dsh-compact-arsenal 0.3.0 → 0.5.0

package/CHANGELOG.md

@@ -4,6 +4,50 @@ All notable changes to this project are documented here. The format follows
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and the project uses
 [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.5.0] — 2026-11-11
+
+### Fixed
+
+- **Reasoning compression could never land on any published DSH release.**
+  Reported in [#1](https://github.com/TacitSay/dsh-compact-arsenal/issues/1):
+  the log enforces two mutually exclusive rules on a surface replacement — every
+  shadowed node must be cited, and `assistant/message` may not cite anything.
+  Rewriting an assistant node in place is therefore impossible, and every
+  reasoning attempt failed with `missing <seq>` while its model call had already
+  been paid for. Reasoning now collapses into a `user/message` checkpoint citing
+  the node it replaces, the same shape the shipped compaction backends use.
+- **`totalPass` cited only the nodes it fed the model.** Nodes filtered out
+  before summarization (system, developer, and turn-less nodes, including this
+  plugin's own earlier checkpoints) still fall inside `[startSeq, endSeq]` and
+  were reported as missing. The whole surface span is cited now.
+
+### Added
+
+- Pre-flight shape check: an impossible replacement is detected before the model
+  call, so a deterministic rejection no longer spends a request.
+- `test/surface-shape.test.mjs` runs against the packaged `dsh-session` and pins
+  the accepted replacement shape for both node types.
+- Source assertions now cover the reasoning shape, the whole-span citation, and
+  the pre-flight check (17 checks).
+
+## [0.4.0] — 2026-11-11
+
+### Fixed
+
+- **The pressure readout never appeared while all sections were disabled.**
+  `rewriteAsync()` short-circuited before measuring, so a plugin with every
+  switch off reported nothing. Measurement now happens first, and the panel
+  endpoint re-reads it from the live agent on every request instead of relying
+  on the once-per-turn hook value.
+
+### Added
+
+- **Version badge in the panel header** (`v0.4.0`), sourced from the running
+  plugin rather than `package.json`, so a stale install is visible at a glance.
+- A disabled **检查更新** button beside it, reserving the seat for a future
+  update check; the market owns package updates, so this stays inert until one
+  exists.
+
 ## [0.3.0] — 2026-11-11
 
 ### Fixed

package/client.js

@@ -270,6 +270,7 @@ window.__ModuleLoader__.load({
       const [archiveRoot, setArchiveRoot] = useState(null)
       const [lastArchivePath, setLastArchivePath] = useState(null)
       const [pressure, setPressure] = useState(null)
+      const [runtimeVersion, setRuntimeVersion] = useState(null)
       const timer = useRef(null)
       const latest = useRef(emptyConfig)
 
@@ -287,6 +288,7 @@ window.__ModuleLoader__.load({
             setArchiveRoot(body.archiveRoot ?? null)
             setLastArchivePath(body.lastArchivePath ?? null)
             setPressure(body.pressure ?? null)
+            setRuntimeVersion(body.version ?? null)
             setStatus(body.editable === false ? 'readonly' : 'ready')
           })
           .catch((error) => {
@@ -317,6 +319,7 @@ window.__ModuleLoader__.load({
               if (body.archiveRoot) setArchiveRoot(body.archiveRoot)
               if (body.lastArchivePath !== undefined) setLastArchivePath(body.lastArchivePath)
               if (body.pressure !== undefined) setPressure(body.pressure)
+              if (body.version) setRuntimeVersion(body.version)
               setTotalApplied(body.totalApplied === undefined ? null : body.totalApplied)
               setStatus('ready')
               setMessage('已保存')
@@ -506,13 +509,49 @@ window.__ModuleLoader__.load({
             : status === 'readonly' ? '配置写入不可用（只读）'
               : (message || '已连接')
 
+      /**
+       * Version badge beside the title.
+       *
+       * The update affordance is deliberately inert for now: the market owns
+       * package updates, so this only reserves the seat and the label. When an
+       * update check exists, it flips `updateAvailable` and the same button
+       * becomes actionable without moving anything.
+       */
+      const versionBadge = h('div', {
+        key: 'version',
+        style: { display: 'flex', alignItems: 'center', gap: 6 },
+      }, [
+        h('span', {
+          key: 'v',
+          title: runtimeVersion ? `运行中：${runtimeVersion}` : '运行中版本未知',
+          style: {
+            fontSize: 11, padding: '1px 7px', borderRadius: 999, color: MUTED,
+            border: `1px solid ${LINE}`, fontVariantNumeric: 'tabular-nums',
+          },
+        }, runtimeVersion ? `v${runtimeVersion}` : 'v?'),
+        h('button', {
+          key: 'check',
+          type: 'button',
+          disabled: true,
+          title: '检查更新（预留：将由插件市场提供）',
+          style: {
+            fontSize: 11, padding: '1px 8px', borderRadius: 999,
+            border: `1px solid ${LINE}`, background: 'transparent',
+            color: MUTED, cursor: 'not-allowed', opacity: 0.55,
+          },
+        }, '检查更新'),
+      ])
+
       const header = h('header', {
         key: 'hdr',
         style: { display: 'flex', flexDirection: 'column', gap: 8 },
       }, [
         h('div', { key: 'row', style: { display: 'flex', alignItems: 'center', justifyContent: 'space-between', gap: 12 } }, [
           h('div', { key: 't', style: { display: 'flex', flexDirection: 'column', gap: 2 } }, [
-            h('h1', { key: 'h1', style: { margin: 0, fontSize: 18, fontWeight: 600 } }, '上下文压缩'),
+            h('div', { key: 'title', style: { display: 'flex', alignItems: 'center', gap: 8 } }, [
+              h('h1', { key: 'h1', style: { margin: 0, fontSize: 18, fontWeight: 600 } }, '上下文压缩'),
+              versionBadge,
+            ]),
             h('p', { key: 'p', style: { margin: 0, fontSize: 12, color: MUTED, lineHeight: 1.6 } },
               '开关即生效，参数改完自动保存。每个板块有自己的触发阈值。'),
           ]),

package/index.js

@@ -27,6 +27,12 @@ import { archiveRewrite, archiveRoot } from './lib/archive.js'
 
 export const name = 'compact-arsenal'
 
+/**
+ * Bumped in lockstep with `package.json` so the panel can show which build is
+ * actually running — a stale install is otherwise invisible.
+ */
+export const PACKAGE_VERSION = '0.5.0'
+
 /** Host services this plugin reads; all ship with dsh-base or dsh-web-app. */
 export const inject = ['sessions', 'tokenMeter', 'webServer', 'configEditor']
 
@@ -174,6 +180,33 @@ function isLegacyModelPrompt(prompt) {
   return typeof prompt === 'string' && prompt.includes('只输出要点') && prompt.includes('不超过 200 字')
 }
 
+/**
+ * Recompute the pressure snapshot from whichever agent is live.
+ *
+ * The step hook updates the same object at most once per turn; a panel read
+ * should reflect the session as it stands now, including when every section is
+ * disabled and the hook short-circuits.
+ *
+ * @param {any} ctx - plugin context.
+ * @param {{ reported: number | null, estimated: number, window: number | null }} pressure - snapshot to update.
+ * @returns {void}
+ */
+function refreshPressure(ctx, pressure) {
+  try {
+    const agents = ctx.get('agents')?.list?.() ?? []
+    const agent = agents[agents.length - 1]
+    if (agent === undefined) return
+    const session = ctx.sessions.get(agent.id)
+    if (session === undefined) return
+    const snapshot = pressureSnapshot(ctx, session)
+    pressure.reported = snapshot.reported ?? null
+    pressure.estimated = snapshot.estimated
+    pressure.window = session.requestContext?.()?.contextWindow ?? null
+  } catch {
+    // A missing agent registry only means "nothing live to measure".
+  }
+}
+
 /**
  * Register the rewrite pass on the per-step hook.
  * @param {any} ctx - plugin context.
@@ -250,7 +283,11 @@ function registerPanelRoutes(ctx, config, stats, archive, pressure) {
 
   route(`${API_PREFIX}/state`, (req, res) => {
     if (req.method !== 'GET') return send(res, 405, { error: 'method-not-allowed' })
+    // Fold any live agent into the snapshot at read time: the step hook only
+    // updates it once per turn, and the panel should not show a stale number.
+    refreshPressure(ctx, pressure)
     send(res, 200, {
+      version: PACKAGE_VERSION,
       config,
       stats,
       editable: configEditor !== undefined,
@@ -393,6 +430,19 @@ async function totalPass(ctx, session, total, turn, signal, stats, archive) {
   if (picked.length < 4) return
 
   const transcript = picked.map((node) => transcriptLine(node.event)).join('\n')
+
+  // The replacement must cite EVERY surface node inside the replaced span, not
+  // just the ones we fed the model: filters above skip system/developer nodes
+  // and nodes without a turn, and those still sit inside [startSeq, endSeq].
+  const startSeq = picked[0].seq
+  const endSeq = picked[picked.length - 1].seq
+  const surface = [...session.surface.nodes]
+  const startIdx = surface.indexOf(startSeq)
+  const endIdx = surface.indexOf(endSeq)
+  if (startIdx < 0 || endIdx < startIdx) return
+  if (startIdx === 0 && session.eventAt(surface[0])?.type === 'system/message') return
+  const shadowed = surface.slice(startIdx, endIdx + 1)
+
   const routed = session.requestContext?.() ?? {}
   const summaryCfg = {
     ...total,
@@ -409,8 +459,6 @@ async function totalPass(ctx, session, total, turn, signal, stats, archive) {
   }
   if (summary === '') return
 
-  const startSeq = picked[0].seq
-  const endSeq = picked[picked.length - 1].seq
   try {
     session.append(
       'user/message',
@@ -422,7 +470,7 @@ async function totalPass(ctx, session, total, turn, signal, stats, archive) {
           text: `This is an automatically generated checkpoint condensing an earlier span of the conversation to free up context. Treat the captured context as established background and build on it without restating it.\n\n<compacted-summary>\n${summary}\n</compacted-summary>`,
         }],
       },
-      { surfaceOp: { op: 'replace', startSeq, endSeq }, sourceEventSeqs: picked.map((node) => node.seq) },
+      { surfaceOp: { op: 'replace', startSeq, endSeq }, sourceEventSeqs: shadowed },
     )
     stats.total = (stats.total ?? 0) + 1
     stats.chars += used
@@ -488,21 +536,22 @@ function rewrite(ctx, config, payload, lastProcessedTurn, stats, archive, pressu
  * @returns {Promise<void>} resolves when every rewrite of this turn landed.
  */
 async function rewriteAsync(ctx, config, payload, lastProcessedTurn, stats, archive, pressure) {
-  if (!config.total.enabled && !config.reasoning.enabled && !config.tools.enabled) return
-
   const session = ctx.sessions.get(payload.agent.id)
   if (session === undefined) return
 
-  if (lastProcessedTurn.get(session.id) === payload.turn) return
-
-  lastProcessedTurn.set(session.id, payload.turn)
-
-  // Snapshot the inputs once per turn so the panel can explain a gate decision.
+  // Measure before any gate: the panel must be able to explain why nothing ran,
+  // so a fully disabled plugin still reports the current pressure.
   const snapshot = pressureSnapshot(ctx, session)
   pressure.reported = snapshot.reported ?? null
   pressure.estimated = snapshot.estimated
   pressure.window = session.requestContext?.()?.contextWindow ?? null
 
+  if (!config.total.enabled && !config.reasoning.enabled && !config.tools.enabled) return
+
+  if (lastProcessedTurn.get(session.id) === payload.turn) return
+
+  lastProcessedTurn.set(session.id, payload.turn)
+
   if (config.total.enabled && underPressure(ctx, session, config.total.pressureRatio)) {
     await totalPass(ctx, session, config.total, payload.turn, payload.signal, stats, archive)
   }
@@ -602,7 +651,16 @@ function currentPosition(session, event, currentTurn) {
 }
 
 /**
- * Replace one event's text with a rewritten version, keeping message identity.
+ * Rewrite one node's text, using the only shape the log accepts for each type.
+ *
+ * `tool/result` may be replaced in place because it can cite the node it
+ * shadows. `assistant/message` cannot: the log rejects any `sourceEventSeqs` on
+ * it ("assistant/message embeds its source stream"), while every replacement
+ * must cite all shadowed nodes — so an in-place assistant rewrite is impossible
+ * on every published version. Reasoning therefore collapses into a
+ * `user/message` checkpoint covering the whole replaced span, which is the same
+ * shape the shipped compaction backends use.
+ *
  * @param {any} session - live session.
  * @param {any} event - the surface event being replaced.
  * @param {number} seq - its surface seq.
@@ -612,26 +670,46 @@ function currentPosition(session, event, currentTurn) {
  * @returns {void}
  */
 function applyRewrittenText(session, event, seq, kind, text, currentTurn) {
-  const blocks = event.data?.message?.content
-  const wanted = kind === 'tools' ? 'text' : 'reasoning'
-  const rich = blocks.filter((block) => block?.type !== wanted)
-  const message = { ...event.data.message, content: [{ type: wanted, text }, ...rich] }
-  const position = currentPosition(session, event, currentTurn)
   if (kind === 'tools') {
+    const blocks = event.data?.message?.content ?? []
+    const rich = blocks.filter((block) => block?.type !== 'text')
+    const message = { ...event.data.message, content: [{ type: 'text', text }, ...rich] }
     session.append('tool/result', {
-      ...position,
+      ...currentPosition(session, event, currentTurn),
       message,
       ...(event.data.error === undefined ? {} : { error: event.data.error }),
       ...(event.data.meta === undefined ? {} : { meta: event.data.meta }),
     }, { surfaceOp: { op: 'replace', startSeq: seq, endSeq: seq }, sourceEventSeqs: [seq] })
     return
   }
-  session.append('assistant/message', {
-    ...position,
-    message,
-    ...(event.data.usage === undefined ? {} : { usage: event.data.usage }),
-    stream: event.data.stream ?? [],
-  }, { surfaceOp: { op: 'replace', startSeq: seq, endSeq: seq } })
+
+  // Reasoning: replace the node with a user-role checkpoint that cites it.
+  session.append('user/message', {
+    role: 'user',
+    source: { kind: 'user' },
+    content: [{ type: 'text', text }],
+  }, { surfaceOp: { op: 'replace', startSeq: seq, endSeq: seq }, sourceEventSeqs: [seq] })
+}
+
+/**
+ * Whether one section can legally replace a node in this session right now.
+ *
+ * The check is deliberately cheap and local: it mirrors the log's own rules
+ * (a replacement must cite the node it shadows, and `assistant/message` may not
+ * cite anything) so an impossible write is detected before a model call.
+ *
+ * @param {any} session - live session.
+ * @param {'reasoning' | 'tools'} kind - section kind.
+ * @param {number | undefined} seq - the node that would be replaced.
+ * @returns {boolean} true when the replacement shape is allowed.
+ */
+function canReplace(session, kind, seq) {
+  if (seq === undefined) return false
+  const surface = session.surface?.nodes
+  if (!Array.isArray(surface) || !surface.includes(seq)) return false
+  if (kind === 'tools') return session.eventAt?.(seq)?.type === 'tool/result'
+  // Reasoning collapses into a user/message checkpoint, which may cite the node.
+  return kind === 'reasoning'
 }
 
 /**
@@ -674,6 +752,14 @@ async function compressSection(ctx, session, policy, kind, seqs, signal, archive
   }
   if (totalChars < minTotal) return { items: 0, chars: 0, failed: 0, archived: 0 }
 
+  // A rejection here is deterministic: the log will refuse the same shape every
+  // time. Probing with one throwaway append keeps a doomed batch from spending a
+  // model call, so `mode: model` cannot burn requests on writes that cannot land.
+  if (!canReplace(session, kind, candidates[0].seq)) {
+    lastFailure = `${kind}: surface shape rejected before any model call`
+    return { items: 0, chars: 0, failed: 1, archived: 0 }
+  }
+
   const routed = session.requestContext?.() ?? {}
   const model = policy.model ?? {}
   const modelCfg = {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-compact-arsenal",
-  "version": "0.3.0",
+  "version": "0.5.0",
   "description": "Configurable context compaction for DeepSeek Harness: three independent sections — whole-context checkpoint, reasoning compression and tool-output pruning, each with its own switch, threshold and method, plus a left-sidebar panel.",
   "icon": "icon.svg",
   "author": "TacitSay",