@lmzhen/dsh-evolution-maintenance 0.32.3 → 0.32.5

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (2) hide show
  1. package/README.md +9 -22
  2. package/package.json +2 -2
package/README.md CHANGED
@@ -1,32 +1,19 @@
1
1
  # @lmzhen/dsh-evolution-maintenance
2
2
 
3
- Skill-library drift-scanning determinism surface for the maintenance subagent: it assembles the
4
- skill snapshot, renders the mechanical-facts block and runs the scan → subagent → validate
5
- chain behind `/evolution maintain` (`evolution-core` `drift-signals`).
3
+ > 手动维护扫描(高级)。
6
4
 
7
- - `snapshotFromLibrary`: assemble a plain skill snapshot from a SkillLibrary-like reader.
8
- - `renderFacts`: canonical `MECHANICAL_FACTS` block (version + joint signature + redaction).
5
+ ## 你会看到
9
6
 
10
- ## Model surface
7
+ `/evolution maintain` 及 status/report/cancel。
11
8
 
12
- - **Model-visible:** the read-only `maintenance_probe` tool schema (`evolution-maintenance/tools`), mounted for every session.
13
- - **Prompt prefix / KV cache:** no system-prompt section of its own; family-level rules: `packages/README.md` §"Model-visible prompt prefix and the KV cache".
14
- - **Mount it?** yes — the `evolution-maintenance/tools` row (host/all/preset); the orchestration is consumed by `evolution-commands` (`/evolution maintain`).
9
+ ## 什么时候会用到
15
10
 
16
- ## Known limitations
11
+ 你想主动扫一遍技能库/生成维护计划时。
17
12
 
18
- - Phase 1-2 expose no service beyond the command surface: the chain (scan → render → subagent → validate) is wired through `/evolution maintain`.
19
- - **The scan's lifetime is the RUN's, not the request's (0.19.0).** `/evolution maintain` starts a run through `evolution-core`'s run registry and answers with a pointer; `evolution-commands` writes the resulting plan to `<evolutionHome>/reports/maintain-<runId>.{json,md}` — this package persists no report of its own. Measured: 949.7 s end to end on the web plane, ~305 s of REQUEST lifetime on the desktop shell (which used to abort the scan; it no longer does). `maintainTimeoutMs` (default 600000) bounds the run, and a run that produced nothing is recorded as such rather than as an empty plan. The published inventory of long runs is `packages/README.md` §"Long-running work", rendered from `evolution-core/src/long-runs.ts`.
20
- - **Plan validation is instance-granular (§3 completeness).** Notes stay id-level: one mention explains away every instance of that id.
21
- - **A budget-capped dedup scan is reported, never silently shortened.**
22
- - `maintenance_probe` is readable by every session; maintenance subagents reach it through the orchestrate `toolFilter` allow-list.
23
- - The model-visible template (`MAINTAIN_PROMPT`) and the subagent output instruction (`MAINTAIN_OUTPUT_INSTRUCTION`, V10 F-16) ship in `evolution-core` `PROMPT_BUNDLE`; the joint-signature mismatch protocol is honored by `renderFacts` callers, not by this package alone.
13
+ ## 默认 / 前提
24
14
 
25
- **Runtime invariant:** No companion is published. The platform auto-assembles nothing and the family mounts no `<pkg>/invariant` cordis row, so a companion here would never execute (v37 S2.1 / I-3).
15
+ 异步运行,返回 run id。
26
16
 
27
- ## Notes and history
17
+ ## 边界
28
18
 
29
- - (design 011: the maintenance-subagent design paper, archived in git history).
30
- - The facts block renders one line per skill, so ONE signal id can be `over` on several skills: every `over` instance (`skill::id`; a library-level signal renders once and keeps the bare id) must be covered by plan evidence or named in a note, so a plan that advises one skill no longer zero-explains the same signal on its siblings (P2-4, 0.3.83).
31
- - `maintenance_probe` appends "note: dedup scan truncated at the pair-comparison budget; groups may be incomplete" and the `dedup_group` signal value carries the same qualification (P2-8, 0.3.83).
32
- - V10 (F-02): orchestrate soft-probes the tools registry before spawning and degrades the filter to `skill`-only (declared in the subagent prompt) when the tool row is not mounted, instead of failing the spawn; an explicit `toolAllow` option bypasses the probe.
19
+ 日常不需要跑;不改技能内容。
package/package.json CHANGED
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "name": "@lmzhen/dsh-evolution-maintenance",
3
3
  "description": "Deterministic maintenance-scan surface: skill-library snapshot assembly, drift signals and mechanical-facts rendering (design 011) (community build)",
4
- "version": "0.32.3",
4
+ "version": "0.32.5",
5
5
  "publishConfig": {
6
6
  "access": "public"
7
7
  },
@@ -31,7 +31,7 @@
31
31
  "license": "MIT",
32
32
  "dependencies": {
33
33
  "@deepseek-ai/schemastery": "^3.18.5",
34
- "@lmzhen/dsh-evolution-core": "^0.32.3"
34
+ "@lmzhen/dsh-evolution-core": "^0.32.5"
35
35
  },
36
36
  "peerDependencies": {
37
37
  "@deepseek-ai/cordis": "^4.0.5",