v2.2.0

DeepDoc 2.2.0 is an infrastructure and planning reliability release focused on data safety, incremental update correctness, and multi-spec API support.

New Features

Persistence & Safety

• Atomic writes across all state files — every .deepdoc/ write now goes through atomic_write_text / atomic_write_json (mkstemp → fsync → os.replace), eliminating partial-write corruption on crash or SIGKILL.
• Exclusive process lock — deepdoc_state_lock uses an fcntl file lock so concurrent generate or update runs on the same repo are detected and rejected immediately with a clear error message instead of silently overwriting each other's state.

Planning

• DocBucket semantic identity — DocBucket gains semantic_id, origin, confidence, evidence_anchors, and planner_schema_version fields. build_bucket_semantic_id() derives a deterministic, slug-stable key from endpoint family, parent slug, or owned-file fingerprint. Serialized to and from the .deepdoc/plan.json ledger.
• Targeted incremental plan merge — SmartUpdater replaces its naive slug-append strategy with _merge_targeted_plan(), which matches replanned buckets against existing ones by slug then semantic ID. Matched buckets are mutated in-place (files, hints, confidence merged) and queued for regeneration only when their ownership actually changed, avoiding unnecessary regeneration of unchanged pages.
• Monorepo service boundary detection — scan_repo now runs _detect_service_boundaries(), which reads an explicit services: list from config or auto-discovers service roots by looking for pyproject.toml, package.json, go.mod, Dockerfile, etc. one level under services/, apps/, and packages/. Every scanned file is tagged with its owning service in RepoScan.file_services.
• Endpoint keyword map is now config-extensible — ENDPOINT_DOMAIN_KEYWORDS moved to planner/common.py and _build_repo_endpoint_keywords() applies a three-layer merge: user endpoint_groups config wins, then families observed by the scanner, then the hardcoded fallback. Honoured in both heuristics.py and endpoint_refs.py.

Pipeline & CLI

• --strict-quality flag — deepdoc generate and deepdoc update now accept --strict-quality, which reads pages_failed, pages_invalid, pages_degraded, and mdx_fallback_slugs from the run stats and raises a ClickException listing all blockers. Designed for CI pipelines that should fail-fast on degraded output.
• Multi-spec OpenAPI staging — stage_openapi_assets now stages all detected specs instead of stopping at the first. Each spec gets a collision-safe deepdoc-openapi-{n}-{stem}-{sha1} filename and a single combined manifest.json is written.
• LLM token usage in quality report — token usage from the LLM client is captured into stats and persisted to generation_quality.json after every run.
• Quality-driven staleness — _doc_quality_requires_regeneration marks any page whose ledger record shows is_valid: false, mdx_fallback_applied, or whose content contains stub: true / deepdoc_status: invalid as stale so it is regenerated automatically on the next incremental run without a --force.
• Duplicate changelog guard — append_changelog_entry skips writing when the incoming commit matches the most recent entry's commit, preventing duplicate entries from rapid pipeline retries on the same SHA.