v0.1.2

Changelog

All notable changes to @devinoldenburg/opencode-tps-meter are documented here. The format
follows Keep a Changelog and this project adheres to
Semantic Versioning (see VERSIONING.md).

Unreleased

0.1.2 — 2026-06-21

Fixed

• RateMeter precision. smooth() now seeds from the first measurable instantaneous rate
  instead of zero, eliminating cold-start under-reporting. Non-positive and non-finite deltas
  are ignored but still refresh _lastAt for accurate active() state. Non-monotonic
  (backwards) timestamps are rejected. The internal _windowSum invariant is now guarded
  with an explicit error throw instead of a silent reset. The active() boundary is exclusive
  (< windowMs), removing a 1 ms false positive at exactly the window boundary. rate() span
  uses the true first-sample timestamp rather than _startedAt, improving precision for streams
  with multiple deltas.
• GenerationTimer fixes. Zero-token deltas no longer consume the post-gap prime marker via a
  _pendingPrime flag, so the first real token after a tool-call gap is correctly classified as
  a resume chunk. Non-monotonic timestamps are rejected. setTokens() now scales _primeTokens
  proportionally when the provider's exact token count differs from the streamed estimate, keeping
  decodeTokens and tps() truly exact at completion.
• Formatter correctness. fmtRate now crosses into "k" format at ≥ 999.5 instead of
  ≥ 1000, so values like 999.7 display as "1k" rather than "1000". fmtTokens guards
  negative inputs and returns "0". fmtMs rounds to the nearest second before splitting into
  minutes, eliminating "Xm60s" overflow at minute boundaries. fmtCost returns "$0" for
  negative and zero values. sparkline clamps max to at least min, handling inverted
  min/max pairs.
• View layer. The empty-segment filter now also catches text: "", removing unnecessary
  empty <span> DOM nodes. The stale-copy DEFAULTS object in view.js is replaced with a
  shared import from config.js, so display config can no longer drift out of sync. Streaming
  messages with history now show both the live-detail and last-message rows instead of one
  hiding the other. The sparkline reappears at turn completion (idle state).
• Config resolution. order now has a 0 floor clamp, matching all other numeric options.
  Boolean options (showSparkline, showSession, showWaits, showTotals, showCost,
  showCache) use a unified bool() helper with consistent opt-in/opt-out semantics and
  environment-variable overrides. Named constants (MIN_ORDER, MIN_POLL_MS, …) replace
  inline magic numbers. OPENCODE_TPS_METER empty string now disables (added to isFalsy).
  isTruthy handles "yes"/"on"/"1" for OPENCODE_TPS_METER_DISABLE. options.enabled=true
  now overrides the env disable flag. num() options correctly fall back to environment
  variables when the option is omitted or nullish, rather than treating explicit 0 as unset.
  The metric env variable (OPENCODE_TPS_METER_METRIC) can now switch back to "generated"
  when the option is set to "output".
• TUI plugin reliability. The per-session partLen map now tracks { length, messageID }
  per part and is cleaned up on message.removed, preventing memory leaks. The chars-per-token
  calibration ratio is now per-model (ratioByModel) rather than a single session-global
  variable, fixing cross-model calibration corruption. Calibration falls back to partLen state
  when state.part() returns empty for completed messages. sessionID comparisons use string
  coercion to avoid type-mismatch silent event drops. onCleanup is registered before
  setInterval, so a component construction throw no longer orphans the interval timer.
  TypeScript annotations added throughout. Delta-fallback no longer inflates token count on
  first sight of already-large part text.
• Installer safety and correctness. Malformed existing tui.json/package.json now causes
  a clear error exit instead of silent overwrite. Rotating backup files (file.bak,
  file.bak.1, …) replace the single .bak that was silently overwritten. --dir followed by
  the next flag now correctly errors instead of capturing the flag name as a directory. The
  --help text is a static constant rather than a fragile source-line-range parse. Tuple-format
  plugin entries ["@devinoldenburg/opencode-tps-meter", { … }] are now detected and migrated
  to @devinoldenburg/opencode-tps-meter/tui preserving their options. --uninstall now also
  removes the package dependency from package.json. A failed npm install now sets a non-zero
  exit code and returns rather than printing a misleading success message.
• Package metadata and exports. The package root (".") now exports plugins/tps/root.js,
  which re-exports all core named exports AND provides a default { id, tui } TUI plugin shim,
  so existing OpenCode configs using ["@devinoldenburg/opencode-tps-meter"] continue to work.
  The explicit TUI path "./tui" is unchanged. TypeScript declaration file (index.d.ts) added
  for the "./core" surface. Peer dependency ranges tightened from "*" to
  @opentui/solid@>=0.4.1 <1 and solid-js@>=1.9.12 <2. package-lock.json is now committed
  for reproducible development/CI installs; the ignore rule was removed. NOTES.md is excluded
  from the published npm tarball.
• Calibration robustness. MIN_RATIO lowered from 1.2 to 0.25 for CJK and other
  high-density languages. calibrateRatio with alpha=0 now emits a diagnostic warning rather
  than silently freezing calibration. The n0() fallback no longer passes negative numbers
  through (returns 0 instead). aggregate now surfaces decodeSource metadata.
• CI, release, and tooling. The test job now runs on ubuntu-latest, macos-latest, and
  windows-latest with Node 20.11, 20, 22, and 24. A TypeScript check step (npx tsc --noEmit)
  was added. npm ci runs before tests. pull_request triggers on all branches. The Bun
  plugin-verification job and release workflow now both run npm run verify:plugin (with Bun
  setup in release). Release publishes prerelease versions under the next tag. Dependabot
  configuration added for npm and GitHub Actions weekly updates. .gitignore covers *.tgz,
  *.bak, and *.bak.*.
• Demo and verification tools. demo.mjs now feeds pre-computed schedule token counts
  directly instead of round-tripping through tokensFromChars, eliminating floating-point
  drift. The --ci tool-gap detection uses the actual TOOL_GAP constant. The CI smoke-test
  assertion checks the first line starts with "TPS" instead of a fragile .includes check.
  The terminal cursor is restored on crash. verify-plugin.mjs wraps the empty-session renderer
  call in a try/catch and uses an exact string match for the expected "No renderer found" error.
  install-peers.mjs version-pins peer packages and copies transitive dependencies without
  overwriting existing packages; the temporary directory is cleaned up.

0.1.1 — 2026-06-20

Fixed

• Documentation accuracy pass. Removed a duplicated "Calibration" section in
  ARCHITECTURE.md; corrected the public exports key notation (. not ./) in
  VERSIONING.md; documented the previously-undocumented showSparkline option and the
  OPENCODE_TPS_METER=0 disable form in README.md; added --help to the documented
  installer flags (README + CHANGELOG); updated the internal NOTES.md to reflect the public,
  published repository.
• Removed a dead, never-read --global flag branch from scripts/install.mjs (no behavior
  change).

0.1.0 — 2026-06-20

Initial release.

Added

• Pure generation TPS — measures tokens ÷ active-generation time, where active time is the
  summed gaps between streamed tokens minus any gap large enough to be a tool call / wait
  (gapMs, default 1500 ms). Tool execution, permission prompts, and provider stalls are
  excluded from the rate and surfaced separately as −Ns wait. Prefill/resume "prime" tokens are
  excluded from the numerator, so a constant-rate stream measures its true rate to the token —
  proven by a server-stream simulation.
• Consistent live + exact — the live headline is the in-flight message's active-generation
  rate; on completion it locks to the provider's exact tokens.{output,reasoning} over the same
  measured time, so the number doesn't jump.
• Measured time-to-first-token, a sparkline (windowed RateMeter) that dips during a tool
  call while the headline holds steady, and pooled session average + peak.
• No native duplication — token totals and cost (shown by OpenCode's Context section) are off
  by default (showTotals / showCost); the meter shows throughput, TTFT, and excluded wait.
• Self-calibrating characters→token ratio, learned per model from each completed message's
  exact token count.
• Additive rendering into the stacking sidebar_content slot (order 150) — never replaces
  native sidebar sections.
• Theme-aware colors with per-tone overrides; configurable metric (default generated),
  gap threshold, detail level, sparkline, and labels, via plugin options or environment variables.
• Installer (scripts/install.mjs, the package bin) — idempotent, reversible, with
  --local / --dir / --no-install / --dry-run / --uninstall / --print / --help modes.
• Pure, framework-free core (plugins/tps/*.js) with 62 unit + integration tests
  asserting hand-computed exact expectations.
• Tooling: a runnable terminal demo (tools/demo.mjs, with a tool-call gap and a
  deterministic --ci mode), a Bun-based plugin verifier (tools/verify-plugin.mjs), and a peer
  installer (tools/install-peers.mjs).

What's Changed

• Fix triaged TPS meter issue batch by @devinoldenburg in #141

New Contributors

• @devinoldenburg made their first contribution in #141

Full Changelog: v0.1.1...v0.1.2