Chimera v0.7.0 — read the front-end + honest benchmarks
The front-end reading + honest-benchmark release. All additive; nothing breaking.
Highlights
🌐 Read the front-end, not just the back-end
The browser tool (Playwright/Chromium) can now read a rendered page, not just drive it:
read_text— the page's full rendered text as clean Markdown (via MarkItDown when thedocumentsextra is present; plain visible text otherwise), with an optionalurlto open+read in one step. Closes the gap where the agent could navigate/click a page but couldn't actually read an article/doc/result that rawhttp_getmisses on JS-heavy sites.find— search the rendered text for a query.- Output stays data-fenced and taints the run (already in
FETCH_TOOLS). NewCHIMERA_BROWSER_HEADLESStoggles headful Chromium for debugging. No new dependency — reuses the existing MarkItDown seam.
🧬 Self-evolution: transfer-holdout gate
A learned change (RFT / GEPA / ACE / skill) must not regress a disjoint, same-capability holdout slice before promotion — not just win its tuned slice. Cites EvoAgentBench (arXiv 2607.05202), which measured ungated evolution producing negative transfer (GEPA −12.3).
💰 Prompt caching, used and measured
Opt-in prompt_cache marks the stable system prefix with a provider cache breakpoint; the gateway captures cache read/write tokens through the budget into delegation receipts — turning the token-economy "dollar caveat" into continuous measurement. Honest status on caching routes documented.
📊 Official Terminal-Bench A/B — published as a pre-registered null
A pre-registered N=40 A/B landed at a variance-dominated floor with no significant difference either way — published as-is, including retracting a wrong intermediate read once the control arm was measured. Null results and self-corrections ship too. See bench/terminal_bench/RESULTS.md.
Plus: curve-driven delegation gate, a 3-axis hierarchy sweep + caching dollar model, and hierarchy code-review fixes.
Full details in CHANGELOG.md.
pip install --upgrade chimera-agent