Adopt hi: record what let should be, as 156 criteria

[0xLeif]

Summary

Adopts hi to record what let should be, as plain sentences with
permanent ids, so tickets and specs can be generated from them instead of written alongside them.

• 156 criteria across 10 families, every one captured with the hi CLI rather than hand-written.
• INTENT.md carries the product prose and a generated feature index; each hi/*.md opens with its own ## Intent.
• CI gains a hi check gate, pinned to --version 0.3.3 --locked.
• intent.html is gitignored.

These are directions, not a status report. A criterion that is false today means the code has
not arrived yet, not that the criterion is wrong. Nothing here records whether anything is built;
that layer belongs to spec-sync. hi check fails only on the six structural problems, so this gate
cannot go red because somebody wrote down a want that is not built yet.

Accuracy audit

Every criterion in this repo was scored against its own evidence (README, docs, CLI surface, tests,
specs), and every line scoring below 95 was re-examined by a second independent auditor.

Mean confidence	96.7
Criteria scored	156
Flagged below 95	28
Confirmed below 85	2

Full triage across all 11 repos: https://claude.ai/artifact/G6TzEz3HHgGsHFH5r2DRHn

Corrected before this PR

• GUARD-1 — rewritten. Was: Chat transcripts are never read out, whatever I ask for.
  
  Now: Finding, showing or opening a chat transcript gives me its size and age, never its contents.
• WIRE-2 — rewritten. Was: Run by a person at a terminal, the same answers read plainly without a JSON parser.
  
  Now: A person reading the terminal gets the answer without piping it through a JSON parser.

Deliberately left out

I left out the development and release machinery, which is about building let rather than using it: the fledge lanes and tasks, spec-sync module contracts, augur risk scoring, attest provenance, the Trust CI workflow, biome/tsconfig/bun lockfile conventions, and the "do not use em-dashes" style rules. I left out the internal architecture as architecture — adapter registries, ScanContext, card factories, id-hash formulas, merge overlay field priority, the path-prefix attribution table — keeping only the person-visible consequences (one entry per checkout, stable identity, credited to the right agent). I left out the design document's unbuilt roadmap (loops, handoffs, learnings, draft-skill generation, SQLite memory schema and migrations, a tools registry, MCP write tools, the index cache, a watcher daemon) because criteria should describe wants for this product, and these are speculative enough that I could not tell which are still wanted. I also skipped the .worktrees/fleet-web checkout, which is a duplicate working copy of the repository rather than product surface, and the npm/GitHub publish gating, which is a release decision rather than a behaviour anyone experiences.

Could not work out the intent for

Superskills are unclear: let super can list, read and generate an example recipe of shell commands, but nothing runs one, and the design hints at step graphs with skill references and validation — I could not tell whether the intent is an executable runtime or deliberately just a readable card, so I wrote it as "write down a recipe and list it later". The corvid host appears in the host list, the attribution table and the worktree bases, but no adapter surfaces anything for it, so I could not tell whether corvid-agent is meant to be a first-class host or a placeholder. Project-scope notes require a git repository, which reads more like an implementation constraint than a want, so I did not encode it. The Fleet dashboard's local process probe is the least settled part: it inspects a bounded descendant process tree and infers operation labels from command lines, and I could not tell how far that inference is supposed to go before it becomes guessing. Finally, let history and let web overlap in purpose (both answer "what has been happening across my machine") and I kept them as separate feature areas, but they may be intended as one idea with two faces.

Test Plan

• hi check passes locally (verified: 156 criteria · 10 families · 10 files)
• CI hi check step passes on a fresh runner
• Existing gates unaffected
• Review the criteria themselves: they are the point, and they are a person's words

🤖 Generated with Claude Code

https://claude.ai/code/session_01UcDK2AAqo8w9vjvwA2yitV

[corvid-agent]

Verdict: REQUEST_CHANGES

Head: c2487d7fa7975e09b57f31ceaa99973c64fd3566 · author 0xLeif

#788 anti-pattern: ~156 criteria reverse-engineered from existing code (“Adopt hi: record what let should be, as 156 criteria”). Permanent ids on unconfirmed sentences; missing wants stay invisible.

Required path: draft → confirm-with-humans → capture only confirmed lines. Prefer the paired seed habit PR (let#10). Please close or rewrite without bulk reverse-engineered criteria.

[corvid-agent]

CoS APPROVE as first-pass draft for human migrate (Leif 2026-09-17).

hi check exit 0 on tip c2487d7 (156 criteria).

Expect human rewrite/retire/replace ids; not treating reverse-engineered AC as final product truth.
No merge by CoS.