# Tamcoder 0.1.55

Move requested fixes out of repeated directory discovery and into targeted implementation.

- After two broad directory listings without an edit/check, retain the discovered paths and temporarily offer targeted reads, edits and checks instead of more broad listings.
- Hand the existing evidence to another eligible route once; plan, capability, permission and budget gates remain in force.
- Refuse listing calls that ignore the narrowed tools without executing another filesystem scan. Restore discovery after a real edit/check.
- Continue/resume keeps the original implementation request, so restarting a turn does not silently disable the action safeguard.
- Preserve Qoder's bounded cached-read recovery. Invalidate caches and failed-command blocks after edits/checks so fixed tests may run again.
- Prevent loop-close notices from scheduling another automatic no-action turn. Manual continuations carry the reason the last loop was closed.
- Ordinary OS session identifiers no longer trigger the vendor-credential preflight check; actual API keys and session tokens still fail closed.
- Headless CLI runs select the shipped model pool when no explicit or user pool is configured, matching interactive startup.
- SDK query handles now support resume and live model changes. Post-tool success/failure hooks run from real completed tool items; assistant messages do not trigger them. The Claude-shaped facade retains these controls.

This is a native Tamcoder engine and SDK update. It does not embed the full vendor Claude Code/Codex runtime, claim universal SDK parity, or guarantee every model can fix every task. Tests use scripted providers and real workspace file edits; no comparative latency claim is made.

The site artifact is independent of npm/PyPI publication. Registry credentials/OTP requirements previously blocked synchronized publication; do not infer registry availability from this site release.