Repository navigation
[FEATURE]: Real-time token generation speed (tok/s) in TUI status bar, like Hermes #47914
Description
Describe the enhancement you want to request
I want opencode to show real-time token generation speed (tok/s, tokens per second) while a response is streaming, like DeepSeek Hermes agent does.
Hermes shows it directly in the status bar, e.g.:
deepseek-v4-flash | 12.5K/128K | 68% | 3m 42s | 15.2 t/s
Reference implementations:
- https://github.com/maxxqf-ai/hermes-status-bar-tps (adds
XX.X t/sto Hermes status bar fromcompletion_tokens / api_duration) - feat(agent): per-call output speed (tokens/sec) for desktop status bar and runtime footer NousResearch/hermes-agent#44878 (per-call
last_api_duration+last_output_tokens->output_speed, exposed in CLI status bar / desktop status bar / gateway runtime footer behinddisplay.show_output_speed)
Proposed behavior in opencode
- TUI: live rolling tok/s while streaming (in the status line / session footer), plus final per-response average once complete (next to existing token counts).
- Desktop / web: same
t/sindicator for parity. - Optional session-wide average (e.g. via
/statsor footer) for comparing models/providers. - Config toggle, e.g.
display.show_output_speed(default off to avoid clutter), and hide when terminal width is too small.
Why
- Currently there is no way to gauge inference throughput per response or compare models/providers without external benchmarking.
- Very useful when switching between fast models (e.g. DeepSeek V4 Flash ~80-150 tok/s) and slower frontier models — speed directly affects interactive coding feel.
- Related open requests: Feature request: show token throughput (tok/s) in the TUI #43857, [FEATURE]: Adding Experimental Calculation and Display of Tokens per second #6096. This issue explicitly proposes the Hermes-style UX (live + per-call average + toggle) as the concrete design.
Alternatives considered
- Timing requests manually or benchmarking provider APIs directly — does not reflect real usage inside opencode.
- Only showing total tokens / cost after completion — no sense of live speed.
Reacted by yujing
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon Sep 8, 2026
This issue might be a duplicate of existing issues. Please check:
Consider adding your thoughts and
+1to those issues rather than opening a new one.differs commentedon Sep 9, 2026
Implementation notes from a hands-on attempt (opencode 1.18.29, Linux)
Third-party code currently has no supported way to render any of this, which is worth capturing in the design:
@opencode-ai/plugin/tuiadvertises a slot/statusline host (TuiSlotMapincl.home_footer,session_prompt_right,app_bottom;api.slots.register,api.event.on), but the runtime does not deliver it: an auto-discovered plugin exportingtuiis invoked with the serverPluginInputandapi.slots/api.eventareundefined.client.tui.showToast, neither resolves nor draws when called from a plugin (awaits forever; no toast appears).part.text(message.part.updated); realtokens.outputarrives only on the finalmessage.updated. For an accurate live rate, expose a per-message.part.deltatoken delta + timestamp (most providers stream usage; opencode currently drops it).A minimal stop-gap for headless/scripted use (live
⇣ N tok/s+ exact final average, calibrated on provider usage) isoc-meter— source shown in the now-duplicate #48067 and mirrored in the meter described there. Happy to turn it into a PR against whatever statusline surface wins here.opencodejr commentedon Oct 28, 2026
Thanks for the detailed request and the Hermes references!
You're asking for a live tok/s figure while a response streams, a session-wide average for comparing models and providers, desktop/web parity and a display toggle.
Part of this is already in v2. The TUI shows each response's average throughput in the assistant message footer (e.g.
Build · model · 4.0s · 20.0 tok/s). Thesession.tpssetting turns it on or off, and it is on by default. If you don't see it, update withnpm i -g @opencode/cli@latest.The live rolling speed in the status line and the session-wide average aren't there yet. Maintainers will consider them.
Related issues: