- manualAgentExampleDryRunEvidence now recognizes executor invocation
envelopes ("kind": "*_invocation" / connect_preview with dry_run) as
invocation previews before the generic request branch, so the 32 devapp
declared tools match their declared preview_kind; pinned with a unit test
covering all invocation kinds plus request/plan precedence.
- TestDevAppWriteGuardRequiresFinalSchemaConfirmation updates devapp wants
to the declared risk grading (reversible writes medium; create/version
create/robot submit high-write; delete/publish destructive) and accepts
contract_final provenance for declared tools while hints-fed tools keep
reviewed_explicit.
Co-authored-by: Cursor <cursoragent@cursor.com>
Align the write-confirmation UX with lark-cli: ConfirmRisk (and the
shortcut confirmRisk) now print the yes/no prompt only when stdin is a
real terminal (ioctl-level check via go-isatty; a char-device stat would
misclassify `< /dev/null`). Non-interactive callers get a clean
structured confirmation_required error on stderr. Piped answers are
still honored for humans/scripts; --yes/--dry-run remain the sanctioned
non-interactive paths.
skills/mono: add the recognition + retry protocol for agents —
identify confirmation_required via error.reason, show action and params,
retry the original command with --yes only after explicit user consent,
never silently append --yes or treat it as a transient error.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Alignment-only changes in runtime_schema.go, schema_contract_model.go,
and devapp_safety_homology_test.go. Remaining make lint findings are in
upstream-owned keychain/transport files untouched by this PR.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
- schemaSafetyFromDecl: drop the now-unreachable nil return; the tier
fill always produces a complete block for a declared Schema
- validateDispatchDecl: panic when ConfirmFirst is set without Risk —
it orders a confirmation that does not exist, and for declared-Schema
writes an empty Risk would silently publish the read safety tier
- RFC: document the boundary that write commands must declare Risk
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Upstream added five reviewed tools (845 total); hashes and counts
refresh. Content of existing tools is unchanged.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Risk (runtime confirmation) and Safety (schema metadata) are two
independent enums composed at embed time: explicit SafetyDecl fields >
CommandSpec.Safety tier > Risk.SafetyDefault(). The tier fill now also
covers idempotency, so an enum-only declaration is self-sufficient and
validateSchemaDecl no longer needs safety completeness checks.
devapp reclassifies its write leaves by reversibility: reversible
mutations declare LeafSafetyWrite (risk high->medium), create/robot
submit/version create declare LeafSafetyHighWrite, and delete/version
publish declare LeafSafetyDestructive (publish effect
write->destructive). Shared hand-written safety constants are deleted.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
- validateSchemaDecl now also requires Safety (effect/risk/confirmation
or the Risk shorthand; Idempotency is declaration-only) and Interface
(mode/availability, plus reason for composite/unavailable), so every
unconditional catalog required key is guaranteed at construction time
- declared dry_run capabilities are indexed by BindEffectiveCommandRegistry
instead of Schema assembly: every process resolving the command tree
gets the reviewed set, removing the hidden "must assemble in-process
first" precondition of the delivery gate
- agent-metadata contract merge now errors when a declared tool has no
canonical CLI projection instead of silently dropping the declaration
Artifacts are byte-identical; full cli/cmdcore/helpers/generator suites pass.
Co-authored-by: Cursor <cursoragent@cursor.com>
- dry_run capabilities declared via cmdcore.SchemaDecl are reviewed by
construction: the Schema pass-through indexes them into the reviewed
capability set, so declared tools no longer need manual entries in
reviewedDryRunCapabilityGroups (31 devapp paths deleted). A conflicting
manual entry for the same canonical is a hard error.
- NewCommand now enforces authoring-time homology for declared commands:
a non-empty Schema without Description/AgentSummary/UseWhen/AvoidWhen/
Examples panics at construction instead of failing later in generated
artifacts or silently drifting from cobra prose.
- --help Example inherits Schema.Selection.Examples when not authored
separately, keeping one authored source for examples.
Catalog and agent metadata artifacts are byte-identical.
Co-authored-by: Cursor <cursoragent@cursor.com>
The agent-metadata generator now merges each registered Contract final
overlay (cmdcore.SchemaDecl) as the top-precedence contract_final
candidate, so declared tools no longer need hint-file rows for
agent_summary/use_when/avoid_when/examples/safety/interface. Selection
eval fixtures and example execution plans synthesize the same assertions
from the declaration, keeping semantic-eval and example coverage intact.
- devapp hint rows deleted from schema_hints/{metadata,selection}/dev.json
(connect_status/connect_stop/search_open_platform_docs_rag kept);
artifact content for all 31 declared leaves is byte-identical, only
provenance now reads contract_final / cmdcore.SchemaDecl
- exact-coverage gates exempt declared tools (hints remain required for
every non-declared command); reviewed-delivery gate accepts
contract_final as the stronger reviewed source
- cmdcore derives effect_source=cmdcore.contract for SchemaDecl-only
safety (read leaves), matching the Risk-shorthand path
Co-authored-by: Cursor <cursoragent@cursor.com>
Every devapp leaf now declares its full final Schema in LeafSpec
(description/safety/interface/selection/dry_run plus Risk, Required
flags and at-least-one Constraints); declaration is the sole final
source, hints no longer shape the published catalog for these tools.
- Hand-written delete/robot submit/robot config migrate to
LeafSpec+RunE with manual cmdcore.ConfirmRisk; robot result becomes
a plain declared leaf. Legacy write guard, runtime_gate annotation
and now-dead helpers are removed; the homology gate is strengthened
to declare-only for the devapp tree.
- New CommandSpec/LeafSpec ConfirmFirst knob reproduces the devapp
guard-first semantics (confirmation_required before parameter
validation) without changing shortcut ordering.
- dry_run is published for all 31 leaves via the reviewed capability
registry (invocation preview, no remote reads).
- Catalog regenerated: dry_run blocks added, unified-app-id/
version-id/member-type/user-ids correctly marked required,
require_one_of constraints published for get/webapp config/security
config, and robot config name corrected to optional (CLI upsert
runtime truth; the remote schema's required was not CLI-accurate).
Co-authored-by: Cursor <cursoragent@cursor.com>
- SchemaDecl on CommandSpec/LeafSpec declares the final ToolSpec payload;
framework converts in-process (no JSON bridge) and Schema assembly
pass-throughs it.
- Assembly fails closed on declared identity mismatched with the bound
entry and on reviewed fields in the declaration payload.
- RFC/homology/AGENTS docs pin declare=final-source, safety precedence
Final > Risk > gate, and light runtime write semantics.
Co-authored-by: Cursor <cursoragent@cursor.com>
Treat EOF/closed stdin as confirmation_required instead of an
interactive decline so agent/CI no longer get exit 0 for writes that
never ran. Align cmdcore.ConfirmRisk the same way.
Co-authored-by: Cursor <cursoragent@cursor.com>
Lift business flags/const params out of Call/PostMount, add typed
flag defaults and policy gates, and realign the RFC acceptance bar to
"no Execute/Call body exists only to assemble params".
Co-authored-by: Cursor <cursoragent@cursor.com>
The single Dispatch hook assumed every command is one MCP call, so the
Shortcut projection could only ever describe a command, never run it. Split
dispatch into Invoke (assembled toolArgs) and Orchestrate (multi-step), keep
RunE as the escape hatch, and reject specs that declare anything other than
exactly one at construction time. Ctx gives both hooks the same typed flag
accessors so orchestration no longer reaches for framework-specific plumbing.
Catalog output stays byte-identical.
Adversarial review confirmed the Phase 1 extraction is a verbatim move (no BLOCKERs) and that cmdcore.BoolFlag is equivalent to both original readers. All fixes below are in the Phase 2 additions.
Correctness: a CommandSpec declaring neither RunE nor Dispatch no longer runs the whole pipeline (write-confirmation prompt included) and silently exits 0 — it now fails with a typed internal error, which also defuses the FromShortcut trap. FromShortcut no longer double-renders the 参数约束 section (shortcutLongHelp already appends it and NewCommand appends ConstraintHelp again): it now maps intent prose only. Flag usage keeps mount()'s flagHelp decoration (必填/可选值) so projected help matches the live shortcut, and the constraint Flags slice is copied instead of aliasing the shortcut registry.
Honesty: the FromShortcut doc block now lists every dropped semantic (Required Changed-vs-effective-value divergence, typed bool/int/slice defaults, Enum, Hidden, Tips to Example, custom constraint, required/enum runtime-schema annotations). cmdcore's package doc no longer claims catalog drift proves runtime behavior — drift covers the build-time projection, unit tests cover the runtime pipeline. leaf.go's stale Phase 1 header updated. Panic messages say command not leaf; doc comments lead with the exported names.
Tests: new mount-equivalence test compares the projected command's flag set/types/usage and rendered Long against live mount(s) — the test that would have caught both bugs above; new root-to-child test exercises the inherited/root-persistent --yes/--dry-run lookup that the leaf-local helper never reached; the nil-dispatch test is inverted to assert the error. docs/architecture.md documents internal/cmdcore and how it differs from internal/cobracmd.
Verified: drift ok (840 tools, identical hashes), make policy pass, skill-command-integrity ok (1033 paths), cmdcore self-coverage 100%, CI-style changed-code coverage 100% (278 statements), full helpers/cmdcore/shortcut suites green.
CI coverage jobs run go test -coverprofile WITHOUT -coverpkg, so each package is measured only by its own tests. cmdcore's logic was exercised only indirectly from internal/helpers, leaving cmdcore self-coverage at 31.5% — failing the CI coverage gate (changed-code 40.5%, overall regression 90.79% to 90.55%) even though the cross-package platform gate reported 100%.
Add direct tests for every cmdcore primitive: flag registration for all four kinds plus hidden aliases/MarkRequired, the explicit-alias-env-default fallback chain incl. Trim/empty skips, integer and slice resolution, required validation, toolArgs assembly incl. Bind/ArgDefault/OmitEmpty/Transform (value, nil-skip, error), constraint declaration panics, constraintProvided (default-not-counted, alias, env, bool, slice main+alias), all three constraint kinds with exact error wording, Risk confirmation (read/--yes/--dry-run/accept/decline), BoolFlag (nil/missing/local/root), schema projection, constraint help, and NewCommand orchestration (order, RunE escape, per-stage abort, decline-cancels, nil dispatch).
cmdcore self-coverage 31.5% to 100%; CI-style changed-code coverage 100%.
Introduce cmdcore.CommandSpec as the single typed leaf definition and
cmdcore.NewCommand as the one orchestration path (flags → constraint decl
checks → Runtime Schema projection → constraint help → PostMount → RunE
escape / generated RunE{required → constraints → Validate → BuildArgs →
ConfirmRisk → Dispatch}). Dispatch becomes a spec property, not a
separate framework.
helpers.NewLeafCommand now delegates to cmdcore.NewCommand(FromLeafSpec),
so every LeafSpec command — including all 27 devapp leaves — flows through
the unified spec. The MCP dispatch (Call / callMCPToolOnServer /
callMCPTool) is captured in the FromLeafSpec closure.
internal/shortcut/adapter.go adds FromShortcut, the typed seam mapping a
Shortcut's shared base (flags of every kind, known constraints, risk,
help identity) into a CommandSpec. It is intentionally NOT wired into the
live mount() path: Shortcut's multi-step Execute, decline-returns-nil
semantics, and Flag.Enum/Hidden/custom-constraint extras are not modeled
by cmdcore yet, so the 376 shipped shortcuts stay byte-identical. Live
wiring is deferred to Phase 3, gated by shortcut-list + schema equivalence.
Commands are provably unaffected: check-generated-drift ok (840 tools,
identical hashes), `dws schema --all` and `shortcut list` unchanged, full
helpers/cmdcore/shortcut suites green, changed-code coverage 100%.
Phase 1 of converging the command frameworks onto one typed base. Extract
LeafSpec's flag registration, alias/env/default effective-value fallback,
required validation, cross-flag constraint declaration checks + runtime
enforcement, Risk-driven write confirmation (--dry-run/--yes/global-flag
aware via a 3-level bool lookup), toolArgs assembly, and Agent Runtime
Schema projection into a new dispatch-agnostic internal/cmdcore package.
internal/helpers/leaf.go now keeps only the LeafSpec shell (with MCP
dispatch fields) and NewLeafCommand orchestration; LeafFlag/LeafFlagKind/
LeafConstraint/LeafConstraintKind/LeafRisk and their constants become
aliases to cmdcore types, so all 27 devapp call sites compile unchanged.
Dispatch (callMCPTool/OnServer/Call) stays in helpers.
Pure extraction, zero behavior change: catalog is byte-identical
(check-generated-drift ok), the leaf unit + risk/constraint tests pass,
and changed-code coverage is 100% (224 statements across both packages).
Only the leaf framework code is touched; Shortcut delegation is deferred
to Phase 2/3.
Close the last capability gap versus the shortcut framework: LeafSpec now
carries a Risk field (read / write / high-risk-write) and enforces the
same pre-dispatch write confirmation as shortcut's confirmRisk. Read (and
empty) risk never prompts; write/high-risk-write prompt unless --yes or
--dry-run, cancelling without dispatch on decline. Prompt wording matches
the shortcut runner verbatim (command path substitutes Service+Command)
so atomic commands and smart shortcuts confirm identically. --yes is read
robustly across local/inherited/root-persistent flags.
Converge the atomic LeafSpec framework toward the shortcut framework's
constraint system so both share one flag-registration + validation base,
differing only in dispatch path (single-step MCP vs multi-step
orchestration).
- Add LeafBool / LeafStringSlice flag kinds (registration, effective-value
detection, required semantics, toolArgs assembly: bool delivers on
Changed incl. explicit false; slice trims and drops empty elements).
- Add LeafConstraint (at_least_one / exactly_one / mutually_exclusive) on
LeafSpec. The framework validates them between required checks and the
Validate hook, with error wording identical to the shortcut runner's
RuntimeContext validators; "provided" reuses LeafSpec's alias/env
fallback chain (registration defaults do not count), which the
shortcut framework's bare Changed check lacks.
- Project constraints to the Agent Runtime Schema (exactly_one =
require_one_of + mutually_exclusive) and render a 参数约束 help section,
matching shortcut leaf help. Declaration errors panic at build time.
玉澜