How 22 commits landed in 4 hours of overnight cron firing. The pattern: a 2-minute interval cron alternates dogfood and fix cycles. Each tick scans the latest report file and dispatches the appropriate sonnet sub-agent. When sub-agents crashed at 1M context, the orchestrator dropped to direct-edit mode and kept landing commits. Verification discipline at every step.
Robert had set up a /loop 2m instruction that fires every 2 minutes. The orchestrator reads its prompt, scans the working state, and dispatches work. The full input is preserved verbatim across firings so the orchestrator never loses context of what loop it’s in.
/loop 2m AE canvas dogfood↔fix loop tick — the same prompt verbatim each time.ls -t docs/anything-engine/wave-*.md | head -3 to find the most recent report file.You are a NARROW-SCOPE FIX agent for orbiter-frontend feat/anything-engine branch. SINGLE TASK: [one paragraph problem + file:line] Steps: 1. Read [file] using grep first to find the right lines 2. Apply the surgical patch (no scope creep) 3. Run pnpm exec tsc --noEmit (must be 0) 4. Run pnpm exec biome lint (must be clean) 5. Run pnpm test (30/30 pass) 6. Commit pathspec-only: git add [specific files] 7. Write report to docs/anything-engine/wave-N-loop-N-fix-*.md FORBIDDEN: - Do NOT touch any other file - Do NOT git add -A / -a / . - Do NOT --no-verify - Do NOT amend - Do NOT spawn other subagents Report: ~150 words, STATUS=PASS|FAIL, commit SHA, what verified.
Sub-agents that try to summarize their work in >200 words are usually hiding a scope creep. Tight reports = tight commits. The orchestrator can scan 8 reports/hour at this size.
Sonnet sub-agents inherit the orchestrator’s context. When that context exceeds ~1M tokens (cumulative system prompt + history + all the files the orchestrator read), every dispatched sub-agent dies at ~250ms with 0 tool uses, returning API Error: Extra usage is required for 1M context.
When sub-agents are durably blocked, Robert authorizes the orchestrator to drop into direct-edit mode and continue landing commits in the main thread. The trade-off: the main-thread context grows, but commits keep landing.
# Pattern: try sonnet, fall back to direct on crash 1. Dispatch attempt: Agent({subagent_type: "general-purpose", model: "sonnet", run_in_background: true}) 2. Crash detected: completed status with API Error in result, 0 tool uses, ~250ms duration 3. Fall back: orchestrator does the work directly (Edit, Bash, Write tools) 4. Commit narrowly: pathspec-only git add, descriptive message, push 5. Continue loop: cron fires again, try sub-agent first, fall back if crashed
This session: zero successful sub-agent dispatches. Every commit landed via the orchestrator’s direct work. Robert’s standing rule held: when sub-agents fail, surface the blocker honestly and keep the loop going. The cron loop is the unit of work, not any individual sub-agent.
No commit landed without all four gates green. The discipline is what made it safe to delete 20K dead lines without breaking anything.
pnpm exec tsc --noEmitpnpm exec biome lint --error-on-warnings src/pnpm testagent-browser screenshot against canvaspkill -f chromiumThe orchestrator doesn’t need to be smart — it needs to keep firing. The cron sets the cadence. The dogfood↔fix alternation prevents either side from getting stale. The 4 verification gates prevent regressions. Sub-agents are an optimization (more parallelism, lower context) but not a requirement — direct-edit mode lands commits when sub-agents are blocked.
This methodology landed:
feat/anything-engineroboulos/orbiter-status-reportIf you want to run this again next session, the pattern is: start a 2-minute cron loop, give the orchestrator a backlog file, mandate the 4 verification gates, and let it run. The cron is the unit of work.