embedded-loop

Embedded debug loop

Execution is the success signal: a fix is done when a checked-in repro passes on-device, never when the diff looks right. For environment-dependent bugs the scarce resource is not execution time but occurrences of the triggering condition — extract maximum information per occurrence (capture-first) and synthesize the condition where possible (fault injection).

Pick the loop shape:

Reproducible (native, injected, or replayed)

hypothesis → edit → harness/loop.sh <repro> → analyze → iterate.

Gates:

• Reproduce-first on-target. No fix edits until repros/issue-N.sh fails reliably under harness/loop.sh. A plausible patch with no failing on-device repro is a guess.
• Self-contained repro. For environment-dependent bugs the repro includes its environment — it invokes fault injection (faults/) or trace replay to create the condition.
• Attempt budget. 2–3 failed fix attempts → write down the ruled-out hypotheses (LESSONS.md) → fresh session with that brief decides patch-or-rewrite.
• Confirm under the real condition. "Passes after" means passes end-to-end, repeatedly, under the triggering condition — not once by hand.

Instrument-and-wait (irreducible environment dependence)

hypothesis → add targeted instrumentation that will confirm or kill it → deploy → wait for a watcher-harvested occurrence (or physically trigger) → analyze the bundle → narrow.

Wall-clock per cycle is long (there may be one or two occurrences per day), so every deployed instrumentation round must be hypothesis-driven, never shotgun logging. The capture-first standing rule applies: the response to any occurrence — lab or field — is harvest, then investigate. No occurrence is ever spent un-mined.

Before any repro run

harness/device-reset.sh restores known state. Agent-driven runs on a stateful device drift (leftover processes, half-deployed binaries); without reset-before-run you chase residue from your own previous attempts.