- drop the algorithm/gate source SHA256 freeze: every ranking change
required a manual manifest hash sync, coupling code evolution to the
manifest
- drop the --require-sealed release gate (sealed qrels hash, dual
signatures, coverage minima, independent-result thresholds) and its
tool-search-independent-data-gate Makefile target
- behavior regression protection stays with the Go evaluation sentinels
that run the real assembled catalog
- default engine: fielded BM25 ensemble (action_v1 remains a shadow candidate)
- add --explain score_breakdown: per-field contributions, structured
multiplier, query classification and avoid_when penalty
- add DWS_TOOL_SEARCH_* env overrides (algorithm, k1/b, field weights,
explain) for release-independent tuning and rollback
- tokenizer: CJK unigram+bigram so single-character intents match; align
isToolSearchCJK with unicode.Han used by the evaluation slices
- structured routing: extend vocabularies with English synonyms,
word-boundary matching for ASCII values, restrict the rerank gate to
technical identifiers only
- avoid_when demotion layer wraps every lexical algorithm (Forbidden@5
proxy 37.55% -> 0); intent-overlap matching rejected after control
showed it demoted gold answers (Recall@5 0.8664 -> 0.7845)
- remove the OA task_id +1.00 fixture-specific multiplier
- build the default engine once per process for long-lived hosts
- evaluation baseline: intent-proxy R@5 990/1123 (88.16%); update
integration sentinels accordingly
Do not expand the default policy gate with mono-multi or skill-commands;
leave them as optional make targets only.
Co-authored-by: Cursor <cursoragent@cursor.com>
Drop the recovery package/commands and related Schema/skill teaching so agents
stop being steered at a dead surface, while keeping mono↔multi content QA work.
Co-authored-by: Cursor <cursoragent@cursor.com>
Schema Catalog now assembles from Contract/ParamDecl/Interface and Cobra only.
Keep fetch-mcp-metadata as an optional diagnostic dump and ban the retired pin path.
Co-authored-by: Cursor <cursoragent@cursor.com>
Phase 3 of identity-deregistry. BuildEffectiveCommandRegistry now builds the
EffectiveCommandRegistry from CollectIdentitySpecs(root) instead of the
embedded reviewed registry, and the registry source is removed atomically:
- delete internal/cli/schema_command_registry/ (registry.json + products/),
schema_command_registry.schema.json, the three //go:embed directives, and
the registry-only loaders/validators (loadReviewedCommandRegistry,
decodeCommandRegistry, ValidateCommandRegistrySource, shard assemble/merge
helpers, ReviewedCommandRegistryMergedJSON/SourceHash, ReviewedCommandSpecs)
- keep CommandSpec/CommandRegistry/EffectiveCommandRegistry,
newEffectiveCommandRegistry/indexCommandSpecs, and SourceHash; the
collector-built effective hash is byte-identical to the reviewed one, so
catalog surface_hash/source_hash are unchanged
- convert the Phase 2 dual-run gate into TestCollectedIdentityIsValidSingleSource:
collected specs non-empty, no missing primaries, effective build succeeds,
SourceHash stable across repeated collection walks
- generators: catalog -surface and agent-metadata -registry/-surface become
fail-closed retired valves; outputguard no longer protects the registry
paths; fetch_mcp_metadata derives interface refs from collected identity
instead of the merged registry JSON
- retire scripts/policy/check-schema-command-registry.sh and its Makefile
invocation; generate-schema/check-generated-drift now fail closed if
schema_command_registry/ reappears
- update AGENTS.md, docs/reference.md, and in-code reviewed-input notes
Finish Catalog/Agent-metadata naming debt so delivery and fixture symbols no
longer imply a retired go:embed Catalog or Hint overlay path.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep declare→ResolveSchemaBuild as the ToolSpec authority, but materialize
map[cli_path]CommandMeta during deliverySchemaCatalog sync.Once so leaf
--help / ResolveMeta are O(1) after the first Schema touch. Defer wire
Catalog/Tools maps, stamp Source as runtime-assembled, drop committed
catalog/gob fixtures, and cover steady-state reuse with app/cli tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
Move schema authority to declare-time ParamDecl/ContractFinal and
ResolveSchemaBuild so CI/runtime assemble instead of shipping JSON
exclusions/meta-index as delivery sources.
Co-authored-by: Cursor <cursoragent@cursor.com>
Publish a compact schema_meta_index.json beside the catalog so help/selection
lookups avoid decoding the full ToolSpec wire on the hot path.
Co-authored-by: Cursor <cursoragent@cursor.com>
Remove schema_hints as a generation input so Catalog delivery depends solely on leaf ContractFinal and ProductDecl; migrate policy and contract tests to embedded catalog introspection and fix publicShortcutCount for chat-list.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add the ⑥ regression gate that replays every reviewed validation_fixture bad case
through the real embedded PreParse pipeline and asserts the canonical outcome
(accepting either semantic rewrite or native real-flag acceptance, failing only
on a genuine unknown-flag hallucination). Add check-param-concepts.sh (dictionary
schema/loader invariants) and check-param-alias-cooccurrence.sh (full-tree
co-occurrence scan), and wire all three into make policy.
Add internal/generator/cmd_param_aliases: reads the reviewed dictionary plus the
live Cobra tree, reduces each concept against a command's real flags (>=2 visible
real flags without a reviewed ambiguous entry fails generation), and emits the
committed internal/cli/param_aliases_generated.go table with lookup helpers.
Extend generate-schema and check-generated-drift.sh to treat the dictionary as a
reviewed input and byte-guard the generated table.
Own safety/gates/parameters in metadata/ and Agent prose in selection/,
drop the monolithic Manual file, and keep confirmation aligned with
per-tool runtime_gate.
Co-authored-by: Cursor <cursoragent@cursor.com>
玉澜