Repository navigation
Context visibility: accurate per-thread context breakdown ('what is eating my context') #317
Description
The single most-upvoted cross-tool feature cluster (opencode 131+100+50 upvotes across three issues; Claude Code's statusline ecosystem exists solely for this). Also: an INACCURATE meter is worse than none — T3's context ring stuck at 5% then jumped to 100%; Conductor's stuck at 0% after model change.
Extend contextRing.ts into a per-thread breakdown (system prompt, memory, tool outputs, transcript, reserve), verify accuracy per provider against the CLI's own usage reading, show rate-limit reset times on hover (Conductor's pattern), and warn before compaction. If a provider cannot be measured accurately, show no number.
Activity
currentbits commentedon Aug 17, 2026
Research note (Conductor 0.75): the density pattern to copy — a hoverable context ring around the composer revealing account rate limits, reset times, session token usage, and a near-limit warning. Ambient info, zero clicks.
currentbits commentedon Aug 17, 2026
Deep-dive detail (Waku): account-level plan usage popover — surface each provider's account quota (Claude OAuth usage endpoint incl. beta header, ChatGPT wham/usage) in one UI, beyond per-thread token counting.
currentbits commentedon Aug 18, 2026
Orchestration plan (thread a8891f43):
Shared contract landed on main first (
SessionUsage.contextTokens= full prompt incl. cache tokens,SessionUsage.contextWindow= CLI-reported window), then two grok workers in parallel:result.usage.input_tokensomitscache_read_input_tokens/cache_creation_input_tokens, so under prompt caching the ring reads ~0-5% then jumps — the exact failure mode this issue names. Also codextotal_token_usage(cumulative) being added as a per-turn delta, andmodel_context_windowplumbed through so a model change can't stale the static catalog. Providers that can't be measured stay unset — no number rather than a wrong one.contextBreakdown.ts(tool output / transcript / system+tooldefs remainder, chars/4 estimate, measured total authoritative), hover popover on the ring, near-limit compaction warning.Deliberately out of this cut: account-level rate-limit/reset popover (the Conductor/Waku comments). It needs provider quota endpoints, not per-thread token accounting — separate slice. The electron worker is reporting whether the CLIs already emit rate-limit fields.
currentbits commentedon Aug 18, 2026
Done on
coder/context-visibility-accurate-per-thread-c-a8891fThe core bug, measured. Captured a verbatim
result.usagefrom the live Claude CLI:{"input_tokens": 10, "cache_creation_input_tokens": 10203, "cache_read_input_tokens": 17748, "output_tokens": 175}The old formula (
input + output) read 185 tokens; the true prompt is 28,136. A 152× under-read — that is the "stuck at 5% then jumps to 100%" failure, in our own code. A mid-session thread in a real transcript showedinput_tokens: 2againstcache_read: 154815.Second bug, found by scanning the live store rather than the tests. Every
claudemodelInfoentry was missingcontextTokens, socontextWindowForreturned null and the ring never rendered for claude threads at all — 161 usage rows with a measured numerator and no denominator. Added the documented windows (1M for Fable 5 / Opus 5 / Sonnet 5, 200k for Haiku 4.5).What landed
contextTokens= input + cache-read + cache-creation + output. Billable in/out counters unchanged.last_token_usageis the turn delta,total_token_usageis session-cumulative and was being added as if it were a delta;model_context_windownow flows intoSessionUsage.contextWindowand beats the static catalog.contextTokensstays unset and the ring hides. No guessed numbers anywhere.contextBreakdown.ts(tool output / transcript / system+tool-defs remainder, chars÷4, labelled as an estimate — the measured total stays authoritative), hover popover on the ring, amber near-limit warning at 85%.Tests: renderer 926 pass / 0 fail; electron 1166 pass / 1 fail (the pre-existing web-integration reconnect failure, baseline on main). Added
test/contextSeam.test.tscovering the electron→renderer join, since the two slices were built in parallel against mocks of each other; it pins the regression at 14% vs the pre-fix 0%.Not in this cut: the account-level rate-limit/reset popover from the Conductor and Waku comments. It needs provider quota endpoints (Claude OAuth usage, ChatGPT wham/usage), not per-thread token accounting — worth its own issue.
Also filed #441:
electron/providers.jscontains a literal NUL byte, which makes grep silently return zero matches on that file. Cost me three false conclusions while working on this.