v4.24.0 — fresh install works without Bun

Fixes a broken fresh install and a red CI that had been failing since 2026-08-14.

Fixed

• The install gate ran doctor --require-llmadapter, so any machine without Bun failed — even though Bun is only needed by the optional llmadapter. With the JSON report redirected to a file, the failure printed nothing: a bare exit 1. The smoke now gates on the portable core, asserts the documented bun_missing degradation, runs the adapter lane only where Bun exists, and prints the report on failure.
• Privacy: removed absolute /Users/<name>/... paths from 11 published benchmark artifacts, which exposed the maintainer's username and unrelated private project names. The guard test now scans all 89 benchmark artifacts instead of an 8-file allowlist that missed them.

Added

• CI runs tests and the neutral install on macOS and Linux, plus a dedicated Bun lane.
• README documents OS/runtime requirements (Python 3.11+, Linux/macOS, WSL2 on Windows) and links the companion skill router.

Changed

• The compact default consumes bounded multi-phase routing: one primary plus at most four supports; single-phase work keeps zero/one routing.
• Tests read the canonical version instead of hardcoding it.

Verified on Linux and macOS runners, with and without Bun.

v3.1.5

Public team profile: Lean runtime plus bounded controller/worker protocol. Removes Superweb and private host-tool references from active public surface. Adds released CLI selection evidence and preserves legacy news config through doctor migration.

v3.1.4

Updates Router companion guidance for v1.2.2: context-mode remains explicit-only and plain test verification stays zero-skill.

v3.1.3

Adds a clear zero-hot commitment, remaining proof obligations, and a dedicated agent-token-saver-skill-router install and routing guide.

Cross-links router v1.2.1, which prevents automatic context-mode handbook loading for ordinary verification tasks.

v3.1.2

Fixes the Claude Code installer to preserve unrelated shared UserPromptSubmit hooks while keeping the managed hook idempotent.

Expands agent-token-saver router metadata for noisy-output compression, CLI benchmarking, and bounded subagent workflows.

v3.1.1 — empty router fallback

Fixed\n\n- Restore the hidden token-saver fallback when an installed companion router returns no valid primary selection for an explicit token/context task.\n- Keep trivial and ordinary prompts silent unless the router selects one primary skill.\n\n## Verified\n\n- 32/32 tests passed.\n- Fresh-HOME neutral install smoke passed.\n- GitHub Actions CI passed on commit 3698e61.

v3.1.0 — zero-hot routing and multi-agent accounting

v3.1.0 — zero-hot routing and multi-agent accounting

This release removes the measurable idle tax that v3.0.1 exposed.

What changed

• Codex and Claude keep the saver skill outside their visible skill catalogs.
• Automatic fuzzy routing is strict, emits nothing when uncertain and activates at most one primary skill.
• A conservative built-in gate still handles explicit token/context tasks when the companion router is absent.
• minimal is truly zero-hot: CLI and ledger only, with no visible skill or prompt hook.
• agent-token-ledger now sums named parent/child usage files and fingerprints duplicate visible context.

Neutral evidence

Clean CODEX_HOME ABBA microbenchmark:

• Baseline median: 11,204 input.
• Lean median: 11,209 input.
• Delta: +5 input tokens / +0.045%.
• Cache and output identical; every answer passed.

This is a no-tool idle-cost result, not an end-to-end savings multiplier. The
146.1x number remains a separate accepted payload-capacity benchmark.

Accepted explicit-RTK Codex probe:

• raw ps aux: 25,210 input
• explicit rtk ps aux: 23,996 input
• 1,214 fewer full input tokens (4.82%), identical 19,968 cached subset

One failed no-command warmup remains recorded and excluded from the accepted
A/B. Transparent Codex RTK rewriting is still not claimed.

agent-token-saver v3.0.1 — neutral-host proof + full-context ledger

v3.0.1 — neutral-host proof and full-context ledger

This patch closes the largest credibility gap in v3.0.0.

• Fresh-HOME universal install now runs on a neutral GitHub-hosted Ubuntu runner.
• doctor distinguishes a usable portable core from a fully populated optional profile.
• agent-token-ledger reconciles provider usage with visible context layers.
• Hidden host overhead is reported as unattributed input, never counted as a saving.
• README labels 146.1x correctly as a dated accepted-payload benchmark.
• Clean Codex full-context control is published even though it is negative:
  +0.68% input on trivial work; the unverified shell arm is rejected.
• Claude uses RTK's native hook. Codex uses skill/CLI guidance until
  unified_exec hook coverage is complete.

Verification:

bash scripts/neutral_install_smoke.sh
bash scripts/remote_bootstrap_smoke.sh
uv run pytest
agent-token-saver doctor --profile lean --json
agent-token-ledger --help

agent-token-saver v3.0.0 — Less noise. Better judgment.

v3.0.0 — agent-token-saver

The former Claude-focused toolkit is now an agent-neutral, measured token
routing stack.

What changed

• One fail-safe installer for Codex CLI, Claude Code, Hermes Agent, GG Coder
  and repo-local agents.
• Native hooks only where the host exposes them; portable skills everywhere
  else.
• Four measured profiles instead of an always-on pile of tools.
• Reproducible matrix: 380,871 raw workload tokens versus 2,768 for the
  CLI-selective accepted path in the dated local benchmark.
• Public memory seam; unreleased Synapse is not required or bundled.
• Security policy, issue templates, CI and benchmark contribution contract.

Upgrade

git pull
./install-universal.sh --profile lean --agent auto --dry-run
./install-universal.sh --profile lean --agent auto
agent-token-saver doctor --profile lean

GitHub redirects the old repository URL. Existing cts entrypoints remain for
compatibility.