dsh-switchman 1.10.3 → 1.11.1

package/cordis.patch.yml

@@ -28,11 +28,13 @@
 #
 # Provenance & maintenance: the four `plugins` lists below are copied verbatim
 # from @deepseek-ai/dsh-web-app presets/{standard,ptc,minimal,cordis}.patch.yml
-# (dsh 0.2.1-rc.1: time-context + tool-schedule + the disabled tool-plugin-market
-# row joined every list, the subagent rows lost `backgroundMode` under the
-# unified activation lifecycles, the ralph/codex/claude-code rows left the
-# presets, minimal gained image-only fs tools and a compaction group, and the
-# plan-mode text was reworded). A Loader override replaces the target row's
+# (dsh 0.2.2-rc.1: tool-fs ships `applyPatch: false`, tool-fs-search ships
+# `sampleOverCapGlobResults: false`, the subagent spawn row carries
+# `modelSelection` + `modelDiscovery` (the 0.2.1 unified-activation shape),
+# tool-plugin-manager/market disabled guards use `!!js "!ctx.get('profileContext')"`
+# expression strings, and no list-agents row ships; dsh 0.2.1-rc.1 had added
+# time-context + tool-schedule, dropped `backgroundMode`, and reworded the
+# plan-mode text). A Loader override replaces the target row's
 # `config` wholesale, so every field is restated here. After a DSH upgrade that
 # changes the shipped presets, re-sync these lists from the new preset files,
 # then reinstall this bundle; scripts/validate.mjs compares this patch against
@@ -45,7 +47,7 @@
 # scope, and a second instance makes session composition throw
 # `tool "list_subagent_models" is already registered in this scope` — breaking
 # creation of ALL new sessions and the recomposition of old ones. The shipped
-# presets (verified against installed dsh-web-app 0.2.0-rc.2, the fork's
+# presets (verified against installed dsh-web-app 0.2.2-rc.1, the fork's
 # 0.2.1-alpha.1, and 0.2.1-rc.1) enable it on the spawn row only; fork-side
 # explicit routing is not a factory capability.
 #
@@ -97,7 +99,7 @@
             - 禁止复活死会话：子智能体达 cap 终止后，用自包含 prompt 新派一个，不 resume 旧壳。
             - 广搜先问：首次整项目范围的搜索先向用户澄清范围，定点检索不受此限。
             - 大额派发先估成本：预计 ≥3 个成员或总消耗 >100k token 时，先报规模与预估请用户确认。
-            - 子路由两通道：显式传 provider/model/effort 任一字段会经 DSH 会话级授权白名单校验（设置页「允许 Agent 为子智能体选择模型」勾选的授权列表，与泳道池无关）；被拒时用 list_subagent_models 查实际可用路由再派，或改为完全不传三字段走隐式继承（不受校验，子智能体用默认值或继承主会话路由）。泳道池只是偏好排序，授权边界以 DSH 设置为准，不试图改写它。
+            - 子路由两通道：派发 subagent/workflow 默认显式传 provider/model/effort（按泳道表逐字段拆填，经 DSH 会话级授权白名单校验——设置页「允许 Agent 为子智能体选择模型」的授权列表，与泳道池无关）；隐式继承（完全不传三字段）不是省事通道，仅当显式尝试被拒且 list_subagent_models 确认无可用路由后才可兜底，并在回报中声明 IMPLICIT-DOWNGRADED——隐式拿到的是主会话路由，泳道模型丢失。泳道池只是偏好排序，授权边界以 DSH 设置为准，不试图改写它。
 
             ## 三、分级验证（质量下限，不可省）
             - 改动 >20 行或涉及多个调用点：交 tester 验证。
@@ -122,6 +124,8 @@
         disabled: !!js process.platform !== 'win32'
       - id: tool-fs
         name: '@deepseek-ai/dsh-tool-fs'
+        config:
+          applyPatch: false
       - id: tool-fs-search
         name: '@deepseek-ai/dsh-tool-fs-search'
         config:
@@ -192,14 +196,13 @@
         config:
           - id: tool-subagent-control
             name: '@deepseek-ai/dsh-tool-subagent-control'
-          - id: tool-subagent-list-agents
-            name: '@deepseek-ai/dsh-tool-subagent-control/list-agents'
           - id: tool-subagent
             name: '@deepseek-ai/dsh-tool-subagent'
             config:
               provider: spawn
               toolName: subagent
-              modelSelectionSettings: true
+              modelSelection: true
+              modelDiscovery: true
           - id: tool-subagent-fork
             name: '@deepseek-ai/dsh-tool-subagent'
             config:
@@ -226,7 +229,7 @@
         name: '@deepseek-ai/dsh-tool-present'
       - id: tool-plugin-manager
         name: '@deepseek-ai/dsh-plugin-manager/tools'
-        disabled: true
+        disabled: !!js "!ctx.get('profileContext')"
       - id: tool-plugin-market
         name: '@deepseek-ai/dsh-plugin-market/tools'
         disabled: true
@@ -260,6 +263,8 @@
         disabled: !!js process.platform !== 'win32'
       - id: tool-fs
         name: '@deepseek-ai/dsh-tool-fs'
+        config:
+          applyPatch: false
       - id: tool-fs-search
         name: '@deepseek-ai/dsh-tool-fs-search'
         config:
@@ -330,14 +335,13 @@
         config:
           - id: tool-subagent-control
             name: '@deepseek-ai/dsh-tool-subagent-control'
-          - id: tool-subagent-list-agents
-            name: '@deepseek-ai/dsh-tool-subagent-control/list-agents'
           - id: tool-subagent
             name: '@deepseek-ai/dsh-tool-subagent'
             config:
               provider: spawn
               toolName: subagent
-              modelSelectionSettings: true
+              modelSelection: true
+              modelDiscovery: true
           - id: tool-subagent-fork
             name: '@deepseek-ai/dsh-tool-subagent'
             config:
@@ -362,10 +366,8 @@
         config:
           fetch: true
           searchTimeoutMs: 60000
-      - id: tool-presentation
-        name: '@deepseek-ai/dsh-agent-tool-presentation'
-        config:
-          mode: ptc
+      - id: tool-executor-ptc
+        name: '@deepseek-ai/dsh-tool-executor-ptc'
       - id: present
         name: '@deepseek-ai/dsh-tool-present'
       - id: tool-plugin-manager
@@ -502,6 +504,8 @@
         disabled: !!js process.platform !== 'win32'
       - id: tool-fs
         name: '@deepseek-ai/dsh-tool-fs'
+        config:
+          applyPatch: false
       - id: tool-fs-search
         name: '@deepseek-ai/dsh-tool-fs-search'
         config:
@@ -568,14 +572,13 @@
         config:
           - id: tool-subagent-control
             name: '@deepseek-ai/dsh-tool-subagent-control'
-          - id: tool-subagent-list-agents
-            name: '@deepseek-ai/dsh-tool-subagent-control/list-agents'
           - id: tool-subagent
             name: '@deepseek-ai/dsh-tool-subagent'
             config:
               provider: spawn
               toolName: subagent
-              modelSelectionSettings: true
+              modelSelection: true
+              modelDiscovery: true
           - id: tool-subagent-fork
             name: '@deepseek-ai/dsh-tool-subagent'
             config:

package/host/config.js

@@ -129,11 +129,21 @@ export const Config = z.object({
 	// --- Phase 3: context watermark control --------------------------------
 	/** Soft watermark: advisory banner + delegation nudge (tokens). */
 	wmSoftTokens: z.number().min(1000).default(80000).volatile(),
-	/** Hard watermark: read budget tightens, wrap-up advised (tokens). */
-	wmHardTokens: z.number().min(2000).default(120000).volatile(),
+	/** Hard watermark: read budget tightens, wrap-up advised (tokens). 95k
+	 *  sits just under the trim pressure line (65% of the 150k window ≈
+	 *  97.5k): past
+	 *  this point the fold engine is already folding, so per-call read
+	 *  budgets must tighten at the SAME boundary — agent sessions grow in
+	 *  PULSES (one large file read = tens of thousands of tokens injected
+	 *  at once, faster than any fold can commit), and the read budget is
+	 *  the only preventive valve on the pulse source. */
+	wmHardTokens: z.number().min(2000).default(95000).volatile(),
 	/** Force watermark: handover document + compaction (tokens). */
 	wmForceTokens: z.number().min(3000).default(150000).volatile(),
-	/** Per-call read budget for read/glob/grep (tokens). */
+	/** Per-TURN read budget (inspection C16 — the name predates the turn
+	 *  accounting): read/glob/grep/bash results accumulate against it within
+	 *  one turn; the account resets after 15 idle minutes and the pre/post
+	 *  gates act at 2.4× this value (CEILING_MULTIPLE in context-watch.js). */
 	wmReadBudgetTokens: z.number().min(200).default(3000).volatile(),
 	/** Hard-tier behavior: cap reads and advise, or deny outright. */
 	wmDenyMode: z.union(["cap", "deny"]).default("cap").volatile(),
@@ -188,7 +198,7 @@ export const Config = z.object({
 	 *  kind fails open to the fallback chain. */
 	intentClassifierEndpoint: z.string().default("").volatile(),
 	/** Model name sent to the decision endpoint for kind "systemone"
-	 *  (free text, e.g. tev1:4b-mxfp8). Independent of the llm provider
+	 *  (free text, e.g. d3:4b-mxfp8). Independent of the llm provider
 	 *  catalog — no provider registration needed. Empty = unset; that
 	 *  kind fails open to the fallback chain. */
 	intentClassifierSystemoneModel: z.string().default("").volatile(),
@@ -238,26 +248,63 @@ export const Config = z.object({
 	trimGrowthTokens: z.number().min(1000).default(22500).volatile(),
 	/** Growth gate companion: the contiguous compressible increment must
 	 *  also reach this many estimated tokens (chars/4 heuristic). */
-	trimIncrementTokens: z.number().min(1000).default(50000).volatile(),
+	trimIncrementTokens: z.number().min(1000).default(25000).volatile(),
 	/** Pressure tier (tier b): fold when view usage reaches this percent of
 	 *  the context window (requires a resolvable window; otherwise skipped). */
-	trimPressurePercent: z.number().min(1).max(100).default(60).volatile(),
+	trimPressurePercent: z.number().min(1).max(100).default(65).volatile(),
 	/** Emergency tier (tier c): forced MECHANICAL fold (no side-model call)
 	 *  at this percent of the window; cooldown-gated. */
 	trimEmergencyPercent: z.number().min(1).max(100).default(95).volatile(),
-	/** Effective-window cap for the tier math: tier points are computed
+	/** [2026-11-07]-[growth/increment constants are absolute tokens
+	 *  calibrated for the default 150k effective window]-[scaling
+	 *  trimWindowTokens down without scaling trimGrowthTokens /
+	 *  trimIncrementTokens proportionally makes the growth tier unreachable
+	 *  (a 30k window can never accumulate a 50k increment; observed live:
+	 *  growth fired once while emergency fired 17 times). Rule of thumb:
+	 *  increment ≈ window/6, growth ≈ window × 0.15 (~25k/22.5k at 150k,
+	 *  2026-11-09 steady-growth recalibration — one fold per ~22.5k of new
+	 *  work instead of one every ~2 minutes; pressure pins the 50-100k band
+	 *  ceiling).]
+	 *  Effective-window cap for the tier math: tier points are computed
 	 *  against min(model-resolved window, cap). 0 = auto (cap follows
 	 *  wmForceTokens, so folds always fire ahead of the handover tier). */
 	trimWindowTokens: z.number().min(0).default(0).volatile(),
 	/** Last N messages of a history are never folded (the live tail the
 	 *  next model call reasons over). */
-	trimProtectedTail: z.number().min(0).max(100).default(6).volatile(),
+	trimProtectedTail: z.number().min(0).max(100).default(10).volatile(),
 	/** Side-model fold cadence cap per rolling hour (emergency mechanical
-	 *  folds are exempt; they have their own cooldown). */
-	trimMaxFoldsPerHour: z.number().min(1).max(60).default(6).volatile(),
+	 *  folds are exempt; they have their own cooldown). 30 sits just above
+	 *  the pure T1 demand of continuous full-rate generation (~27 folds/hour
+	 *  at the 22.5k growth cadence and ~170 tok/s output), so it only bites
+	 *  on runaway churn; at the old cap of 16 growth consumed the whole
+	 *  budget and the ladder starved. */
+	trimMaxFoldsPerHour: z.number().min(1).max(60).default(30).volatile(),
 	/** Route for the side-model summarizer (L3). Empty = inherit the
 	 *  session's current model. */
 	trimSideModel: IntentRouteSchema.default({ provider: "", model: "" }).volatile(),
+	/** Route for the fold-fallback summarizer. When the primary side route
+	 *  (trimSideModel or the inherited session model) fails, raw folds retry
+	 *  on this route before degrading to a mechanical fold; ladder folds
+	 *  retry but never go mechanical. Pick a route with headroom on cooldown,
+	 *  risk-control, and concurrency limits. Empty = no fallback leg. */
+	trimFoldFallbackModel: IntentRouteSchema.default({ provider: "", model: "" }).volatile(),
+	/** T1/T2/T3 ladder folding (docs/ladder-fold.md): master switch for the
+	 *  T2/T3 distillation tiers. T1 raw-range folds are governed by
+	 *  trimMode; the ladder only adds the summary-maintenance tiers. */
+	trimLadderEnabled: z.boolean().default(true).volatile(),
+	/** Ladder T2 gate: minimum active-T1 summary token mass before a T2
+	 *  distillation fires (the need signal — bili #379: block count alone
+	 *  is not a need signal). */
+	trimLadderT2Tokens: z.number().min(1000).default(8000).volatile(),
+	/** Ladder T2 companion floor: a T2 distillation merges at least this
+	 *  many active T1 blocks (filters degenerate two-block rewrites). */
+	trimLadderT2MinBlocks: z.number().min(2).max(200).default(3).volatile(),
+	/** Ladder T3 gate: minimum active-T2 summary token mass before a T3
+	 *  condensation fires. */
+	trimLadderT3Tokens: z.number().min(1000).default(6000).volatile(),
+	/** Ladder T3 companion floor: a T3 condensation merges at least this
+	 *  many active T2 blocks. */
+	trimLadderT3MinBlocks: z.number().min(2).max(200).default(2).volatile(),
 	/** Explicit ledger state directory. Empty = <dshHome>/.dsh-switchman/
 	 *  state/trim/. */
 	trimStateDir: z.string().default("").volatile(),

package/host/context-trim/fork.js

@@ -159,7 +159,11 @@ export function resolveInheritance(parentLedger, parentId, childId, childMessage
 		version: LEDGER_VERSION,
 		sessionId: childId,
 		blocks: blocks.map((block) => {
-			const clone = { ...block, refRange: { ...block.refRange } };
+			// [2026-11-07]-[v2 ladder blocks carry refRange: null]-[spreading
+			// null produced {} on the clone, which slipped past the view's
+			// `refRange !== null` T1 filter and inflated the unlocated count
+			// (a phantom epoch signal); ladder clones must keep null]
+			const clone = { ...block, refRange: block.refRange !== null && block.refRange !== undefined ? { ...block.refRange } : null };
 			delete clone.surfaceSeqs;
 			delete clone.surfaceReplaced;
 			delete clone.surfaceSkip;