Loop directory

Take story P2-2 from spec through implementation, Codex dev, and T11 code review until it commits green to fidelity-perception-arc-2026-06-19.

A second, read-only loop that babysits a first one. Every ten minutes it reads the run's state file and latest log, runs the test suite so its evidence is an exit code rather than a self-report, and applies four explicit escalation triggers (no progress across two checks, repeated identical stack trace, budget exceeded, merge-queue conflict). On any trigger it prints ESCALATE and stops so a human looks; it never restarts, edits, merges, or pushes. Exits when the state file reports PASS or DONE, and after twelve checks regardless. Adapted from ECC's loop-operator agent (affaan-m/ECC, MIT), which lists the same triggers but has no cadence, cap, check command, or exit.

Write a one-page brief where every claim has three or more sources and every link is opened and confirmed to support the claim — the loop that catches hallucinated citations a single prompt never can.

Read every loop's run ledger and find the one root problem showing up across multiple workflows.

Make every live option declare the single fact that would disqualify it upfront, then hunt for that evidence.

Send every question to a cheap model first; escalate to Fable only on a logged failure.

Digest only what materially changed in the sources you must track, each with a plain-language 'so what.'

A cheap model watches your tracked sources nightly; the expensive model writes only the monthly synthesis.

Keep an aging ledger of unpaid invoices and draft the next nudge at each threshold — you send it.

Stay silent on normal days and page you with a pre-investigated cause only when a metric breaks its expected band.

Draft the repeated 80% of a proposal or RFP from your answer library and flag the new 20% for a human.

Each week, compare how work actually happened against the written SOP and flag where they've drifted apart.

Categorize transactions against history and hand a human a clean exception list — it never files anything.

Turn every unread email into a decision — do, delegate, defer, or drop — with the top reply already drafted.

Let analytics find the exact moment users give up, then rewrite the words on that one screen.

Compile app-store reviews and support into a roadmap ranked by how much pain each issue causes.

Sweep where people talk about you each week, cluster the complaints and asks, and deliver the loudest one as a drafted implementation plan.

Turn the pre-purchase questions people ask in DMs, comments, and emails into a reusable, deduplicated answer bank.

Start from 'I thought it did X' support tickets and walk back to the exact page sentence that over-promised.

Draft the next batch of ad creative before your current set fatigues — nothing launches without you.

Watch a curated list on X each morning, rank what's actually working, and draft your take in your own voice ready to review.

Keep your content pipeline stocked with fully-specified briefs mined from your own search traffic, so writers never start from a blank page.

Read competitors' newly published articles and the keywords they target, then plan your moves against the gaps they leave open.

Every week, ask the AI engines the same 'best [category]' questions your buyers ask, log where your brand ranks, and track the trend over time instead of guessing.

Catch thin or duplicate template-generated pages and flag them for a human before Google penalizes the set.

Each week, find the buyer questions people ask ChatGPT and Perplexity that your site doesn't answer yet, and close exactly one of those gaps.

Inventory checkable claims, verify them against primary sources, repair high-risk mismatches, and log what stays unresolved.

Build a seven-section LaTeX preprint with native figures, traceable claims, repeated compilation, and stated weaknesses.

Six weeks, one variable changed per post; measure replies, saves, and questions, and end with a winner or an honest null.

Chase the most important evidence gaps until a memo, brief, spec, or page is genuinely ready to use.