@devflow-core/dsh-devflow 0.6.0 → 0.6.2

package/assets/presets/devflow-2/README.md

@@ -1,20 +1,21 @@
 # DevFlow 2.0 — DeepSeek Harness Agent Preset
 
 A DSH agent preset that anchors the FIRST model request on the Minimal
-surface, then promotes into a full DevFlow session: Code Mode (PTC) on the
-wire, the complete DevFlow persona, and the DevFlow lifecycle guaranteed.
+surface, then promotes into a full DevFlow session: the native tool surface,
+the complete DevFlow persona, and the DevFlow lifecycle guaranteed.
 
 ## What it is
 
 - `agent.cordis.yml` — two-phase composition (adapted from the `liangshen`
   preset). Phase 1 shows only the one-line Minimal persona
-  ("You are a helpful software engineer assistant."), persistent
+  ("You are a helpful software engineer assistant."),
   `bash` + `str_replace_editor`, no runtime contexts, no injected pre-step
-  messages, and a 1024 output budget. The anchor gate (first minimal-like
-  reasoning block, 4-step fallback, or first-response release) then promotes
-  the session: the wire switches to Code Mode (a single `run_code` tool), the
-  full DevFlow persona is restored, and workspace instructions (AGENTS.md
-  digest) plus the skill catalog are deferred one step.
+  messages, and — since 2026-09-10 — no output-token cap. The anchor gate
+  (first minimal-like reasoning block, 4-step fallback, or first-response
+  release) then promotes the session: the wire switches to the deployment
+  default presentation (`native` — 33 tool schemas, no `run_code`, no generated
+  SDK section), the full DevFlow persona is restored, and workspace
+  instructions (AGENTS.md digest) plus the skill catalog are deferred one step.
 - `tool-bootstrap.mjs` — the two-phase bootstrap plugin (from
   `xiaobright/dsh-anchored-standard`, MIT, extended by `dsh-liangshen`),
   with a `phase1Persona` swap so the DevFlow persona only appears after
@@ -49,18 +50,48 @@ cp dsh/agent-presets/devflow-2/NOTICE                ~/.dsh/.agent-presets/devfl
 Start a new session in the web UI and pick **DevFlow 2.0** in the
 preset picker. The roster may need a refresh or restart to show a newly added
 preset. Existing sessions keep their old phase; the two-phase behavior applies
-to new sessions.
+to new sessions. The same holds for the 2026-09-10 prompt-budget changes: a
+running session keeps the policy it mounted with, so start a new session to
+pick them up.
 
 ## Caveats
 
 - `~/.dsh/.agent-presets/` is plugin-maintained: a DSH plugin upgrade may
   regenerate it. Re-run `npm run install:user -- --home ~/.dsh --write --force`
   (or re-copy) after upgrades.
-- Phase 1 uses the persistent (PTY-backed) Minimal `bash` shell. On Windows
-  builds where DSH's PTY backend is unavailable, phase-1 `bash` calls fail but
-  `str_replace_editor` still works, and the first-response release /
-  composition-drift guard keep the session usable (Code Mode `run_code` does
-  not need the PTY).
+- Phase 1 uses the Minimal `bash`: the persistent PTY-backed shell on
+  linux/darwin, or the stateless Git-Bash-backed `custom-bash.mjs` on Windows,
+  where the PTY backend is unavailable. `str_replace_editor` is the phase-1
+  editor on both paths, and the first-response release / composition-drift
+  guard keep the session usable if either bootstrap tool is missing.
+
+## Changes 2026-09-10 (prompt budget)
+
+Three independent defects were measured on real session logs and fixed:
+
+- **Phase-1 output cap removed.** `bootstrapMaxTokens: 1024` capped every
+  pre-promotion request, and output tokens INCLUDE reasoning tokens: under
+  `reasoningEffort: max` the first request was hard-truncated at 1024
+  (`finish_reason=length`, reasoning-only blocks with no text and no tool call),
+  which the web UI reports as "已达到输出 token 上限 / 回答被截断". The cap was
+  the anchor's forcing function for the short We-need block, not a safety limit;
+  the gate still promotes through its other three paths.
+- **`promotedPresentation: code` → `native`.** `code` does NOT mean "PTC only".
+  In `@deepseek-ai/dsh-tools`, `presentAs(mode)` attaches the collapse + SDK
+  sections for every mode except `native`, while `wireSchemas()` narrows the
+  wire to `run_code` only for `ptc`; `code` falls through to the default branch.
+  `code` therefore behaves as `both`: native schemas AND `run_code` AND the
+  generated SDK section — measured at 33,805 + 35,333 chars, i.e. one 36-tool
+  catalog paid for twice. `native` keeps the tools and drops the duplicate
+  (~-48% prompt per request). Use `ptc` instead to keep PTC batching at about
+  the same saving, at the cost of routing every call through a program.
+- **Three never-used tool families disabled** (0 calls across 30 recorded
+  sessions): `tool-jobs` (`job_kill`/`job_list`/`job_output`, 1,508 B + a
+  384-char guidance section that the same plugin contributes), the `workflow`
+  rows (4,067 B — the largest single schema), and `tool-ralph` (845 B). Each is
+  commented out in place with its restore note. The `subagent` family is kept
+  (it has real calls); `herdr_*` is kept because this deployment installs
+  `@deepseek-ai/dsh-tool-herdr` deliberately via `~/.dsh/cordis.patch.yml`.
 
 ## Customize
 

package/assets/presets/devflow-2/agent.cordis.yml

@@ -4,13 +4,18 @@
 #
 # TWO-PHASE PERSONA: the FIRST model request sees only the one-line Minimal
 # persona ("You are a helpful software engineer assistant."), the minimal tool
-# pair (persistent `bash` + `str_replace_editor`), NO runtime contexts and NO
+# pair (`bash` + `str_replace_editor`), NO runtime contexts and NO
 # auto-injected pre-step messages. After the anchor gate passes (first
 # minimal-like reasoning block, 4-step fallback, or first-response release) the
-# session promotes: the wire switches to Code Mode (PTC — a single `run_code`
-# tool), the FULL DevFlow persona below is restored (with {{model}}/{{cwd}}),
-# and workspace instructions (AGENTS.md digest) + the skill catalog are
-# deferred one step, then stay current.
+# session promotes: the wire switches to the deployment default presentation
+# (`native` — all tool schemas, no `run_code`, no generated SDK section), the
+# FULL DevFlow persona below is restored (with {{model}}/{{cwd}}), and
+# workspace instructions (AGENTS.md digest) + the skill catalog are deferred
+# one step, then stay current.
+#
+# The pre-promotion requests carry NO output-token cap (2026-09-10): the
+# former `bootstrapMaxTokens: 1024` truncated the first request under
+# `reasoningEffort: max`, because output tokens include reasoning tokens.
 #
 # This row MUST stay the FIRST row: the pre-step strip relies on waterfall
 # reverse-registration order so request #1 stays minimal-exact.
@@ -24,12 +29,33 @@
     anchorGate: true
     maxBootstrapSteps: 4
     promoteAfterFirstResponse: true
-    bootstrapMaxTokens: 1024
+    # DISABLED 2026-09-10: the phase-1 output cap collides with
+    # `reasoningEffort: max`. Output tokens INCLUDE reasoning tokens, so a
+    # 1024 cap hard-truncates the first request (finish_reason=length) and the
+    # web UI reports "已达到输出 token 上限 / 回答被截断". The cap is the
+    # anchor's forcing function for the short We-need block, not a safety
+    # limit: without it the gate still promotes via the anchor heuristic, the
+    # `maxBootstrapSteps` fallback, or `promoteAfterFirstResponse`.
+    # Re-enable only alongside a reasoningEffort low enough to fit under it.
+    # bootstrapMaxTokens: 1024
     # Core work set after a compaction, before re-promotion.
     compactionTools: [read, write, edit, glob, grep, todo_write, ask_user_question]
     deferredSources: [agent-instructions, skill-catalog]
     deferredGraceSteps: 1
-    promotedPresentation: code
+    # native (changed 2026-09-10, was `code`): `code` does NOT mean "PTC only".
+    # In @deepseek-ai/dsh-tools, presentAs(mode) attaches the collapse + SDK
+    # sections for every mode except `native`, while wireSchemas() narrows the
+    # wire to `run_code` ONLY for `ptc`; `code` falls through to the default
+    # branch, so it behaves as `both` — native schemas AND run_code AND the
+    # generated SDK section. Measured on a real session: 36 tool schemas
+    # (33,805 chars) plus a 35,333-char SDK section in the system prompt, i.e.
+    # one 36-tool catalog paid for twice (76,968 chars/request). `native` keeps
+    # the same 36 native tools and drops the duplicate SDK section
+    # (~40,065 chars/request, -48%). Trade-off: `run_code` disappears.
+    # Use `ptc` instead if the PTC batching surface is wanted: it saves about
+    # the same amount (drops the 33,805-char schema array instead) but forces
+    # every tool call through a program.
+    promotedPresentation: native
     # Aligned with liangshen 0.2.8 (issue #388): after promotion, replace the
     # full-text AGENTS.md injection with one non-imperative reference hint.
     instructionHint: true
@@ -66,7 +92,7 @@
 - id: persona
   name: '@deepseek-ai/dsh-persona'
   config:
-    text: |-
+    prefix: |-
       You are a coding agent powered by the {{model}} model. Your working directory is {{cwd}}.
 
       You run DevFlow. Before any development work — creating features, building components, adding functionality, modifying behavior, or defining an unapproved problem-directed change — load the `devflow-core` skill with the skill tool and read its `core-methods.md` reference, then load only the selected lifecycle owner skill. The route table and hard boundaries live in that skill and in AGENTS.md; do not invent lifecycle details. Pure Q&A, lookup, verification, and investigation-only reports need no lifecycle.
@@ -157,8 +183,11 @@
 # registry is keyed by owning agent anyway. With the persistent shell, long
 # commands run through `&`; these controls still expose whatever the host
 # registry tracks.
-- id: tool-jobs
-  name: '@deepseek-ai/dsh-tool-jobs'
+# DISABLED 2026-09-10 (prompt-budget trim): `job_kill` / `job_list` /
+# `job_output` were never called in 30 recorded sessions. Uncomment to restore
+# background-job control. Long commands still work through `&` in the shell.
+# - id: tool-jobs
+#   name: '@deepseek-ai/dsh-tool-jobs'
 
 # ── skills ──────────────────────────────────────────────────────────────────
 
@@ -307,19 +336,25 @@
         enableRunInBackground: false
         maxDepth: provider-managed
 
-    - id: workflow-worker-thread
-      name: '@deepseek-ai/dsh-workflow-worker-thread'
-      config:
-        provider: spawn
-
-    - id: tool-workflow
-      name: '@deepseek-ai/dsh-tool-workflow'
-
-    - id: tool-ralph
-      name: '@deepseek-ai/dsh-tool-ralph'
-      config:
-        subagentProvider: spawn
-        maxRounds: 64
+    # DISABLED 2026-09-10 (prompt-budget trim): the `workflow` schema is the
+    # single largest tool description in this catalog (4,067 chars) and was
+    # never called in 30 recorded sessions. Uncomment both rows to restore
+    # large multi-agent orchestration.
+    # - id: workflow-worker-thread
+    #   name: '@deepseek-ai/dsh-workflow-worker-thread'
+    #   config:
+    #     provider: spawn
+    #
+    # - id: tool-workflow
+    #   name: '@deepseek-ai/dsh-tool-workflow'
+
+    # DISABLED 2026-09-10 (prompt-budget trim): `ralph` was never called in 30
+    # recorded sessions. Uncomment to restore fresh-agent Ralph loops.
+    # - id: tool-ralph
+    #   name: '@deepseek-ai/dsh-tool-ralph'
+    #   config:
+    #     subagentProvider: spawn
+    #     maxRounds: 64
 
 # ── remaining model-facing rows ─────────────────────────────────────────────
 

package/assets/presets/devflow-2/preset.yml

@@ -1,3 +1,3 @@
 name: DevFlow 2.0
-description: DevFlow 开发工作流 + 梁神式两阶段锚定：首个请求只暴露一行 persona（You are a helpful software engineer assistant.）+ 最小工具对（持久 bash + str_replace_editor），无运行时上下文与任何注入；anchorGate 门控（首个 minimal-like 推理块或 4 步兜底）+ 1024 输出预算稳定锚定后，晋升为 Code Mode（单一 run_code），恢复完整 DevFlow 人设并自动激活 devflow-core 生命周期（Brainstorm → Spec/Cut → Plan → Build → Prove）与验证完成闭环；workspace 指令与技能目录延迟一步注入。
+description: DevFlow 开发工作流 + 梁神式两阶段锚定：首个请求只暴露一行 persona（You are a helpful software engineer assistant.）+ 最小工具对（bash + str_replace_editor），无运行时上下文、无任何注入、无输出 token 上限；anchorGate 门控（首个 minimal-like 推理块或 4 步兜底）晋升后切换为 native 工具面（33 个工具，无 run_code、无 SDK 重复表达），恢复完整 DevFlow 人设并自动激活 devflow-core 生命周期（Brainstorm → Spec/Cut → Plan → Build → Prove）与验证完成闭环；workspace 指令与技能目录延迟一步注入。
 order: 6

package/assets/presets/devflow-2/tool-bootstrap.mjs

@@ -29,10 +29,16 @@
  * `we` and no `let me`) or the `maxBootstrapSteps` fallback.
  * `promoteAfterFirstResponse` promotes a tool-less first response once it has
  * responded, and also releases an anchor-gated session when its first turn
- * ends (`turn/end`). With `promotedPresentation: code` the promoted catalog
- * is presented as Code Mode (PTC): the wire shows a single `run_code` tool
- * backed by the generated SDK, switched at the step boundary so the current
- * step's native calls are never interrupted. `deferredSources` and
+ * ends (`turn/end`). `promotedPresentation` is passed verbatim to
+ * `tools.presentAs(mode)`, so the promoted wire depends on which mode string
+ * the preset names: `native` = every tool schema and no SDK section; `ptc` =
+ * the `run_code` transport plus the generated SDK; `both`, and any other
+ * string — including `code`, which this preset used before 2026-09-10 and
+ * which `dsh-tools` falls through to the default branch — = every tool schema
+ * AND `run_code` AND the SDK section, i.e. one catalog paid for twice. Only
+ * `ptc` narrows `wireSchemas()` to the `run_code` transport. The switch
+ * happens at the step boundary so the current step's native calls are never
+ * interrupted. `deferredSources` and
  * `deferredGraceSteps` delay selected injected message kinds (workspace
  * instructions, skill catalog) for a few steps after promotion.
  *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@devflow-core/dsh-devflow",
-  "version": "0.6.0",
+  "version": "0.6.2",
   "description": "DevFlow for DeepSeek Harness: devflow-2 agent preset + skills + commands + verification scripts, synced into ~/.dsh on host startup.",
   "type": "module",
   "engines": {