slash-tokens 1.5.7 → 1.6.1

package/CHANGELOG.md

@@ -0,0 +1,249 @@
+# Changelog
+
+## [1.6.1] — The Fixed Deal Edition
+
+*2026-08-24*
+
+Team/Pro is the live cart; 10% of savings shows on the weekly number as the comparison, not as what Team is charged.
+
+1.6.0 put `counterfactual_fee_usd` on npm without Doc Gate (the tarball README still headlined v1.4). This patch is the public identity for that ship. Trial still meters 10% on the $5 credit. Team `fee_usd` is 0. slashtokens.com copy is unchanged.
+
+**Not claimed:** a walked Team checkout (cart unbought). 10% is postponed until-further-notice, not dropped. CLI scanner “10% gate efficiency” is a different 10% (salvage rate).
+
+**Verification:** `cd npm && bun test` (157 pass, 2 skip). Worker: `npx vitest run tests/wjttc/slash-usage-counterfactual.test.ts tests/wjttc/slash-statement-comparison.test.ts`.
+
+## 1.6.0 (2026-08-24)
+
+### Fixed Deal — 10% on the report, not the Team invoice
+
+Team/Pro is the live cart ($39/mo or $390/yr). 10% of measured savings is postponed UFN; it still appears on the weekly number as the comparison (what the meter would have been vs the plan). Trial keys still pay 10%. `report()` may return `counterfactual_fee_usd` from the worker. Site copy unchanged.
+
+## 1.5.7 (2026-08-24)
+
+Punch-list cleanup from the full-codebase review after calibration shipped.
+
+## 1.5.6 (2026-08-23)
+
+`preflightRoute()` respects `init({ route, models })`.
+
+## 1.5.5 (2026-08-23)
+
+CLI scanner/report calibrate and price per-provider, not one flat rate.
+
+## 1.5.4 (2026-08-23)
+
+Claude Sonnet 5 pricing corrected to $2/$10. Doc drift.
+
+## 1.5.3 (2026-08-23)
+
+WASM no longer silently truncates input above ~1.06 MB.
+
+## 1.5.2 (2026-08-23)
+
+Gemini and Grok factors re-verified against the 29-sample corpus.
+
+## 1.5.1 (2026-08-23)
+
+Corpus 9→29 samples. Claude factor 1.85→2.05 (Spanish under-report).
+
+## 1.5.0 (2026-08-23)
+
+Per-model calibration wired into every real call site. Gemini and Grok calibrated. Retired API IDs fixed.
+
+## 1.4.1 (2026-04-20)
+
+### Polish release — no code changes
+
+**FAF 100% 🏆 · synced with a lean CLAUDE.md.** Documentation and brand consistency pass ahead of wider distribution. No SDK surface changes; all code identical to 1.4.0.
+
+`project.faf` now scores a full 🏆 Trophy (12/12 slots). CLAUDE.md is intentionally lean — a one-line `<!-- faf: ... -->` meta-stamp and a bi-sync status line, nothing more. Every piece of context has exactly one home: structured facts in `project.faf`, narrative in `README.md`, the sync pointer in `CLAUDE.md`. Token-frugal by design — which is on-brand for a token optimization SDK.
+
+### Changed
+- **License/Brand language rewritten.** Merged the prior separate `Brand` and `License` sections into one. Leads with "Code: MIT. Fork it, ship it, change it, show it, share it, sell it." — six verbs, because *show* and *share* are community drivers. Brand protection follows in plain language ("don't represent your app as Slash") instead of the prior "violators will get a takedown request" phrasing. Consistent across the ecosystem (slash-nextjs, RadioFAF).
+- **v1.4 and v1.3 callouts now stacked** at the top of both READMEs — newest first, with the 🆕 badge on v1.4. The Opus 4.7 story (v1.3) keeps its visibility, not erased by the newer release.
+- **"Know the cost before the call leaves your machine."** subtitle added to `npm/README.md` so both READMEs read identically below the headline.
+- **WASM badge now shows 4.8 KB** (the raw WASM engine size, matching every other surface) instead of the 5.1 KiB minzipped package total. Both numbers are true; the brand claim ("4.8 KB WASM") is the one users recognize.
+- **Signature sign-off added:** 🏎️ *Don't go to the Corner Shop in a Ferrari.* — the line we've been using everywhere else now appears on both READMEs.
+- **shields.io badge URLs refreshed** to the newer (`?style=flat`) format so GitHub's camo image proxy re-fetches under fresh URLs (fixes the intermittent "rate limited by upstream service" render on the bundle-size badge).
+
+### Why
+The npm package page bakes the README into the tarball at publish time. After v1.4.0 shipped, several improvements landed on GitHub but couldn't reach the npm landing page without a new publish. With the launch X post about to fire, the page visitors land on first needs to reflect the current brand and language.
+
+---
+
+## 1.4.0 (2026-04-19)
+
+### The Single-Source-of-Truth Edition
+
+Extracted `PROVIDER_MODELS` to a shared module and added `preflightRoute()` — the exact routing decision the Slash proxy makes (same-provider only). Fixes a class of downstream bug where consumers used `preflight().options[0]` as a routing decision (cross-provider analysis) instead of the actual routing. Also caught a latent bug: `claude-opus-4.7` was in `MODELS` but missing from the Anthropic provider group, making it an orphan to the routing algorithm.
+
+### New
+- **`preflightRoute(content, model)`** — returns the single cheapest same-provider `Alternative | null`. Matches mcpaas-cf proxy's `findCheapestRoute` semantics exactly. This is what downstream consumers should use to show "what Slash would route to," not `preflight().options[0]`.
+- **`PROVIDER_MODELS`** and **`providerOf(model)`** — exported from the new `providers.ts` shared module. Single source of truth for same-provider grouping.
+
+### Changed
+- **`intercept.ts` refactored** — no longer maintains its own `PROVIDER_MODELS` copy; imports from `providers.ts`. Removes drift risk between `preflight.ts` and `intercept.ts`.
+- **`claude-opus-4.7` added to Anthropic provider group** — was already in `MODELS` but orphaned from `PROVIDER_MODELS`, caught by the new test suite.
+
+### Clarified (not changed)
+- **`preflight().options` is cross-provider analysis** — returns cheapest models globally, sorted by price. Use it to *show alternatives*. Do not use it as a routing decision.
+- **`preflightRoute()` is the routing decision** — same-provider only. Use it to show *"what would Slash route to right now."*
+
+### Testing
+- 127 pass, 0 fail, 432 assertions (up from 118 / 386 in v1.3.0)
+- 7 new TIER 1 BRAKE tests asserting the same-provider invariant, null return semantics, and `providerOf` hygiene
+- See `TEST-NOTES.md` for the full test matrix and WJTTC tier plan
+
+### Docs
+- `TEST-NOTES.md` — new living doc collecting invariants and WJTTC tier plan
+
+### Non-breaking
+- `preflight()` behavior unchanged
+- All existing exports retained
+- Only additions + one internal refactor
+
+---
+
+## 1.3.0 (2026-04-16)
+
+### The Opus 4.7 Edition
+
+Same-day support for Claude Opus 4.7 — the new flagship — with measured token calibration and proxy-ready benchmarking. Plus a Gemini URL-extraction fix.
+
+### New
+- **Opus 4.7 support** — pricing + calibration layer. Measured 1.16–1.51x token increase vs prior Opus generations
+- **Benchmark harness** — now supports custom proxy + auth headers, so you can benchmark Slash against direct upstream
+- **Gemini URL model extraction** fix — proxy path correctly identifies model from `/v1beta/models/:model:generateContent`
+- **Cross-repo integration** — Gemini proxy path covered by integration tests
+
+### Model Registry
+- `claude-opus-4.7` with confirmed API ID `claude-opus-4-7`
+- All 4 providers routable through Slash proxy (Anthropic, OpenAI, xAI, Google)
+
+### Testing
+- 118 pass, 0 fail, 386 assertions
+
+### Docs
+- Honest production analysis — $75.98 saved, $7.60 collected (real numbers, not projections)
+
+---
+
+## 1.2.1 (2026-04-15)
+
+### Fixes
+- Add `pass` action type to `ReportOptions` — SDK types now match server (`prevented | routed | pass`)
+
+### Testing
+- Cross-repo integration suite: 17 tests verifying SDK ↔ mcpaas-cf contract
+- Self-registering test keys, no secrets needed
+- TIER 5 proxy-path tests (Anthropic gateway, OpenAI, txn counting)
+- 118 tests total, 374 assertions
+
+---
+
+## 1.2.0 (2026-04-14)
+
+### The Next.js Edition
+
+Session summary on exit — every time you stop your dev server, Slash prints how much you saved that session:
+```
+[slash] Session: 47 calls | 23 routed | $0.89 salvaged — The more you build, the more you save
+```
+Fires on `exit`, `SIGINT` (Ctrl+C), and `SIGTERM`. Zero calls = silent.
+
+### New
+- `auto` mode: session summary on process exit
+- Dashboard: 90-day transaction history (History panel, show all)
+- Dashboard: live feed now scrollable, shows `$saved` instead of raw token counts
+
+### Testing
+- 99 pass, 0 fail, 326 assertions
+
+---
+
+## 1.1.2 (2026-04-13)
+
+### Fixes
+- Fix Haiku model ID — `claude-haiku-4-5-20251001` (was wrong, routing would fail)
+- Action types aligned to Gate vocabulary — `prevented | routed` (was `skipped | reduced | routed`)
+
+### Testing
+- 99 pass, 0 fail, 326 assertions
+
+## 1.1.1 (2026-04-08)
+
+Token Optimization for Context Engineers.
+
+### Auto Mode — Routes to Cheapest Model
+- `import 'slash-tokens/auto'` — one import, every call optimized
+- Routes within same provider: Opus → Haiku, GPT-5.4 → Nano, Grok → Fast
+- Rewrites request body before fetch — the API gets the cheaper model
+- Verbose by default — new users see every call, every route, every save
+- Nudges at $5/$10/$25/$50/$100 thresholds: `⚡ $5 salvaged — keep building`
+
+### Configuration
+- `init({ route: true })` — route to cheapest model (default)
+- `init({ route: false })` — track only, don't touch my models
+- `init({ models: ['claude-haiku', 'claude-sonnet'] })` — model-level control
+- `init({ quiet: true })` — nudges only, no per-call logging
+
+### CLI
+- Evaluator doesn't charge — `bunx slash-tokens` is free
+- "Token optimization analysis" (not "burn analysis")
+- "TOKENS SALVAGED" (not "savings")
+- "DAILY TOKEN VOLUME" (not "daily burn")
+- Single "Next" link to setup page
+
+### API
+- `salvaged` / `salvagePercent` (replaces `savings` / `savingsPercent`)
+- npm description: "Token Optimization for Context Engineers"
+
+### Testing
+- 99 pass, 0 fail, 326 assertions
+
+## 1.1.0 (2026-04-07)
+
+Token Optimization for Context Engineering.
+
+### Features
+- `preflight(content, model)` — tokens, cost, fits, cheaper alternatives
+- `import 'slash-tokens/auto'` — one import, every LLM call pre-flighted
+- 10 models with verified April 2026 pricing (Claude, Grok, Gemini, GPT)
+- `report()` — metered savings reporting, 10% of what you save
+- `init()` — key configuration with 3-layer resolution
+- `MODELS` — full pricing table exported for direct access
+- `listModels()` — enumerate all supported models
+
+### Models (April 2026)
+- Anthropic: claude-opus ($5/M), claude-sonnet ($3/M), claude-haiku ($1/M)
+- xAI: grok-4.20 ($2/M), grok-4-1-fast ($0.20/M)
+- Google: gemini-3.1-pro ($2/M), gemini-2.5-flash ($0.30/M)
+- OpenAI: gpt-5.4 ($2.50/M), gpt-5.4-mini ($0.75/M), gpt-5.4-nano ($0.20/M)
+
+### Auto mode
+- Intercepts fetch() to Anthropic, OpenAI, xAI, Google endpoints
+- Model name normalization (handles full API IDs + legacy names)
+- Non-blocking, sub-millisecond
+
+### Testing
+- 323 tests: 172 Zig, 101 TypeScript, 50 API
+
+## 1.0.0 (2026-04-07)
+
+Initial v1 release. Superseded by 1.1.0 (pricing corrections).
+
+## 0.3.0 (2026-04-06)
+
+Deprecated. Added preflight + auto mode with outdated pricing.
+
+## 0.2.0 (2026-04-06)
+
+Deprecated. Added report() + init() SDK integration.
+
+## 0.1.0 (2026-04-04)
+
+Deprecated. Token estimation + CLI scanner.
+
+## 0.0.1 (2026-04-02)
+
+Initial placeholder.

package/README.md

@@ -13,10 +13,20 @@ For anyone building with LLMs. 4.8 KB WASM. Sub-millisecond. Zero dependencies.
 
 Know the cost before the call leaves your machine.
 
-> 🆕 **v1.4 — The Single-Source-of-Truth Edition.** New `preflightRoute()` matches the Slash proxy exactly — same-provider only, zero routing drift. `PROVIDER_MODELS` is now the single source of truth across the SDK.
+> 🆕 **v1.6.1 — The Fixed Deal Edition.** Team/Pro is the live cart; 10% of savings shows on the weekly number as the comparison, not as what Team is charged.
+>
+> **v1.5 — The Calibration Fix Edition.** Per-model calibration is wired into every call path. All four providers checked against the same 29-sample corpus.
+>
+> **v1.4 — The Single-Source-of-Truth Edition.** New `preflightRoute()` matches the Slash proxy exactly — same-provider only, zero routing drift. `PROVIDER_MODELS` is now the single source of truth across the SDK.
 >
 > **v1.3 — The Opus 4.7 Edition.** Same-day support for Claude Opus 4.7 with measured token calibration (1.16–1.51x). Plus Gemini proxy fix and benchmark harness for any upstream.
 
+## v1.6.1 — The Fixed Deal Edition
+
+Team/Pro is the live cart; 10% of savings shows on the weekly number as the comparison, not as what Team is charged.
+
+The library stays free. Trial still takes 10% of measured savings on the $5 meter. Team is $39/month or $390/year: the weekly number shows what 10% would have been next to the plan. Team is not charged that 10%.
+
 ```bash
 npm install slash-tokens
 ```

package/dist/transact.d.ts

@@ -11,5 +11,7 @@ export interface ReportResult {
     fee_usd: number;
     balance_remaining_usd: number;
     timestamp: string;
+    /** 10% of this call's savings — comparison only. Charged fee is fee_usd (0 on Team). */
+    counterfactual_fee_usd?: number;
 }
 export declare function report(opts: ReportOptions): Promise<ReportResult>;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "slash-tokens",
-  "version": "1.5.7",
+  "version": "1.6.1",
   "description": "Token Optimization for Context Engineers. 4.8 KB WASM. Sub-millisecond. Zero dependencies.",
   "main": "dist/index.js",
   "module": "dist/index.js",
@@ -20,7 +20,8 @@
     }
   },
   "files": [
-    "dist"
+    "dist",
+    "CHANGELOG.md"
   ],
   "sideEffects": false,
   "scripts": {