Scenario: Daily-driver terminal agent work
Compiled by pkoskinen
Provisionally: Claude Code remains the stronger daily driver for most people — the harness tuning shows up in fewer stalls, better file-edit hygiene, and instruction files that just work. OpenCode is the right choice if you need provider flexibility (local models, mixed providers, cost routing) or want to see exactly what your agent sends. Evidence volume here is thin and OpenCode ships weekly, so treat every specific complaint below as dated the month it was filed.
Confidence is labeled honestly: weakly supported — based on a small sample; treat as directional.
191 practitioners weighed in for the “Daily-driver terminal agent work” scenario specifically
Strong evidence gets promoted into the record above.
For infra work the transparency isn't a preference, it's a requirement. I can produce the exact bytes OpenCode sent when auditors ask what the agent saw before it touched Terraform state. No closed harness can answer that today. WEAK support level is fair for general use but for regulated environments this comparison isn't close.
Tried the mixed-routing setup from the cost point on Flutter work. Qwen3 Coder local was fine for Dart boilerplate, genuinely bad at anything touching platform channels — silently wrong, which is the expensive kind. Routing by task type sounds elegant until the router is you, guessing.
This matches the submitter's caveat — 'acceptable on internal tooling' was doing a lot of work in that sentence. Added a note to the point. Mixed routing needs a quality floor per route, not just a cost ceiling.
Filed as evidence rather than a comment, but flagging here: the edit-retry rate in point three improved noticeably in 0.6.2 versus 0.6.0 on my repos — roughly one manual retry per day now, not per session. The 'dated the month it was filed' warning in the recommendation is earning its keep.
Sign in to join the discussion, vote, and verify fixes.
Related records
Six weeks on OpenCode as my only agent, coming from Claude Code. Honest summary: I miss the polish daily and I'm not going back. The edit-retry annoyance is real. But last week I traced a bad behavior to a tool description in my own config in ten minutes. With Claude Code that would've been a shrug and a workaround. Owning the failure modes is worth an occasional retry to me.
Rude and possibly accurate. Check back in a year.
"Owning the failure modes" is what we said about self-hosted Jenkins, and I eventually paid someone to take it away. Check back in a year.