philosopherstable

Imported from Ethiventure/philosopherstable (AGENTS.md). Install upstream with npx skills add Ethiventure/philosopherstable. Copyright stays with the author.

graphify

This project has a knowledge graph at graphify-out/ with god nodes, community structure, and cross-file relationships.

When the user types /graphify, use the installed graphify skill or instructions before doing anything else.

Rules:

• For codebase questions, first run graphify query "<question>" when graphify-out/graph.json exists. Use graphify path "<A>" "<B>" for relationships and graphify explain "<concept>" for focused concepts. These return a scoped subgraph, usually much smaller than GRAPH_REPORT.md or raw grep output.
• Dirty graphify-out/ files are expected after hooks or incremental updates; dirty graph files are not a reason to skip graphify. Only skip graphify if the task is about stale or incorrect graph output, or the user explicitly says not to use it.
• If graphify-out/wiki/index.md exists, use it for broad navigation instead of raw source browsing.
• Read graphify-out/GRAPH_REPORT.md only for broad architecture review or when query/path/explain do not surface enough context.
• After modifying code, run graphify update . to keep the graph current (AST-only, no API cost).

genealogy diagram

Rebuilding the influence map (or any node-and-debt diagram in this project's style): follow docs/diagram-style.md — vertical road system (Recipe B): time-ordered seats, bundled channel threads, one cubic per debt rim to rim, reciprocals drawn once, solid direct / dashed indirect, cradled end circles, full edge-list fallback. Content-agnostic: substitute any entities.

working rules (shared project norms)

• Plan before building anything non-trivial. Accessibility is in the initial plan, never a later pass — keep the house style, layer opt-in adjustments.
• Truthfulness: verify against live sources (API model lists, rate tables, link checks) before claiming. HMR/stale tabs lie — hard-refresh + fresh session before trusting a test. Verify in layers: lint + typecheck + build
  • retrieval-eval + live-transcript evidence.
• Provenance: mocks reference only what exists in the repo; never invent resources. Licence every corpus source; verify against the work's own copyright page, never an uploader's label. Imported code carries a note saying where it came from.
• Public app, no login, zero identifying data. Keys never in repo/bundle.
• No OpenAI models for generation, ever. Prefer good-value open-weight models; cost paid picks per session before recommending.
• Provider model IDs rot in days: free-text ID field + live key check, never a pinned dropdown; dead IDs fall back to the default on load.
• Debts table (CABINET_DEBTS) is truth; heirs derive. Model verdicts go in docs/models-tried.md — corpses stay buried, never re-test without new evidence. PLAN.md (agents) + README.md (humans) stay current. Structural lessons for the owner accumulate in docs/things-taught.md (plain words, newest first) — add one entry per lesson.
• Fail-soft everywhere with visible reasons; every failure preserves progress and offers recovery. UI copy keeps one speaker per surface.
• Voice and tone: chats in plain everyday language (~IELTS 6), jargon unpacked; no superlatives, no flattery, no emotional validation. UI copy uses domain diction with banned stock verbs — homogenised prose is a defect. When the evidence is good, public-facing text states the claim plainly (no hedging); caveats live in code comments, never on the surface.
• Beauty is a requirement: adjustments speak the design's palette and type; generic accessible-minimalism is a failure mode. Dark surfaces keep tested font/button/accent pairs — never a lone text-colour toggle.
• Quota-honest UX: per-user caps stated up front, visible halt states with recovery paths, provider/model switches that preserve progress.
• Model transparency in plain text: every provider and model option in Settings names its exact model ID and why it is picked (price, speed, voice, licence — one plain line each), rendered as visible text, never tooltip-only. The export names the model behind each turn. Source of truth lives in DEEPINFRA_PRIMARIES-style structures so UI copy and code cannot drift.
• Difficulty levels that hold (Low/Medium/High): define ALL levels, govern the simple one by removing difficulty machinery (not adding rules), every level gets a reader test + worked example + failure condition.
• First-run card: modal on first visit (what it is, honest quota picture, keys-never-leave-browser), "don't show again" default checked, Esc dismisses, header button reopens.

graphify workflow

First setup: graphify opencode install, then graphify . (--code-only without an LLM key). After edits: graphify update . (AST-only, no API cost). Install once per machine; per-project graph lives in graphify-out/ (dirty files after hooks are expected — not a reason to skip it).

free TTS recipe (Web Speech, no key)

Feature-detect, hide UI if absent. Set lang + rate only (device-default voice, no picker). Chunk ~200 chars at sentence boundaries, chain onend/onerror through one shared queue (generation counter vs overlap), cancel on unmount/clear/pagehide. Chrome silence is always voices-not-loaded or paused-synth: await voiceschanged (1.5s cap) + resume() in the click gesture. aria-live announces position, never full text.

RAG runbook (this repo)

• Truth: data/sources.json manifest (explicit rights gate — rights_status: approved with dated owner act, importer refuses the rest) → local data/rag.sqlite (gitignored) → shipped per-author shards in public/rag/ (+ manifest.json with schema/chunker versions).
• Corpus entries live in src/data/corpus-sources.ts; rag-manifest.mjs exports new ones as pending (never auto-approved). Licence every source (licence_status is free text: CC_BY_NC_ND_4_0, IN_COPYRIGHT…); verify against the work's copyright page, never an uploader's label.
• Fetch modes: remote HTML/text URLs, archive.org /stream/<id>/<file>_djvu.txt (HTML <pre>, indexable — /download/ raw text extracts zero paragraphs), and local_path for committed OCR/converted text whose canonical URL serves no fetchable text (scans: PyMuPDF render at 300dpi → tesseract over stdin pipe, commit text under data/local-text-*, canonical URL stays source_url).
• Chunking: heading-aware ~350w chunks, idempotent stable passage IDs; non-author front matter filtered before chunking. Eval (rag-eval.mjs, 12/12) records its chunker version and refuses stale contracts.
• Fail-soft per source with visible reasons; thin/link-dense/stub pages refused, never forced. Ship full text only with owner approval.