skills(objectstack-data): `rules/field-types.md` carries a ~1800-byte ASCII decision tree that duplicates the docs' mermaid page, in a file 17 tokens under its ceiling

[claude]

Observed by the os-dev flight on #16083 (PR #16187) while paying the token ratchet for two contract sentences, and handed to the seat to file because both dedup channels available to that flight were down (REST /search/issues 403, the one MCP search_issues refused by the rate limit). The seat's dedup: every open domain:skills card read for decision tree / field-type-decision; the only hit is #16083 itself, the card whose delivery surfaced this.

The finding

skills/objectstack-data/rules/field-types.md carries a ~1800-byte ASCII "Field Type Decision Tree" block (around lines 147–216 at 5a30844272) that re-lists every field type already given in the same file's tables. content/docs/data-modeling/field-type-decision-tree.mdx carries a mermaid version of the same tree, for the same reader. Two carriers of one tree, in a rules file whose token ceiling is 3584 and whose headroom after PR #16187 is 17 tokens.

Why it is a card and not a note

The ratchet made a two-sentence contract fix on #16083 pay with an unrelated deletion. The next author who needs one row in this file hits the same wall. Removing the duplicated tree frees roughly 450 tokens — real headroom, paid once, in the file's own density rather than in whichever unrelated sentence the next flight can spare.

The rules-only ruling for the skill corpus points the same way: a decision tree that only re-lists the tables is not a rule the client agent cannot get elsewhere in the same file.

Acceptance

• The ASCII tree block is removed from rules/field-types.md; the tables it re-lists are untouched.
• If the file needs a pointer to the tree, it is one line naming the docs page, not a second rendering.
• check-skills-token-ratchet exit 0 with the new reading recorded; the per-file ceiling is NOT lowered in the same PR (a ceiling move is its own ruling).
• Governed surface (skills/**): draft PR, seat review, os-zhuang + hotlong requested, human merge.

Refs #16083, PR #16187.

---

Generated by Claude Code

[claude]

Claim: PM loop round 8 (continued) — flight J, one governed skills card: the ~1800-byte ASCII "Field Type Decision Tree" block in skills/objectstack-data/rules/field-types.md re-lists the file's own tables and duplicates the mermaid page in content/docs; it goes, the tables stay, and the file gets real headroom under its unchanged token ceiling (the deletion PR #16187 had to make to fit two contract sentences is the measured cost this card removes).
Session: session_019RfFHiRCSs3JXLK4cwcfox
Branch: claude/issue-16324-field-types-decision-tree
Worktree: objectstack-issue-16324
Domain: domain:skills
File surface: skills/objectstack-data/rules/field-types.md (the ## Field Type Decision Tree heading and its fenced block; every table untouched; at most one pointer line to the docs page). ⛔ No other file; ⛔ the token ceiling in scripts/check-skills-token-ratchet.mjs is not lowered in this PR — a ceiling move is its own ruling.
Container & model: XS, mode:subagent, opus — dispatch-gates.mjs --tier: no path-derived mandate (derived at 3e270d4); prose deletion on a governed surface, no contract accept/reject moves.
Clause-②: no
Serial: PR #16187 (flight D, 5a30844272, the two tree rows in this same file) sits at the governed terminal awaiting a human merge; this flight branches from origin/main, touches none of D's lines, and proves git merge-tree --write-tree clean against 5a30844272 before opening its PR (the shape #16003 used against PR #15946). GOVERNED (skills/**) ⇒ draft PR, in-seat review, os-zhuang + hotlong, human merge, skip-changeset; at ACCEPT the seat adds needs-user-decision + the 维护者速读. The PR carries Fixes #16324.

---

Generated by Claude Code

[claude]

os-dev-report

{
  "issue": 16324,
  "status": "done",
  "branch": "claude/issue-16324-field-types-decision-tree",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/16369",
  "premise_still_valid": true,
  "summary": "Deleted the `## Field Type Decision Tree` heading and its fenced ASCII block from skills/objectstack-data/rules/field-types.md — one file, a pure 71-line deletion, no additions. Every table, the top-of-file blockquote, the Relational `tree` row and the Common Field Configurations / Incorrect vs Correct sections are untouched. The file goes 3578 -> 3042 tokens against an unchanged 3584 ceiling, so headroom goes 6 -> 542; the ceiling row in scripts/check-skills-token-ratchet.mjs is deliberately NOT lowered, and the PR body states the reading so the seat can rule the lowering separately. I added NO pointer line, which was the flight's one discretionary call: git grep shows ZERO published skill files cite any content/docs path (the bundle is self-contained and a customer agent holding only the bundle cannot open a repo path), and the docs page that would have been cited routes `Secret value` to `password` in its mermaid flowchart, contradicting this file's own `password` row, the same page's quick-reference table and ADR-0100 — filed separately as #16366. The block also taught nothing the tables do not and was strictly less complete than them (`slider` appeared in no branch). Draft PR #16369, `skip-changeset` applied additively with a comparison read-back, governed surface so no ready-flip / enqueue / approve / auto-merge.",
  "tests": "All readings taken AFTER the final commit, on head bf3e5690a2 (quoted in the PR body too). (1) dispatch-gates: `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` derived 20 families from the committed change set (1 path, three-dot vs merge base a87163c89); every one run with its exit captured BEFORE any pipe (redirect-then-capture, never a pipe), all exit 0; reconciled with `--ran`: 'Run reconciliation — 20 derived, 20 run, 0 NOT-MEASURED, 0 UNRUN.' / '✓ dispatch-gates --ran: 20 derived famil(ies) accounted for — 20 run, 0 NOT-MEASURED.' exit 0. (2) One family needed a second measurement: `pnpm --filter @objectstack/lint run check:doc-formula-expressions` first returned exit 3 'PREREQUISITE NOT MET' (@objectstack/formula and @objectstack/lint unbuilt) — its own text says nothing was measured, so it is NOT a red gate; after `turbo run build --filter=@objectstack/formula --filter=@objectstack/lint` it re-ran at exit 0. (3) check-skills-token-ratchet: --self-test exit 0, live exit 0, verdict line '✓ check-skills-token-ratchet: skills/objectstack-data/rules/field-types.md is 3042 tokens (ceiling 3584; headroom 542).' and '✓ ... 36 authored bundle file(s) within their ceilings'; ratcheted authored total 140395 -> 139859, bundle total 151041 -> 150505. (4) check-nul-bytes exit 0, plus a self-scan `grep -naP` over the edited file for control characters: clean. (5) `node scripts/pm/check-governed-merges.mjs --test skills/objectstack-data/rules/field-types.md` exit 3, printing '⛔ GOVERNED — a human merge is the review record for this PR' / 'skills/** ×1 — the published skills catalog'. (6) check-skill-id-lint does NOT reach this diff and is not applicable: its corpus is SCAN_ROOT '.claude/skills/pm-dispatch' plus EXTRA_FILES ['.claude/agents/os-dev.md','AGENTS.md'] — the published skills/** catalog is outside it, which is why the derivation does not name it either. (7) Whole-repo `pnpm lint` (eslint . --no-inline-config) through `bash scripts/pm/os-verify-lock.sh` with OS_VERIFY_LOCK_SLOT=issue-16324, acquired once with the full 540s budget: 'ACQUIRED after 1s', 'VERDICT command-exit 0 · held the lock 139s'. Exit 0, zero findings. NOT narrowed, so there is no narrowing to declare; the 139s is a shared-box reading per the lock's own verdict line. (8) Merge-tree proof before opening the PR: `git merge-tree --write-tree HEAD 5a30844272` exit 0, 0 CONFLICT lines, merged tree 0217dc4b5b472ac009ab73b3dc81ebae8fa2523d — PR #16187 touches lines 91 and 106-110 of this file, this deletion started at 147, disjoint. (9) Mutation confirmed on disk before committing, not by an editor exit code: blob 80bbe19a41 -> b66f40f68f, 14309 -> 12167 bytes, 396 -> 325 lines, and the two anchor greps ('Field Type Decision Tree', 'What kind of data?') both went 1 -> 0. No ablation was performed — this is a pure prose deletion with no runtime behaviour to ablate. (10) Reader grep over skills/, scripts/, content/docs/, package test sources and every must_contain token in the tree: zero readers go red. 'Field Type Decision Tree' = 4 hits (3 are the docs page's own title and two links to that page); 'Decision Tree' case-insensitive = 20 hits, of which only skills/objectstack-data/SKILL.md:45 names this file's block, in prose no gate reads; 'What kind of data?' = 2 (deleted line + the docs mermaid's independent START node); 'Owned child' = 2 (deleted line + unrelated TSDoc at packages/spec/src/data/field.zod.ts:1251); 'Name-keyed map ->' = 1 and 'Single choice, always visible' = 1 (deleted line only); must_contain tokens = 0 (skills/objectstack-data/evals holds only README.md; the tree's one evals JSON with must_contain belongs to the objectstack-ui bundle); scripts/ = 1, check-skills-token-ratchet.mjs:420 naming the FILE's ceiling row, never the block. (11) CI polled in the foreground to convergence on bf3e5690a2: 34 check runs, 23 success, 11 skipped, ZERO failures; required contexts Lint & Repo Gates, TypeScript Type Check and Test Core (all six shards) success, Build Core / Dogfood shards / Temporal Conformance skipped by path filter. (12) PR body read back in full after creation: the sent body is an exact byte prefix of the stored body, only the platform's own appended footer differs — no sanitizer damage. Label read-back was comparative: union(target, write-result) minus read-back set = empty, and the concurrently-added `size/s` survived, proving the additive write stripped nothing.",
  "mcp_calls": "0 — the whole run went through repo-scoped REST (probe returned HTTP 200 on this seat) and git; zero MCP GitHub calls, zero GraphQL pool spend. /search/issues is refused by the egress proxy (HTTP 403, 'sessions are bound to their configured repositories'), so dedup used the sanctioned fallback: REST list endpoints plus local grep.",
  "open_questions": [
    {
      "question": "Should rules/field-types.md carry the one pointer line to the docs decision tree that ruling 2 permits? I shipped WITHOUT one; this is the call I would most want overruled if the seat disagrees.",
      "options": [
        "A — no pointer line (shipped). The tables already answer the tree's every question; zero published skill file cites content/docs (measured 0 hits), so the citation would be the corpus's only one and would name a repo path the customer's bundle cannot open; and the page it names currently routes `Secret value` to `password`, contradicting this very file.",
        "B — add the dispatch's candidate line naming content/docs/data-modeling/field-type-decision-tree.mdx. Costs roughly 25-30 tokens out of 542 headroom and keeps SKILL.md:45's 'with decision tree' clause nominally true, but points at a repo path and at a page that currently teaches the wrong credential type.",
        "C — add a pointer line only after #16366 lands, and phrase it as a site URL (https://objectstack.ai/docs/...) rather than a repo path, matching the only docs URL that exists anywhere in this bundle."
      ],
      "recommendation": "A, on all four axes. Real business need: nobody consumes a repo path from a published bundle — the measurement is 0 cross-references in the entire skill corpus, so B serves a reader that does not exist. Long-term soundness: the skill bundle is self-contained by construction, and the first outbound path reference is the precedent that erodes that. Anti-AI-error: this is the decisive axis — B would send an authoring agent to a flowchart that routes credentials to `password`, a type this file's own table warns stores plaintext at rest and that packages/spec/src/data/object.test.ts:2367 pins a build warning against; a pointer to a contradicting source is exactly the lenient second contract the repo forbids. Startup focus: the card asked for headroom, not for a new cross-reference surface; C is the staged option and the 2026-08-27 ruling says staged only with named external-user evidence, which nobody has here. If the seat prefers a pointer anyway, C over B — but only after #16366 lands."
    },
    {
      "question": "skills/objectstack-data/SKILL.md:45 reads 'All 49 field types with decision tree and configs' and is stale the moment this PR lands. Ruling 1 pinned me to ONE file, so I did not touch it. Where does the four-word fix go?",
      "options": [
        "A — fold it into the ceiling-lowering PR the seat will open on this file anyway (same bundle, same governed terminal, one human merge instead of two).",
        "B — a separate one-line PR now.",
        "C — leave it; the clause is loose enough that an agent opening the file finds the tables and is not misled."
      ],
      "recommendation": "A. It is the same bundle and the same governed surface, so folding it in costs one human merge rather than two, and the ceiling PR is already the place where this file's post-deletion reading is being ruled on. It does not meet the maintainer's finding threshold on its own (no reproducible defect, no contract violation, no metadata-authoring trap — the agent that opens the file finds the tables), which is why it is noted rather than filed. Reported in the PR body's 验收备注 as well, so it does not depend on this report being read."
    }
  ],
  "out_of_scope_findings": [
    "filed as #16366: content/docs/data-modeling/field-type-decision-tree.mdx line 36 routes `Secret value` to `password` in its mermaid flowchart, contradicting the same page's own quick-reference table (lines 138-139), its own correct `STRUCT` edge at line 78, ADR-0100, and the shipped skill rule — an authoring trap with a data-at-rest consequence, pinned on the runtime side by packages/spec/src/data/object.test.ts:2367 ('warns once when a password field is declared on a generic object'). Filed unassigned, no labels, for PM triage. Dedup: /search/issues is 403 on this seat, so I enumerated 669 open + 505 recently-closed issues by REST page walk (open exhausted at a short page 7; closed bounded to 12 pages since 2026-08-01, declared) and grepped locally — a control term (#16324) confirmed the corpus was real; nearest neighbours #14901 and #14940 are the same docs tree but different defects, both already closed.",
    "noted, not filed: skills/objectstack-data/SKILL.md:45 ('All 49 field types with decision tree and configs') goes stale in prose when this PR lands. Below the maintainer's threshold and out of the ONE-file ruling; recommended above for the seat's ceiling-lowering PR. check:corpus-claim-drift is green over it, so no gate catches it.",
    "noted, not filed: the deleted ASCII tree was already drifting from the tables it duplicated — `slider` appeared in no branch, and it compressed the `password` / `secret` distinction into one parenthetical. Evidence for the card's premise, not a separate defect."
  ]
}

Generated by Claude Code

---

Generated by Claude Code