Index of every page on this site: llms.txt
All skills
Run it yourself

End-to-end autonomous delivery loop for large, multi-step build requests in this repo - discover, plan, get the plan independently audited, execute one task at a time with a fresh-evidence verifier, run the full quality gate, deploy to production, get the live deploy independently verified, and hand back a plain-English testing guide. State persists to a run directory so it survives context compaction or a new session. Invoke for large multi-step builds, end-to-end features, migrations over many files, quality loops that must iterate until a measurable bar passes, or any request to plan+build+test+deploy+verify+hand off "without stopping." Trigger phrases: "use Loop Architect," "run this end to end," "keep going until it works," "build and deploy this," "complete the whole plan without stopping." Explicit invocation: `/loop-architect <task>`.

Requires /baseline-ui
  • 8 files
  • 45.2 KB
  • Updated last week
  • GitHub
Use this Skill: https://skilld.dev/gh/colaberryintern/colaberryenterprise_ai_leadershipaccelerator/loop-architect

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈218 tokens for metadata: the name and description. ≈2.7k when used: this file. ≈8.7k more on demand in 7 files.

Description uses 10.9% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈871 characters, or 10.9%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

loop-architect

A governed autonomous build loop for this repo. It does not replace CLAUDE.md - it operationalizes it: the plan-auditor rubric is the confidence-scoring gate, the task-verifier retry cap is Stall Detection (3 identical failures), the hard-stop list is the Escalation Protocol, and every task the loop marks complete still needs an entry in this session's docs/sessions/CC-<id>.md log under CLAUDE.md's hard gate. Read CLAUDE.md in full before running this skill - it governs the run, this skill just sequences it.

Relationship to other skills in this repo: remediate-pr is the narrow maker/verifier loop for fixing one open PR; loop-architect is the general-purpose loop for building a whole feature end to end (plan -> build -> deploy -> verify -> handoff). telemetry-emission still fires after non-trivial builds; screenshot-review still runs before shipping user-facing portal UI; story-build still owns turning a requirements doc into a Basecamp project. Use those directly when only their narrow job is needed - use loop-architect when the ask is the whole pipeline.

Phases (detail in references/)

Phase What happens Reference
A. DISCOVER Read CLAUDE.md + nested CLAUDE.md, inspect code/tests/CI/deploy config/git status, classify the request (finite tasks / repeated loop / quality loop / one-shot) -
B. PREFLIGHT Ask ONE consolidated question set only if something is genuinely undiscoverable and blocking; otherwise assume + log references/execution-contract.md
C. EXECUTION CONTRACT Write request.md + execution-contract.md to the run directory references/execution-contract.md
D. PLAN Small, dependency-aware, verifiable tasks in plan.md references/worked-examples.md
E. PLAN AUDIT loop-plan-auditor subagent scores 10 dimensions, 0-2 each; PASS >= 18/20, no zero; up to 3 cycles. On PASS: fires the Kickoff dashboard. references/plan-rubric.md
F. EXECUTE One task at a time; smallest complete change; loop-task-verifier grades fresh evidence, 6 criteria, 0-2 each; PASS >= 11/12, no zero; up to 3 attempts/task. Fires the Halfway dashboard the first time passed-tasks cross 50%. references/task-verification-rubric.md
G. QUALITY GATE Full repo checks - tsc --noEmit (both stacks), jest, Playwright where present, lint, git diff review for unrelated changes. On green: fires the Shipping dashboard. -
H. DEPLOY Documented prod deploy mechanism only; capture command, commit SHA, timestamp, URL in deployment-log.md references/production-verification.md
I. PROD VERIFY loop-production-verifier subagent checks the live release, not the local build; up to 2 fix/deploy/verify cycles. On PASS: fires the Live dashboard alongside handoff.md. references/production-verification.md
J. HANDOFF handoff.md - production link, plain-English test scenarios, regressions, limitations, troubleshooting, rollback references/worked-examples.md

State ledger schema and resume behavior: references/state-ledger-schema.md.

Progress is visual, not a chat pause (references/milestone-dashboard.md)

This loop never pauses mid-run to narrate progress in chat, and it never asks for routine approval between phases. Instead, at up to 5 fixed milestones (Kickoff -> Halfway -> Shipping -> Live, plus a conditional Blocked variant) it writes one self-contained HTML dashboard into the run directory. .claude/hooks/open-html.sh already auto-opens any .html file this loop writes - just writing the file is enough, no separate "open browser" step. Every dashboard leads with a progress bar (n of N tasks, phase name), an ETA, and a confidence level (mapped to CLAUDE.md's own >0.80 / 0.65-0.80 / <0.65 tiers), followed by what's next. Style matches /baseline-ui tokens and the docs/POST_DEPLOY_WALKTHROUGH.html hero/summary-bar/card pattern - this is "the right design" already established in this repo, reused rather than reinvented. This changes how progress is reported, not whether the loop stops for real - genuine hard-stops (below) still halt the run; they just render as the Blocked dashboard instead of chat prose.

Repository facts this skill assumes (verify at DISCOVER time - these drift)

  • Deploy mechanism: ssh root@95.216.199.47, then ./scripts/deploy-prod.sh backend nginx from /opt/colaberry-accelerator. Never a bare docker compose up: the script holds a flock (concurrent deploys race -> Cloudflare 521 / shared Postgres OOM cascade) and proves the containers are running afterwards. Prod is in registry mode: it pulls the image CI published for main's tip, so a deploy before that image exists is a silent no-op, and rolling back by SHA needs DEPLOY_MODE=build. Backend takes ~60-90s to bind port 3001 - a 502 in that window is timing, not failure. Several sessions deploy main: before deploying, check whether the merge is already live (git merge-base --is-ancestor <merge-sha> HEAD on the box, and container Created later than the merge). If it is, record that in deployment-log.md and go straight to Phase I - redeploying the same code only adds race risk.
  • Remote commands from Windows: PowerShell rewrites quotes and runs $(...) locally before ssh sends it. Anything beyond a one-word command goes in a script file: scp it to /root, run it with bash, delete it. Do not debug quoting.
  • Protected path: work happens on a workstream/* branch, merges to main via PR (1 review + 4 required checks; frontend typecheck is the authoritative frontend gate). Only main deploys to production.
  • Quality gates that exist today: tsc --noEmit (backend and frontend, mandatory minimum), jest under backend/src/**/__tests__/, Playwright under /tests/systemV2 (target coverage, not yet universal - if it doesn't cover the changed surface, say so rather than claiming E2E coverage that doesn't exist).
  • Progress-log hard gate (CLAUDE.md, non-negotiable): every task this loop marks passed needs an entry in docs/sessions/CC-<id>.md with verification evidence before the run's Phase G is considered satisfied. This loop does not substitute for that gate - it feeds it.
  • Escalation boundary (CLAUDE.md Autonomy Model): routine app deploys (scripts/deploy-prod.sh on an already-approved plan) are PROCEED-tier. Changes to docker-compose*.yml, nginx/, DB engine/schema redesign, new paid external dependencies, or anything in CLAUDE.md's Strategic Decisions list are ESCALATE-tier and become a hard-stop for this loop, not a task it executes silently.
  • Runtime state: lives in .loop-architect/runs/<timestamp>-<task-slug>/, gitignored. Never put secrets in it; deploy/prod checks read credentials from the environment (VPS env vars), never from files this loop writes. Gitignored means one machine. The operator switches computers mid-run; when that happened (hire.refactored.ai, 2026-11-09) the ledger, deploy log and handoff were stranded on the old one. So at every milestone, mirror the ledger summary and a resume pointer into the committed session log docs/sessions/CC-<id>.md, and commit handoff.md to docs/ with the run's last PR - never let the only copy of run state be a gitignored folder.

Access preflight (DISCOVER, not DEPLOY)

Check at Phase A, on the machine the run is on, that every tool Phases H-I need works: ssh -o BatchMode=yes root@95.216.199.47 true, gh auth status, the tracker token, and a browser for Playwright. Claude cannot grant itself any of these (the safety classifier blocks minting keys or adding access, correctly), so a gap found at deploy time stalls the run on the operator. Found at DISCOVER, it goes in the single preflight question set with copy-paste steps, and the build proceeds while the operator fixes it. Re-run the check on resume: a resumed session may be on a different machine.

Keep the plan and the tracker honest

  • Superseded, not stale. When an owner decision makes a planned task moot (Ali cut the real domain over directly, so the preview-host tasks never ran), mark the task superseded in the ledger with the decision that replaced it, comment that on its ticket, and re-point any "done when" check at what actually exists. A task left open because the world moved past it reads as unfinished work.
  • Reconcile ledger and tracker at every milestone. Before firing a dashboard, every passed task's ticket must be ticked with its evidence comment, and every ticked ticket must be passed in the ledger. The hire run shipped M1 with 25 passed tasks and their tickets still open; the tracker is what the owner reads.

Assumption defaults when the repo is silent

  • Iteration caps: 3 plan-audit cycles, 3 attempts per task, 2 production fix cycles (matches CLAUDE.md Stall Detection's "same failure 3 times").
  • No time/cost cap is configured repo-wide; if the request doesn't set one, log that as an assumption and proceed - do not block on it.
  • Rollback default: redeploy the last known-good commit SHA with the same documented deploy command; git revert on main if the bad commit must not remain in history.

Running it

/loop-architect <describe the build> or trigger it implicitly with the phrases above. On invocation:

  1. Mint or reuse this session's CC-<YYYYMMDD>-<id> per CLAUDE.md's Session Start Protocol (this loop's tasks land in docs/sessions/CC-<id>.md, the file named for that ID).
  2. Check .loop-architect/runs/ for an incomplete run matching this request's hash before starting a new one (see references/state-ledger-schema.md Resume Behavior).
  3. Proceed through the phase table above, invoking loop-plan-auditor, loop-task-verifier, and loop-production-verifier (in .claude/agents/) as separate graders - never grade your own plan, task, or deployment. Do not pause for routine approval between phases; fire the milestone dashboards (references/milestone-dashboard.md) instead of narrating progress in chat.
  4. Stop on any hard-stop condition (see references/execution-contract.md), fire the Blocked dashboard, and report exactly what's blocked and the exact resume command.

Final report (every run, pass or blocked)

Plan audit score + cycles used; tasks passed/failed/blocked/skipped/superseded + retries; ledger-vs-tracker reconcile result; quality gate result; deploy environment + commit SHA + URL; production-verification result; path to handoff.md; which milestone dashboards fired and where they live; why the loop stopped; the exact next action, if any.

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No third-party reports yet.

Signed by skilld at 7f168b8. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated last week

README badge

README badge for colaberryintern/colaberryenterprise_ai_leadershipaccelerator/loop-architect