bbrysonelite-max The Loop
Ten gates (0–9). In order. No gate skipped, no gate merged into its neighbor. The loop's exit feeds its entrance: step 9's write-down is step 0's next audit.
Brent's canonical 8-step statement (dictated 2026-08-24, maps onto the gates): 1 ground-truth read-only audit (dispatched agent) → 2 spec amended only from audit receipts → 3 plan + independent test-writer, sabotage RED-proven → 4 PR + CI → 5 independent adversarial review, verdicts with own receipts → 6 merge = BRENT'S call, each on his word → 7 prove it after merge (browser walk, or bounce + live eval for daemons) → 8 loop: the running audit is step 1 of the next cycle.
The gates
- GROUND-TRUTH AUDIT — read-only, before anything. DISPATCHED as its own read-only audit agent per slice (never the builder, never the orchestrator's guess). Read the real thing in the real place NOW: live logs, running containers, actual files, live queries. No memory, no docs-as-proof, no "right shape" guessing. Absence is a measurement — a zero needs the query that produced it. Output: what is actually true, with command + output + UTC.
- SPEC — one file, simple artifact first, AMENDED ONLY FROM AUDIT RECEIPTS — never from memory or intention. What changes, what never changes, what "working" means (observable behavior, not vibes). Name the one thing most likely to make it wrong.
- PLAN — slices, not timelines. Each slice: smallest provable unit, its prove-it receipt named in advance, its rollback named in advance (SAFE ROAD: backup + TESTED rollback before anything live).
- INDEPENDENT TESTS — written by an agent that did NOT and will NOT write the implementation, from the behavior contract only. Persisted in the repo's test suite, never ad-hoc shell history. Sabotage check: delete or break the code — tests MUST go red. A test that can't fail is a lie.
- PR — branch, never main. Stage by path, never
git add -A. One slice per PR. - CI — green on the PR, watched to completion. Skipped/gated tests are not green — say so.
- INDEPENDENT REVIEW — a different agent than the builder, read-only tools, own context, ADVERSARIAL by instruction. Verdicts carry the reviewer's OWN receipts (its own commands + output + UTC), never a restatement of the builder's. Builder applies findings; reviewer re-passes the delta. No agent ever grades its own work.
- MERGE = BRENT'S CALL — present the PR + receipts, one plain yes/no. Never merge, never push to main, never "while I was at it."
- PROVE IT LIVE — after merge/deploy, walk the real path with real eyes.
Two variants by project class, same bar:
- UI product (Tiger, sites): real browser, real click path, live endpoint SHA-matched to the merge.
- Daemon/service (Jumbo, bridges, runners): BOUNCE the service (bootout/bootstrap — never kickstart where that law applies), then a LIVE EVAL against the bounced process, then wire/read its probes. Receipt = command + observed output + UTC. Anything not walked is labeled UNVERIFIED, out loud.
- WRITE IT DOWN — reconcile living docs, capture durable decisions to memory (local + Jumbo), append the session/flight record. What you wrote becomes the next loop's ground truth. Not written = will be re-derived wrong later.
Role separation (hard)
| Role | Rule |
|---|---|
| Orchestrator (Fable) | Runs the loop, writes ZERO code |
| Builder (Opus) | Writes code, never tests it, never reviews it |
| Test author | Third agent, blind to implementation, contract only |
| Reviewer | Read-only tools, separate context, re-reviews every fix delta |
| Brent | Merges, decides spend, adjudicates — the only human gate |
Debugging is the same loop
"Vibe debugging" = fixing before auditing. A bug enters at gate 0: reproduce it live and read the actual failure (the exact log line, the exit code) before naming a cause. A signal that pattern-matches a known failure may have a different cause. Then spec the fix and run the remaining gates. A designed refusal (a cap, a guard) is not a bug — it is the system asking a human a question; route it to Brent as one plain decision.
Standing laws the loop inherits
- Prove-it gate: command + output + UTC, or the word UNVERIFIED.
- Fail loud: every non-healthy state names the missing fact and the action.
- Spend and irreversibles halt for Brent:
DECISION REQUIRED: [what] — [A] vs [B] — [recommendation]. Everything reversible proceeds and reports. - Flag, don't fix: wrong things outside the slice get one line, not a detour.
- Never trust a subagent's report — diff the disk.