Skip to content

CHANGELOG.md

Latest commit

 

History

History
206 lines (170 loc) · 26.9 KB

CHANGELOG.md

File metadata and controls

206 lines (170 loc) · 26.9 KB
Raw

Changelog

[1.11.1] - 2026-08-15

Security

  • Cleared all 10 open Dependabot alerts. In tools/mcp-cost-server (all four are transitive dependencies of @modelcontextprotocol/sdk): ip-address 10.2.0 -> 10.5.0 (three SSRF / trust-boundary bypasses -- octal decoding of leading-zero octets, IPv4-mapped/NAT64 misclassification, and a CIDR suffix suppressing special-use classification), hono 4.12.31 -> 4.13.2 (CORS ReDoS, language-middleware algorithmic DoS, memo() retaining SSR output across requests, and the proxy helper leaking Connection-listed headers), @hono/node-server 1.19.14 -> 2.1.0 (serve-static path traversal on Windows via encoded backslash), and fast-uri 3.1.4 -> 4.1.2 (host confusion via backslash authority introducer). In site: react-router / react-router-dom 7.18.1 -> 7.18.2 (RSC-mode CSRF bypass executing actions before the 400 response), plus nanoid -> 3.3.18 from an npm audit fix in the same pass. pnpm audit and npm audit both report zero known vulnerabilities.
  • Fixed the root cause behind the mcp-cost-server drift: pnpm 11 silently ignores the pnpm field in package.json. The ip-address and hono overrides added in 1.8.0 lived there, so once the toolchain reached pnpm 11 they stopped applying (pnpm only warns "the following keys were ignored") and the pinned transitive versions rolled back to vulnerable ranges. Moved them to tools/mcp-cost-server/pnpm-workspace.yaml, which is the supported home for the setting, raised each floor to the first patched version, and added fast-uri and @hono/node-server. Verified the overrides now take effect: pnpm rejected the old lockfile with ERR_PNPM_LOCKFILE_CONFIG_MISMATCH before regenerating it. Removed the dead pnpm block from package.json so it cannot mislead again.
  • Closed SonarCloud code-scanning alert #8 (tssecurity:S8476, high, "Client-side requests should not be vulnerable to forging attacks") in site/src/utils/repoAnalyzer.ts. Request URLs are no longer assembled by string concatenation: buildApiUrl now encodes each path segment individually, resolves them against a fixed base with the URL constructor, and drops any request whose parsed origin is not https://api.github.com -- a backstop that holds even if a segment smuggles in a scheme or a protocol-relative //host. Empty, ., and .. segments are rejected outright. Verified by driving analyzeRepo directly with hostile input (evil.test, .., a/../../evil, and a ../../../other branch): zero off-origin requests were issued, hostile owner/repo values produced no request at all, and the traversal branch fell back to main.

Removed

  • Deleted the stray untracked site/pnpm-lock.yaml. The site is an npm project -- it tracks site/package-lock.json and CI builds it with npm ci -- so a second pnpm lockfile would have drifted from the one actually used on the next dependency bump.

[1.11.0] - 2026-07-25

Added

  • Claude Opus 5 support (claude-opus-5, GA 2026-07-24) across all pricing tables, guides, tools, templates, and the web calculator. Opus 5 is the new Opus-tier flagship at $5/$25 per 1M input/output -- exactly what Opus 4.8 cost, so posted rates do not move. What moves is the shape of the bill: adaptive thinking is on by default, reasoning tokens bill as output at $25/1M, and max_tokens caps thinking plus text. 1M context at standard rates, 128K max output (300K on Batch via beta), cache hit $0.50, 5m-write $6.25, 1h-write $10, Batch $2.50/$12.50. Knowledge cutoff May 2026. Earliest retirement 2027-07-24. Bedrock anthropic.claude-opus-5 (legacy InvokeModel/Converse: us.anthropic.claude-opus-5), Vertex AI claude-opus-5.
  • Minimum cacheable prompt length is now first-class, not just prose. A cache_control block on a prefix below the model's floor is silently ignored -- no error, no cache_creation_input_tokens, and full input price on every turn. Opus 5 halves the Opus 4.8 floor from 1,024 to 512 tokens, so prefixes that never cached on 4.8 start caching on 5. Added a 13-model threshold table to guide 01, a 14-row table plus usage-field detection to guide 08, a new section to guides/diagrams.md, and rows to CLAUDE.md and the tool READMEs. The MCP cost server now carries a typed minCacheTokens field per model and emits a cache_warning when your prefix falls short; the VS Code extension exports MIN_CACHE_TOKENS and adds the same note to its CLAUDE.md per-turn report. Haiku 4.5's 4,096 floor is 8x Opus 5's, so a Haiku-first routing tier can quietly pay full price on a prefix that caches fine on Opus.
  • Opus 5 migration guidance in README and guide 03: output_config.effort (defaults high) as the primary cost dial, thinking: {type: "disabled"} being legal only at effort high or below (xhigh/max + disabled returns 400), cybersecurity classifiers returning stop_reason: "refusal" with the server-side fallbacks param (beta server-side-fallback-2026-07-01) as the recovery path, and beta mid-conversation-tool-changes-2026-07-01 for changing tool definitions between turns without busting the cache.
  • New "The Opus 5 Default-Thinking Effect" (guide 09) and "Opus 5 and the Default-On Thinking Tax" (guide 11) sections: why an unchanged workload costs more on Opus 5 than Opus 4.8 despite identical per-token rates, and which levers claw it back.
  • Tool-token overhead figures in guide 07, which previously omitted them entirely: tool-use system prompt 286 (auto/none) / 406 (any/tool) on Opus 5, plus the bash and text-editor per-tool costs.

Changed

  • Fast Mode economics rewritten everywhere -- the 6x tier no longer exists. Fast Mode is now Opus 5 and Opus 4.8 only, both at a flat 2x ($10/$50). Opus 4.7 errors on speed: "fast"; Opus 4.6 silently runs standard and reports usage.speed: "standard" rather than failing, which is the dangerous case -- check that field instead of assuming you got what you paid for. Removed the fast_mode_legacy ($30/$150) entry from the token-estimator, deleted the $5 x 6 = $30.00 row from the diagrams stacking table, and recomputed every figure that was derived from the dead rate (cache-hit floor $3.00 -> $1.00, cache rewrite $37.50 -> $12.50).
  • opus alias now resolves to Opus 5 in the site calculator, MCP cost server, VS Code extension, token-estimator, usage-analyzer, and claude-rate. Opus 4.8 is flagged legacy and gained its own explicit key (opus-4.8 / opus_4_8) in every tool so a pinned 4.8 can still be priced. claude-rate positions Opus 5 as flagship directly after Fable 5.
  • Opus 4.8 repositioned from flagship to "previous flagship" -- still the right pin if prompts are tuned to it or you need thinking off at xhigh/max, and the fallback target for Opus 5 refusals.
  • Corrected two stale figures in CLAUDE.md that predate this release: the bash tool adds 325 tokens on Opus 5 / 4.8 / 4.7 and 244 on Opus 4.6 and earlier (was a flat "+245"), both verified live against Anthropic's bash-tool and tool-use pricing docs on 2026-07-25.
  • Added Sonnet 5 (claude-sonnet-5) rows alongside Opus 5: $3/$15 standard with an introductory $2/$10 through 2026-08-31, earliest retirement 2027-06-30. Tool tables use the standard rate so projections stay valid past the intro window.
  • Mythos Preview moved to past tense (retired 2026-06-30); Sonnet 4 and Opus 4 moved to the retired list (2026-06-15). Opus 4.1 is now the next retirement due (2026-08-05) and the only $15/$75 model left.
  • guides/diagrams.md Mermaid charts rebuilt: Opus 5 and Sonnet 5 nodes added, legacy snapshots restyled, the decision tree gained a cost-sensitive branch pointing at effort, and the pricing-modifier stack now shows all three Fast Mode outcomes (2x, error, silent standard).
  • usage-analyzer detect_model() recognizes opus-5 / claude-opus-5 / the Bedrock-prefixed IDs, and distinguishes opus-4-8 from the generic opus catch-all. Its README gained a "Models and Pricing" section -- it previously documented no pricing at all, so code and docs were silently out of sync.
  • CLAUDE.md sync checklist expanded to name every file carrying a pricing table, including the per-tool sibling READMEs that were being missed.

[1.10.0] - 2026-07-10

Added

  • Guide 11: Speed vs Cost (guides/11-speed-vs-cost.md) -- making Claude faster without burning money. Speed levers ranked by cost (cache warmth, shorter context, Haiku/Sonnet routing, effort control, then Fast Mode last), Fast Mode economics (2x on Opus 4.8 vs 6x on 4.7/4.6, the OTPS-vs-TTFT caveat, cache-pool invalidation on speed switch, compatibility matrix), deadline math for when the premium pays off, and the Batch API as the inverse lever. Every figure adversarially verified against the repo pricing block.
  • Interactive charts on the Cost Calculator (site/src/components/charts.tsx): per-turn cost curve (line + area, crosshair hover), current-vs-optimized savings breakdown (paired bars with legend), and same-settings-on-other-models comparison (emphasis bars). Hand-rolled SVG, no new dependencies; palette validated for lightness band, chroma, CVD separation, and contrast against the site surface.
  • "How Far Can You Actually Go?" README section -- per-lever savings table with primary sources (Anthropic prompt caching up to 90%, Batch flat 50%, RouteLLM routing, context management 84%, subscription economics) and the honest caveats separating typical results from stacked ceilings.
  • SPA fallback for GitHub Pages (site/scripts/spa-404.mjs copies index.html to 404.html at build) so deep links like /calculator load directly.

Changed

  • Headline savings claim reframed from "30-60%" to the "30-90%" band (30-60% typical for a mixed workload; up to 90% ceiling when every lever stacks against an unoptimized all-Opus baseline) across README, site title/meta/hero, and awesome-list submission entries. Backed by a deep-research pass: 103 agents, 20 sources, 25 claims adversarially verified (2 aggressive community claims refuted and excluded). The cost-mode skill's own claim stays 30-60% -- the skill alone does output reduction and routing hints, not batch/caching/subscription.
  • Calculator, Badge Checker, and Repo Analyzer pages are now routed and live -- App.tsx previously routed every path to the landing page; added routes plus navbar links, and removed the README "Not live yet" notice.
  • Guide counts updated to 12 across README, CLAUDE.md, site copy, cheatsheet links, awesome-list PR descriptions (which were stale at 7), and diagrams.md related-guides list.

[1.9.0] - 2026-06-12

Added

  • Claude Fable 5 support (claude-fable-5) across all pricing tables, guides, tools, and the web calculator. Fable 5 is Anthropic's most capable widely released model -- a new Mythos-class tier above Opus at $10/$50 per 1M input/output (2x Opus 4.8). 1M context at standard rates, 128K max output, cache hit $1, 5m-write $12.50, 1h-write $20, Batch $5/$25. Always-on adaptive thinking (thinking: disabled not supported; depth via effort). Safety classifiers can decline requests: HTTP 200 + stop_reason: "refusal", pre-output refusals unbilled, beta fallbacks param + fallback credit for retries. No Fast Mode. Requires 30-day data retention. GA 2026-06-09 on Claude API, Claude Platform on AWS, Bedrock (anthropic.claude-fable-5), Vertex AI, and Microsoft Foundry.
  • Claude Mythos 5 entry (claude-mythos-5): same specs and pricing as Fable 5 but without the safety classifiers; limited availability to approved Project Glasswing customers. Successor to Mythos Preview. Flagged inviteOnly in the site registry (reference tables only, hidden from calculator selection).
  • fable model alias in token-estimator, usage-analyzer, mcp-cost-server (tool enums + compare output), VS Code extension, and claude-rate cost projections.
  • Fable 5 routing guidance: cheatsheet decision tree, guide 03 quick-reference card, guide 10 "Tier 2+" block, cost-mode skill model-routing table (downshift suggestion when running routine work on Fable 5), calculator and usage-analyzer recommendations for Fable-heavy usage.

Changed

  • Mythos Preview marked retiring 2026-06-30 (announced with the Mythos 5 launch) -- lifecycle flipped to legacy, pricing rows annotated, retirement tables updated across README, cheatsheet, and CLAUDE.md. The token-estimator mythos alias now prices Mythos 5 ($10/$50) instead of Mythos Preview ($25/$125).
  • Opus 4.8 repositioned from "most capable model" to "Opus-tier flagship" in copy across README, cheatsheet, guides, and the site registry; "most capable" now refers to Fable 5.
  • usage-analyzer model detection extended: recognizes fable/mythos and distinguishes opus-4.7 / opus-4.6 from the opus alias; the Opus-share hotspot check now matches all Opus variants.
  • All pricing references re-verified against Anthropic docs on 2026-06-12; "verified" dates bumped from 2026-06-06.
  • Plugin-identity versions bumped to 1.9.0; plugin distribution copy of the cost-mode skill re-synced.

[1.8.0] - 2026-06-06

Added

  • Claude Opus 4.8 support (claude-opus-4-8) across all pricing tables, guides, tools, and the web calculator. Opus 4.8 is Anthropic's new flagship and most capable model at $5/$25 per 1M input/output (same posted price as Opus 4.7 / 4.6). 1M context at standard rates (200K on Microsoft Foundry), 128K max output, adaptive thinking only, effort defaults to high on all surfaces, knowledge cutoff Jan 2026. Earliest retirement 2027-05-28. Bedrock ID anthropic.claude-opus-4-8.
  • Per-model Fast Mode pricing. Promoted the single FAST_MODE_MULTIPLIER constant to a per-model fastModeMultiplier field on ModelPricing (site) and per-model entries in the standalone tool tables. Opus 4.8 Fast Mode is 2x ($10/$50); Opus 4.7 / 4.6 remain 6x ($30/$150). Opus 4.6 Fast Mode is deprecated as of the 4.8 launch (removed ~30 days later, then falls back to standard speed). Opus 4.8 Fast Mode is Claude API + Managed Agents only.
  • opus-4-7 model entry in site/src/utils/pricing.ts and the standalone tool tables.

Changed

  • Default opus alias now resolves to Opus 4.8 in the site calculator, MCP cost server, VS Code extension, token-estimator, usage-analyzer, and claude-rate. Opus 4.7 / 4.6 / 4.5 and Sonnet 4.5 are now flagged legacy.
  • Renamed the site ModelId opus-legacy key to opus-4-6 and added an explicit opus-4-7 entry; the Fast Mode UI label and savings recommendation are now computed from the per-model multiplier instead of a hardcoded "6x / 83%".
  • Per-model tool-use system-prompt overhead documented in CLAUDE.md (Opus 4.8 = 290/410, Opus 4.7 = 675/804, Opus 4.6 + Sonnet 4.6 = 497/589) -- replaces the stale flat "346 / 313 tokens".
  • Batch API extended-output note added (up to 300K output on Opus 4.8/4.7/4.6 + Sonnet 4.6 via output-300k-2026-03-24).
  • Opus 4.1 marked deprecated with a firm 2026-08-05 retirement date (announced 2026-06-05). Migration targets across README + cheatsheet updated to Opus 4.8.
  • All pricing references re-verified against Anthropic docs on 2026-06-06; "verified" dates bumped from 2026-05-22.
  • Reconciled plugin-identity versions to 1.8.0. .claude-plugin/plugin.json, .claude-plugin/marketplace.json, and plugins/cost-mode/.codex-plugin/plugin.json were stuck at 1.5.0 while the changelog had advanced to 1.7.0 -- bumped all three to match this release.

Security

  • Updated react-router-dom to 7.17.0 in site/ (was resolving to 7.14.0), clearing 3 Dependabot alerts: turbo-stream deserialization RCE (high), __manifest DoS (high), and protocol-relative open redirect (moderate).
  • Updated transitive hono to 4.12.23 and qs to 6.15.2 in tools/mcp-cost-server/ (via pnpm update, both within existing semver ranges), clearing 5 Dependabot alerts: Set-Cookie injection, mount-prefix routing, IPv6 deny-rule bypass, JWT scheme acceptance (all moderate), and qs.stringify DoS (moderate). npm audit and pnpm audit both report 0 vulnerabilities.

[1.7.0] - 2026-05-22

Added

  • claude-rate CLI at tools/claude-rate/. Local rater that scans a project directory and grades the Claude/AI setup across 7 categories (CLAUDE.md size discipline, .claudeignore coverage, settings.json model+budget+permissions, MCP server count, hooks, security/secrets hygiene, optimizer tooling). Returns 0-100 score, A+ to F letter grade, per-category breakdown with text bar charts, monthly-cost projection on every active model tier, copy-pasteable fix suggestions, and a shields.io badge URL.
    • Three runners: npx -y @sagargupta16/claude-rate ., curl | sh one-shot, persistent curl | sh -- --install. Plus direct python rate.py.
    • Stdlib only (Python 3.10+). No pip install needed for the rater itself.
    • Flags: --fix (copy-pasteable fix list), --strict (CI gate, exits 1 below grade B), --json (machine-readable output), --version.
    • Why local? Inspects things the deployed web analyzer can't see: real MCP server count from .mcp.json, hooks in settings.json, .claudeignore coverage gaps vs files actually on disk, accidentally-committed secrets (sk-..., AKIA..., ghp_...), missing .env entries in .gitignore, cost-mode skill installation status, custom slash command count.
    • Files: tools/claude-rate/rate.py (single-file Python CLI), tools/claude-rate/bin/claude-rate.js (npx shim that locates Python and forwards args), tools/claude-rate/install.sh (POSIX one-shot/persistent installer), tools/claude-rate/package.json (npm distribution metadata), tools/claude-rate/README.md (full docs with example output and CI snippet).
  • README "Rate your setup" section -- new top-level entry point. Local claude-rate is the recommended path; web tools (analyzer, calculator, badge) remain available for browser-only / public-repo flows.
  • Legacy & Retired Models section in README. Self-contained migration reference covering recently-retired models (Opus 3, Sonnet 3.7, Haiku 3 / 3.5, Sonnet 3.5 v1/v2, Sonnet 3, Claude 2.x, Claude 1.x, Instant 1.x), deprecated-soon models (Sonnet 4 / Opus 4 retiring 2026-06-15), still-callable older snapshots (Opus 4.5, Opus 4.1, Sonnet 4.5), and historical pricing patterns no longer in effect (the obsolete "2x over 200K" long-context premium, single-endpoint Bedrock, ARN-versioned-only model IDs). Cheatsheet got an expanded version of this in 1.6.0; this PR adds it to README too with last-known pricing for every retired tier.
  • tools/README.md quick-reference table updated to lead with claude-rate as the recommended starting point.

[1.5.0] - 2026-05-20

Added

  • Added repository-level SECURITY.md and .github/pull_request_template.md to align with community standard practices.

Changed

  • Resolved 3-way version drift across repository files, standardizing on version 1.5.0.

Fixed

  • Fixed uncommitted CLAUDE.md duplicate H1 header by merging headers and keeping the stacking blockquote intact.

[1.4.0] - 2026-04-17

Added

  • Claude Opus 4.7 support across all tools, guides, pricing tables, and the web calculator. Opus 4.7 is Anthropic's current flagship (April 2026) at $5/$25 per 1M input/output -- same posted pricing as legacy Opus 4.6.
  • Claude Mythos Preview entry in pricing tables, token-estimator, and the site's model registry. $25/$125 per MTok. Invite-only via Project Glasswing for defensive cybersecurity research. Added inviteOnly flag on ModelPricing so Mythos appears in reference tables but is filtered out of the calculator/analyzer selection UIs.
  • opus-legacy model option in site/src/utils/pricing.ts (maps to Opus 4.6) to keep Fast Mode accessible for users who need it.
  • fastModeCapable flag on ModelPricing type; calculator now scopes Fast Mode UI to the model that actually supports it (Opus 4.6) instead of hardcoding to model === 'opus'.
  • Tokenizer overhead warning: Opus 4.7 introduced a new tokenizer that may use up to 35% more tokens for the same source text. Guides now recommend budgeting 20-35% higher for Opus 4.7 tasks vs Opus 4.6.
  • Bedrock model ID reference table in guides/06 (includes us.anthropic.claude-opus-4-7 cross-region profile + research-preview caveat).
  • New Mermaid diagrams in guides/diagrams.md: "Claude Model Family (April 2026)" showing all GA + research-preview models with cost tiers, and "Pricing Modifier Stack" showing how cache/batch/regional/fast-mode multipliers compose.
  • Thinking modes table in cheatsheet: clarifies Opus 4.7 uses adaptive thinking only (no extended thinking), Sonnet 4.6 supports both, Haiku 4.5 is extended-only.
  • Upcoming retirements table in cheatsheet and README: Haiku 3 retires 2026-04-20 (corrected from April 19 in 1.6.0), Sonnet 4 / Opus 4 retire 2026-06-15.
  • Published benchmark numbers (CyberGym, SWE-bench Verified/Pro, Terminal-Bench) in guides/03 comparing Opus 4.6 vs Mythos Preview.

Changed

  • 1M context pricing corrected: Opus 4.7, Opus 4.6, and Sonnet 4.6 now bill the full 1M window at standard per-token rates. The earlier "2x input, 1.5x output over 200K" documentation was obsolete and applied only to Opus 4.1 and older. Removed the "Long Context Threshold Trap" anti-pattern section since it no longer applies.
  • All pricing tables updated to April 2026 (was March 2026).
  • Max Output column for Opus updated to 128K (was incorrectly listed as 32K).
  • Default opus alias in the site calculator now refers to Opus 4.7; the calculator surface supports both.
  • Regional endpoint +10% premium clarified to apply specifically to Sonnet 4.5+ and Haiku 4.5+ (global pricing structure changed with that generation).
  • Data residency +10% multiplier clarified as inference_geo: us-only on Opus 4.7 and newer, not a blanket "Opus and above" rule.

Fixed

  • Token estimator Python CLI: removed obsolete opus_4.6_1m entry; added legacy opus_4_6 explicitly.
  • Cheatsheet: corrected max output per turn values and the historical Haiku-to-Opus cost ratio explanation.
  • Benchmarks headers now include explicit "measured on Opus 4.6" note since Opus 4.7 benchmarks are not yet collected.
  • Comprehensive factual sweep: updated every remaining "Opus 4.6" default-recommendation reference to "Opus 4.7" across guides/09-subscription-value.md (plan model lineup, allowance table, Batch API table), guides/08-prompt-caching.md (cache pricing table), guides/06-access-methods-pricing.md ("stay under 200K" tip rewritten for current standard-rate 1M context, March→April 2026 date), benchmarks/context-size-impact.md (file read budget table, context window reference), benchmarks/model-comparison.md (decision tree, recommendation labels), and benchmarks/leaderboard.md (pricing note).
  • Diagrams rewritten (guides/diagrams.md): replaced \n line breaks with <br/> inside quoted labels (GitHub renders these correctly, \n showed as literal text), dropped inline <b> tags that GitHub's sanitizer strips, switched from inline style to classDef + class syntax, wrapped model groupings in subgraph blocks for proper boxed regions. All 5 diagrams validated via @mermaid-js/mermaid-cli@11.12.0.
  • Issue templates (.github/ISSUE_TEMPLATE/leaderboard-entry.md, case-study.md): added Opus 4.7 option.
  • Case study template (case-studies/TEMPLATE.md): added Opus 4.7 to example model list.
  • Context size benchmark: corrected the "200K token context window" claim to reflect that Opus 4.7/4.6/Sonnet 4.6 are 1M and Haiku 4.5 is 200K.

[1.3.0] - 2026-04-06

Added

  • Installable cost-mode skill: npx skills add Sagargupta16/claude-cost-optimizer then /cost-mode
  • Plugin structure (.claude-plugin/, plugins/, skills/, .agents/) for Claude Code marketplace
  • Three intensity levels for cost-mode: lite (20-40% output reduction), standard (40-60%), strict (60-70%)
  • Guide 00: Getting Started in 5 Minutes -- zero to optimized in 5 steps
  • Guide 10: Three-Tier Task Routing -- skip LLM for Tier 0, Haiku for Tier 1, Opus for Tier 2
  • Repo Analyzer page -- paste a GitHub URL to get full cost audit, grade, and recommendations
  • "Try it on our repo" prefilled demo on Analyzer page
  • cost-logger.sh hook -- logs estimated tokens and cost per tool call
  • Output Token Optimization section in cheatsheet (5 strategies + arXiv reference)
  • Community Tools section in README referencing caveman project
  • Before/after cost comparison example in README (61% savings)
  • Star History chart in README with dark/light mode support
  • SEO meta tags (Open Graph, Twitter Card, canonical URL, keywords)
  • tools/README.md with quick reference table
  • GitHub Discussions enabled on the repository

Changed

  • Repo transformed from docs-only to installable skill + docs
  • README install command featured at the top
  • Home page hero links to Repo Analyzer
  • Repo Analyzer shows results for repos with no config files (was showing error)
  • Cheatsheet links table updated with all new guides and tools
  • CLAUDE.md updated with new file structure and skill directories
  • Home page guide count updated to 10

[1.2.0] - 2026-04-03

Added

  • Guide 08: Prompt Caching Deep Dive - cache mechanics, TTL economics, ROI math
  • Guide 09: Maximizing Subscription Value - plan comparison, upgrade/downgrade signals
  • Visual decision tree diagrams (Mermaid) for model selection, session optimization, cost tiers
  • React site (Vite + React 19 + TypeScript) with cost calculator and badge checker for GitHub Pages
  • MCP cost estimation server with estimate_cost, session_estimate, compare_models tools
  • Claude Code budget enforcement hooks (budget-tracker, session-summary)
  • Efficiency badge generator (A+ to F grading, shields.io badge output)
  • VS Code extension for token count and cost estimation in status bar
  • GitHub Action for automated cost auditing on PRs
  • /optimize custom command for project cost-efficiency analysis
  • Case studies directory with submission template and issue template
  • Community benchmark leaderboard with seed data
  • 5 new stack-specific CLAUDE.md templates: Go, Rust, Django, Rails, Java Spring Boot
  • Leaderboard entry issue template
  • Case study issue template
  • Awesome-list submission preparation guide
  • GitHub Pages deployment workflow (deploy-site.yml)
  • Issue template config with quick-link cards

Changed

  • Simplified CONTRIBUTING.md with contribution ladder (Level 1-6) and "Your First PR in 5 Steps"
  • Updated CLAUDE.md file size guidance with precise limits (4K chars/file, 12K total) based on community research
  • Updated compaction docs with thresholds (10K tokens trigger, 4 messages preserved)
  • Updated prompt caching guide with static/dynamic boundary explanation
  • Updated hooks with correct JSON payload format and exit code semantics
  • Updated cheatsheet with 7 new entries (character limits, compaction, output caps, token estimation)
  • Cleaned up duplicate calculator files (replaced vanilla JS with React site)

[1.1.0] - 2026-03-16

  • Add regional pricing, PPP note, and cloud discount info
  • Add off-peak 2x usage documentation
  • Update 1M context from beta to native, add README badges

[1.0.0] - 2026-03-08

  • Add 1M context pricing, Fast Mode, and access methods guide
  • Update all pricing to March 2026 (Opus 4.6, Haiku 4.5)
  • Initial release: Claude Cost Optimizer