Why: I2 acceptance - the evaluator must emit metrics over a
multiturn fixture no smaller than the existing corpus, not just
print branch names.
Evidence: go vet clean; full suite ok; intentctl eval prints
precision=1.000 recall=1.000 (TP=12 FP=0 FN=0) over the 11-case
multiturn fixture; MetricsOf unit test pins TP/FP/FN counting
including wrong-match counting as both.
Impact: CorpusMetrics + MetricsOf (expectation_hit counts as a
hit); intentctl eval prints the metrics line after the summary;
corpus_multiturn.json expanded to 11 cases (chained expectation,
tie-break, ambiguous, fallback, handler action, clarify paths).
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md I2
Why: I2 layering - between rules and the L2 classifier there is now
a deterministic host-scored semantic layer (rules then semantic then
model), and fallback can be rescued by a validated command
generator instead of going straight to the default channel.
Evidence: go vet clean; full suite ok, including ten new tests
(rerank resolve/below-threshold/tie/error-then-L2, generator
valid/invalid/error, catalog top-k).
Impact: IntentSpec gains Examples; Options gains Semantic (with
threshold), Generator; ambiguous candidates rerank through host
scores before L2; fallback validates generated commands
(action/slots/kind fail-closed) and synthesizes Top with slots;
generated commands are no longer overwritten by deriveCommands;
CatalogTopK by priority.
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md I2
Why: I1 WP6 - eval needs per-turn command expectations to carry the
SetSlot+Start acceptance case, and corpora must reproduce standalone
in intentctl (the guard case has been red there since before I1
because the guard spec only existed as a test builtin).
Evidence: go vet clean; go test ./... all ok; intentctl eval passes
11/11 on the v1 corpus (guard case fixed via self-declaration) and
1/1 on the new two-turn SetSlot+Start corpus.
Impact: new intent/corpus.go (CorpusCase/CorpusTurn/CommandExpect,
RunCorpusCase with chained expectations); intentctl eval consumes
turns and corpus-declared builtin_specs and domains; report.generate
spec and corpus_multiturn.json fixtures added. v1 corpus format and
conformance unchanged.
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md I1 WP6
Why: I1 WP2 - one explicit action field outranks the four legacy
action keys (Handler, metadata.blueprint, command-id, workflow
IntentID) so hosts migrate to a single derivation path.
Evidence: go vet clean; go test ./... all ok, including declared
action winning the Start command and kind/id validation rejections.
Impact: IntentSpec gains Action (command/workflow/blueprint/handler);
Candidate carries it through; actionOf prefers it, then blueprint,
then Handler, then auto. Legacy keys keep working.
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md I1 WP2
Why: I1 increment one - recognition output becomes a command list
for the host router, same-score ties resolve by priority (I5), and
CJK substring false positives get an opt-in boundary mode (I3).
Evidence: go vet clean; go test ./... all ok, including new command
(6), boundary (3), and tie-break tests; intentctl lint smoke run
shows boundary hints on short unbounded terms. Existing classifier
and ambiguity fixtures repointed to an equal-priority pair; the
greet corpus case now expects the I5 tie-break outcome.
Impact: Result gains Commands; recognizer derives them at one point
(guard/fallback stay empty); MatchSpec gains Boundary (default off,
legacy behavior unchanged); lint gains boundary hints. Top/Branch
stay for compatibility.
Refs: Documents proposal 2026-10-07-agent-stack-n1q1i1-plans.md I1 WP1/WP4/WP5
Why: the intent library has to publish the shape it owns, without sharing a Go type with the other libraries.
Evidence: GOWORK=off go test -count=1 ./schema/ ./intent/ passed.
Impact: schema/ is new. Recognition behavior is unchanged.
Refs: D1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: fullwidth text missed case-folded terms, SpecByID shared maps, the catalog prompt printed handlers and private metadata, overlapping slots kept both values, and negation was easy to mistake for semantics.
Evidence: GOWORK=off go vet ./... && go test -count=1 -race ./... passed, including TestR2NFKCFoldsFullwidthButNotTraditional, TestR2SpecByIDDeepCopiesMetadata, TestR2CatalogHidesPrivateMetadataAndHandlerInMarkdown, TestR2EvalRulesDocumentsMaxSpecs, TestR2OverlappingSlotsAreAmbiguous, and TestR2NegationIsOnlyExcludesAny.
Impact: NFKC is opt-in and does not convert traditional Chinese; catalog markdown omits handlers and non-allowlisted metadata; overlapping slot spans are invalid.
Refs: I4, I7, I8, I10, I11, I12
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The runner cannot reach github.com, so cloning actions/checkout timed out before any test ran.
Evidence: mirrors.aliyun.com/golang/go1.27.1.linux-amd64.tar.gz returned HTTP 200. The workflow now fetches GITHUB_SHA with the job token and installs Go from that mirror.
Impact: Submodule checkout still requires SUBMODULE_TOKEN. A missing secret fails the job with that message.
Refs: feature/agent-stack-r0-r1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The runner has no SSH key, so recursive checkout failed, and tests had no module proxy or race detector.
Evidence: Workflow matches PROJECT.md vet, test, and race commands, with setup-go 1.27 and GOPROXY=https://goproxy.cn,direct.
Impact: vet-and-test runs on ubuntu-latest. SUBMODULE_TOKEN must be set by the owner before checkout can succeed.
Refs: feature/agent-stack-r0-r1
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Refine could swap handlers or emit out-of-range scores, Ambiguous tops skipped slot extraction, and the process-global validator map coupled independent registries.
Evidence: GOWORK=off go vet ./... && go test -count=1 ./... && go test -race ./... passed, including TestRecognizeClassifierRejectsBadRefinement, TestRecognizeClassifierRebuildsTamperedHandler, TestRecognizeClassifierRefinedTopGetsSlots, and TestScopedSlotValidatorsDoNotInterfere. error-code-system and logging-system advanced to a3ee67b and 49b5d8b (docs-only).
Impact: illegal Refine results stay Ambiguous; a resolved Top is slotted; Options.SlotValidators wins over the deprecated global table.
Refs: I1, I2, I6
Co-authored-by: Cursor <cursoragent@cursor.com>
weichen
