@mstar-harness/opencode 2.1.1 → 2.3.0

package/CHANGELOG.md

@@ -6,6 +6,29 @@ The monorepo root [CHANGELOG.md](../../CHANGELOG.md) summarizes cross-surface re
 
 ## [Unreleased]
 
+## [2.3.0] - 2026-08-16
+
+### Harness
+
+- Internal `@mstar-harness/engine` devDependencies in cli/opencode/dsh now use the `workspace:*` protocol; the release-prep engine-spec sync step was removed.
+
+- Version alignment with harness **2.3.0** (no OpenCode package API change).
+
+See root [CHANGELOG.md](../../CHANGELOG.md) **2.3.0**.
+
+## [2.2.0] - 2026-08-13
+
+### Bundled harness skills (`harness-skills/` at publish)
+
+- Added the **dsh host reference** to `mstar-host`: a detect-table row for dsh's `subagent` delegation tool and `references/dsh.md` (tool map, in-process gates/enforcement, bundled commands, PM dispatch, gotchas).
+- **Sync upstream v2.1.1**: merged the upstream `mstar-harness` v2.1.1 line into the dev-dsh branch — adds the `code-reviewer` role (read-only L2 SDD task reviewer / audit executor; replaces `generalPurpose` as the SDD per-task review seat, with generic fallback only when the role agent is absent on the host), ships the canonical default-ignore harness `.gitignore` format (`.mstar/**` + tracked re-includes `AGENTS.md` / `knowledge/` / `specs/`) across the engine, CLI `init` fence and bundled skills, and aligns all 11 version surfaces to 2.1.1.
+- **engine**: `emitGitignoreSnippet` / `validateGitignore` / `HARNESS_PROCESS_GITIGNORE` now emit the default-ignore + re-include format instead of the flat per-directory ignore list; `ROLE_MAPPING` grows to 14 ids with `code-reviewer`.
+- **bundle-assets**: re-synced `packages/dsh/harness-skills` / `harness-commands` from the merged `skills/` tree — the 6+ upstream-touched bundled skills and all `mstar-host/references/*.md` host adapters (cursor/kimi/omp/opencode/zcode) now carry the v2.1.1 wording (SDD task reviewer → `code-reviewer`).
+
+- Version alignment with harness **2.2.0** (no OpenCode package API change).
+
+See root [CHANGELOG.md](../../CHANGELOG.md) **2.2.0**.
+
 ## [2.1.1] - 2026-08-12
 
 ### Harness

package/dist/mstar.js

@@ -1,6 +1,7 @@
 // ../engine/dist/engine.js
 import { existsSync, mkdirSync, readFileSync, renameSync, unlinkSync, writeFileSync } from "node:fs";
 import { mkdirSync as mkdirSync2, readdirSync, readFileSync as readFileSync2, statSync } from "node:fs";
+import { execFileSync } from "node:child_process";
 import { basename as basename2, dirname as dirname2, isAbsolute, join as join2, relative, resolve as resolve2 } from "node:path";
 import { existsSync as existsSync2, readFileSync as readFileSync3, readdirSync as readdirSync2 } from "node:fs";
 import { join as join4, resolve as resolve4 } from "node:path";
@@ -26,18 +27,45 @@ function resolveHarnessDir(startDir = process.cwd(), opts = {}) {
   const explicit = opts.harnessDir ?? process.env.MSTAR_HARNESS_DIR;
   if (explicit)
     return resolve2(start, explicit);
+  const boundary = resolve2(start, opts.workspaceRoot ?? defaultWorkspaceRoot(start));
   let dir = start;
   for (;; ) {
+    if (!isAtOrBelow(dir, boundary))
+      return null;
     for (const candidate of [join2(dir, ".mstar"), join2(dir, ".agents"), join2(dir, ".plans"), join2(dir, "plans")]) {
       if (isDirectory(candidate))
         return candidate;
     }
+    if (dir === boundary)
+      return null;
     const parent = dirname2(dir);
     if (parent === dir)
       return null;
     dir = parent;
   }
 }
+function defaultWorkspaceRoot(startDir) {
+  try {
+    const cdup = execFileSync("git", ["rev-parse", "--show-cdup"], {
+      cwd: startDir,
+      encoding: "utf8",
+      stdio: ["ignore", "pipe", "ignore"]
+    }).trim();
+    if (!cdup)
+      return startDir;
+    let boundary = startDir;
+    for (const segment of cdup.split(/[\\/]/)) {
+      if (segment && segment !== ".")
+        boundary = dirname2(boundary);
+    }
+    return resolve2(boundary);
+  } catch {}
+  return startDir;
+}
+function isAtOrBelow(dir, root) {
+  const rel = relative(root, dir);
+  return rel === "" || !rel.startsWith("..") && !isAbsolute(rel);
+}
 function resolveIterationDir(harnessDir) {
   return join2(resolve2(harnessDir), "iterations");
 }
@@ -167,7 +195,7 @@ function parseAssignmentBranchForms(assignmentText) {
   if (fields.branchPolicy !== undefined && fields.branchPolicy !== "") {
     const direct = fields.branchPolicy.match(/^direct\s+on\s+(\S+)/i);
     if (direct) {
-      const strict = fields.branchPolicy.match(/^direct\s+on\s+(\S+)(?:\s*(?:[—–]|--|-)\s*(.+))?$/);
+      const strict = fields.branchPolicy.match(/^direct\s+on\s+(\S+)(?:\s*(?:[\u2014\u2013]|--|-)\s*(.+))?$/);
       forms.directOn = { branch: direct[1].trim(), reason: strict ? (strict[2] ?? "").trim() : "" };
     }
   }

package/harness-commands/codebase-audit.md

@@ -2,6 +2,7 @@
 name: codebase-audit
 description: Survey a codebase as a senior advisor and produce prioritized, self-contained improvement plans. Read-only on source code. Use standalone before iteration-start to discover what's worth doing, or independently to build a prioritized backlog.
 agent: project-manager
+input: "[simplify]"
 ---
 
 # Audit Codebase
@@ -32,7 +33,7 @@ The audit is **advisory** — it does not enter the per-plan state machine (`Tod
 
 ## Execute
 
-Execute **`mstar-audit`** end to end（SKILL.md：Recon → Audit → Vet & prioritize → Write plans；effort `quick` / `standard` / `deep`；scope variants `security` / `perf` / `tests` / `branch` / `next` / `roadmap`）。Plans → `{PLAN_DIR}/audit-<YYYY-MM-DD>/NNN-<slug>.md` + `README.md` index，per **`mstar-plan-artifacts/references/plan-quality-bar.md`**。
+Execute **`mstar-audit`** end to end（SKILL.md：Recon → Audit → Vet & prioritize → Write plans；effort `quick` / `standard` / `deep`；scope variants `security` / `perf` / `tests` / `branch` / `next` / `roadmap` / `simplify`）。Plans → `{PLAN_DIR}/audit-<YYYY-MM-DD>/NNN-<slug>.md` + `README.md` index，per **`mstar-plan-artifacts/references/plan-quality-bar.md`**。
 
 Executor: PM dispatches `@code-reviewer`；大型仓库 scout 扇出经 Assignment `Delegation: allowed (scout/explore only, read-only)`（Routing 表）。
 

package/harness-commands/iteration-drive.md

@@ -2,6 +2,7 @@
 name: iteration-drive
 description: Drive the active iteration to completion — Phase 2 Autonomous Execute, Phase 3 iteration-close, Phase 4 Create PR, Phase 5 PR merge-ready loop (prefer babysit/*-babysit; optional greploop when repo has it; else CI fallback) until mergeable. Not Done until Phase 5 exit checklist passes.
 agent: project-manager
+input: "[no args]"
 ---
 
 # Drive Iteration
@@ -78,7 +79,7 @@ command -v mstar-harness >/dev/null 2>&1 && mstar-harness dispatch validate "<la
 if command -v mstar-harness >/dev/null 2>&1; then mstar-harness dispatch validate "<latest-assignment-file>" || exit 1; fi
 ```
 
-> 路径必须加引号且替换为具体文件（如最新 `{SDD_DIR}/task-N-brief.md`，勿留尖括号）——agent 代入的路径不得进入 shell 无引号展开（qc2 W-2）。
+> 路径必须加引号且替换为具体文件（如最新 `{SDD_DIR}/task-N-brief.md`，勿留尖括号）——agent 代入的路径不得进入 shell 无引号展开。
 
 ## Phase 3: iteration-close
 

package/harness-commands/iteration-loop.md

@@ -2,6 +2,7 @@
 name: iteration-loop
 description: "Autonomous full iteration loop for cloud agents — Phase 1 (code-first auto direction lock + compass/plans + Review & Edit chain) through Phase 2–5 (execute → close → PR → merge-ready). Optional args: direction, scale (S|M|L|XL, default M). Not Done until Phase 5 exit checklist passes. Minimal human intervention; no grill-me."
 agent: project-manager
+input: "[direction] [scale]"
 ---
 
 # Iteration Loop
@@ -129,7 +130,7 @@ command -v mstar-harness >/dev/null 2>&1 && mstar-harness dispatch validate "<la
 if command -v mstar-harness >/dev/null 2>&1; then mstar-harness dispatch validate "<latest-assignment-file>" || exit 1; fi
 ```
 
-> 路径必须加引号且替换为具体文件（如最新 `{SDD_DIR}/task-N-brief.md`，勿留尖括号）——agent 代入的路径不得进入 shell 无引号展开（qc2 W-2）。
+> 路径必须加引号且替换为具体文件（如最新 `{SDD_DIR}/task-N-brief.md`，勿留尖括号）——agent 代入的路径不得进入 shell 无引号展开。
 
 **Loop 特有**：Phase 5 push cadence（HARD）→ **`mstar-iteration` §5.1a**；exit checklist → **`mstar-iteration` §5.2**（`references/phase-4-5-pr-delivery.md` §5.2）。
 

package/harness-commands/iteration-start.md

@@ -2,6 +2,7 @@
 name: iteration-start
 description: "Start a new harness iteration — optional direction hint, research, grill-me, compass/plans, Review & Edit chain (long-lived {SPECS_DIR}/ + {ITERATION_DIR}/<id>/ package; compound promotes package at close only), PM lock, integration branch; then auto-continue Phase 2→5 (execute → close → PR → merge-ready) unless `pause` arg given."
 agent: project-manager
+input: "[direction] [pause]"
 ---
 
 # Start Iteration
@@ -111,7 +112,7 @@ command -v mstar-harness >/dev/null 2>&1 && mstar-harness dispatch validate "<la
 if command -v mstar-harness >/dev/null 2>&1; then mstar-harness dispatch validate "<latest-assignment-file>" || exit 1; fi
 ```
 
-> 路径必须加引号且替换为具体文件（如最新 `{SDD_DIR}/task-N-brief.md`，勿留尖括号）——agent 代入的路径不得进入 shell 无引号展开（qc2 W-2）。
+> 路径必须加引号且替换为具体文件（如最新 `{SDD_DIR}/task-N-brief.md`，勿留尖括号）——agent 代入的路径不得进入 shell 无引号展开。
 
 **Prepare gate (per plan in compass)**:
 

package/harness-skills/mstar-audit/SKILL.md

@@ -70,6 +70,21 @@ Every finding follows **`references/finding-format.md`** — read it before the
 
 ### Phase 3 — Vet, prioritize, confirm
 
+**Attack before vet — claims must survive an adversarial pass first.** Take the top candidate findings (by leverage; scale the count to finding volume — attack the whole table when small, the head when large) and run a three-way attack on each:
+
+1. **Counter-example** — find a boundary case that makes the claim not hold.
+2. **Simpler explanation** — does a simpler explanation cover the same evidence?
+3. **Evidence verifiability** — open the cited `file:line` and check it actually supports the claim.
+
+Dispose per the five-state rule (single-pass version — the four dispositions below implement the survey's five-state semantics: uncovered-keep == 未提及保留; never-drop == 全空/null 回流):
+
+- **Survived** — passes to vet unchanged.
+- **Refuted** — drop, and record in the index's "considered and rejected" section: `- <finding>: not worth doing because <one line>`.
+- **Hallucinated** — the attack surfaced a claim never in the original finding set: discard it and log a red-team record line in the index (never into the findings table; it does not occupy a "considered and rejected" slot — it was never a finding). Disambiguation: hallucinated = a claim the attack itself produced; fabricated/unsupported evidence inside the original finding goes to **Refuted** via the evidence-verifiability axis, not to Hallucinated.
+- **Uncovered** — the attack did not reach a finding: treat as unreviewed and keep for vet. Never drop a finding just because the attack missed it.
+
+Where the attack step decides whether a claim stands on its face, vet below confirms the code itself — opening cited files and disposing by-design / mis-attribution / duplicate cases; survived findings hand to vet below.
+
 **Vet before presenting — subagents over-report.** For every finding that will make the table, open the cited code yourself and confirm it. Three failure classes to expect:
 
 1. **By-design behavior** reported as a bug or vulnerability (e.g. honoring `https_proxy` flagged as SSRF — standard proxy convention; or a tradeoff explicitly recorded in an ADR).
@@ -112,6 +127,7 @@ If an audit directory from a previous run exists, **reconcile, don't duplicate**
 | Category focus (`security`, `perf`, `tests`, ...) | Recon, then that category only, then plan | Useful for targeted sweeps |
 | `branch` | Current branch changes only | Files changed since merge-base with default branch + their direct importers. Tag every finding `introduced` or `pre-existing` |
 | `next` / `roadmap` | Direction category only, in depth | 4–6 grounded suggestions; selected ones become design/spike plans |
+| `simplify` | DEBT-focused deep pass: dead / duplicated / speculative / over-built / added-then-removed / hand-rolled-where-a-dependency-exists surfaces | Prove-or-reject per playbook §5; findings use Category DEBT; tiny-real items → "considered and rejected" rows, never inline TODOs (Hard Rule 1) |
 
 ## Output format
 
@@ -138,6 +154,10 @@ If an audit directory from a previous run exists, **reconcile, don't duplicate**
 ## Findings considered and rejected
 
 - <finding>: not worth doing because <one line>.
+
+## Red-team dispositions
+
+- <finding>: <survived / refuted / hallucination-dropped / uncovered-kept>, <one-line reason>
 ```
 
 Status values: `TODO` | `IN PROGRESS` | `DONE` | `BLOCKED` | `REJECTED`

package/harness-skills/mstar-audit/references/audit-playbook.md

@@ -18,6 +18,8 @@ The highest-trust category — real bugs found by reading, not speculation.
 - Concurrency: check-then-act on shared resources, missing transactions around multi-write operations, idempotency of retried operations (webhooks, queues).
 - Type escape hatches: `any` / `as` casts / `@ts-ignore` clusters — each one is a place the compiler was overruled.
 - Resource leaks: unclosed handles, connections, subscriptions; missing `finally`.
+- Derived-state drift: every cache, replay, projection, denormalized copy, or UI echo must trace to an authoritative source and an invalidation point; flag retained state with neither.
+- Bounds covering the final operation: who owns the complete emitted/retained result (wrappers and metadata included)? Probe tiny/exact limits, oversized single chunks, and multibyte text against byte limits.
 
 ## 2. Security
 
@@ -34,6 +36,7 @@ Review only what is directly supported by code evidence. Keep findings framed as
 - Dependency posture: run the ecosystem's audit command (`npm audit`, `pip-audit`, `cargo audit`) in read-only mode. Report only critical/high advisories that affect reachable runtime code.
 - Production configuration: overly broad CORS where credentials are allowed, missing response-hardening headers (e.g. CSP), cookies missing appropriate `HttpOnly`/`Secure`/`SameSite` attributes, debug/verbose behavior enabled in production.
 - Data minimization: PII or sensitive operational data in logs, stack traces returned to clients, internal error details exposed through API responses.
+- Enforcement bypass: for every validation/rejection point, look for alternate callers that route around it — direct calls, wrappers, facades, schema-less paths, listener ordering.
 
 ## 3. Performance
 
@@ -56,6 +59,9 @@ The goal is not a percentage — it's *which untested code is dangerous*.
 - Existing test quality: tests that assert nothing meaningful, heavy mocking that tests the mocks, snapshot tests nobody reads, flaky patterns (real timers, real network, order dependence).
 - Missing test layers: unit-only suites with zero integration coverage on API boundaries, or the inverse.
 - Verification infrastructure: is there a one-command way to know the codebase works? If not, that's finding #1 and a prerequisite plan for any risky change.
+- Real entry path: do tests exercise the shipped entry (CLI, loader, plugin boot) rather than a hand-mounted equivalent?
+- Externally observable state: assertions verify logs, events, files, exit codes — never implementation restatement or agent-reported success.
+- User-visible output is behavior (conditional): in repos shipping UI copy, CLI output, API error shapes, or prompt text, wording is behavior — snapshot or e2e coverage should pin it.
 
 ## 5. Tech Debt & Architecture
 
@@ -65,6 +71,27 @@ The goal is not a percentage — it's *which untested code is dangerous*.
 - God objects/modules: files an order of magnitude larger than the repo median that everything touches; functions with double-digit parameters or deep conditional nesting.
 - Inconsistent patterns: three ways of doing data fetching / error handling / styling — pick the winner (the one the team converged on most recently) and plan the consolidation.
 - Abstraction mismatches: premature abstractions with a single implementation, or missing abstractions where the same change always requires touching N files in lockstep.
+- Public-but-one-caller: a public method on a generic service with a single internal caller is a private-capability closure candidate.
+- Unjustified defaults/public options: flag defaults or public operations/formats with no current-consumer evidence or prior art.
+
+### Prove-or-reject before reporting DEBT (`simplify` scope)
+
+**Prove before reporting dead code.** Classify consumers first — production corpus / tests-docs-only / ambiguous (examples, scripts: inspect, don't assume). Grep the exact symbol, plus event/field/config names, both quoted and bare. Read the call sites — a grep hit is a lead, not a verdict. "Tests are the only consumer" is a finding-enabler when the pinned behavior is non-load-bearing; "an invariant/test existing only to protect an unused API" is itself the signal.
+
+**Hand-rolled vs dependency swap bar.** Name the exact surface the package covers — residual semantics count against the swap. Health-check the dependency honestly. A recorded decision (ADR/knowledge doc) beats the swap claim — re-litigating a settled tradeoff needs new evidence. Weigh net deletion: implementation + dedicated tests + docs − remaining glue. A wrapper that relocates the same complexity is not a win.
+
+**Mirrored-fact test.** When several mechanisms track the same liveness/settlement fact, propose one controller — but preserve machinery protecting publication, rollback, callback containment, and first-terminal arbitration.
+
+**Strong-candidate families** (one line per family):
+- Symbols with no production consumer.
+- Tests/docs-only consumers pinning non-load-bearing behavior.
+- Two representations mirroring the same fact.
+- Seam methods fully implemented for support but with zero consumption.
+- Speculative product generality.
+- Invariants/rollbacks/expected-outputs existing only to protect an unused API.
+- Hand-rolled where a dependency exists.
+
+**Guards.** A production caller exists → feature decision, not cleanup (reject). A recorded seam/ADR rationale → new evidence must beat it. Tiny-but-real items → "considered and rejected" rows in the index, never inline TODOs (Hard Rule 1).
 
 ## 6. Dependencies & Migrations
 

package/harness-skills/mstar-branch-worktree/SKILL.md

@@ -150,7 +150,7 @@ Default process artifacts (`plans/`, `iterations/`, `status.json`, `sdd/`, `note
 **Naming conventions (PM / ops; examples only — paths MUST be canonical absolute)**
 
 1. **Control worktree** — usually the primary checkout or a PM-designated path on `spec_integration_branch`; record once in `metadata.control_worktree_path`.
-2. **Feature worktree (per plan)** — one distinct sibling directory per active `plan_id`, e.g. `<repo-parent>/worktrees/<plan-id>` or team `.worktrees/<plan-id>`; Assignment **`Worktree path`** must match lease `worktree_path`.
+2. **Feature worktree (per plan)** — one distinct subdirectory under the workspace root **`.worktrees/`** per active `plan_id` (e.g. `.worktrees/<plan-id>-<slug>`; AGENTS.md「Local scratch layout」), gitignored by the repo convention; Assignment **`Worktree path`** must match lease `worktree_path`.
 3. **L2 track worktrees (within-plan)** — additional distinct directories per parallel implement track under the **same** plan (see **`references/parallel-writable-pre-dispatch.md`**), each with its own PM-approved **`Working branch`**.
 
 > **Engine check (when available):** run `mstar worktree check <plan-id>` (L1) / `mstar worktree check --l2 --tracks <json>` (L2) (or `import { l1PreDispatchCheck, l2PreDispatchCheck, assertControlVsFeaturePath, assertBranchAlignment } from "@mstar-harness/engine"` in a host hook) to verify the L1/L2 isolation rules above (lease worktree ≠ control path; checked-out branch matches `Working branch`). On `fail` -> do not proceed; fix and re-run. Skill text below remains authoritative when the runtime is absent.
@@ -217,3 +217,10 @@ Default process artifacts (`plans/`, `iterations/`, `status.json`, `sdd/`, `note
 - 若 **QA 与同仓其他可写角色并发**提交测试代码，仍须遵守上文「同仓并发写入」**worktree** 规则（可为 QA 单开一条写入 worktree，**同一 `Working branch`**，由 PM 在 Assignment 写明）。
 
 派发前清单与常见反模式 → **`references/parallel-writable-pre-dispatch.md`**。
+
+## History rewrite 与推送安全
+
+- 已推送分支的任何 history rewrite：先 `git fetch` 记录远端**精确 OID**，发布用 `--force-with-lease=<branch>:<observed-oid>`；**禁止**裸 `--force`。
+- Rewrite 推送后：重新 fetch heads；rewrite 前的 review threads / approvals / check 结果**不再是当前证据** — merge 结论前须重审（commit hash 与 inline-comment anchor 已失效）。
+- 证据最窄原则（audit / QA Assignment 场景）：选择会在目标回归上失败的**最窄**检查；不因「push 在即」重跑已通过的检查。
+- 本节只管 rewrite / lease / 证据失效面；CI / review 波次 push 门禁（时序）SSOT → `mstar-iteration` §5.1a。

package/harness-skills/mstar-coding-behavior/SKILL.md

@@ -92,6 +92,11 @@ When something does not work, investigate; do not guess.
 - **Run existing tests before and after changes.** If they passed before and fail after, you broke something. If they were already failing, say so.
 - **If stuck, say so.** "I tried X and Y; neither worked. I'm seeing Z. I think it might be W but am not sure" is infinitely more useful than silently trying random things for 20 iterations.
 
+**Dataflow-directed diagnosis — how to locate, verify, and falsify** (for where to fix, see Surgical Changes · bug=root-cause; for the pre-fix repro test, see the bullets above):
+- **Map the data flow before judging.** Trace input → processing → storage → output, noting who writes and who reads at each step. A bug is a state deviation from expectation at some point in that flow (expected state vs observed state) — locate it from the data, not by static code reading.
+- **Four verifiable cross-checks.** Every hypothesis must be cross-checkable: re-run the repro / log comparison / input-output comparison / dual-path comparison. A hypothesis that cannot be verified on the spot is not a conclusion.
+- **Falsify the fix.** After the fix, re-run the original repro and compare output with expectation. If the problem did not disappear, the root cause was wrong — report "verification failed" explicitly; never pretend success.
+
 ## 5) Goal-Driven Execution
 
 Convert vague requests into verifiable outcomes and iterate until verified. Define concrete success criteria before major edits; use brief `Step -> verify` checkpoints for multi-step tasks; for split delivery, maintain a durable roadmap (current slice, later slices, dependencies, owner/trigger, completion condition); prefer evidence-backed completion (tests, command output, reproducible checks). If verification fails, loop on diagnosis and fix before declaring completion. Do not finish with "next plan / later / follow-up" only in prose — remaining work must be written to the plan/status artifact or the task reports `Partial` / `Blocked`.

package/harness-skills/mstar-compound/references/compound-workflow.md

@@ -113,6 +113,22 @@ After writing, validate the YAML frontmatter:
 - `date` is YYYY-MM-DD format
 - Array fields use `[item1, item2]` syntax
 
+### 3.5 知识文档质量门：HEAD-resolvability & prose hygiene
+
+写作完成、进入 Phase 4 前，对文档运行本质量门（writing-specialist 编辑 durable harness artifacts 时适用同一 rubric — 见 `mstar-roles/references/writing-specialist.md` Output Guidance 指针）：
+
+- **The one test**：读者在 HEAD（无 chat transcript / dispatch prompt / 未合并草稿访问权）能否解析每个引用并验证每个声明？
+- **mstar 泄漏分类**（vantage 来自 authoring session 而非仓库现状；修剪或重述）：
+  - dead session citations：chat-only decision 序号 → 引 plan/knowledge/roadmap 属主路径，或独立重述事实
+  - durable docs 中的 change narration：knowledge/roadmap/README 陈述现状；已修回归 → counterfactual-present（「without X, Y happens」）
+  - review choreography：谁在哪轮确认 → 平实事实；finding id 留在 review bundle（其 sanctioned genre）
+  - reviewer-addressed justification（向评审自辩的措辞）
+  - control-flow narration（过程/控制流叙述）
+  - hedges without markers → 既有 `simplify:` / `temporary` 标记约定（`mstar-coding-behavior`；引用不重复）
+  - authoring-language slips：双语对之外的 zh/en 混杂
+- **Keep 规则（mstar-sanctioned）**：review bundle / QC report 内的 R# 与 finding id（durable 文档引用 QC 证据的 sanctioned pattern 是经 review bundle 路径引用：finding id 留在 bundle 内，durable 文档只引用 bundle 路径本身）；issue 引用；带 provenance 词的 measured bounds；runtime old/new 生命周期态；Alternatives-considered genre 节；HEAD 持久工件中的 iteration/plan id
+- **过度修正陷阱**：义务↛背书翻转；hypothetical 保持标记；共享一句时删子句不删句。「修剪前枚举命题」→ complete-proposition rule（`mstar-roles/references/writing-specialist.md` Output Guidance — SSOT，不在本节重述）
+
 ## Phase 4: Discoverability check
 
 Check whether the project's `AGENTS.md` or `CLAUDE.md` would lead a future agent to discover `{KNOWLEDGE_DIR}`.

package/harness-skills/mstar-compound-refresh/SKILL.md

@@ -50,6 +50,9 @@ For each candidate document, classify into one of five outcomes:
 | **Replace** | Old doc is misleading; known better replacement exists | Create trustworthy successor, then delete old |
 | **Delete** | No longer useful, applicable, or distinct | Delete — git history preserves it |
 
+> **Guardrail — rejected approaches:** a captured rejected approach stays only while the losing idea remains a tempting, meaningful mistake; delete it once obsolete or no longer plausible.
+> **Calibration:** first real run of this axis — 9 size-divergent knowledge docs (36–78 lines) under `{KNOWLEDGE_DIR}/architecture-patterns/` classified at fold time, all kept by future value (evidence: `.mstar/plans/20260815-mstar-knowledge-prose-folds.md` `## Review Gate Summary` → Task evidence); length proved not to be a criterion.
+
 ## Core rules
 
 1. **Evidence over opinion.** Signals are inputs, not a scorecard. Use engineering judgment.
@@ -58,6 +61,7 @@ For each candidate document, classify into one of five outcomes:
 4. **Be decisive.** When evidence is clear (file renamed, class moved), apply. Only ask PM when genuinely ambiguous.
 5. **Avoid low-value churn.** Don't edit for typos, polish, or cosmetic changes that don't improve accuracy.
 6. **Delete, don't archive** — **except** formal **iteration-start** §1.6 corpus hygiene (`mstar-iteration/references/iteration-corpus-hygiene.md`), which **moves** superseded/redundant knowledge/specs to `{HARNESS_DIR}/archived/knowledge|specs/`. Outside that gate, git history is the archive; `git log --diff-filter=D -- <path>` finds deleted docs.
+   > **Not adopted:** frozen-archive seal machinery, `Archived:` triplets, or a verifier — they contradict rule 6 (delete, don't archive; git history is the archive); do not reintroduce.
 7. **Evaluate document-set design.** Check whether two+ docs overlap and should be consolidated. Redundant docs silently drift apart.
 
 ## Scope selection
@@ -92,6 +96,7 @@ For each doc, check:
 2. **Referenced conventions still match?** — check against current `AGENTS.md`, `CONCEPTS.md`, lint configs.
 3. **Solution still the recommended approach?** — has a newer pattern superseded it?
 4. **Overlap with other docs?** — search for same module/tags to find duplicates.
+5. **Future decision value?** — rationale / alternatives considered / negative guarantees / reintroduction conditions that would still guide a future change → **Keep, regardless of length**. A completed decision whose body has no future leverage left (one-off fix detail, superseded mechanics) enters the Delete candidate pool. Word count and age remain discovery aids, never criteria.
 
 ### Phase 3: Classify and act
 

package/harness-skills/mstar-host/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: mstar-host
-description: Morning Star host adapter (OpenCode, Cursor, Codex, Kimi, ZCode, omp). Use after mstar-harness-core whenever host entry, clarify, dispatch, or plan UX differs by platform - OpenCode question/task-tool subagent invoke, Cursor /pm and CreatePlan/SwitchMode dual-write and Task parallel QC, Codex plugin skills plus Plan/Goal Mode, Kimi Agent/AgentSwarm with built-in subagent types only (coder/explore/plan) and role-in-prompt binding, ZCode Agent/AskUserQuestion/EnterPlanMode with built-in subagent types and role-in-prompt binding, omp task/ask/hub preferring live-schema role agents (agents/*.md) with C5b skill-load binding (generic task/scout only as fallback), sandboxed tools, and tool discovery. Auto-detect host from session tools; then Read references/<host>.md. Always load after mstar-harness-core.
+description: Morning Star host adapter (OpenCode, Cursor, Codex, Kimi, ZCode, omp, dsh). Use after mstar-harness-core whenever host entry, clarify, dispatch, or plan UX differs by platform - OpenCode question/task-tool subagent invoke, Cursor /pm and CreatePlan/SwitchMode dual-write and Task parallel QC, Codex plugin skills plus Plan/Goal Mode, Kimi Agent/AgentSwarm with built-in subagent types only (coder/explore/plan) and role-in-prompt binding, ZCode Agent/AskUserQuestion/EnterPlanMode with built-in subagent types and role-in-prompt binding, omp task/ask/hub preferring live-schema role agents (agents/*.md) with C5b skill-load binding (generic task/scout only as fallback), dsh (DeepSeek Harness) subagent tool with in-process engine gates and bundled mstar commands, sandboxed tools, and tool discovery. Auto-detect host from session tools; then Read references/<host>.md. Always load after mstar-harness-core.
 ---
 
 # Morning Star Host Adapter
@@ -30,12 +30,13 @@ Detect from **session tool shapes and available commands** — not from plugin m
 | **`subagent_type`** param on the Task tool (plus **CreatePlan**/**SwitchMode** when Plan mode is active) | `cursor` | `references/cursor.md`; Plan mode also `references/cursor-plan-mode-bridge.md` |
 | **`question`** tool, or **`task`** tool with **`subagent`** (singular) — no `tasks[]` batch | `opencode` | `references/opencode.md` |
 | **`task`** tool with **`agent`** / **`tasks[]`** batch, **`ask`**, **`hub`** (omp also exposes `/goal`; goal rule is host-agnostic per below) | `omp` | `references/omp.md`; Plan mode also `references/omp-plan-mode-bridge.md` |
+| **`subagent`** tool (dsh's model-facing delegation tool — `@deepseek-ai/dsh-tool-subagent` default `toolName`) | `dsh` | `references/dsh.md` |
 | **`Agent`** / **`AskUserQuestion`** / **`EnterPlanMode`** + **`AgentSwarm`** (Kimi-only) | `kimi` | `references/kimi.md`; Plan mode also `references/kimi-plan-mode-bridge.md` |
 | **`Agent`** / **`AskUserQuestion`** / **`EnterPlanMode`** / **`TodoWrite`**, **no `AgentSwarm`** | `zcode` | `references/zcode.md`; Plan mode also `references/zcode-plan-mode-bridge.md` |
 | `/plan`, `/goal` slash commands; **Goal tools**; `functions.*` / `codex_app.*` tool namespaces; `tool_search`; Browser plugin tools | `codex` | `references/codex.md`; Plan mode also `references/_shared/plan-mode-bridge-core.md` |
-| Still ambiguous | - | Read sections in **`cursor.md`**, **`opencode.md`**, **`codex.md`**, **`kimi.md`**, **`zcode.md`**, and **`omp.md`** that match tools you have; **`mstar-harness-core` wins** on conflict |
+| Still ambiguous | - | Read sections in **`cursor.md`**, **`opencode.md`**, **`codex.md`**, **`kimi.md`**, **`zcode.md`**, **`omp.md`**, and **`dsh.md`** that match tools you have; **`mstar-harness-core` wins** on conflict |
 
-Order matters: check `cursor` → `opencode` → `omp` → `kimi` → `zcode` → `codex`. `subagent_type` (Cursor) vs `subagent` (OpenCode) vs `agent`/`tasks[]` (omp) is the sharpest split among the Task-based hosts.
+Order matters: check `cursor` → `opencode` → `omp` → `dsh` → `kimi` → `zcode` → `codex`. `subagent_type` (Cursor) vs `subagent` (OpenCode) vs `agent`/`tasks[]` (omp) is the sharpest split among the Task-based hosts; dsh's `subagent` tool collides with no other row, so it sits with the agent-tool hosts.
 
 > **Engine check (when available):** run `mstar host detect --signals <comma-list>` (or `import { detectHost } from "@mstar-harness/engine"` in a host hook) to resolve the detection table above from session tool shapes (prints the host id, or `ambiguous` to fall back on the table + judgment). On `fail` -> do not proceed; fix and re-run. Skill text below remains authoritative when the runtime is absent.
 
@@ -62,6 +63,7 @@ Docs name assets as skill **`<name>`** → `scripts/…` / `references/…`. **R
 | **Cursor** | Skill **name** via plugin skills | Global `~/.cursor/plugins/local/morning-star-harness/skills/<name>/`; project `.cursor/plugins/morning-star-harness/skills/<name>/` |
 | **Codex** | Skill **name** via plugin | Plugin-mounted `skills/<name>/`; project command skills under `.agents/skills/<name>/` |
 | **OpenCode** | Skill **name** via `@mstar-harness/opencode` | Package-internal `harness-skills/<name>/` — never `process.cwd()/skills/` |
+| **dsh** | Skill **name** via the mstar skill-local provider (`providerName: mstar`) | `$DSH_BUNDLED_SKILL_DIR/<name>[/<rel>]` — the packaged `harness-skills/` mirror mounted package-relative by `@mstar-harness/dsh`; never app cwd |
 | **Kimi / ZCode** | Skill **name** / `/skill:<name>` | Plugin mount `./skills/<name>/` from the installed plugin root |
 
 Authoring convention: **`mstar-skill-authoring`** § Skill-relative script and asset paths. Per-host URI / mount detail: `references/<host>.md`.