/now
What I'm doing now
A snapshot of what has my attention — in the spirit of a now page. The numbers below are read live from the framework's own metrics file, not typed by hand.
Last updated: 2026-10-11
Framework pulse
fiser_framework_ia v3.42.0
Source: assets/data/framework-stats.json, regenerated from the vault by scripts/gen_stats.py. Read how the framework works →
Building
Where the effort is going
Control over prose
Written rules kept relapsing, so the fix is mechanical: deny-only hooks that block credential dumps and unsafe writes, each backed by regression tests built from real incidents — and tested again under real session conditions.
Production observability
A headless watcher over automation run history that pushes an alert only for genuinely new failures. A script, not an agent — nothing in it needs reasoning.
A leaner context budget
Measuring what an agent has to load before it can start, removing duplicated identity and distilled-rules weight, and checking recall with fixed queries so a smaller index never means a worse one.
Right-model routing
Scout, Builder and Advisor roles instead of one big model for everything — provider-agnostic rules, so the model can change without rewriting the process.
Multi-AI design reviews
Structured debates between agents from different vendors plus an outside reviewer, forced to converge into a table of decisions with an owner, a metric and a date — and a list of what was rejected.
Enterprise data & integration
Moving analytics extraction onto vendor-sanctioned replication paths ahead of a hard compliance deadline, and keeping Power Platform automations governed and observable.
Changelog
Recent framework milestones
- Enforcement hook for secrets, production watcher, single identity loadA deny-only hook stops credentials reaching the screen; a headless watcher covers production automations; the identity file loads once instead of four times (−17 KB per session).
- v3.42.0 — P7 Right-Model RoutingScout / Builder / Advisor roles become the seventh core principle; no new skill or MCP was added.
- v3.41.0 — framework audit, 5 skills retiredUnused skills removed under the complexity-budget rule (every addition needs a compensating removal); health checks stopped producing false positives.
- One write surface for distilled rulesCross-agent rules consolidated into a single hand-maintained file with a 6-month vigency rule, instead of building a memory engine.