[finding] .claude/agents/os-dev.md:190 tells every dispatched dev that gen:schema rewrites authorable-surface.base.json "as an expected artifact" — measured false, and backwards for a deletion-gate anchor #16201

Surfaced by the os-dev seat working #14612 (PR #16199) and independently re-measured by the domain:spec PM seat (#6017, session_01T6HeZvT9wdSJD1ZxJb5Eno) before filing, because the subject is a governed file (.claude/**) and a dev's report is not sufficient evidence to change one. Unassigned, no domain:* — routing and grading are triage's.

The two readings, both on origin/main

The instruction — .claude/agents/os-dev.md:190:

packages/spec:gen:schema 会重写 authorable-surface.base.json,这是预期产物。

The code — packages/spec/scripts/build-schemas.ts:108:

--update-base gen:authorable-surface-base — the ONLY mode that writes the …

⇒ Since #5358, a plain gen:schema does not write that file; only --update-base (gen:authorable-surface-base) does. The #14612 dev corroborated it behaviourally: a full @objectstack/spec build left git status empty.

Why this is worth a card rather than a shrug

The clause is not merely stale — it points the wrong way for this particular file, and this file is unusually unforgiving:

• authorable-surface.base.json is the pinned anchor for the deletion gate. Its whole job is to be hard to change: readCommittedSurfaceBase() byte-compares the committed file against serializeSurfaceBase(), and check:authorable-surface fails on any difference it did not generate. A PR that edits it to hide a deletion is supposed to go red.
• Telling a dev that churn in that file is "预期产物" (an expected artifact) invites exactly the two failures the gate exists to catch: committing a rewrite it did not intend, or not investigating when the file moves. For a gate whose failure direction is a silently-hidden deletion, "expected artifact" is the most costly possible framing.
• It also mis-sets expectations in the other direction: a dev who edits the JSON alone and sees no rewrite may conclude the tooling is broken, when in fact the one-sided edit is fatal by construction ([finding] packages/spec/authorable-surface.base.json is a pinned deletion-gate anchor that lags the live ratchet by 532 keys — three documents name it as if it were the live authorable surface #14612 measured this — the constant SURFACE_BASE_DESCRIPTION in build-schemas.ts is the only producer of that string, so the JSON cannot be edited on its own at all).

⚠️ The clause's operative half is fine and should survive any rewrite: never revert the file, never hand-edit it to satisfy an equality, and the assertion that counts is check:authorable-surface green. #14612's dev obeyed that half and said so. It is only the causal claim — which command writes it, and whether seeing it move is expected — that is false.

Suggested shape (⛔ not a decision)

Narrow the clause to what is true: the file is written by build-schemas.ts only in --update-base mode (gen:authorable-surface-base), never by a plain build; if it moves on a plain build, that is a finding, not an artifact. Keep the operative prohibitions verbatim.

⚠️ Whoever takes it should know that #14612's PR (#16199) corrects the same causal claim in build-schemas.ts's own doc comment, so the wording to converge on already exists in that diff — read it rather than inventing a second phrasing.

Expected lane (for triage, not a routing claim): .claude/agents/os-dev.md is named in pm-dispatch/SKILL.md as part of the 协议语义面 ⇒ reads as domain:skills, and any PR touching it is a governed surface with the human-merge endgame.

Dedup

search_issues for this clause returned 0 — and that zero is backed by a control run in the same session: authorable-surface anchor returned #13713 (the card whose 532-key figure #14612 re-measured) and #11979, so the instrument answers. No existing card names the os-dev.md clause.

Refs: #14612 / PR #16199 (where it was found and where the sibling correction lives) · #5358 (the change that made --update-base the only writer) · #5235 / #4650 (the anchor and its gate).