bochen2029-pixel

Everywhy — Development skill for Claude Code

Development community

Content x relation for coding agents: a phrase in, a budgeted read plan out -- pointers, never prose -- so a session knows where to look before it spends a token reading.

How to install Everywhy

This entry records only its repository, not the path inside it, so there is no exact command to give. Open bochen2029-pixel/everywhy and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Everywhy does

Content x relation for coding agents: a phrase in, a budgeted read plan out -- pointers, never prose -- so a session knows where to look before it spends a token reading. Precompute the adjacency, never the answers. The question that cost 493,355 subagent tokens by hand now returns a ~1,400-token plan.

Alternatives in Development

Maestro Learn — User-invoked learning toolkit — guided reading, investigation, pattern extraction, or second opinions 530 ★
Obsidian ADR — Generate a decision record when the vault structure changes — the vault knows why it knows what it does 277 ★
CI Mockup Figure — Read and follow the skill definition 239 ★

README

everywhy — why these things belong together

**The organ of content × relation. A phrase in, a read plan out, in milliseconds — so an agent knows *where to look* before it spends a token reading.**

**Status: R1 "the reflex" — ✓ PASSING (2026-09-09).** 9/9 binding gates green via `python everywhy.py` + `gate.py`. **126,156 chunks** indexed (sessions 62,847 · subagents 31,663 · docs 31,646) from ~1,795 sessions · `r0` 3,017,366 directed edges · 80,898 artifact edges · `fire` ~580 ms inner. **The motivating A/B:** the question that cost **493,355 subagent tokens** by hand now returns a **~1,400-token plan** that surfaces both the known session *and* an origin session the manual sweep missed. Not yet built: `r5` provenance · incremental `--since` in anger · GUI · the native C/C++ rung. The contract is `EVERYWHY_ORGAN_SPEC_proposed_by_opus5_2026-09-04.md`; this README is kept truthful against what actually runs. Store repaired 2026-09-09 (`docs/BOOTSTRAP_LOG.md`, `docs/DECISIONS.md`): a killed `index` had left `r0` addressing a smaller store, so `fire` ran with **no graph at all** for three days while `doctor` reported healthy. Both the defect and the health checks that now catch it are recorded there.

everything = names · everywhere = bytes · everywhen = content × time · everywho = who, now
                        everywhy = content × relation

The gap

None of the other organs answers *"pull this idea and bring everything connected to it."* Today that is composed by hand, per question, at frontier prices. **Measured on this box, 2026-09-04:** one question — *find the sessions about the Gaussian-splat viewer* — cost **493,355 subagent tokens** and ~12 lexical queries, several returning zero because they guessed the wrong vocabulary. At 100 questions that is the same corpus read 100 times.