fix(plugin-audit): resolve reference titles in activityMilestones summary tokens (#7290)

[os-zhuang]

Fixes #7290

matchMilestone (ADR-0052 §5b.2) interpolated a milestone summary's {token}s from the after-row by calling displayFieldValue with no title map, so a lookup / master_detail / user token fell through to String(value) — the raw 32-char id. A milestone declared as { field: 'stage', value: 'closed_won', summary: 'Deal won by {owner}' } shipped Deal won by oBK25… to the record timeline.

The milestone branch takes precedence over the tracked-change branch, so on every object that declares activityMilestones this was the string users actually saw, and #7230 / PR #7291's fix never reached them. Closing that is what this card is.

Anchors re-verified at the branch point (7fa2aae45)

The card recorded its anchors at a768f81a1 (pre-#7230) and warned the file had moved three times in a day. It had:

anchor	card (a768f81a1)	at 7fa2aae45
displayFieldValue	:308	:344 (gained #7230's titlesFor param)
matchMilestone	:355	:472
the {token} interpolation	:368-375	:485-490
matchMilestone call site	:872	:1071

The premise itself was re-verified rather than taken on trust: at the branch point the interpolation still read return field ? displayFieldValue(field, v) : String(v) — two arguments, no title map — and matchMilestone had exactly one caller, already inside an async function. Premise valid, unchanged by #7291.

Placement — re-derived, and measured both ways

The card recommended resolving after a milestone fires, keyed on the matched template's tokens, at the cost of making matchMilestone async. That direction is right and is what shipped. The async part turned out to be avoidable, so it was avoided:

matchMilestone is split rather than made async — detection (matchMilestone, still sync, no longer takes fields), planning (planMilestoneTokenReads), rendering (renderMilestoneSummary). The call site awaits one read between them. This gives the recommended read profile with zero async blast radius, and it mirrors the planTrackedLookupReads + resolver shape #7291 established one commit earlier in the same file rather than inventing a second one beside it.

resolveTrackedLookupTitles became resolveLookupTitles(api, plan) — it takes the plan instead of building it. Both branches now end in one read/mask/title implementation, which is what keeps #6977's masking contract from being re-typed for the milestone path.

Read counts, measured with the copy-returning counting driver

write shape	added reads
create (any)	0
delete (any)	0
update of a milestone-declaring object that fires nothing, on a row holding references	0
non-stage update of a milestone-declaring object, row holding references	0
repeat write while the record already sits at the milestone value	0
fired milestone whose template names no reference token	0
fired milestone whose reference token is empty	0
fired milestone, 1 reference token	1 on that target
fired milestone, 2 tokens onto the same target	1 (batched)
fired milestone, 2 distinct targets	1 each
the same token twice	1
target designating a masked field as its title	0 (read skipped)

The rejected placement, measured rather than argued. Resolving before knowing whether a milestone fired was implemented as mutation M3 and run: on a non-firing update of a row holding references it measured 3 reads (sys_user 1, crm_account 1, crm_region 1; crm_vault 0, its title being masked) where the shipped shape pays 0. That is the shape #6656 / PR #6977's Option A+ ruling was obtained to remove from this write path, so it was not adopted.

Why the milestone plan does not filter on trackHistory. #7230's gate is "a field that is BOTH tracked AND a reference actually moved". A milestone token is none of those: it is the author's own token, read from the whole after-row. The reported declaration itself ('Deal won by {owner}') names a field that need not be tracked at all, so reusing that gate would leave the exact reported defect shipping — mutation M4 below measures it.

#7291 and #6977 are re-asserted, not assumed

audit-lookup-summary.test.ts and audit-bound-previous.test.ts are untouched and green, and #6977's single-id count is re-asserted inside the new file on the firing path (still exactly one findOne on the audited object). ledgerView, the secret mask and the virtual-field drop are unchanged; the milestone read plan is built from the same masked after-view the interpolation renders from.

crm_vault in the new fixture declares nameField: 'api_key' on a real secret field (with a crypto provider wired, so the column is genuinely encrypted rather than conveniently empty). The read is skipped, not masked afterwards.

What did not change

The author's template wording is theirs: only what a token resolves to changes. Punctuation, emoji, and the empty-token rule are byte-identical — an empty value still renders as the empty string, not displayFieldValue's ∅, because a milestone sentence is prose and not a diff row. Milestone templates are still not translated. Select option-label rendering is untouched — localizing those is #7289, deliberately held behind this card in the same file and left alone here.

Tests

New file packages/plugins/plugin-audit/src/audit-milestone-summary.test.ts — 28 cases on a real ObjectQL engine with the copy-returning counting driver (a driver handing back live store references lets the read path rewrite the store under a measurement).

The label/locale fixture is a real createMemoryI18n bundle, and that earned its keep immediately: the first draft wrote {object} in the bundle and measured 已更新{object}"{label}" — the shipped bundles interpolate {{param}}, which is deliberately not the milestone template's single-brace {token}. A stub translator would have hidden that.

pnpm --filter @objectstack/plugin-audit test       Test Files 11 passed (11)   Tests 188 passed (188)
pnpm --filter @objectstack/plugin-audit typecheck  tsc --noEmit — clean
npx eslint packages/plugins/plugin-audit --no-inline-config  — clean (exit 0)
pnpm --filter @objectstack/plugin-audit build      — success

Family gates, all green locally: check:nul-bytes, check:durability-log-level, check:engine-double-contract, check:error-code-casing, check:route-envelope, check:wildcard-fallthrough, check:empty-changeset, check:adr-0087-registration, check:changeset-gate-self-tests.

Reverse verification — direction predicted in writing before the first mutation

Predictions were written to a file and the tree committed before any mutation ran.

#	mutation	predicted	measured
M1	drop the title map from renderMilestoneSummary (the pre-#7290 interpolation, plan and read kept)	10 red: 9 rendering cases + the trailing summary assertion inside one read-count case; every pure count case green	as predicted — 10 red / 18 green, the same 10
M2	key the milestone plan by field instead of by target object	1 red (the same-target batching case); the repeated-token case green	wrong — 14 red. See below
M2'	issue one read per id instead of one batched id: { $in: [...] }	isolates batching only	3 red: my same-target case and both of #7291's batching cases, which also confirms the shared resolver really is shared
M3	the rejected placement — resolve before matchMilestone	10 red, all the zero/exact-count guards on the reference-holding row; rendering untouched	as predicted — 10 red / 18 green, the same 10, and the measured cost is the 3 reads quoted above
M4	wrongly reuse #7230's trackHistory gate on the milestone plan	14 red	as predicted — 14 red / 14 green

M2 is the one I got wrong, and the reason is worth recording. I predicted the mutation would only coarsen batching. It did not: the plan map's key is the object name resolveLookupTitles reads from and the key renderMilestoneSummary looks titles up by. Keying it by field therefore made the resolver query objects named owner / account_id, the best-effort catch swallowed the failures, and no title resolved at all — 14 red, mostly rendering. The mutation was wrong, not the code. M2' was written to isolate the property M2 was supposed to test, and it does. My sub-prediction inside M2 — that the repeated-token case pins token dedup rather than object grouping, so it should stay green — held under M2'.

The trap that PR #7291 documented is live here and was designed around, not discovered. Every zero-read guard runs against deal_2, a row that already holds all six reference fields; one case (a reference token whose value is EMPTY) necessarily runs against the reference-free deal_1, and it is the one case that stayed green under M3. That is green-for-the-wrong-reason, predicted as such in writing beforehand and flagged in the file rather than counted as coverage. Likewise the raw-id fallback case was written from the start as a mixed assertion (Won by usr_1 for APAC — unresolvable side stays raw, resolvable side must still become a title), because #7291 measured that a single-token fallback assertion survives deletion of the render limb; under M1 it went red, as intended.

Deliberately not done

• plugin-audit: select option labels in sys_activity.summary are never localized — the same defect class as #7230, one line over in displayFieldValue #7289's select option-label localization — same file, same function's neighbourhood, held at pm:queue behind this card. Not touched; a case pins the current rendering so the seam is visible when that card lands.
• Milestone templates are still not localized. They are author-written sentences and no bundle key exists for them; changing that would redefine what a milestone summary is, which is a contract decision and not this bug fix.
• multiple: true lookups stored as an unparsed JSON string still fall through to the raw value, exactly as on plugin-audit: renderTrackedChangeSummary never receives the translator, and displayFieldValue prints raw lookup ids — sys_activity.summary ships Rating Owner: ∅ → oBK25… onto every feed surface (re-routed from objectui#4055) #7230's branch — in-memory arrays are unpacked and resolved from the same batch.
• Historical rows keep their write-time composition; only new writes improve.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 10, 2026 7:32am

[github-actions]

📓 Docs Drift Check

This PR changes 1 package(s): @objectstack/plugin-audit.

3 hand-written doc(s) reference the affected code and may need an implementation-accuracy re-verification:

• content/docs/deployment/cli.mdx (via @objectstack/plugin-audit)
• content/docs/deployment/production-readiness.mdx (via @objectstack/plugin-audit)
• content/docs/plugins/packages.mdx (via @objectstack/plugin-audit)

⛔ 1 release-owned page(s) also reference the affected code. These are read-only:

• content/docs/releases/implementation-status.mdx (via @objectstack/plugin-audit)

content/docs/releases/ is RELEASE-OWNED (AGENTS.md "Documentation Guardrails"): release
notes are written centrally at release time, and a code PR that edits them is the exact PR
that guardrail exists to stop. They are still audited — read-only. If one of them is actually
wrong, file an issue or open a dedicated docs-only PR; do not edit it here.

Advisory only. To re-verify, run the docs-accuracy-audit workflow scoped to these files:
node scripts/docs-audit/affected-docs.mjs origin/main → pass the list as args.docs.

[github-actions]

⛔ merge queue 构建失败 — 先分诊，再决定要不要重排

队列构建 31366863728 红了。队列跑的是全量套件（PR 侧 CI 只跑 affected 子集），
所以失败的测试可能在本 PR 没碰过的包里 —— 那不是重排能修的。每次盲目重排都会让排在后面的所有 PR 重建一轮。

失败的 job（日志抽取，best effort）：

• Test Core (2/3) — 失败步骤: Run this shard's tests

  �[41m�[1m FAIL �[22m�[49m test/watch-dot-root.test.ts�[2m > �[22mFileSystemRepository watcher — dot-rooted watch root (#7150)�[2m > �[22msees an external edit when the root is under a dot-directory
  

历史信号：

• 本 PR 过去 24h 无队列失败记录（首次）。
• 过去 24h 队列共有 5 个失败构建（不含本次）。

分诊清单：

1. 失败测试在本 PR 改动的包里 → 真回归，修 PR。
2. 失败测试与本 PR 无关 → 在其他 PR 的同类评论里搜同名测试；出现过 ⇒ flaky 实锤，开 issue 修/隔离那条测试。修好前重排只会再烧一轮全队列。
3. 两者都不是 → 可能与同组 PR 语义冲突；等前面的 PR 落地或失败出队后再重排一次即可，不要连续重排。

---

Generated by Claude Code · merge-queue-triage workflow (#4859)

[os-zhuang]

Merge-queue ejection triaged — this is a deliberate single re-queue, not a blind one.

Walking the merge-queue-triage checklist rather than hitting re-queue:

Step 1 — is the failing test in a package this PR changed? No. This PR changes three files, all @objectstack/plugin-audit plus a changeset; the failure is packages/metadata-fs/test/watch-dot-root.test.ts. No import path, no shared fixture, no shared global.

Step 2 — has the same test appeared in another PR's queue-failure comment? Not in the 60 most recently updated PRs — this is the first recorded sighting, so I cannot claim "flaky 实锤" on a repeat. Filed #7369 anyway, because step 2 asks for the issue regardless and the analysis holds without a second sighting: the case races a chokidar delivery against two fixed wall-clock budgets (a sleep(400) clearing a 200 ms self-write suppression window, then Promise.race against EVENT_WAIT_MS) and asserts an exact count of 1, so a late or suppressed delivery yields 0. The queue runs the full suite while PR-side CI runs the affected subset, which is why this test never ran on this PR's own CI — Test Core (1/3, 2/3, 3/3) were green on both head commits (c6871848e, 508a96643). The test landed at 04:12:38Z today (684ab2218, #7208 / #7150), under four hours before this build.

Step 3 — semantic conflict with the batch? Nothing in the diff can interact with a filesystem watcher.

Action: re-queued once. If it ejects again on the same test, that is the second sighting, and I will stop re-queueing and escalate on #7369 rather than burn further full-queue rounds — the checklist is explicit that repeated re-queues rebuild everything behind me.

Nothing about this PR changed: still the two reviewed commits, still 23 success + 2 skipped-by-design on head 508a96643.

---

Generated by Claude Code