v5.10.1 — Continuity & avoided-search measurement + cross-platform parity

What's new

Smarter measurement of what Token Optimizer saves you:

• Calibrated token estimator — more accurate token counts across read-cache, structure-map, and telemetry (replaces the coarse bytes/4 heuristic).
• Checkpoint-restore credit — when a session resumes from a checkpoint, you're credited for the working set you skip re-reading (conservatively floored and capped).
• Avoided-search measurement — when a proactive prior-session hint sends an agent straight to the right file, the exploratory search it spared is now measured (estimated tier, observed trigger).
• Dashboard surfaces the new continuity and avoided-search lines, keeping measured and estimated clearly distinct.

Cross-platform parity: Codex (inherited), Hermes, OpenClaw (2.4.7), and OpenCode (1.0.7) now share the same measurement surface.

All numbers stay honest: estimated savings are never folded into the metered total.

v5.10.0

Full Changelog: v5.9.3...v5.10.0

v5.9.3

Nudge follow-through now credits heeded compaction recoveries across the UserPromptSubmit→PostCompact hook boundary. The credit is idempotent (consume-then-credit, gated on the persist succeeding) and degrades safely on a corrupt cache. Patch release.

v5.9.2 — Accurate mixed-model savings

Prices every model in your before/after mix at its real weight. An unpriced model id (an unrecognized or local model) is now estimated at the runtime-default rate instead of being dropped, so the savings headline stays accurate when your model mix includes something the rate card doesn't recognize. Routing savings reports an honest "unavailable" when there's no data to measure.

v5.9.1 — Codex/GPT support for the savings transformation

The before/after savings transformation is now model-aware, so it works on Codex (GPT models + OpenAI pricing), not just Anthropic.

What changed

• Codex/GPT support for the transformation. Each era's typical session is now priced at its real model mix (blending per-model rates over the measured mix) instead of an Opus/Sonnet binary. Anthropic blends Opus/Sonnet/Haiku; Codex blends gpt-5-codex / gpt-5.x / mini by their measured shares.
• Provider-correct labels. The Before/Now columns show the dominant model in the mix, e.g. "95% Opus" or "80% GPT-5.4", rather than a hardcoded "% Opus".
• Generalized routing lever. The "where it comes from" routing line now reflects any model-mix shift, and reports a shift to costlier models honestly as added cost.
• No more hidden panel for Codex. Replaced the Opus-share gate that hid the transformation for any non-Opus (Codex / 0%-Opus) baseline.

Anthropic numbers are unchanged within noise; Codex now renders a correct, GPT-priced transformation.

v5.9.0 — Honest savings, end to end

Token Optimizer's Savings view now leads with a clear before/after transformation: what your workload would have cost the heavy way versus what it costs now, with every number documented and defensible.

Highlights

• Before/after transformation headline with a plain-language breakdown of where the savings come from.
• New methodology doc (docs/METHODOLOGY.md): the exact source, formula, and conservative assumptions behind every number, written to hold up to a skeptical reader.
• Three honest tiers kept distinct: measured (directly metered), estimated (grounded in your own behavior, with sample sizes), and opportunity (reclaimable, never counted as realized).
• Fair, sample-size-independent baseline math so the before/after comparison holds whether you have a handful of recent sessions or hundreds.
• Readability pass on the Savings panel: collapsible explainers, larger type, clearer copy.
• Cross-session continuity recovery for OpenCode and OpenClaw.

v5.8.10

Reliable OpenCode session rollup + token usage capture, and a quiet quality-bar self-heal for plugin installs.

OpenCode dashboard now populates from live sessions (#54)

• Sessions roll up into trends.db on session.idle, on eviction, and right before the dashboard renders — no longer dependent on the rarely-emitted session.deleted event.
• Per-response token usage and cost are captured from message.updated, so Sessions / Daily Stats show real tokens, cost, and model.

Quieter quality bar on plugin installs (#53)

• The cache hook ships in the plugin's hooks.json, so the settings.json self-heal is now skipped for plugin installs and runs silently elsewhere. The status-line integration tip is reserved for explicit setup-quality-bar runs.

Also: OpenCode plugin bumped to 1.0.4.

v5.8.9

What's Changed

• fix(codex): resolve marketplace plugin install failure on codex-cli 0.136.0 (#51) by @alexgreensh in #52

Full Changelog: v5.8.8...v5.8.9

v5.8.8

Context fill is now correct on 1M-context sessions. measure.py writes the resolved context window and per-session usage tokens into the quality cache; the VS Code companion mirrors the same model→window logic. Opus/Sonnet 1M sessions now show true fill instead of a ~5x-inflated estimate, and sessions with heavy cached context no longer fall back to a 0% char-length guess — fixed across the terminal status line, HTML dashboard, and the VS Code extension.

Leaner VS Code extension (0.1.9): usage limits come straight from the status-line sidecar.

Refreshed dashboard footer: brighter social icons, an X (Twitter) link, and a 'Star on GitHub' call-to-action with a live star count (Claude/Codex, OpenClaw, OpenCode dashboards; matching links in the VS Code panel).

VSIX attached for VS Code / Cursor / Windsurf: code --install-extension token-optimizer-statusline-0.1.9.vsix

v5.8.7 — per-session VS Code extension, fairer efficiency, self-healing dashboard

VS Code extension (attached .vsix v0.1.8)

• Status bar now resolves to this window's Claude session (scoped to the workspace folder), instead of whichever session ran most recently anywhere.
• Context fill is computed from the transcript's token count divided by the real context window — correct for both 200k and 1M-context sessions.
• New expanded status panel (click the status bar) with live context, ContextQ, Eff, warnings, compactions, duration, agents, and 5h/7d limits.
• Honest states for no-folder / warming-up; brighter grey for secondary info; the obscure "Regime" indicator removed.
• Install: code --install-extension token-optimizer-statusline-0.1.8.vsix (works in Cursor and Windsurf too).

Efficiency metric

• A brand-new session no longer scores low for not having made decisions yet — decision density stays neutral until there's enough conversation to judge. Established low-density sessions are still scored normally.

Dashboard daemon

• Self-heals a stale tombstone instead of staying dead until re-setup, so the bookmarkable localhost dashboard keeps working across updates.
• Regenerates the dashboard in the background when opened stale, so it reflects recent activity.

Terminal status line

• Shows your 5-hour and 7-day usage limits with reset times; brighter grey; Regime removed.