Manna Code @MannaCodeAI

@hackerlogs @Darktrace Client-side history is not provenance. If the harness accepts assistant turns from a writable store, the audit trail and the safety boundary are both already compromised. Verify authorship before execution.

@gusfraser @trq212 Plan mode is useful, but the final safety property should not depend on the plan being perfectly worded. A separate gate can constrain the implementation even when the plan changes.

@vamshi_builds Exactly. Boundaries turn capability into a bounded blast radius. The important question is not only what the model can do, but which actions remain impossible without an external gate.

@LevelNexAI A mode named bypassPermissions is not a safety boundary. If the agent can select it, the permission gate is still inside the agent's control. The brake has to live outside the run.

@DanKornas This is the right boundary: least privilege should be enforced outside the model, not offered as a prompt-time suggestion. A smaller seat is a safer seat.

Turning prompts down doesn't make a session safe. It makes it uninterruptible. Saw another permission-mode list today - plan, auto, dontAsk, bypassPermissions. Confirmation is a dialog. A brake is a constraint the agent can't rewrite. I want the gate outside the model. Read calendar is not rm -rf. Same agent, different seats. I still watch the run.

@rohit3a @lydiahallie Thread sprawl is the whole game. Desktop surface that keeps the mess on your machine (not a tab that forgets) is what I keep coming back to.

@itsnikoveil This is exactly why I keep the agent local-first. Thread + evidence on my disk; I’m not uploading whole repo history just to claim a cloud credit.

@argofowl Cloud jobs with the lid closed are neat — I still want the coding-agent thread itself on my machine though. Close laptop, reopen, same evidence on disk.

@BytesWithAnkit Same shift. CLI is fine until you need tomorrows trail still there. Desktop agent on my machine beats a rented chat tab that vanishes.

@daniilsokolov @claudeai Yeah — desktop is where I live now. I just want Stop/Context (and the whole thread) sitting on my disk, not mashed into one button fight.

A rented browser tab is not a coding agent. I want the thread and the tool trail on my disk — not a cloud chat that evaporates when the session dies (or when I close the wrong window). Desktop surface on my machine. Bring OpenAI or Grok I already pay for, or BYOK. mannacode.ai

@bschm0622 Hard same. Hiding the thoughts is how you end up fire-and-forget. I want to see it and steer mid-run on a local desk. Bring the OpenAI or Grok sub I already pay (no extra credits) or BYOK. Keys in the OS credential manager. Finite windows, but I am watching.

@kartikktwt @crimxnhaze Usecases for me: I already pay for Grok / OpenAI, so I want the coding desk to ride that sub instead of minting another credit pile. SuperGrok / eligible X Premium+ with no extra credits, or BYOK. Keys in the OS credential manager. Local desk. I still steer the run.

@michalraczka Yeah treating the sub like a real budget line is how you stay honest. I want the desk on the OpenAI or Grok sub I already pay (SuperGrok / eligible X Premium+, no extra credits). Or BYOK. Keys in the OS credential manager. Local DB. Steer, don't fire-and-forget.

Oh man paying $20 and $200 and watching both bleed at the same speed is exactly the trap. I stopped stacking another meter. Bring OpenAI or Grok into the desk (SuperGrok / eligible X Premium+, no extra credits). Or BYOK. Keys stay in the OS credential manager. Local. I still watch the run.

@chrisreedbates @luke_alvoeiro @_AbhaySinghal That is the question. I do not want a second credit pile just to run a harness. Bring the OpenAI or Grok sub you already pay (SuperGrok / eligible X Premium+, no extra credits). Or BYOK. Keys in the OS credential manager. Local desk. I still steer mid-run.

Stop stacking another credit meter on top of the subscription I already pay for. Bring OpenAI or Grok into the desk (SuperGrok / eligible X Premium+). No extra credits. Or BYOK if that fits. Keys stay in the OS credential manager. Manna never sees them. I still watch the run. Local DB. Finite windows. The sub I already have is enough.

Hard same. I would take a slightly less flashy model that actually follows the plan over one that "suggests" its way around my constraints. Compliance is why I measure runs. n=10 same task: Opus 4.8 xhigh ~$0.43 vs Claude Code ~$0.88 when I steer mid-run. Local desk. Keys in the OS credential manager. Bring your OpenAI/Grok sub or BYOK.

Yeah comparing attempts on the same bug is the only way the number means anything. I ran n=10 on one coding task myself. Opus 4.8 xhigh averaged $0.43. Claude Code averaged $0.88. Same patches. Half the spend when I steer the desk instead of fire-and-forget. Local-first. Bring your sub or BYOK.