Issues

Order-dependent parser-test failures: pytest tests/parsers fails 12 on a clean master while every file, pair, and the full suite pass

#627

· gadievron opened

on Sep 14, 2026

No prompt caching on any phase, and the tracker cannot price cached input (an enhancement: the provider-aware cache accounting prerequisite ships before any cache control)

#626

· gadievron opened

on Sep 13, 2026

No request-side thinking control, and the response translation discards unsupported thinking blocks while forwarding their usage (an enhancement: expose the policy, keep the default)

#625

· gadievron opened

on Sep 13, 2026

The per-phase stderr summary's "N API calls" counts tracker records, not requests — and the record unit differs by phase (enhance counts units; single-shot counts requests)

#624

· gadievron opened

on Sep 13, 2026

The JSON rescue schema offers a verdict enum the analysis prompt never does, and INSUFFICIENT_CONTEXT hits four consumers that disagree (errors / completed / dropped / absent)

#623

· gadievron opened

on Sep 13, 2026

Stage-2 disagreements corrected to protected fold into the safe metric — the protected destination category is lost from every summary (#509's recorded open question)

#622

· gadievron opened

on Sep 13, 2026

The built-in Stage-2 verification persona contradicts untrusted-input application contexts: a browser-only attacker, an ungated local-access exclusion, and trust boundaries the verifier never sees

#621

· gadievron opened

on Sep 13, 2026

The agentic enhance give-up classes are conflated under one 'incomplete' classification — expose exit_kind diagnostics (the iteration-limit budget exhaustion differs in kind and remedy)

#615

· gadievron opened

on Sep 13, 2026

Agentic enhance stores a completed paid classification, then overwrites it with an error dict when post-result assembly fails (unchecked include_functions finish items)

#614

· gadievron opened

on Sep 13, 2026

Enhancement non-verdicts leak into analysis hints, disappear under --exploitable filtering, and survive re-enhancement in analyze checkpoints

#611

· gadievron opened

on Sep 13, 2026

Catalogue lead: five SDK-listed Anthropic model literals have no registry record under any provider

#610

· gadievron opened

on Sep 13, 2026

Enhancement's running summary omits unsuccessful-attempt spend (the error-path agent_state usage is wiped by the retry) — summaries and the terminal report cannot reconcile

#609

· gadievron opened

on Sep 13, 2026

Navigation Menu

Uh oh!

All issues

Order-dependent parser-test failures: pytest tests/parsers fails 12 on a clean master while every file, pair, and the full suite pass

No prompt caching on any phase, and the tracker cannot price cached input (an enhancement: the provider-aware cache accounting prerequisite ships before any cache control)

No request-side thinking control, and the response translation discards unsupported thinking blocks while forwarding their usage (an enhancement: expose the policy, keep the default)

The per-phase stderr summary's "N API calls" counts tracker records, not requests — and the record unit differs by phase (enhance counts units; single-shot counts requests)

The JSON rescue schema offers a verdict enum the analysis prompt never does, and INSUFFICIENT_CONTEXT hits four consumers that disagree (errors / completed / dropped / absent)

Stage-2 disagreements corrected to protected fold into the safe metric — the protected destination category is lost from every summary (#509's recorded open question)

The built-in Stage-2 verification persona contradicts untrusted-input application contexts: a browser-only attacker, an ungated local-access exclusion, and trust boundaries the verifier never sees

The agentic enhance give-up classes are conflated under one 'incomplete' classification — expose exit_kind diagnostics (the iteration-limit budget exhaustion differs in kind and remedy)

Agentic enhance stores a completed paid classification, then overwrites it with an error dict when post-result assembly fails (unchecked include_functions finish items)

Enhancement non-verdicts leak into analysis hints, disappear under --exploitable filtering, and survive re-enhancement in analyze checkpoints

Catalogue lead: five SDK-listed Anthropic model literals have no registry record under any provider

Enhancement's running summary omits unsuccessful-attempt spend (the error-path agent_state usage is wiped by the retry) — summaries and the terminal report cannot reconcile

Uh oh!

Search results