-
Sync runtime development branch: provider compat, evidence-gate engine, tool-result compaction, sandbox policy
frostbyte_neo 发布于
2026-07-08 01:48:41 +00:00 Merge a long-running runtime development branch into main:
- Provider compatibility: by_alias tool schemas + reserved-keyword strip
for strict OpenAI-compatible gateways; request-proof budgeting with
deterministic compaction (inline compaction markers replace the JSON
envelope); DashScope provider profile, prompt-cache markers, and
thinking-mode plumbing (thinking_budget omitted unless explicitly
configured). - Provider retry defaults: length-capped responses with no visible text
are treated as reasoning-only and retried for visible content;
thinking-related provider errors and large-context invalid-response
retries disable thinking for the next call only (one-shot). - Context compaction defaults: read_file/git_diff results preserved
verbatim; results already shown in full are never retroactively
compacted; compaction placeholders carry preview_complete plus
retrieval hints. - Tool dispatch: preflight validation pipeline; invalid tool arguments
reported as retryable; write_file destructive-overwrite refusal;
grep_search header/paging/VCS-exclusion; edit_file single-edit
shorthand + near-miss recovery; apply_patch @@ hunks; patch-hygiene
and masked-pipeline advisory notes on shell/git output; coding
profile enforces fresh workspace reads. - Evidence-gate engine modules (all default-off, env-lever controlled):
finalize evidence gate, final diff contract, patch evidence ledger,
source-diff preservation/candidate modes, progress watchdog,
tool-repeat nudge, identical-request loop break, text-only tool
recovery. - Tool-result store: gzip compression and compaction with
backward-compatible reads. - Sandbox policy: sandbox-descriptor integration for filesystem tools,
dispatch-level policy checks, write-policy hardening (deny patterns
now workspace-contained only). - Security: secret redaction now masks assignment values that start
with a quote (previously missed on the always-on memory path). - LLM trace recorder: optional full request/response capture with
Authorization-header redaction. - Developer tooling for run-delivery verification under
scripts/experiments/ (not shipped in the package);
OPENSQUILLA_EXPERIMENT_LEDGER_ROOT with old-name fallback alias
OPENSQUILLA_SWE_EXPERIMENT_LEDGER_ROOT.
See PR #504 for the itemized list of intentional default-behavior
changes and the changelog entry.Co-Authored-By: Claude Fable 5 noreply@anthropic.com
下载附件
- Provider compatibility: by_alias tool schemas + reserved-keyword strip