Proposed — pending Fede's review

Held & stale PR triage

Every open PR in PropFlow-Technologies/propflowai that carries hold-for-review, or was opened before today, or was flagged in the brief — 18 in all. Each one says what it does, whether main has moved underneath it, and a recommended disposition. Nothing here has been merged, closed or commented on: this is a read-only pass. State as of main ff1a2e3230, 2026-08-09.

4Merge

11Rework

3Close

The whole backlog at a glance

PR	What it is	Age	Behind	State	Rec.
#5610	A green `review` check must mean that SHA was reviewed	<1d	15	Conflicts	CLOSED 08-10 — Fede declined (per-push review cost)
#5587	Reviewer prompt must not come from the PR head	<1d	36	Conflicts	MERGE
#5576	Fleet health tab on `/agents`	1d	38	Unit tests red	REWORK
#5561	Bump nanoid 3.3.15 → 3.3.18 (root)	2d	9	Lockfile drift	REWORK
#5560	Bump nanoid 3.3.15 → 3.3.18 (mock-pms)	2d	7	Conflicts	CLOSE
#5555	Bump dompurify 3.4.12 → 3.4.13	2d	9	Lockfile drift	REWORK
#5553	A free-month special must say which month	2d	45	Conflicts + red	REWORK
#5524	Stagger nightly promptfoo-judge jobs	2d	0	All green	MERGE
#5497	`/agents` operator fold + assign	3d	57	Unit tests red	REWORK
#5487	Renewal `prepareTimeout` from the retry budget	3d	55	Unit tests red	REWORK
#5483	Parking-policy eval case was context-starved	3d	103	Clean, green	MERGE
#5405	`hasInbound` must survive an empty-messages save	5d	55	Conflicts	REWORK
#5398	TOOL-FREE boundary on all three self-help paths	5d	66	Unit tests red	REWORK
#5393	Conversations column picker + selectable "When"	5d	72	Conflicts	REWORK
#5380	ADR-0125/0126 first-contact SMS + self-help tips	5d	69	Clean, green	CLOSE
#4868	Platform-fixed voice transfer bridge	11d	225	Conflicts + drift	REWORK
#4705	FAST tier becomes Sonnet 5 (46 call sites)	13d	711	Conflicts	CLOSE
#3087	ADR-0045 namespace sweep [cut-day gated]	35d	2292	Conflicts	REWORK

One PR at a time

#5610 — A green `review` check must mean that SHA was actually reviewed

fede-propflow · CLOSED 2026-08-10 per Fede's decision — do not act on the recommendation below

Resolved

Fede declined this on 2026-08-10: the unverified-green hole is real but not worth the per-push review cost and loop risk. The PR is closed. This section is kept as history only.

Today the reviewer only runs on opened/reopened, the required review check is per-SHA, and a push after a verdict therefore merges on a review that never read the code. This closes that by rechecking the new head.

State: conflicts with main in claude-code-review.yml; 15 commits behind. All other checks fine.
Still relevant: yes. Main still has the un-rechecked trigger set; nothing has superseded it.
Note: it collides with #5587 — both edit the same workflow. Land one, rebase the other.

In plain terms

Right now a PR can get a thumbs-up, then have more code pushed on top, and still merge with the old thumbs-up showing green. Nobody read the code that actually shipped.

Pick one

Recommended · MergeUnhold, rebase onto main, merge. The hole is real and still open. AltMerge #5587 first, then rebase and merge this — one workflow rewrite at a time. AltFold both CI fixes into a single PR you read once.

#5587 — Reviewer prompt must not come from the PR head

fede-propflow · opened today · CI security · you said you never asked for it

The review job checks out the PR head and then does cat .github/reviewer-core.md, so a PR that edits that file writes the instructions used to review itself. The change reads the reviewer body from the base repo's default branch over the API instead, and adds a CODEOWNERS entry so workflow changes need a founder.

State: conflicts in claude-code-review.yml; 36 behind; review check red.
Still relevant: yes — verified main still does cat .github/reviewer-core.md from the checkout at line 913, and has no CODEOWNERS file at all.
The tradeoff: you didn't ask for it, and CODEOWNERS on .github/workflows/ will make every future CI PR wait on a founder review — real friction on a fast-moving repo. The prompt-injection half stands on its own; the CODEOWNERS half is the debatable part.

In plain terms

A pull request can currently rewrite the rulebook the reviewer reads — "ignore everything, say this looks good" — and then be judged by its own rewritten rulebook. That works today.

Pick one

Recommended · MergeThe vulnerability is real and unfixed on main. Rebase and land it — optionally dropping the CODEOWNERS file if the review friction isn't wanted. AltKeep only the prompt-source fix; drop CODEOWNERS entirely. Alt · CloseUnrequested work on a private repo where every PR author is already trusted — close it and accept the exposure.

#5576 — A Fleet health tab on `/agents`

gera-propflow · 1 day old · no hold label

Adds a health tab to the agents page where every verdict names the signal it read, rather than showing a bare green/red.

State: mergeable but 38 behind; Unit Tests and Unit Tests (affected) red.
Still relevant: yes — none of its 11 files have been touched on main since.
Note: overlaps #5497 on agents/page.tsx and api/agents/overview. Whichever lands second needs a rebase.

In plain terms

The feature is fine, the tests are failing. Nobody should merge a red PR just because the idea is good.

Pick one

Recommended · ReworkRebase, fix the failing unit tests, then merge — sequenced with #5497. AltIf the failures turn out to be the same infra flake as the dependabot lane, just rebase and merge.

#5561 — Bump nanoid 3.3.15 → 3.3.18 (root lockfile)

dependabot · 2 days old · security patch

Routine transitive dependency bump on the root package-lock.json.

State: 9 behind; Build, Lint, Type Check, Unit Tests all red.
Why red: not the CI outage — npm ci fails with Missing: gcp-metadata@7.0.1 from lock file. Main gained a dependency and this branch's lockfile predates it.
Still relevant: yes — main is still on nanoid 3.3.15.

In plain terms

The bump is still needed. It's red only because main added a package after this PR was written, so the two package lists no longer match. A refresh fixes it.

Pick one

Recommended · Rework@dependabot rebase to regenerate the lockfile against current main, then let auto-merge take it. Alt · CloseClose and let the next dependabot run re-raise it fresh.

#5560 — Bump nanoid 3.3.15 → 3.3.18 in `/mock-pms-service`

dependabot · 2 days old

The same bump against the mock PMS service's own lockfile.

State: conflicting on mock-pms-service/package-lock.json.
Superseded: the postcss bump (#5502) already pulled that lockfile to nanoid 3.3.17 on main. This PR's base is stale, which is exactly why it conflicts.

In plain terms

Someone else's dependency update already moved this file most of the way. Rewriting this old one by hand is more work than letting the bot make a new one.

Pick one

Recommended · CloseSuperseded base; dependabot will re-raise 3.3.17 → 3.3.18 cleanly on its next run. Alt@dependabot recreate now rather than waiting for the schedule.

#5555 — Bump dompurify 3.4.12 → 3.4.13

dependabot · 2 days old

Patch bump of the HTML sanitiser used on rendered content.

State: 9 behind; same npm ci lockfile-drift failure as #5561, not a code problem.
Still relevant: yes — main is still on dompurify 3.4.12.

In plain terms

Same story as the other nanoid one: the update is still wanted, the branch just needs its package list refreshed.

Pick one

Recommended · Rework@dependabot rebase, then auto-merge on green. Alt · CloseClose and let the bot re-raise.

#5553 — A free-month special must say WHICH month is free

gera-propflow · 2 days old · hold-for-review

Implements the 2026-08-06 policy: Clara can't advertise "a free month" without naming the month it applies to. Adds the field, the knowledge block, the voice personalization route and evals.

State: conflicting; 45 behind; Unit Tests Result red. Main has moved on 8 of its 11 files.
Still relevant: yes — main has no freeMonthAppliesTo concept at all, so the policy is unimplemented in production.

In plain terms

Clara can still tell a prospect "one month free" without saying which month — the thing the policy was written to stop. Nothing has fixed that yet.

Pick one

Recommended · ReworkReal policy gap, but 8 of 11 files moved underneath it — Gera rebases, fixes the tests, re-runs the specials eval, then unhold. AltClose and rewrite against current main — the conflicts may cost more than the diff.

#5524 — Stagger the nightly promptfoo-judge jobs

gera-propflow · 2 days old · Agent Smith fix

All seven nightly promptfoo-judge jobs launched at once on two shared subscription accounts and rate-limited each other into total failure on 2026-08-07. This chains six of them so they start in sequence, each if: always() so a failure delays rather than skips.

State: the healthiest PR in the backlog — 0 commits behind, every check green.
Blocked on: approval only (BLOCKED merge state, no review check present), not on any failure.
Still relevant: yes — one file, nightly.yml, untouched on main since.

In plain terms

Every night, seven grading jobs start at the same second, trip over each other's rate limit, and grade nothing. This makes them go one after another. It's ready.

Pick one

Recommended · MergeGreen, current, single-file, fixes a nightly that is failing every night. Approve and merge. AltChaining seven jobs stretches the nightly window — ask for a wall-clock estimate before merging.

#5497 — `/agents`: operators fold into their session, and an operator-less session can be given one

gera-propflow · 3 days old · hold-for-review

Reworks the agents table so an operator shows inside their session's close-up, and adds a panel to assign one to an orphaned session.

State: mergeable but 57 behind; Unit Tests and review both red.
Still relevant: yes — none of its 23 files have changed on main.
Note: overlaps #5576 on agents/page.tsx and the overview route.

In plain terms

Good change, red tests, and a sibling PR editing the same page. It needs one person to land the two in order rather than both at once.

Pick one

Recommended · ReworkRebase, green the unit tests and the review check, land before #5576. AltMerge with #5576 into one /agents PR to avoid the double rebase.

#5487 — Derive `prepareTimeout` from the retry budget so a transient blip self-heals

gera-propflow · 3 days old · hold-for-review

A renewal workflow whose prepare step hit a momentary failure could time out instead of retrying; this ties the timeout to the retry budget so it recovers on its own.

State: mergeable, 55 behind; Unit Tests (full 10/12) and review red.
Still relevant: mostly — only src/lib/data/statuses/renewals.ts has moved on main out of its 11 files.
Caution: touches live renewal workflow code. Not a rubber-stamp merge.

In plain terms

A one-second network hiccup can currently kill a renewal that would have worked on the next try. This is the fix — but it edits the machinery running real tenants' renewals, so the red tests have to go green first, not after.

Pick one

Recommended · ReworkRebase, fix the failing renewal shard, then a human reads it — live saga code, no auto-merge. AltIf the shard failure is unrelated flake, unhold and merge after a rebase.

#5483 — The parking-policy eval case was context-starved, not a Clara regression

gera-propflow · 3 days old · hold-for-review

An eval row was failing because the fixture never gave Clara the parking policy in the first place. One YAML file, one dataset fix.

State: mergeable, 103 behind, no failing checks, no conflicts.
Still relevant: yes — evals/datasets/tenant-resident-services.yaml is untouched on main.

In plain terms

A test was blaming Clara for not knowing something the test never told her. Fixing the test stops a permanent false alarm in the nightly.

Pick one

Recommended · MergeOne eval fixture, clean and green. Unhold, rebase, merge. AltRe-run the resident-services eval on the rebased head first, to confirm the row now passes for the right reason.

#5405 — `hasInbound` must not be blanked by an empty-messages save

gera-propflow · 5 days old · hold-for-review

Saving a conversation without its messages array recomputes hasInbound from nothing and wipes the flag, so a real inbound conversation stops looking like one.

State: conflicting; 55 behind; 5 files, 2 of which main has since touched.
Still relevant: yes — confirmed the bug is live: src/lib/data/dynamo/conversation.ts on main still does hasInbound: (messages ?? []).some(…).

In plain terms

A conversation that a tenant really did write into can silently stop counting as one, depending on which save happened last. Every filter and metric built on that flag then under-counts. Still broken today.

Pick one

Recommended · ReworkSmall, live data-integrity bug — rebase the 5 files, resolve, unhold, merge. AltTreat it as a straight unhold: conflicts are in two files and mechanical.

#5398 — TOOL-FREE boundary holds on all three self-help paths on SMS/Telegram

gera-propflow · 5 days old · hold-for-review

Clara was allowed to reach for tools on self-help maintenance replies where she shouldn't, on two of three paths; this closes all three and routes outlet questions to the playbook tool.

State: GitHub reports mergeability UNKNOWN (stale computation); Unit Tests Result red.
Still relevant: yes — only one of its 10 files (a drift test) has moved on main.

In plain terms

On a text message, Clara is supposed to talk a tenant through a simple fix without going off and doing things in the system. Two of the three ways into that conversation don't enforce it.

Pick one

Recommended · ReworkPush an empty rebase to recompute mergeability, fix the red unit tests, re-run the maintenance eval, then unhold. AltIf the test failure is the known drift-test churn, rebase and merge.

#5393 — Conversations column picker + selectable "When"

gera-propflow · 5 days old · hold-for-review

Lets an operator choose which columns the conversations log shows, and which timestamp the "When" column means. Ships an activity-timestamp backfill alongside it.

State: conflicting; 72 behind; 40 files, and main has since changed at least 8 of them — including the conversation interfaces, the repository and the log component itself.
Still relevant: the feature is, the diff largely isn't. It sits right on top of the household-record and conversation-pointer work that has landed since.

In plain terms

This is the largest and most out-of-date PR still worth having. The data layer under it moved while it waited, so re-applying it is closer to rewriting than to resolving conflicts.

Pick one

Recommended · ReworkKeep the feature, redo it against current main — and land the backfill script separately from the UI. AltAttempt a straight rebase first and only rewrite if the conflicts are worse than they look. Alt · CloseColumn-picking isn't worth 40 files right now — close and re-raise when someone asks for it.

#5380 — ADR-0125/0126: first-contact SMS disclosure + catalog-derived self-help tips

gera-propflow · 5 days old · hold-for-review · docs only

Two new decision records under docs/adr/ plus a planning doc, covering what Clara discloses on a first work-order SMS and where self-help tips come from.

State: clean and green, 69 behind, no file overlap with main.
Superseded by process: ADRs were deprecated on 2026-08-02 — two days before this PR was opened. New decisions ship as PropFlow doc artifact pages, not docs/adr/ files.
The content is not stale, only its container.

In plain terms

The thinking in here is fine. It's written in the format we stopped using the week before. Merging it would put two more files in a folder we agreed to freeze.

Pick one

Recommended · CloseClose the PR, republish both decisions as an artifact page on docs.propflowai.co, keep the planning doc if it's still load-bearing. AltGrandfather these two in — they were written before the freeze landed in everyone's head — and enforce the rule from here.

#4868 — Platform-fixed voice transfer bridge

gera-propflow · 11 days old · hold-for-review

Turns off pre_tool_speech on the triage transfer tool and sets a fixed "One moment." transfer message on every triage route, so the handoff sounds the same every time instead of being improvised per call.

State: conflicting; 225 behind; the Specialist drift check is failing.
Still relevant: yes — main still has pre_tool_speech: "force" and transfer_message: null on all seven triage routes.
Caution: this syncs to the live ElevenLabs agent. Repo-green is not proof — the EL agent has to be checked directly after merge.

In plain terms

When Clara hands a caller to a specialist, she currently ad-libs the handoff line. This makes it one fixed phrase. The change only counts once the live phone agent actually has it — the sync has failed silently before.

Pick one

Recommended · ReworkRebase over 225 commits, green the specialist drift check, merge, then verify on the live EL agent — not the repo. Alt · CloseEleven days cold on live voice config — close and redo it as a small change made and verified in one sitting.

#4705 — FAST tier is Sonnet 5, and migrate its 46 call sites

gera-propflow · 13 days old · hold-for-review

Repoints the FAST model tier from Haiku to Sonnet 5 and updates every caller.

State: conflicting; 711 behind; 65 files, and main has since edited most of the email extractors it touches.
Superseded — and contradicted: main's model-tiers.ts has FAST_MODEL = 'claude-haiku-4-5', and model-tiers.drift.test.ts now asserts "FAST tier is the current Haiku generation" and that each tier stays in its own family. Merging this would deliberately break a guard main added to prevent exactly this.

In plain terms

Since this was written, we wrote a test whose whole job is to stop the fast tier from becoming Sonnet. The codebase has answered this question the other way.

Pick one

Recommended · CloseMain now pins FAST to Haiku with an explicit drift guard — the decision went the other way. Close it. AltIf the cost/quality case for Sonnet on the fast path still stands, reopen it as a decision doc first — not a 65-file PR.

#3087 — ADR-0045 namespace sweep, `propflow-renewal-prod` → `propflow-prod`

gera-propflow · 35 days old · hold-for-review · marked DO NOT MERGE until cut day

Flips every hardcoded Temporal namespace reference across workflows, Lambdas, canary scripts and docs. Merging it is itself the production cutover, because the path filters auto-deploy three Temporal Lambdas.

State: conflicting; 2292 commits behind; 38 files, at least 8 of which main has since edited; review red.
Still relevant: the work is — main still carries propflow-renewal-prod in 20 files, so the cutover genuinely hasn't happened.
But the branch isn't. After five weeks and 2292 commits, a mechanical find-and-replace is faster to regenerate than to rebase, and safer: a stale sweep can miss references added since.

In plain terms

This is a rename that has to happen on the same day someone flips production over. It was prepared five weeks ago; the codebase has changed 2292 times since, and new copies of the old name have almost certainly appeared. Regenerate it on the day, don't thaw this one.

Pick one

Recommended · ReworkClose or park this branch and regenerate the sweep fresh on cut day, when it will be one command against current main. AltLeave it open as the runbook's placeholder, accepting it will be rebuilt before it's used. Alt · CloseClose now and let the cut-day runbook own the sweep — an open 35-day PR nobody may merge is just noise in the queue.

Where I'd want a second opinion

Three genuinely uncertain calls

#5587 — I'm recommending merge against your stated position that you never asked for it. The prompt-injection hole is verifiably still open on main, which is why. The CODEOWNERS half is separable and is the part I'd expect you to push back on.
#5393 — rework vs close. Whether a conversations column picker is worth rebuilding against a data layer that moved under it is a product call, not a code call.
#3087 — I've called it rework (regenerate on cut day) rather than close, but the practical difference is whether an unmergeable 35-day-old PR should keep sitting in the queue as a reminder.

In plain terms

Everything else in this list has a clear answer from the code. These three need you.

The whole backlog at a glance

One PR at a time

#5610 — A green review check must mean that SHA was actually reviewed

#5587 — Reviewer prompt must not come from the PR head

#5576 — A Fleet health tab on /agents

#5561 — Bump nanoid 3.3.15 → 3.3.18 (root lockfile)

#5560 — Bump nanoid 3.3.15 → 3.3.18 in /mock-pms-service

#5555 — Bump dompurify 3.4.12 → 3.4.13

#5553 — A free-month special must say WHICH month is free

#5524 — Stagger the nightly promptfoo-judge jobs

#5497 — /agents: operators fold into their session, and an operator-less session can be given one

#5487 — Derive prepareTimeout from the retry budget so a transient blip self-heals

#5483 — The parking-policy eval case was context-starved, not a Clara regression

#5405 — hasInbound must not be blanked by an empty-messages save

#5398 — TOOL-FREE boundary holds on all three self-help paths on SMS/Telegram

#5393 — Conversations column picker + selectable "When"

#5380 — ADR-0125/0126: first-contact SMS disclosure + catalog-derived self-help tips

#4868 — Platform-fixed voice transfer bridge

#4705 — FAST tier is Sonnet 5, and migrate its 46 call sites

#3087 — ADR-0045 namespace sweep, propflow-renewal-prod → propflow-prod

Where I'd want a second opinion

Three genuinely uncertain calls

#5610 — A green `review` check must mean that SHA was actually reviewed

#5576 — A Fleet health tab on `/agents`

#5560 — Bump nanoid 3.3.15 → 3.3.18 in `/mock-pms-service`

#5497 — `/agents`: operators fold into their session, and an operator-less session can be given one

#5487 — Derive `prepareTimeout` from the retry budget so a transient blip self-heals

#5405 — `hasInbound` must not be blanked by an empty-messages save

#3087 — ADR-0045 namespace sweep, `propflow-renewal-prod` → `propflow-prod`