fix(flows): unbreak main — contract_renewal throws; rewrite the two tests #565 turned red

[yinlianghui]

Description

main is currently RED. Both failures are tests #563 introduced, and this PR fixes them. It also fixes a live regression #565 left behind.

#565 landed the same CEL-envelope conversion this branch originally carried, using the P tagged-template helper — a cleaner notation, so the flow files here take main's version wholesale. This PR is now the remainder: the defect #565 made reachable, the two tests its merge turned red, and the guards that keep the class fixed.

Type of Change

• Bug fix (non-breaking change which fixes an issue)

Related Issues

Follow-up to #495 / #563. Fixes the fallout of #565.

1. contract_renewal throws on every sweep (live regression on main)

#565 correctly wrapped the notice-window gate as a CEL envelope so it finally evaluates. That exposed a second defect sitting underneath:

timestamp(currentContract.end_date) <= daysFromNow(int(currentContract.renewal_notice_days))

end_date is a DATE field and arrives as YYYY-MM-DD, but CEL's timestamp() accepts only a full ISO 8601 datetime — it throws timestamp() requires a string in ISO 8601 format.

While the condition was a bare string it was never evaluated, so this sat latent. Making the envelope real makes it throw mid-sweep. Net effect on main today: contract_renewal still books nothing, having traded a silent no-op for an exception.

Fixed by appending T00:00:00Z. Verified load-bearing — reverting just that one change fails four contract_renewal tests:

× books a renewal task and notifies the owner inside the notice window
× honours each contract's own renewal_notice_days, not a shared constant
× opens a pre-filled renewal opportunity only when auto_renewal is on
× is idempotent across repeated sweeps within the same window

2. The two tests #565 turned red

Both came from #563 and both fired exactly as designed — they existed to go red the moment someone fixed the behaviour they pinned.

test	why it failed	fix
flow-scheduled → KNOWN DEFECT	asserted the stagnation gate stays shut; #565 opened it, so it failed with the message written for this moment: "gate now opens — rewrite this test"	rewritten to assert real behaviour
flow-record-change → opportunity_approval	asserted typeof startCondition === 'string'; #565 made it an envelope	accepts either form, checks a non-empty expression exists

The second one is worth calling out as a flaw in my original test: it pinned the notation rather than the thing that matters. A legitimate refactor should not have broken it. It now asserts what it always meant — that a gate exists and carries an expression.

The rewritten scheduled tests assert real behaviour for both sweeps: nudge task and notification created, repeated sweeps stay idempotent, the sweep re-arms once the previous stall task completes, and for renewals the task/notification/auto_renewal opportunity, the per-contract notice window, and no duplicate renewal deal.

3. task_recurrence overflowed month-end instead of clamping

Independent of the above. advanceDate stepped months with Date.setMonth, which rolls a nonexistent day forward: a task due Jan 31 spawned its next occurrence on Mar 3, drifting deeper into the following month every time and skipping February outright.

Month and year steps now move to the 1st before the arithmetic, then clamp: Jan 31 → Feb 28 (Feb 29 in a leap year), Mar 31 → Apr 30, Feb 29 2028 → Feb 28 2029.

advanceDate also moves inside the handler — it was the last module-level helper in src/objects/, and every sibling hook documents why that is unsafe (L2 hook bodies run body-only in the QuickJS sandbox, where module scope is not available).

Residual behaviour, asserted rather than left to be discovered: each occurrence is computed from the previous due date, not an anchor day, so a month-end series settles on the shorter day (Jan 31 → Feb 28 → Mar 28). Preserving the anchor needs a schema change; what this replaces drifted forward without bound, so clamping is a strict improvement.

What keeps this fixed

Two guards in test/flow-scheduled.test.ts:

1. The engine asymmetry is pinned — a bare string still evaluates false where an envelope evaluates true. If that flips upstream, it surfaces as a deliberate review rather than a silent behaviour change.
2. Every registered flow is walked, including nested loops, failing if any loop body reintroduces a bare string condition. It asserts the walk found something, so it cannot pass vacuously.

campaign_enrollment gains a runtime suite (eligibility, opt-out, cross-campaign dedupe, closed-campaign refusal) and leaves PENDING_FLOWS — now down to case_csat_followup and demo_bootstrap.

⚠️ Behaviour change

With both defects fixed, these sweeps start writing records they never wrote before:

• The first run after deploy will nudge every currently-stalled deal and open renewal tasks for every contract already inside its notice window — a one-time burst proportional to the backlog.
• The idempotency gates that stop a daily sweep from piling up duplicates have never actually run in production, because they were the broken part. Each is now covered by a repeated-sweep test asserting exactly one task / notification / opportunity across two consecutive sweeps.

Testing

Run on the merged tree (main incl. #565):

• pnpm verify exits 0 — 382 passed / 21 files (2 skipped, pre-existing)
• pnpm test:e2e — 11 passed against a cold database
• pnpm typecheck, pnpm validate, pnpm lint all clean
• Confirmed the timestamp() fix is load-bearing by reverting it (4 failures)
• Confirmed both main-red tests fail on origin/main and pass here

Checklist

• I have added a changeset (two — one per defect)
• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• My changes generate no new warnings
• I have added tests that prove my fix is effective
• New and existing unit tests pass locally with my changes

Note on CI

pull_request-triggered workflows have not started for any PR in this repo since ~04:27 UTC (push-triggered runs on main still work). The head commit here may therefore show no checks through no fault of the branch — worth a look independently of this PR.

[vercel]

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

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
hotcrm	Ignored		Jul 31, 2026 6:51am