Limitations.

LinkedIn / Twitter / Press wire publishing

Demo OAuth, PublishedPost rows, demo URLs, and EngagementSnapshot metrics exist for local testing. Real provider OAuth and external API posting are not wired.

Feature gap

Stripe billing

Subscription tier is stored in the database but no payment flow. The /settings page lets you switch tiers as a no-op.

Feature gap

Email / Slack outbound

Notification rows are created in-app, but no real email/Slack delivery. SMTP and webhook code paths exist but are unconfigured.

Feature gap

Variant body parsing extracts wrong section

Issue #4 from the lessons-learned deck. Writer wraps body in --- delimiters; regex parser occasionally captures the 'Craft Notes' paragraph instead of the body. Fix: replace with submit_variants tool call.

Critical bug

Timezone display assumes UTC

Scheduled times shown in calendar are UTC, not the user's local time. Date can appear one day off for late-night users.

Polish

Evaluator-optimizer loop

Editor's feedback doesn't drive a Writer revision. Just one pass through the pipeline. The 'multi-agent' framing earns its keep only when there's a real feedback loop — adding this is the highest-leverage product change.

Tech debt

Model differentiation per agent

All 4 agents currently use the default Sonnet model. Strategist would benefit from Opus + extended thinking; Analyst could use Haiku for routine queries. ~30% cheaper AND smarter when differentiated, but not yet wired.

Tech debt

Parallel variant review

Editor receives all 3 variants in one prompt and emits one combined review. Could be 3 parallel calls for ~3x faster wall-clock latency.

Tech debt

LLM-based intent router

Current routing uses keyword regex which fails on ambiguous briefs. A Haiku-call router (~$0.0002/call) would handle composite intents correctly.

Tech debt

Retrieval has no calibrated abstain threshold

Semantic memory is real cosine similarity (recall@5 0.87 on a labelled 34-doc corpus, vs 0.28 for the ILIKE keyword search it replaced), but raw cosine is not a confidence. A query with no answer in the corpus still scores 0.54 against a mean of 0.47 for irrelevant documents, so the crew cannot decide 'I don't know' from the score. The eval gate checks the no-answer query doesn't collapse; it deliberately does not claim the number is calibrated.

Tech debt

Vector scan runs in Python on SQLite

Postgres computes similarity in SQL against an HNSW index and returns only k rows. SQLite pulls the org's corpus into Python and scores it there, which crosses a 100 ms per-retrieval budget at roughly 2,500 memories per organization. Measured, not guessed — and a floor, since the measurement is in-memory SQLite with no network. The fix past that point is Postgres, not faster Python.

Tech debt

One known retrieval failure, kept rather than relabelled

'How much did we raise and from whom?' does not retrieve 'closed a $40M round led by Kestrel Partners' — neither the semantic retriever nor the keyword baseline finds it. It stays in the eval suite as a failing case, because a suite where every case passes has stopped measuring anything.

Tech debt

External approvers without seats

The wedge feature in the product positioning. Tokenized review links for legal / board / outside counsel. Data model supports it; UI doesn't.

Feature gap

Multi-stage approval chains (legal AND exec)

Today: single linear approver. Tomorrow: parallel approval chains with auto-escalation.

Feature gap

Audit log export

Compliance teams in regulated industries need timestamped approval trails. Data is captured; export endpoint isn't built.

Feature gap

Website scan depth

Onboarding can scan one submitted URL, extract voice/facts, and store the cleaned copy as memory. It does not yet crawl sitemaps, docs, blogs, or social profiles.

Feature gap

Calendar drag-and-drop

Items scheduled only via modal or the '+ Schedule' button. No drag from one day to another.

Polish

Test user is shared

test@commscrew.ai is the only seeded account. Anyone using the demo logs in as the same user. Multi-user testing requires registering new accounts.

Polish

Walkthrough completion is per-browser

localStorage flag, not server-side. Different browser = walkthrough fires again.

Polish

No SOC 2, no SSO, no enterprise procurement

This is a showcase project, not an enterprise vendor. If you need those things, use Writer.com.

Polish