-agentic-sessions]] (concept pages) + the 7 surviving cond-* pages' Related sections — all redirected to [[conductor]] or plain text; frontmatter sources: arrays cleaned of deleted slugs.
• Updated: [[conductor]] entity (frontmatter sources 42→9, consolidation banner, severed dangling links, updated 2026-05-27); [[index.md]] (sources 84→51; Conductor block 40 lines → 7 + collapse note); README (~84/~169 → ~51/~138).
• Verification: grep confirms 0 remaining wikilinks to any deleted slug and 0 deleted slugs in any frontmatter sources: across wiki/.
• Rationale (compounds back into schema thinking): the wiki is the navigable surface; raw/ is storage. Mirroring a vendor's docs site into the navigable layer violated that split. Going forward, ingest product docs into the relevant entity/concept pages and keep only the pages that feed durable concepts as standalone sources.

[2026-05-27] ingest | pocock-dictionary — Dictionary of AI Coding (62 entries)

Single-source ingest, large surface area. The whole mattpocock/dictionary-of-ai-coding GitHub repo: 62 vocabulary entries in a 7-section curriculum + the published aihero.dev/ai-coding-dictionary site.

• Raw: copied repo (minus .git) to raw/repos/dictionary-of-ai-coding/ (CLAUDE.md, README.md, dictionary/.md, internal/.md, package.json, .prettierrc.json).
• Created sources (1): [[pocock-dictionary]] — comprehensive source page with curriculum structure, 13 key claims, notable quotes, open questions, and relationship to existing wiki pages.
• Created concepts (55): the canonical AI-coding operational vocabulary, organized to mirror Pocock's curriculum:
  • Model & mechanics (14): [[model]], [[parameters]], [[training]], [[inference]], [[token]], [[next-token-prediction]], [[non-determinism]], [[model-provider]], [[harness]], [[model-provider-request]], [[input-tokens]], [[output-tokens]], [[prefix-cache]], [[cache-tokens]].
  • Sessions, context, turns (7): [[context-noun]], [[stateless]], [[stateful]], [[agent]], [[system-prompt]], [[session]], [[turn]].
  • Tools & environment (9): [[environment]], [[filesystem]], [[tool]], [[tool-call]], [[tool-result]], [[permission-request]], [[permission-mode]], [[agent-mode]], [[sandbox]].
  • Failure modes (9): [[sycophancy]], [[hallucination]], [[parametric-knowledge]], [[contextual-knowledge]], [[knowledge-cutoff]], [[attention-relationship]], [[attention-budget]], [[attention-degradation]], [[smart-zone]].
  • Handoffs (6): [[clearing]], [[compaction]], [[autocompact]], [[handoff-artifact]], [[spec]], [[ticket]].
  • Memory & steering (4): [[memory-system]], [[agents-md]], [[progressive-disclosure]], [[context-pointer]].
  • Patterns of work (6): [[human-in-the-loop]], [[automated-check]], [[automated-review]], [[human-review]], [[design-concept]], [[grilling]] (bonus — promotes the named technique behind /grill-me / /grill-with-docs to its own concept).
• Updated existing pages (10):
  • [[vibe-coding]] — added Pocock's sharpening: "names the review stance, not the resulting code quality." Explicitly rejects the "low-quality" connotation.
  • [[mcp]] — added Pocock's mechanical "agent never calls MCP; harness happens to have gotten the tool from MCP" framing.
  • [[subagents]] — added Pocock's "tree is one level deep" rule and the subagent-vs-handoff distinction (return path vs no return path).
  • [[agent-skills]] — added Pocock's skill-vs-tool clarification ("a tool is what the agent calls; a skill is instructions it reads").
  • [[handoff]] — restructured to use Pocock's named hierarchy: clearing < compaction < handoff-artifact < spec < ticket vs subagent (the only one with a return path).
  • [[dumb-zone]] — added the mechanistic chain attention-relationship → attention-budget → attention-degradation → dumb-zone; two-flavor symptom diagnosis.
  • [[afk-agents]] — added Pocock's "AFK centers the human pattern, not the machine" naming defense against "background agent."
  • [[agent-sandboxing]] — added cross-reference to the new short [[sandbox]] entry.
  • [[prompt-caching]] — paired with new [[prefix-cache]] (mechanism) and [[cache-tokens]] (billing slice) concept pages.
  • [[claude-md]] — added the [[agents-md]] umbrella-vs-implementation relationship.
  • [[matt-pocock]] (entity) — added the dictionary as a third workstream (skills + sandcastle + dictionary).
  • [[grill-me]] — linked to new [[grilling]] technique and [[design-concept]] target.
• Updated: [[index.md]] (sources 51→52; concepts 55→110, new vocabulary supersection across 7 thematic groups); [[matt-pocock]] entity (added dictionary workstream + 2026-05-27 update).
• Contradictions: none across sources. Several sharpenings of existing pages: vibe-coding had a softer definition; subagents lacked the "one level deep" rule; AGENTS.md was previously treated as a side note under [[claude-md]] but is properly the umbrella convention.
• Notable quotes captured: "Same model, same parameters, same harness — just more mouths to feed from the same plate." / "The model never 'calls MCP'; it calls a tool, and the harness happens to have gotten that tool from an MCP server." / "A flaky test is a broken check, not a non-check." / "AFK centers the human pattern; 'background agent' centers the machine."
• Net new vocabulary the dictionary introduces to the wiki: [[harness]] (was implicit in [[thin-harness-fat-skills]] only); the attention-mechanism chain ([[attention-relationship]], [[attention-budget]], [[attention-degradation]]); the handoff hierarchy ([[handoff-artifact]] / [[spec]] / [[ticket]]); the verification-stance triad ([[automated-check]] / [[automated-review]] / [[human-review]]); [[design-concept]] (Brooks's term); [[grilling]] (technique name distinct from skills); [[agents-md]] as the canonical umbrella; [[prefix-cache]] / [[cache-tokens]] as the mechanism-level cache vocabulary.
• Follow-up: dictionary doesn't cover [[hooks]] / [[plugins]] / agent-loop / [[parallel-agentic-sessions]] / [[mcp]] security — scope is mechanical, not architectural. Worth checking if Pocock publishes follow-up sections.