PuneetSingh⚡️

I would love to change the world but they won't give me the source code

[PuneetSingh⚡️ @iPuneetSingh]

@madebygps This closes a painful gap; I’d also make the attachment command emit a stable URL in CI logs, so agent-created issues stay reproducible.

[PuneetSingh⚡️ @iPuneetSingh]

@bkdgiffug Centralising skills is useful; I’d add a lockfile with source+version and a dry-run diff, otherwise CI can still drift between agents.

[PuneetSingh⚡️ @iPuneetSingh]

@omarsar0 The hidden metric for these systems is failure recovery: can one bad tool output be isolated without poisoning shared memory? I’d instrument that first.

[PuneetSingh⚡️ @iPuneetSingh]

@pauliusztin_ I’ve found a compact summary plus retrieved snippets beats dumping history; logging what was dropped makes production misses explainable.

[PuneetSingh⚡️ @iPuneetSingh]

@mohtashamdotdev I’d still keep one external harness for retries and audit logs; chat is great for iteration, but production runs need replayable state.

[PuneetSingh⚡️ @iPuneetSingh]

@HARNESSROUTER I’d test one failure path across each harness before standardising—tool retries and context carryover expose the real integration cost.

[PuneetSingh⚡️ @iPuneetSingh]

default covers committed + uncommitted vs main; ask uncommitted-only for a narrower pass. try-this: local review, push unchanged, confirm the PR comment says already reviewed instead of a second billable run.

[PuneetSingh⚡️ @iPuneetSingh]

paying Bugbot twice for the same diff is a silent tax. run /review-bugbot locally first — it stores the git patch-id. open the PR with that same diff and SCM Bugbot skips the re-review.

[PuneetSingh⚡️ @iPuneetSingh]

patch-id hashes change content, not commit SHA, so a rebase that keeps the same edits still matches. bugbot run / @cursor review uses the same key. fix findings after local review and the PR correctly gets a fresh pass.

[PuneetSingh⚡️ @iPuneetSingh]

@GitHub_Daily This is the bit most framework demos skip: build the loop once, then add tools and memory; you can debug each layer before orchestration gets fancy.

[PuneetSingh⚡️ @iPuneetSingh]

@leodev Good orchestration only helps if I can see what each agent is doing; a calmer activity view would make parallel runs usable for real repo work.

[PuneetSingh⚡️ @iPuneetSingh]

@unicodef1wn I do the same in agent runs: keep the task narrow, save the working flow as a skill, and test it on one small job before making it routine.

[PuneetSingh⚡️ @iPuneetSingh]

@every I've found the small personal benchmark beats generic evals: one real task, one expected outcome, then compare models before wiring it into a routine.

[PuneetSingh⚡️ @iPuneetSingh]

pass a trimmed chat_ctx (exclude_instructions + exclude_function_call) so the collector does not inherit the whole support prompt. Python-only for now. try speak 3 digits then press the rest.

[PuneetSingh⚡️ @iPuneetSingh]

IVR that only hears spoken digits fails when someone mashes the keypad. GetDtmfTask takes DTMF or spoken digits, interrupts on tone, returns user_input as the digit string.

[PuneetSingh⚡️ @iPuneetSingh]

num_digits=10 for a phone capture, ask_for_confirmation=True to read back. dtmf_input_timeout default 4s per digit. stop event defaults to # (pound).

[PuneetSingh⚡️ @iPuneetSingh]

@dexhorthy @vaibcode @tbrownio @linear @colemurray The discuss/do/compound loop plus per-session traces sounds like the useful bit; how are you measuring when a task should compound vs restart?

[PuneetSingh⚡️ @iPuneetSingh]

@sydneyrunkle I’d route on task shape first, then keep a cheap fallback when context grows. Logging route + outcome makes the savings believable.

[PuneetSingh⚡️ @iPuneetSingh]

@Docker @rishabincloud @krisvelkov @khushbooverma_ The sandbox angle is the right baseline—I’d also log tool and network calls per run, otherwise a blocked secret leak is hard to prove later.

[PuneetSingh⚡️ @iPuneetSingh]

@AIWorksLabon9h @Rahatcodes This is the bit I was missing—policy shapes the run, but human stage approval must be the only artifact promotion path. No silent auto-advance.