v1.2.0

Feature-heavy minor release. Everything is backward-compatible by default — every new input defaults to the previous behavior; existing workflows keep working unchanged.

Features

• Structured findings (#181): the model now returns a normalized findings[] array (severity, file, line, message) exposed as a findings output, plus verdict_policy: findings_severity_gated to derive the verdict deterministically from findings (request_changes iff any blocker-severity finding; falls back to the model verdict when no findings).
• Inline review comments (#182): opt-in inline_findings: true attaches diff-anchorable findings as native line-anchored review comments (inline_findings_max, default 20). Ignored for publish_mode: comment.
• Fast/smart model routing (#186): review_routing_mode: auto routes low-risk PRs to a fast model and risk-flagged PRs (auth, secrets, migrations, security issues, …) to a smart model via ai_fast_* / ai_smart_* inputs. Fast defaults to the primary config, smart to the fallback config.
• Review escalation (#187): when the fast review looks insufficient (request_changes verdict, unaddressed required checks, low confidence, blocker signals), the review is automatically re-run on the smart model; escalate_on_* inputs control the triggers. New outputs: review_route, escalation_reason.
• Required-check completeness validation (#185): the classifier emits must_check items for risky PRs and the review is validated against them (validate_required_checks, default auto; required_check_validation_mode: warn/fail/metadata_only). New output: required_checks.
• Deterministic PR classification (#163) now steers the review prompt with pr_kind, risk_flags, and must_check (#178) — injection-safe, derived from file patterns rather than PR text.
• Token-aware context budgets (#170): model_context_tokens sizes the corpus to the model's real context window — important for 8k–32k local models that the named modes silently overflowed.
• OpenAI-compat API knobs (#168): ai_response_format (json_object/json_schema — strongly recommended for small local models), ai_temperature, ai_tokens_param (max_completion_tokens support).
• Graceful model-failure handling (#174): on_model_failure: notice posts a visible request_changes notice instead of failing the check with nothing on the PR; plus a step-summary table (engine, verdict, route, token usage, truncation).

Fixes

• Skip-if-unchanged fingerprinting and incremental review actually work; fingerprint includes prompt/config/rules hashes (#165).
• wait_for_ci.sh works on Checks-API-only repos and excludes its own run from the wait (#176).
• Response parsing and SSE reassembly hardened for local models (#166); model-call failure classes distinguished with HTTP error bodies captured (#167, #169).
• Tool harness planner contract fixed; executor tolerance and parse retry added (#171).
• Classifier: renovate mixed-PR mislabeling fixed; multi-language risk patterns (#172).
• Audit closeouts: #164 (tool harness/precheck/corpus correctness + security) and #184 (all 11 findings of audit #183).
• Post-merge cleanup of routing/escalation rough edges (#188).

Performance

• Parallelized context-collection I/O, deduplicated PR object/diff fetches, precheck short-circuit (#175).
• Low-value bytes trimmed from the review corpus and planner context (#177); corpus reordered so high-value evidence survives truncation (#170).
• Linked-source fetching parallelized with a bounded enrichment time budget.

Security

• Model API keys and GitHub tokens passed via 0600 curl --config files instead of argv (#179).
• @-mentions neutralized in posted reviews; redact.py Python 3.9 compat (#173).
• Random heredoc delimiters prevent model-controlled text from injecting step outputs.

Docs

• Local-model troubleshooting guide, smoke-test repo fix, run_command documentation (#180); review guidance tuning (#161).

New inputs

verdict_policy, inline_findings, inline_findings_max, review_routing_mode, ai_fast_model/ai_fast_base_url/ai_fast_api_format/ai_fast_api_key, ai_smart_model/ai_smart_base_url/ai_smart_api_format/ai_smart_api_key, escalate_on_risk_flags, escalate_on_incomplete_required_checks, escalate_on_fast_request_changes, escalate_on_fast_low_confidence, escalate_on_tool_or_evidence_blockers, validate_required_checks, required_check_validation_mode, model_context_tokens, ai_response_format, ai_temperature, ai_tokens_param, on_model_failure, ci_status_check/ci_timeout_sec/ci_interval_sec/ci_skip_on_timeout, review_scope.

New outputs: findings, verdict_source, required_checks, review_route, escalation_reason, effective_review_scope, ci_status_skipped, ci_status_final.

Behavior notes

• review_scope defaults to auto: full review first, then incremental reviews on later pushes when safe — a meaningful token saving on long-lived PRs.
• tool_planning_timeout_sec documentation now matches the actual default (60s).