v1.1.0: bench, image route

Added

• bench/: a routine-task bench for the maintainer. bench.py runs one model CLI (codex,
  claude or agy) on ten tasks; a hidden check grades a temp copy of each result, and pass,
  seconds and tokens go to bench/results.tsv. --selfcheck proves the checks of tasks 6–10
  fail on the untouched files and pass on a reference answer, with no model call. --task N runs one task.
• Tasks 6 to 10 are hard on purpose: a bug fix whose causes sit in other files, a git bisect
  where some commits can't run, reading a table and a chart from a screenshot, a long
  version-range spec (73 hidden cases), and a signature change through indirect calls.
• An image route: screenshots, scanned pages and charts go to GPT-6.1 Terra (codex exec -i),
  with Gemini 3.8 Flash as fallback.

Changed

• ROUTING.md: benchmarks pick a new model's first route; the bench decides whether it keeps it.
  This replaces "don't run trial calls to confirm them".
• ROUTING.md lists Claude Haiku 5.5 (claude-haiku-5-5) as not routed, replacing Haiku 4.5,
  and names GPT-6.1 Terra, so the watcher counts both as covered.
• The offline check now includes python3 bench/bench.py --selfcheck.

Bench result, 2026-10-27: GPT-6.1 Sol high and GPT-6.1 Terra medium both passed 10/10.
Code stays with Sol: the bench doesn't yet test the long agentic work where trackers put Sol
ahead.

delegate-router v1.0.1

Docs only. No code change since v1.0.0, so its verification and known limits still apply.

• Requirements: the README now says what's supported. Your main agent must be Claude Code or Codex CLI, plus at least one more of Codex, Gemini (agy) or Claude. DeepSeek and local models aren't supported yet.
• README: new title, intro and keywords, picked for how well they match what the repo does.

delegate-router v1.0.0

First release of delegate-router, an agent skill for Claude Code and OpenAI Codex CLI. It hands bulky, already-decided coding work to another model you already pay for: Codex (GPT-6.1 Sol, GPT-6 Luna, GPT-6 Astra), Gemini 3.8 Flash through the Antigravity agy CLI, or a Claude Opus 5.5 subagent.

What's in it

• The /delegate skill: four gates (bulk, decided, verifiable, contained), a never-delegate list, spec rules, and host-side verification of every diff.
• ROUTING.md: a task-to-model map with reasons, effort levels and a council protocol for when two models disagree.
• model_watch.py: notices new model releases in your CLIs and the vendors' release notes. It can optionally update the CLI and probe the new model with one read-only prompt, and it stops every step at 15 minutes.
• Interactive install.sh: links the skill into Claude Code and/or Codex and remembers your answers. It never overwrites anything it didn't create.
• Optional Jev (TypeSafe) release-note triage and file scouting, with a data gate.

Install

git clone https://github.com/pvjbuilds/delegate-router ~/.delegate-router
~/.delegate-router/install.sh

Needs Claude Code or Codex CLI as your main agent, plus at least one more of Codex, Gemini (agy) or Claude.

Verified

• 44 offline tests: the watcher on Python 3.9, the installer, Jev and scout. They use stub CLIs and make no network calls.
• Live checks: the Claude and agy probes answer ok. A Codex implementer running with user config, rules and apps off writes its file. A Codex host runs the Claude delegate once network access is on.
• Cross-provider review: GPT-6 Astra at medium effort, three rounds, with all findings fixed.

Known limits

• The round 3 fixes were not re-reviewed by another model.
• Tested on macOS only. Linux should work but is unverified.
• A Claude Code host's subagent runs with the host's permissions.
• agy has no switch to turn off integrations.
• Codex plugins and admin-managed config still apply.
• *_BIN paths are used by the watcher only.
• status may start the 12-hour background check.

The tag predates two README-only commits on main: the requirements line and the keywords.