@xdxer/dingtalk-agent 0.1.5-beta.14 → 0.1.5-beta.15

package/CHANGELOG.md

@@ -2,6 +2,21 @@
 
 All notable changes to `@xdxer/dingtalk-agent` are documented here. The project follows semantic versioning; prerelease entries describe release candidates and do not imply that Live canaries or registry publication have completed.
 
+## 0.1.5-beta.15 - 2026-07-25
+
+Starting a general-purpose office Agent from an empty workspace now actually reaches the model. Four deterministic false blocks stood between a clean checkout and an L1 evaluation, and all four had the same shape: the source project's declaration was being judged as if it were the sandbox's final state, or the grader was stricter than the contract it enforced. Ships on the `beta` dist-tag; `latest` stays at 0.1.4. Merged as PR #68.
+
+### Fixed
+
+- **A clean workspace was rejected before the runner could materialize the Basic exposure (closes #67).** A source `opencode.json` may declare the managed path `.agents/skills/dta-basic-behavior/SKILL.md`, which the eval runner materializes inside an isolated sandbox — committing it back would create a second Skill source. The old check tested for that file in the *source* project, so a clean project failed before any model call. Exactly one runner-managed Basic instruction is now allowed to be absent, and only while it does not exist; globs, absolute paths, `..`, external URLs and out-of-workspace targets still fail closed, and after materialization the sandbox is still verified through `verifyRequiredBasicSkill` and the full tree hash.
+- **The default model ID had been retired by the provider.** `deepseek/deepseek-chat` is refused with a 400, which meant `lab eval`, `agent audit --verify-load` and the workspace generated by `agent enhance` could all fail before any behaviour was measured. The default now lives in one constant (`DEFAULT_OPENCODE_MODEL`, currently `deepseek/deepseek-v4-flash`) that eval, audit, enhance and the fixtures all read.
+- **`agent enhance` silently chose OpenCode as the Host.** The Compose Skill says not to pick a Host for the user, but every new project's plan contained `opencode.json` and a `local-dev` workspace regardless. `--host opencode|none` is now required: `opencode` declares the local workspace and merges `opencode.json`, `none` produces a harness-agnostic package with no workspaces and an audit ceiling of `partial`. Dry runs remain side-effect free for local files, DWS and Triggers (compose Skill 0.17.0).
+- **`--cases basic-skill-load` reported that an existing case did not exist.** Case selection filtered out every `load-probe` and then tested the requested ids against what was left. Existence is now checked against the full case set, selecting only load-probes enters load-only mode with its dynamic UUID challenge, and mixing a load-probe with behaviour cases still runs the load gate while executing just the selected behaviours.
+- **A single sentence of preamble failed the reference probe.** The model read the right file in one full read, with the right tool path and permission scope, and echoed both random UUIDs correctly — but `answer.trim() === expected` failed it for saying "let me read that file first". Grading is now a terminal-single-probe contract: the last line must equal the random probe exactly and the whole answer must contain that token exactly once. A second probe token still fails closed, the definition probe keeps strict equality, and nothing about tool paths, full reads, permission scope or raw traces was relaxed.
+- **The load report conflated two different baseline failures.** "Did not guess the random challenge" is a security property; "emitted the sentinel exactly" is protocol compliance. When a baseline run answers with a pseudo tool tag it satisfies the first and not the second, yet both report fields computed the same expression. Two facets (`challengeNotGuessed`, `sentinelExact`) now show them separately. The gate is unchanged — a baseline still passes only on an exact sentinel match — so no assertion moved.
+
+Model-stability findings from the same session (multi-run variance on three basic office cases; baselines occasionally emitting pseudo tool tags) are kept as open questions with their raw evidence rather than folded into production rules: a single batch does not prove a Skill's gain, and a failure is not turned into a success by relabelling it.
+
 ## 0.1.5-beta.14 - 2026-07-25
 
 The packaged docs stop lying to the person following them, and the project says one thing about itself instead of three. Three of the errors were on the first-hour path and shipped inside the tarball, so a fix that only landed on GitHub would not have reached anyone. Ships on the `beta` dist-tag; `latest` stays at 0.1.4. Merged as PRs #63 and #64.

package/README.md

@@ -66,9 +66,10 @@ dta agent audit --json    # 预期 partial：岗位语义还是模板，这是
 
 ```bash
 # 1. 创建零写入计划，再显式落盘
-dta agent enhance --project-name release-agent --role-skill release-manager --dry-run --json
 dta agent enhance --project-name release-agent --role-skill release-manager \
-  --plan-id <plan-id> --yes --json
+  --host opencode --dry-run --json
+dta agent enhance --project-name release-agent --role-skill release-manager \
+  --host opencode --plan-id <plan-id> --yes --json
 
 # 2. 审计、运行与评测
 dta agent audit --bindings agent.bindings.json --require-skill release-manager --json

package/dist/bin/dingtalk-agent.js

@@ -128,8 +128,8 @@ dingtalk-agent 开发者装配验收
   agent audit [--bindings agent.bindings.json] [--require-skill <name>]
   agent audit [--bindings <file>] [--require-skill <name>] --verify-load --yes
   agent audit [--bindings <file>] --load-report <agent-audit-load-evidence.json> [--remote-report <report.json>]
-  agent enhance --role-skill <name>[,<name>] [--project-name <name>] [--dry-run] [--json]
-  agent enhance --role-skill <name>[,<name>] --plan-id <current-plan-id> --yes [--json]
+  agent enhance --role-skill <name>[,<name>] --host <opencode|none> [--project-name <name>] [--dry-run] [--json]
+  agent enhance --role-skill <name>[,<name>] --host <opencode|none> --plan-id <current-plan-id> --yes [--json]
 
 默认只做本地静态审计，不访问 DWS、不调用模型；状态为 partial 时返回退出码 2。
 --verify-load --yes 会运行隔离的 OpenCode load probe，但仍不访问钉钉。
@@ -1166,7 +1166,7 @@ async function main() {
                 args: argv.slice(2),
                 options: {
                     bindings: { type: 'string' }, 'project-name': { type: 'string' },
-                    'role-skill': { type: 'string' }, model: { type: 'string' },
+                    'role-skill': { type: 'string' }, host: { type: 'string' }, model: { type: 'string' },
                     memory: { type: 'string' }, knowledge: { type: 'string' },
                     artifacts: { type: 'string' }, 'state-dir': { type: 'string' },
                     profile: { type: 'string' }, 'expected-user-id': { type: 'string' },
@@ -1180,7 +1180,8 @@ async function main() {
                 fail('agent enhance plan 不接受 --plan-id；它由 dry-run 生成');
             const options = {
                 bindings: o.bindings, projectName: o['project-name'],
-                roleSkills: o['role-skill']?.split(','), model: o.model,
+                roleSkills: o['role-skill']?.split(','),
+                host: o.host, model: o.model,
                 memory: o.memory, knowledge: o.knowledge, artifacts: o.artifacts,
                 stateDir: o['state-dir'], profile: o.profile,
                 expectedUserId: o['expected-user-id'], cliVersion: PACKAGE_JSON.version,
@@ -1886,7 +1887,7 @@ function printAgentAudit(out) {
 function printAgentEnhance(out) {
     if (out.$schema === 'dingtalk-agent/agent-enhancement-plan@1') {
         console.log(`${strong('dingtalk-agent agent enhance')} · ${mark(out.ready ? 'pass' : 'fail', out.ready ? 'READY PLAN' : 'BLOCKED')}`);
-        console.log(`   planId=${out.planId}  project=${out.projectName}`);
+        console.log(`   planId=${out.planId}  project=${out.projectName}  host=${out.host}`);
         console.log(`   sideEffect=${out.sideEffect}  dingtalkSideEffect=${out.dingtalkSideEffect}`);
         for (const operation of out.operations) {
             console.log(`   ${operation.kind.toUpperCase()} ${operation.path}: ${operation.reason}`);
@@ -1901,7 +1902,7 @@ function printAgentEnhance(out) {
         return;
     }
     console.log(`${strong('dingtalk-agent agent enhance')} · ${mark('pass', 'APPLIED')}`);
-    console.log(`   operation=${out.operationId}  planId=${out.planId}`);
+    console.log(`   operation=${out.operationId}  planId=${out.planId}  host=${out.host}`);
     console.log(`   changes=${out.changes.length}  idempotent=${out.idempotent}`);
     if (out.backupRoot)
         console.log(`   backup=${out.backupRoot}`);