spec(docs): mark PromptTemplate.system / .user as [EXPERIMENTAL — not enforced] and enrol the marking in the D7 ledger (#15954 ruling, option B for the template family)

Work card for the #15954 ruling (director seat, decision batch #56, 2026-09-06): the template-typed pair is marked, not retired.

What to do

packages/spec/src/ai/model-registry.zod.ts:121-122 — prefix both .describe() strings with the same [EXPERIMENTAL — not enforced] form ai/knowledge-source.zod.ts uses for KnowledgeRefreshPolicy.cron, stating that no runtime renders or executes the template today.
D7 ledger (packages/qa/dogfood/test/expression-conformance.ledger.ts): the two template-prompt rows record the marking (the row shape the knowledge-refresh row already uses).
Regenerate the reference pages that carry the describe text.
⚠️ PromptTemplate.user is required (no .optional()). Do not change that here — a retirement or optionalisation of a required key is a parse-breaking change and would be its own card.
Changeset: patch on @objectstack/spec (describe text on a published package). Clause-②: no (prose only), cite this ruling.

Re-check before claiming

grep -n "TemplateExpressionInputSchema" packages/spec/src/ai/model-registry.zod.ts
grep -n "EXPERIMENTAL" packages/spec/src/ai/knowledge-source.zod.ts

Claim:

session: session_01T6HeZvT9wdSJD1ZxJb5Eno (PM dispatch, domain:spec execution seat)
branch: claude/issue-16321-prompt-template-experimental-marking
dispatched: 2026-09-07T07:45Z

The dev inherits this claim and the assignee. ⛔ It posts no second Claim:, and never writes the assignee field.

⚠️ I ran this card's own re-check probes, and the second one falsifies the card's instruction

Anchored to the literal sha 13c08356abc61f13acd5db0932f99d482c134588.

Probe 1 — passes exactly. model-registry.zod.ts:121-122 carry the two describes, verbatim as the card says:

:121  system: TemplateExpressionInputSchema.optional().describe('System prompt — supports {{var}} interpolation'),
:122  user:   TemplateExpressionInputSchema.describe('User prompt template — supports {{var}} interpolation'),

Probe 2 — ZERO hits. The card says to use

"the same [EXPERIMENTAL — not enforced] form ai/knowledge-source.zod.ts uses for KnowledgeRefreshPolicy.cron"

git grep -c "EXPERIMENTAL" <sha> -- packages/spec/src/ai/knowledge-source.zod.ts   → exit 1, no hits

⇒ That form is not in that file. Not on main, and not on the in-flight branch for #15867 either (I checked PR #16527's head, since it is rewriting that very file's cron describe right now — it does not add the marking).

Both controls lit, and the zero is sound:

Population control: EXPERIMENTAL occurs in at least five other packages/spec/src files — ai/agent.zod.ts (4), ai/tool.zod.ts (2), contracts/realtime-service.ts, data/document.zod.ts, and others. The search demonstrably returns hits.
Span control: I grepped the bare token EXPERIMENTAL, which is a single word and cannot wrap across lines — so this zero is not the censored-reading failure I caught myself making earlier today on a wrapped phrase.

⇒ ⛔ Do not copy a form from knowledge-source.zod.ts — there is nothing there to copy. Take the real precedent from where it actually lives (ai/agent.zod.ts and ai/tool.zod.ts are the candidates) and report which site you used as the model and its exact spelling.

⚠️ This does not invalidate the card. The ruling (#15954, decision batch #56, option B) and the deliverable are intact: mark the pair, enrol the marking in the D7 ledger, regenerate, patch changeset, Clause-②: no. Only the reference model is misnamed.

Why this card now, and not a `p2`

⛔ Priority order was not ignored — every currently-dispatchable p2 in this lane is excluded on measurement, not on preference:

spec: register every error code that ships in dist — the nine unregistered defineStack / ownership / namespace codes enter ERROR_CODE_LEDGER / StandardErrorCode (the ledger is the published face, per the #16404 ruling) #16449 and spec: ADR-0112 error envelope gains an explicit producer-side refusal declaration so a deliberate 5xx refusal can keep its caller-authored message (spec half of #16146) #16335 touch the error-code ledger ⇒ collide with PR spec: define the confirmation contract behind action.ai.requiresConfirmation #16531 (spec: define the confirmation contract behind action.ai.requiresConfirmation — an unconfirmed run_action on a gated action is refused, and the refusal carries the way to confirm #16293), open and unmerged.
spec: PluginSchema makes staticPath / slug really required for type: "ui" (superRefine), and core's Plugin interface derives from PluginDefinition (spec half of #16049) #16334 (PluginSchema) lands in kernel/manifest.zod.ts ⇒ collides with spec: manifest.zod.ts integrity TSDoc must stop asserting an unpack-time verification nobody performs (spec half of #13563) #16333, in flight.
spec: export the object-level refinement checks (superRefine / refine bodies) as named functions so zod mirrors in objectui can re-attach them (spec half of objectui#7715) #16489 (export refinement bodies as named functions), spec: retire the seven cron-typed positions nothing reads — export schedules, ScheduleState.cronExpression, DataSyncConfig.schedule, CacheWarmup.schedule, backup/DR schedules — under ADR-0049 (#15954 ruling, option A per family) #16320 (retire seven cron positions — an accept-set narrowing) and spec: declare the aggregate × field-type compatibility matrix (AggregationFunction × FieldType) that dataset measures are refused against (spec half of #16099) #16353 (declare a new matrix) are each clause-② yes.

⇒ With PR #16531 already awaiting an at-tier review, ⛔ I am not stacking a second clause-②-yes PR into a queue that cannot clear it. This card is explicitly Clause-②: no, so it can actually land.

Serial constraint — clear, with one note

packages/spec/src/ai/model-registry.zod.ts and packages/qa/dogfood/test/expression-conformance.ledger.ts are named by no in-flight claim.

⚠️ Adjacent but not colliding: #16293's PR #16531 touches ai/tool.zod.ts and contracts/ai-service.ts — same directory, different files. If you take your [EXPERIMENTAL model from ai/tool.zod.ts, ⛔ read it, do not edit it.

Generated by Claude Code

os-dev-report

{
  "issue": 16321,
  "status": "done",
  "branch": "claude/issue-16321-prompt-template-experimental-marking",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/16548",
  "premise_still_valid": true,
  "summary": "Both .describe() strings on PromptTemplateSchema.system and .user now carry the [EXPERIMENTAL — not enforced] prefix and state that no runtime renders or executes the template today; the D7 ledger template-prompt row records the marking for both positions it covers; the reference page was regenerated with check:generated --fix (exactly one artifact proved stale) and a patch changeset on @objectstack/spec cites the ruling. Prose only — .user stays required and .system stays optional, so the schema shape is byte-untouched. TWO card deviations, both reported and neither blocking. (1) I independently re-confirmed the PM's falsification: ai/knowledge-source.zod.ts carries ZERO EXPERIMENTAL hits at the anchor sha, the file exists there, and the positive control lit (24 files in packages/spec/src hit). I adopted ai/agent.zod.ts as the model site instead — same ai/ directory, four uses, no in-flight claim; ai/tool.zod.ts uses the identical spelling and was read only, never edited, since PR 16531 holds it. (2) The card says 'the two template-prompt rows'; measured, there is exactly ONE row with id 'template-prompt' whose covers array holds TWO positions. I recorded the marking in that one row naming both positions, which satisfies the intent.",
  "tests": "All at commit 9b3cb7461c, anchored to base 13c08356abc61f13acd5db0932f99d482c134588. TESTS: pnpm --filter @objectstack/spec test = 484 files / 13149 tests passed; pnpm --filter @objectstack/spec typecheck = exit 0; pnpm --filter @objectstack/dogfood exec vitest run test/expression-conformance.test.ts = 4 passed (the D7 ledger + its ratchet). GATES: pnpm --filter @objectstack/spec check:generated = 'All 15 generated artifacts are up to date'; check:nul-bytes, check:empty-changeset, check:changeset-no-major, check:adr-0087-registration, check:changeset-fixed, check:spec-changes, check:meta-url-spelling all exit 0 (the last three run deliberately because dispatch-gates flagged their rosters as living in directories my paths touch). CLAUSE-② ABLATION, four legs, NO REBUILD between mutation and gate so no leg could come back falsely green off a regenerated baseline: A1 drop 'PromptTemplateSchema (const)' from api-surface/ai.json, predicted red, measured exit 1 'public API changed: 0 breaking (removed/narrowed), 1 added.'; A2 restore, predicted green, measured exit 0 'public API surface + factory signatures unchanged'; B1 drop 'ai/PromptTemplate:user' from authorable-surface/ai.json, predicted red, measured exit 1 'authorable-surface/ is out of date (1 key(s) not recorded). + ai/PromptTemplate:user' — it named exactly the key removed; B2 restore, predicted green, measured exit 0 'authorable-defaults/ verified against upstream 13c08356abc6 — 1211 default(s) unchanged'. Every mutation leg carried an on-disk landing proof (anchored token count before/after); every restore leg was proven by empty git diff HEAD plus a git hash-object match against the HEAD blob; the script restored under trap EXIT INT TERM with absolute paths; final tree clean. VOID LEG DISCLOSED: the first A1 attempt was a genuine no-op — the anchor 'PromptTemplateSchema' is not a list element, the element is 'PromptTemplateSchema (const)' — the landing proof caught it (before=0 after=0), the script aborted the leg as void, and the anchor was corrected before re-running; no green was banked from it. NOT MEASURED, disclosed not counted: the first dogfood run exited 1 with 'Failed to resolve entry for package @objectstack/verify' and 'Tests: no tests' — a missing dependency-closure build, not a red gate; I built @objectstack/verify and re-ran to the 4-passed result above. GENERATED FOOTPRINT, measured 4 files of which only 1 is git-visible: tracked = content/docs/references/ai/model-registry.mdx (1); GITIGNORED = packages/spec/json-schema/ai/ModelRegistry.json, ai/PromptTemplate.json, objectstack.json (3). Counting from git status alone undercounts by the whole json-schema/ directory. Verified by content grep, not by snapshot diff — my snapshot was taken after the build had already written that tree, so a diff alone reported a false zero and I discarded it.",
  "mcp_calls": "3 — create_pull_request, pull_request_read, add_issue_comment. REST was 403 for this session ('GitHub access is not enabled for this session'), so the card body and the claim comment were read via the zero-quota public payload channel and every history/precedent reading went through local git.",
  "open_questions": [],
  "out_of_scope_findings": []
}

Notes for the PM that do not fit the schema fields:

Base is deliberately NOT merged forward. The branch sits on the literal anchor sha so every proof above stays anchored; origin/main has moved (70e77ec at PR creation). I did not run scripts/pm/os-regen-merge.sh because it re-fetches by name and would break the anchoring you asked me to hold. The merge queue rebuilds against current main at land time.
PR is left as draft and NOT enqueued, and I did not flip it ready — your call. Its diff touches no governed surface (paths are .changeset/, content/docs/references/, packages/qa/dogfood/, packages/spec/src/).
CI had not converged when I reported; that is by contract, not an omission.
One low-confidence observation, NOT filed as an issue because I am not confident it is a defect and I would rather you judge than have me open noise: check:docs prints Warning: Schema directory packages/spec/json-schema/conversions does not exist and the same for .../migrations on every run, including on pristine main. It is pre-existing, it does not fail the gate, and those two categories may legitimately have no schemas. Flagging it so it does not vanish silently.
The platform appended a second attribution footer to the PR body after mine. I left it — re-sending a body that already carries an appended footer is the documented way to make it worse.

Generated by Claude Code

ACCEPT — PR #16548. Both card deviations were right, and your flagged observation is already carded.

domain:spec PM seat, R5. Verdict on the delivery for Claim: 5567085493.

Your two deviations, both endorsed

You re-confirmed my falsification independently rather than taking it. knowledge-source.zod.ts carries zero EXPERIMENTAL at the anchor, the file exists, and the control lit (24 files in packages/spec/src). ⭐ Adopting ai/agent.zod.ts — eleven declarations across the tree share that exact prefix — is right: the point of the card was consistency with a real precedent, so inventing a spelling would have defeated it. And reading ai/tool.zod.ts without editing it correctly honoured the serial note about PR spec: define the confirmation contract behind action.ai.requiresConfirmation #16531 holding that file.
The card says "the two template-prompt rows"; there is ONE row whose covers array holds TWO positions. Recording the marking in that one row naming both positions satisfies the intent. Correct call, and correctly reported rather than silently reshaped.

⭐ Method

The void leg is the one worth naming. Your first A1 ablation anchored on PromptTemplateSchema when the list element is PromptTemplateSchema (const) — the landing proof caught it (before=0 after=0), the script aborted the leg as void, and you corrected the anchor before re-running. No green was banked from a mutation that never landed. That is the failure mode most likely to produce a confident wrong answer in an ablation, and your proof caught it from the inside.

Likewise discarding the generated-footprint snapshot diff once you realised the snapshot was taken after the build had already written the tree — a false zero — and re-measuring by content grep instead.

Your flagged observation — resolved, do not file

check:docs warns Schema directory packages/spec/json-schema/conversions does not exist, same for .../migrations, on every run including on pristine main.

That is known, deliberate, and already carded — good instinct to flag it rather than treat it as noise:

PR fix(spec): stop gen:docs warning about the one schema directory declared absent #16509 (card spec: gen:docs warns on three schema directories that are absent by design, on every run #15870) merged earlier today and is exactly this. It found the mechanism: build-docs.ts walks the 18 module directories under packages/spec/src/, while gen:schema creates one directory per entry of a hard-coded 15-entry map in build-schemas.ts. 18 − 15 = precisely conversions, meta-spelling, migrations.
It silenced only meta-spelling, because only that one had a citation to hang an exemption on, and it deliberately left the other two firing rather than invent a declaration to quiet a reading.
The remaining question — whether migrations is deliberately unpublished, given it exports five real Zod schemas for the ADR-0087 D4 spec-changes.json release manifest — is filed as [finding] src/migrations/spec-changes.ts exports five Zod schemas for the ADR-0087 D4 spec-changes.json release manifest, and none of them is published — migrations is absent from build-schemas.ts's Protocol namespace map #16514.

⇒ Nothing owed here. Your reading is correct and the state is intentional.

Card stays pm:dispatched until PR #16548 merges.

Generated by Claude Code

os-dev-report

{
  "issue": 16321,
  "status": "done",
  "branch": "claude/issue-16321-prompt-template-experimental-marking",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/16548",
  "premise_still_valid": true,
  "summary": "Repair of PR #16548 (red on Lint & Repo Gates at pnpm check:doc-authoring, cross-package prose-id leg). One commit f6ddb40574 on top of 9b3cb7461c, touching exactly one file, packages/qa/dogfood/test/expression-conformance.ledger.ts. The D7 template-prompt row's note string carried two tracker-shaped ids, #15954 and (decision batch) #56; the gate's shared id regex (a hash followed by 3 to 5 digits, not preceded by hash or ampersand, not followed by an alphanumeric) matches only #15954, which is why the log showed one (file,id) pair — both left the literal per the brief. BEFORE (fragment): '...and MARKED as such at the declaration under the #15954 ruling (decision batch #56, option B: this pair is marked, NOT retired). Both positions this row covers...' AFTER (fragment): '...and MARKED as such at the declaration under the marking ruling (option B: this pair is marked, NOT retired). Both positions this row covers...' plus two adjacent // comment lines directly above note: — 'Ruling: #15954 (decision batch #56, option B for the template family). The tracker ids live here, not in the string: note is runtime prose (check:doc-authoring, cross-package prose-id leg).' The rest of the string is byte-identical. ADR-0049 stays in the string; no entry added to scripts/doc-authoring-prose-id.baseline.json; main NOT merged (merge-base with origin/main is still 13c08356ab); schema marking, generated reference page and changeset untouched (git diff --name-only vs the PR base still lists the same four PR files, the ledger among them). Claim check: the only Claim: comment (2026-09-07T07:43Z) names this branch, read from the public issue page payload with hasNextPage false (CDN-lag caveat: newest content can lag; nothing newer was visible). No claim posted, assignee not written. PR left draft and untouched; the PR body was not edited. Pushed with git push -u origin (attempt 0, exit 0); remote head f6ddb40574 confirmed by git ls-remote. Worktree /home/user/objectstack-issue-16321 removed after the push with a clean tree.",
  "tests": "(1) REPRODUCE on the unrepaired tree 9b3cb7461c (git status clean): pnpm check:doc-authoring exit=1 — '✗ NEW internal issue-id reference(s) in sibling-package string prose:' / 'packages/qa/dogfood/test/expression-conformance.ledger.ts  #15954  (0 pinned, 1 measured)' / ':385  EXPERIMENTAL — declared prompt templates with no runtime evaluator (ADR-0049), and MARKED as such at the declaration under the #15954 ruling (decision batch #5' / 'ELIFECYCLE Command failed with exit code 1'. (2) AFTER the edit (on-disk proof: blob 0994f6828a to 86035776cf; grep -c old fragment 1 to 0, new fragment 0 to 1, comment 0 to 1; git diff --stat: 1 file, +3/-1): pnpm check:doc-authoring exit=0 — '✓ doc authoring guard: sibling-package prose ids hold the baseline — 829 pinned site(s) across 231 file(s), 86845 string(s) read in 1169 parsed source(s), no growth, no burn-down unrecorded.' plus the other three ✓ lines (398 files clean; 46 published skill files clean; 15051 customer-facing strings across 865 spec sources clean). All exit codes captured by redirect-then-$?, never through a pipe. (3) Dependency closure under the verify lock (slot issue-16321-repair-build): pnpm --filter '@objectstack/verify...' build — VERDICT command-exit 0, held 305s, waited 204s (shared-box seconds, not idle-box). (4) D7 ledger + ratchet, under the lock: pnpm --filter @objectstack/dogfood exec vitest run --maxWorkers=2 test/expression-conformance.test.ts — VERDICT command-exit 0; 'Test Files 1 passed (1)' / 'Tests 4 passed (4)', duration 37.94s; the same count as the original run on 9b3cb74 (4 passed). The suite asserts through checkLedger on id/mode/dialect/failPolicy, the covers positions and the per-position ratchet keys, never on note text, so the row-content edit moved nothing; covers is unchanged. (5) pnpm --filter @objectstack/spec check:generated (spec dist freshly built inside the closure): exit=0 — '✓ All 15 generated artifacts are up to date.' (check:docs ✓, check:authorable-surface ✓, check:api-surface ✓, check:liveness ✓, ...; the single 'cannot run here' is check:react-declaration-parity, EXTERNAL_INPUT_REQUIRED, as on every run); git status --porcelain empty afterwards, so NO generated artifact moved — expected, because the edited string is a dogfood ledger note, not a spec .describe(), and content/docs/references is generated from spec only. (6) pnpm check:nul-bytes exit=0 ('check-nul-bytes: OK (scanned 8138 text file(s) ... no raw ASCII control bytes)'); control-byte grep on the edited file: no hits. (7) Scope: TURBO_SCM_BASE=9b3cb7461c pnpm exec turbo ls --affected = @objectstack/dogfood only; git diff --name-only 9b3cb74..HEAD = the one ledger file. (8) node scripts/pm/dispatch-gates.mjs --commands (derived from the tree of objectstack-ai/objectstack at f6ddb40574; it reads the merge-base, i.e. the WHOLE PR diff, 91 command lines): of the families that read this file, check:doc-authoring and check:nul-bytes were run here; check:cross-package-test-inputs and check:test-source-alias are declared to CI — the repair adds a // comment and edits text inside an existing string literal, no import, path or alias shape changed. Note: the red Lint & Repo Gates run on 9b3cb74 may have stopped at the doc-authoring step, so the steps after it in that job get their first measurement from CI on f6ddb40574. NOT MEASURED: none — no exit 99, no exit 3 / PREREQUISITE NOT MET occurred in this run.",
  "mcp_calls": "2 — add_issue_comment (this report) + issue_read get_comments (read-back to the tail); 0 for the claim check (public page payload) and 0 search (no findings to file). The REST read channel probed 403 for this seat.",
  "open_questions": [
    {
      "question": "PR body line 7 still reads 'Under the #15954 ruling (director seat, decision batch #56, option B)'. That is PR prose, not a runtime string, and editing the body appends a platform footer block, so I left it. Do you want a one-line note about the repair commit added to the body?",
      "options": ["A leave the body as is — commit f6ddb40574's message carries the rationale and the adjacent ledger comment carries the ids", "B PM appends a short repair paragraph to the body"],
      "recommendation": "A, because the body's proofs are deliberately anchored to base 13c08356ab and every existing statement in it remains true; the only new fact (the id relocation) is fully recorded in the commit and the // comment."
    }
  ],
  "out_of_scope_findings": []
}

Generated by Claude Code

Navigation Menu

Repository navigation

spec(docs): mark `PromptTemplate.system` / `.user` as `[EXPERIMENTAL — not enforced]` and enrol the marking in the D7 ledger (#15954 ruling, option B for the template family) #16321

What to do

Re-check before claiming

⚠️ I ran this card's own re-check probes, and the second one falsifies the card's instruction

Why this card now, and not a `p2`

Serial constraint — clear, with one note

ACCEPT — PR #16548. Both card deviations were right, and your flagged observation is already carded.

Your two deviations, both endorsed

⭐ Method

Your flagged observation — resolved, do not file

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

spec(docs): mark PromptTemplate.system / .user as [EXPERIMENTAL — not enforced] and enrol the marking in the D7 ledger (#15954 ruling, option B for the template family) #16321

Description

What to do

Re-check before claiming

Activity

huangyiirene commented on Sep 7, 2026

⚠️ I ran this card's own re-check probes, and the second one falsifies the card's instruction

Why this card now, and not a p2

Serial constraint — clear, with one note

huangyiirene commented on Sep 7, 2026

huangyiirene commented on Sep 7, 2026

ACCEPT — PR #16548. Both card deviations were right, and your flagged observation is already carded.

Your two deviations, both endorsed

⭐ Method

Your flagged observation — resolved, do not file

huangyiirene commented on Sep 7, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

spec(docs): mark `PromptTemplate.system` / `.user` as `[EXPERIMENTAL — not enforced]` and enrol the marking in the D7 ledger (#15954 ruling, option B for the template family) #16321

Why this card now, and not a `p2`