CCF v1.7.1

[1.7.1] — 2026-06-17

Added

• Windows one-click install: install.ps1 rewritten into a full bootstrap — detects winget
  (or Chocolatey), auto-installs every missing requirement (Git for Windows → bash+curl, jq,
  Python 3), refreshes PATH in-session so tools work without a reboot, verifies, installs CCF,
  and launches the setup wizard. -Yes for fully unattended.
• install-windows.bat — double-click installer (handles execution policy, pauses on output).
• docs/WINDOWS.md — Windows quickstart + troubleshooting table (winget/App Installer,
  execution policy, PATH refresh, proxy, WSL fallback, manual install).

CCF v1.7.0

[1.7.0] — 2026-06-17

Added

• Cross-platform onboarding (ccf-onboard, Python stdlib): replaces the bash-only wizard for
  setup. Works the same on macOS, Linux, and native Windows — no bash or jq needed to onboard.
  Dependency doctor (OS-specific install hints), unified provider menu (default panel + 20-provider
  catalog), hidden key entry / browser-login (Codex) / keyless (Ollama), per-provider model choice,
  live probe, idempotent writes with backups. Modes: interactive, --check, --list,
  --add P [MODEL], --set-key ENV, --enable/--disable.
• /fusion-setup — Claude-Code-driven setup. Say it (or "set up CCF" + the repo link) and the
  agent installs if needed, runs the dependency doctor, asks which providers + models you want
  (buttons, no secrets in chat), and walks you through key entry privately. No file editing.
• Windows: install.ps1 now auto-installs jq via winget (the runtime needs it) and points to
  the Python wizard / /fusion-setup.

Changed

• install.sh end-of-install offer now prefers the cross-platform Python wizard (falls back to the
  bash onboarder). README onboarding section rewritten around the two new paths.

CCF v1.6.3

[1.6.3] — 2026-06-17

Added

• README hero banner + Solo-vs-Fusion benchmark section with results image and badges.
• benchmark/RESULTS.md — published benchmark report (Fusion 196 vs Solo 192 / 200), blind-spot
  coverage, the deepseek truncation finding, and an explicit judge/solo bias disclosure.
• assets/ — banner + solo-vs-fusion marketing images.

Fixed

• run-benchmark.sh: replaced mapfile (bash 4+) with a bash-3.2-compatible read loop so the no-arg
  "run all tasks" path works on macOS (default bash is 3.2).

CCF v1.6.2

[1.6.2] — 2026-06-17

Fixed

• deepseek / opencode-go usability: raised opencode-go max_tokens 8192 → 64000. deepseek-v4-pro
  is a reasoning model — on complex tasks its internal reasoning consumed the entire 8192 output
  budget and the answer came back truncated (finish_reason=length). 8192 was a hard ceiling for
  every user on hard prompts. max_tokens is a cap (successful answers stop early), so the higher
  value only prevents truncation — no latency/cost penalty on a flat-rate sub. Verified: two
  benchmark tasks that returned 0.4KB truncated now return full 13–16KB answers.

CCF v1.6.1

[1.6.1] — 2026-06-17

Changed

• Benchmark concurrency: tasks now run sequentially (one at a time, never overlapping) while
  panelists within a task run in parallel (launched together, batch-awaited before judging).
  Previously every call was fully serial with a 3s gap between calls. run-benchmark.sh, the
  /fusion-benchmark command, and the benchmark README all updated. Verified: a 3-panelist task
  finishes in ~slowest-panelist time (85s) instead of the sum.
• benchmark/results/ is now gitignored (generated output).

CCF v1.6.0

[1.6.0] — 2026-06-17

Added

• Browser login for Codex (ccf-codex-login + /fusion-codex-login): users no longer need the
  codex CLI or a pre-existing ~/.codex/auth.json. Runs the OpenAI OAuth 2.0 + PKCE flow (same
  official client the codex CLI uses) on a local callback server, exchanges the code, derives the
  ChatGPT account id from the token JWT, and writes ~/.codex/auth.json (chmod 600). Flow ported
  from NousResearch/hermes-agent's tested constants. Python 3 stdlib only (login-time; runtime stays bash).
  --no-browser for headless/SSH, --refresh to refresh without re-login.
• Auto-refresh in fusion-call: when the Codex access token is within 120s of expiry and a
  refresh_token exists, it refreshes silently before the call — no more mid-session expiry failures.

CCF v1.5.1

[1.5.1] — 2026-06-17

Fixed

• Codex auth (#10): fusion-call read the access token + account ID from the wrong path —
  the codex CLI nests them under .tokens.access_token / .tokens.account_id, not top-level.
  Codex auth was failing for everyone ("no Codex token found"); now reads .tokens.* with a
  top-level fallback. Verified end-to-end against the live ChatGPT Codex backend.

CCF v1.5.0

[1.5.0] — 2026-06-17

Benchmark slash commands + fair comparison design.

Added

• /fusion-benchmark slash command: runs SOLO (orchestrator answers directly)
  vs FUSION (panel → orchestrator judges → synthesizes) on 5 coding tasks.
  Sequential only — no parallel panelist calls, respects rate limits. Outputs
  markdown files per task to benchmark/results/.
• /fusion-benchmark-report slash command: generates REPORT.md from
  benchmark results. Grades both arms on the rubric (correctness/completeness/
  blind spots/code quality), produces blind-spot analysis table, cost comparison,
  and honest verdict with caveats about orchestrator/judge bias.
• benchmark/run-benchmark.sh: standalone sequential data collector. Calls
  each enabled panelist one at a time with identical prompts, saves raw responses.
  Use without Claude Code for CI or batch collection.
• Fair comparison design: SOLO = orchestrator answers alone (no panelists
  called). FUSION = orchestrator calls panelists → judges. This isolates the
  panel's value-add. Documents the bias warning: if orchestrator = same model
  family as a panelist, the test is not fair.

[1.4.0] — 2026-06-17

Five new features: Codex/GPT-5.5 panelist (#10), Tavily web search (#11), fusion verification
fixes (#12), analytics dashboard (#13), benchmark suite (#14).

Added

• Codex/GPT-5.5 panelist (#10): new codex-responses transport. Auto-reads ~/.codex/auth.json,
  extracts ChatGPT-Account-ID from JWT, injects Cloudflare headers (User-Agent: codex_cli_rs/0.0.0,
  originator, ChatGPT-Account-ID). Requires SSE streaming (stream:true), store:false, and
  instructions field. Does NOT support max_output_tokens. Disabled by default — enable in panel.json.
• Tavily web search (#11): --search flag on fusion-call enables multi-turn tool calling.
  Panelists request searches via standard function calling → CCF calls Tavily locally → results fed
  back. Max 3 rounds (override with MAX_SEARCH_TURNS). Zero change without --search.
• Analytics dashboard (#13): /fusion-analytics slash command. Pure bash + jq text dashboard
  reading fusion.log. Shows total runs, success rate, per-panelist latency + ok%, cost saved vs
  OpenRouter Fusion ($0.50/call), recent errors.
• Benchmark suite (#14): 5 real coding tasks (bug fix, security, refactor, architecture,
  concurrency) with 0-100 grading rubric. Two new slash commands: /fusion-benchmark runs
  SOLO (orchestrator answers directly) vs FUSION (panel → orchestrator judges) sequentially;
  /fusion-benchmark-report generates comparison REPORT.md with blind-spot analysis. Standalone
  benchmark/run-benchmark.sh collects raw panelist responses without AI judgment. Sequential
  only — no parallel calls, respects rate limits.

Changed

• Fusion verification (#12): Judge JSON schema now matches OpenRouter (partial_coverage,
  unique_insights). FUSION CONFIRMED: N/N line is mandatory. failed_models tracked. Judge
  instructed to verify claims via tools. Warning when <2 panelists respond.

CCF v1.3.0

Roadmap items: dispatcher hardening (#3), retry/fallback (#4), logging (#5), progress (#6),
panel presets (#8), test suite + CI (#1).

Added

• Retry/fallback (#4): per-provider max_retries (default 0) with linear backoff; transient
  network failures retried, full-budget timeouts not.
• Telemetry log (#5): each call appends JSONL to ~/.claude/fusion/fusion.log
  (panelist, provider, model, status, latency_s, bytes); auto-rotates at ~5 MB.
• Progress (#6): ~/.claude/fusion/.last-call written each call; /fusion-status --history
  shows recent runs + per-panelist avg latency and success rate.
• Dispatcher hardening (#3): curl exit-code → actionable messages (DNS 6, connect 7, SSL 35/51/60,
  timeout 28); unparseable (non-JSON) responses reported instead of crashing.
• Panel presets (#8): coding, research, writing, budget, max in config/presets/;
  /fusion-config load-preset / save-preset (backs up, never touches secrets).
• Test suite + CI (#1): bats offline tests for the dispatcher error paths (unknown panelist/provider,
  empty prompt, missing key, context guard) + JSON/preset validity; GitHub Actions runs them on
  Ubuntu + macOS with a secret scan.

CCF v1.2.0

Added

• Provider catalog expanded to 20 (from 13): added Kimi-Code, MiniMax, Novita, HuggingFace,
  Xiaomi MiMo, Ollama (cloud), and Ollama (local, keyless). Updated Gemini (3.x, 2M ctx), Groq
  (Llama-4 Scout), and OpenRouter model lists to current IDs. Endpoints are full chat-completion URLs.
• Keyless providers: fusion-call now supports providers with an empty/none key_env
  (no Authorization header) — enables self-hosted Ollama (localhost:11434) for air-gapped,
  zero-retention panels.
• ccf-models — live models.dev registry browser/installer: the curated catalog is a subset;
  ccf-models reads https://models.dev/api.json (121+ providers, every model with context + cost,
  the same DB Hermes uses) and can providers / models <p> / show <p> / add <p> <model>
  (registers the provider + adds a panelist). Cached 24h, honors CLAUDE_HOME.

Notes

• Catalog model IDs are current-best starters with per-provider docs links; verify the latest at the
  linked docs. OAuth-only providers (Qwen portal) and endpoint-less providers (Arcee) are omitted —
  the dispatcher needs an endpoint plus bearer/x-api-key auth or a keyless host.