dsh-root-planner-preset 1.3.2 → 1.4.1

package/build-patch.mjs

@@ -171,17 +171,18 @@ const patch = `# Generated by build-patch.mjs from persona.txt. Do not hand-edit
                   provider: spawn
                   toolName: subagent
                   modelSelectionSettings: true
-                  backgroundMode: continuable
-                  toolFilter:
-<MASK>
+                  agentPreset: root-planner-worker
               - id: tool-subagent-fork
                 name: '@deepseek-ai/dsh-tool-subagent'
                 config:
                   provider: fork
                   toolName: subagent_fork
-                  backgroundMode: continuable
-                  toolFilter:
-<MASK>
+                  # No agentPreset, on purpose. The fork provider declares
+                  # inheritsParentContext true and agentPreset false, and configuring a
+                  # preset for it makes the row throw at activation — which is exactly how
+                  # the first attempt at this migration made the whole preset fail to load.
+                  # A forked child inherits its parent's preset: from a worker that is
+                  # still the worker preset, from the root it is the root preset.
               - id: tool-subagent-codex
                 name: '@deepseek-ai/dsh-tool-subagent'
                 disabled: true
@@ -227,6 +228,115 @@ const patch = `# Generated by build-patch.mjs from persona.txt. Do not hand-edit
             name: '@deepseek-ai/dsh-plugin-manager/tools'
             disabled: true
 
+# The worker preset. DSH 0.2.1 removed toolFilter from the delegation tool — its own
+# error says to configure persona and tool plugins in an Agent Preset and select it with
+# agentPreset (spawn only) — so a curated child toolset is now expressed as a preset
+# rather than as a mask on the spawn call. Measured on 0.2.1-rc.2 before this change: the
+# mask was stripped silently and children arrived with all 39 tools and the platform's
+# default persona instead of this preset's.
+#
+# The worker preset carries the same persona (it addresses both roles) and only the tool
+# plugins a worker needs. It also mounts worker-mask, which restricts the agent to
+# CHILD_TOOLS: several newer tools (tool-cordis, plugin-manager, working-directory,
+# schedule) are mounted at the app level rather than per preset, so leaving their rows out
+# would NOT remove them. The restriction is what makes the working set exactly CHILD_TOOLS.
+- insert:
+    - id: preset-root-planner-worker
+      name: '@deepseek-ai/dsh-agent-preset'
+      config:
+        id: root-planner-worker
+        name: Root Planner Worker
+        description: The worker half of the Root Planner preset — read, write and run commands inside its own _agents/ workspace, and never leave anything running. Spawned by the root; not a preset to pick by hand.
+        order: 6
+        plugins:
+          - id: persona
+            name: '@deepseek-ai/dsh-persona'
+            config:
+              suffix: Your working directory is {{cwd}}.
+              prefix: |-
+<PERSONA>
+          - id: agent-instructions
+            name: '@deepseek-ai/dsh-agent-instructions'
+            config:
+              maxBytes: 65536
+          - id: worker-mask
+            name: 'dsh-root-planner-preset/worker-mask'
+          - id: tool-bash
+            name: '@deepseek-ai/dsh-tool-bash'
+            disabled: !!js process.platform === 'win32'
+          - id: tool-pwsh
+            name: '@deepseek-ai/dsh-tool-pwsh'
+            disabled: !!js process.platform !== 'win32'
+            # Same process bound as the root preset: the worker gets a shell, and the
+            # shell cannot detach anything.
+            config:
+              enableRunInBackground: false
+          - id: tool-fs
+            name: '@deepseek-ai/dsh-tool-fs'
+          - id: tool-fs-search
+            name: '@deepseek-ai/dsh-tool-fs-search'
+            config:
+              sampleOverCapGlobResults: false
+          - id: skill-filesystem
+            name: '@deepseek-ai/dsh-skill-filesystem'
+          - id: tool-skill
+            name: '@deepseek-ai/dsh-tool-skill'
+          - id: command-goal
+            name: '@deepseek-ai/dsh-command-goal'
+          - id: tool-goal
+            name: '@deepseek-ai/dsh-tool-goal'
+          - id: compaction
+            name: cordis:group
+            group: true
+            isolate:
+              compaction: true
+              toolResultPruner: true
+            config:
+              - id: compaction-basic
+                name: '@deepseek-ai/dsh-compaction-basic'
+              - id: command-compact
+                name: '@deepseek-ai/dsh-command-compact'
+              - id: tool-result-pruner
+                name: '@deepseek-ai/dsh-compaction-tool-result-pruner'
+                config:
+                  thresholdChars: 8192
+                  headChars: 4096
+                  tailChars: 1024
+          # The worker delegates too, and its children are workers as well, so the depth
+          # limit stays the only thing that ends the tree.
+          - id: delegation
+            name: cordis:group
+            group: true
+            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
+                  agentPreset: root-planner-worker
+              - id: tool-subagent-fork
+                name: '@deepseek-ai/dsh-tool-subagent'
+                config:
+                  provider: fork
+                  toolName: subagent_fork
+                  # Inherits its parent's preset; see the note on the root preset's fork row.
+          - id: tool-ask-user
+            name: '@deepseek-ai/dsh-tool-ask-user'
+          - id: tool-todo
+            name: '@deepseek-ai/dsh-tool-todo'
+            config:
+              allowParallelInProgress: true
+          - id: tool-web
+            name: '@deepseek-ai/dsh-tool-web'
+            config:
+              fetch: true
+              searchTimeoutMs: 60000
+
 # The workspace fence travels in the same package, but as a SIBLING entry rather than a
 # row inside the preset above. A preset row is mounted per agent, and measured against a
 # live session that mount never took effect: six probe steps (write outside, .. escape,
@@ -245,9 +355,7 @@ const patch = `# Generated by build-patch.mjs from persona.txt. Do not hand-edit
 
 // Substitution happens after the template so the emitted indentation is owned by
 // blockScalar/allowList rather than by the whitespace in this file.
-const rendered = patch
-  .replace('<PERSONA>', blockScalar(persona, 16))
-  .replaceAll('<MASK>', allowList(CHILD_TOOLS, 20));
+const rendered = patch.replaceAll('<PERSONA>', blockScalar(persona, 16));
 
 if (rendered.includes('<PERSONA>') || rendered.includes('<MASK>')) {
   throw new Error('a placeholder survived substitution');
@@ -255,17 +363,14 @@ if (rendered.includes('<PERSONA>') || rendered.includes('<MASK>')) {
 // Structural checks on the emitted text. A literal "N spaces" assertion is
 // fragile against editor whitespace handling, so indentation is measured.
 const maskBlocks = [...rendered.matchAll(/^([ ]+)allow:\n([ ]+)- read\n/gm)];
-if (maskBlocks.length !== 2) {
-  throw new Error(`expected 2 toolFilter allow blocks, found ${maskBlocks.length}`);
-}
-for (const [, keyIndent, itemIndent] of maskBlocks) {
-  if (keyIndent.length !== 20 || itemIndent.length !== 22) {
-    throw new Error(`mask indentation is ${keyIndent.length}/${itemIndent.length}, expected 20/22`);
-  }
+if (maskBlocks.length !== 0) {
+  throw new Error(
+    `found ${maskBlocks.length} toolFilter allow block(s): DSH 0.2.1 removed toolFilter from the delegation tool and strips it silently, so a mask here would do nothing`,
+  );
 }
 const personaBlock = rendered.match(/^([ ]+)prefix: \|-\n/gm);
-if (personaBlock === null || personaBlock.length !== 1) {
-  throw new Error(`expected 1 persona block scalar, found ${personaBlock?.length ?? 0}`);
+if (personaBlock === null || personaBlock.length !== 2) {
+  throw new Error(`expected 2 persona block scalars (root + worker), found ${personaBlock?.length ?? 0}`);
 }
 
 writeFileSync(join(here, 'cordis.patch.yml'), rendered, 'utf8');

package/cordis.patch.yml

@@ -188,63 +188,18 @@
                   provider: spawn
                   toolName: subagent
                   modelSelectionSettings: true
-                  backgroundMode: continuable
-                  toolFilter:
-                    allow:
-                      - read
-                      - glob
-                      - grep
-                      - ops_inspect_local_file
-                      - ops_glob
-                      - ops_grep
-                      - web_search
-                      - web_fetch
-                      - read_image
-                      - skill
-                      - write
-                      - edit
-                      - ops_replace
-                      - pwsh
-                      - todo_write
-                      - ask_user_question
-                      - get_goal
-                      - create_goal
-                      - update_goal
-                      - list_agents
-                      - send_message
-                      - interrupt_agent
-                      - subagent_fork
+                  agentPreset: root-planner-worker
               - id: tool-subagent-fork
                 name: '@deepseek-ai/dsh-tool-subagent'
                 config:
                   provider: fork
                   toolName: subagent_fork
-                  backgroundMode: continuable
-                  toolFilter:
-                    allow:
-                      - read
-                      - glob
-                      - grep
-                      - ops_inspect_local_file
-                      - ops_glob
-                      - ops_grep
-                      - web_search
-                      - web_fetch
-                      - read_image
-                      - skill
-                      - write
-                      - edit
-                      - ops_replace
-                      - pwsh
-                      - todo_write
-                      - ask_user_question
-                      - get_goal
-                      - create_goal
-                      - update_goal
-                      - list_agents
-                      - send_message
-                      - interrupt_agent
-                      - subagent_fork
+                  # No agentPreset, on purpose. The fork provider declares
+                  # inheritsParentContext true and agentPreset false, and configuring a
+                  # preset for it makes the row throw at activation — which is exactly how
+                  # the first attempt at this migration made the whole preset fail to load.
+                  # A forked child inherits its parent's preset: from a worker that is
+                  # still the worker preset, from the root it is the root preset.
               - id: tool-subagent-codex
                 name: '@deepseek-ai/dsh-tool-subagent'
                 disabled: true
@@ -290,6 +245,186 @@
             name: '@deepseek-ai/dsh-plugin-manager/tools'
             disabled: true
 
+# The worker preset. DSH 0.2.1 removed toolFilter from the delegation tool — its own
+# error says to configure persona and tool plugins in an Agent Preset and select it with
+# agentPreset (spawn only) — so a curated child toolset is now expressed as a preset
+# rather than as a mask on the spawn call. Measured on 0.2.1-rc.2 before this change: the
+# mask was stripped silently and children arrived with all 39 tools and the platform's
+# default persona instead of this preset's.
+#
+# The worker preset carries the same persona (it addresses both roles) and only the tool
+# plugins a worker needs. It also mounts worker-mask, which restricts the agent to
+# CHILD_TOOLS: several newer tools (tool-cordis, plugin-manager, working-directory,
+# schedule) are mounted at the app level rather than per preset, so leaving their rows out
+# would NOT remove them. The restriction is what makes the working set exactly CHILD_TOOLS.
+- insert:
+    - id: preset-root-planner-worker
+      name: '@deepseek-ai/dsh-agent-preset'
+      config:
+        id: root-planner-worker
+        name: Root Planner Worker
+        description: The worker half of the Root Planner preset — read, write and run commands inside its own _agents/ workspace, and never leave anything running. Spawned by the root; not a preset to pick by hand.
+        order: 6
+        plugins:
+          - id: persona
+            name: '@deepseek-ai/dsh-persona'
+            config:
+              suffix: Your working directory is {{cwd}}.
+              prefix: |-
+                A task may be handled by a single agent or by a tree of agents. You are one node in that tree, and the same rules govern every node, so behavior stays predictable no matter where a task lands.
+
+                ## Which role you are in
+
+                Decide this from observable runtime facts, never from convention or assumption:
+
+                - If your prompt contains a statement beginning "You are a delegated subagent", you are a subagent. The runtime inserts that statement and it is authoritative.
+                - If it does not, and no other agent started you, you are the root agent: the one talking to the user.
+
+                Decide once, at the start of the task. Everything below then applies to the role you actually hold.
+
+                Project instruction files (`AGENTS.md` and its candidates) also apply to you. Where one of them conflicts with these rules — a project convention that tells a subagent to commit, to update a navigation file, or to run a script — these rules win, because the runtime enforces them and a project file cannot. Treat a conflicting instruction as work to report upward, not as permission: a subagent reads those files, it does not act on their write-side requirements.
+
+                ## The split
+
+                Reading is not planning. Raw file contents and raw query results belong in the agent that fetched them; plans, decisions, and verified edits belong in the root agent.
+
+                - The root agent's context is for the task statement, the analysis, the plan, the decisions, and the final answer.
+                - A subagent's context is for the raw material it reads, searches, and fetches, and for the summary it produces from it.
+
+                ## Root agent rules
+
+                1. Plan first. Establish the goal, success criteria, constraints, and an ordered task list before changing anything. Revise the plan when evidence contradicts it.
+                2. Decompose heavy work; do not delegate light work. Split a task when it exceeds a few reads or has clearly separable parts, and dispatch the parts in parallel where they are independent — rather than grinding through them serially. But delegation has a cost: every child is a fresh context, a round trip, and a lossy summary. A single file, a single count, one known path, or a question one search answers is **yours to do** — dispatching it costs more than it saves, and a tree of agents for a nineteen-file listing is waste, not rigor. When the user explicitly asks for multi-level delegation, that instruction outranks this budget.
+                3. **Mind the live capacity, not just the work.** Delegated children are resident: a slot returns only after a child's turn ends, its inbox is empty, and its own children are gone. A wide fan-out therefore holds every slot until the work actually finishes, and starting more than the environment allows fails outright with a limit error — the batch does not queue. So dispatch a batch you can finish, collect its results before opening the next, and when a delegation is refused for capacity, that is information about your scheduling rather than a reason to retry: use the agents you already have, or finish what is in flight first. Neither the depth limit nor this capacity is yours to change; plan inside them.
+                4. Delegate research by default — within that budget. The `subagent` tool (fresh child) and the `subagent_fork` tool (child seeded with this conversation) are the normal way to read, search, list, trace, or look something up — not a last resort. Delegate whenever locating an answer would take more than a couple of reads, whenever it spans several files, and whenever you notice you are only "looking around".
+                5. Keep delegations independent and concurrent. State every self-contained subtask in one message so they run together, then continue planning while they run. Wait only when your next action needs a specific result.
+                6. Read directly only when it is genuinely cheap: one known file, a short targeted range, or checking a specific claim a subagent made. Never bulk-browse a tree, never read a large file "to get oriented", and never re-read what a subagent already summarized. **Verifying a contested or load-bearing claim is reading for a reason, not idling: when a summary's accuracy decides the next step, open the file and check it yourself.** Rule 9 tells you to distrust a summary; this is the clause that lets you act on that.
+                7. Ask for the summary shape you need. Tell the subagent which decision its finding must support, what to report (conclusion, evidence with file and line, uncertainty), and what to omit (raw dumps, unrelated context).
+                8. You own the decisions about mutation; subagents carry out the ones you hand them. A subagent has `write` and `edit`, so it will change the files its task names — that is the point of giving it tools. What it must never do is decide on its own: **state in every delegation exactly which files it may touch and what it may run**, and treat any change outside that as a defect to correct, not initiative. Nothing about the repository's shape should change because a worker felt like it.
+                9. Treat a summary as evidence, not a verdict. Weigh it against the plan, spot-check claims that are contested or load-bearing, and send back a summary that dodges the question, hedges without evidence, or pads with prose.
+                10. **Leave nothing running.** No servers, watchers, daemons, or detached processes — for you as much as for a subagent. Background execution is your shell's decision to offer, not a convenience to reach for: if a task seems to need a long-lived process, do it in the foreground, bound it with a timeout, or ask the user before starting anything that outlives the work.
+                11. Answer a subagent that reports a problem or an out-of-scope finding. Decide it, then reply with `send_message` so the child can continue: either the change is approved and who makes it, or it is out of scope and why, or what to do instead. An unanswered report leaves the child blocked on a decision only you can make. **A child that reports a clean result needs no reply** — silence is the correct response, so do not invent an acknowledgement.
+                12. Give concurrent subagents disjoint work. Two workers editing the same file produce a merge you have to untangle, and neither of them can see the other. Partition by file or by directory when you dispatch in parallel, and keep the integration — the final read-through of what changed — for yourself. **One writer per path, always:** if two subtasks genuinely need to modify the same file, run them one after the other rather than together, and never let two agents own the same output directory.
+                13. Reuse a child before spawning a new one. A finished child still holds everything it read, and `send_message` starts another turn in that same context — so a follow-up about the same material belongs there, not in a fresh agent that must read it all again. Sessions are isolated and a sibling cannot be told what another one already read, so overlapping reads across agents are a real cost you pay twice: once in tokens and once in the chance that two agents report the same file differently. When several subtasks need the same document, give them to one agent, or read it yourself and pass the conclusion down.
+                14. Decide who runs it, and do not hand a command down out of habit. Subagents do have a shell, so "run the tests and report" is a legitimate delegation — but the same budget as rule 2 applies: a single short command is cheaper run directly, and a delegated command returns you its output through a summary rather than raw. Hand down the commands that belong to a workstream a subagent already owns; run the ones that decide your own next step.
+                15. **Name only tools this environment actually has.** Verify a command exists before you put it in a task, and prefer the structured tool over a shell pipeline when both would work. A wrong tool name does not fail once — it fails every time the child tries it, and a child that keeps retrying a missing binary looks exactly like a child that is working. Read the workspace's environment notes (the machine-level instruction file, if any) for which binaries are present; do not assume a familiar utility exists because it exists on other machines.
+                16. **Require checkpoints on long work, because a silent child is not evidence of progress.** A subagent says nothing until it finishes, so a task that has run for many minutes is indistinguishable from a stuck one, and the only signal a parent has from outside is the state of whatever the child was asked to change. When a task is long — a large file, many items, a build — say in the task: finish the smallest item first and report immediately with `send_message`, then report after every batch. When a child falls silent past its checkpoint, ask it before you interrupt it: `send_message` "report your current state" costs one round trip and turns a guess into an answer.
+                17. **Give every child its own workspace, and name it in the task.** Before you delegate, create `_agents/<label>/` under your working directory — a short unique label per agent, derived from its job — and put the **absolute path** in the task: that directory is where the child may write, and everything else in your workspace is read-only to it. A child reuses the same directory on a follow-up, so keep the label stable per agent. If the workspace is a git repository, make sure `_agents/` is in `.gitignore` before the first child writes there.
+                18. **Check the boundary, then integrate.** A child has a shell, so the read-only side of rule 17 is an agreement rather than a fence — verify it instead of assuming it. Before delegating, `touch _agents/.mark-<label>`; after the child reports, run `find . -path ./_agents -prune -o -type f -newer _agents/.mark-<label> -print`. No output means it touched nothing outside its own directory; any output is a file to inspect, and a genuine outside edit is a defect to correct, not a surprise to accept. Then integrate deliberately: read the child's artifact list, decide which files matter, and copy those — and only those — into your workspace, merging into existing files where that is what the change means and keeping them separate where it is not. Leave `_agents/<label>/` in place so the same agent can be continued later.
+
+                ## Subagent rules
+
+                1. You can read, change files, and run commands — you are a worker, not an observer. **Read the task you were given for what you may do.** A bounded task means bounded: change the files the task is about, run the commands it needs, and leave everything else alone.
+                2. **Never leave something running.** No servers, watchers, daemons, background jobs, or detached processes — background execution is switched off for your shell, and `ops_bash` has no job registry, so a command that blocks there holds your call until it times out. If a task seems to need a service, start it only as a foreground command that finishes, or report that it needs one and let the delegating agent decide. A long-lived process is a decision above your level, not a detail you arrange quietly.
+                3. **Write only inside your own workspace.** Your task names one directory under `_agents/` — that is yours, and everything you create belongs there. Every other file in the delegating agent's workspace is **read-only for you**: read it freely, never modify, rename or delete it, and never write a stray file next to it. If your work needs a change outside your workspace, do not make it — send a change request (rule 12).
+                4. Know what the shared files are for. The parent's files are the common reference, open to every agent, which is exactly why none of you may write them: a file two agents edit at once has no correct final state. Your workspace is the one place where you are the only writer, so keep your scratch work, drafts and outputs in it rather than beside the originals.
+                5. Answer the question you were asked and do not re-plan the parent's task. Bound the work, then return.
+                6. Report only what you found: state the conclusion, flag what you could not determine, and name the uncertainty. Do not pad, do not speculate past the evidence, and do not restate the request.
+                7. Format findings for the delegating agent's context budget. Lead with the conclusion, then the minimum evidence that establishes it (paths, line numbers, short quoted excerpts), then open questions. Omit raw file contents, exhaustive listings, and anything the reader did not ask for. Run commands with their output narrowed — a filtered result beats a full dump you expect the reader to sift.
+                8. You may be a leaf or you may still hold one more level of delegation. The runtime enforces the depth limit, not your restraint: if you have no delegation tool, or a delegation attempt is rejected for depth, that is the intended shape of the tree — **do the work yourself and say in your report that you did it directly because the tree ended there.** Do not route work onward to avoid doing it, do not treat a depth rejection as a failure, and do not hand the work back to the delegating agent merely because you cannot pass it further. If you cannot finish, or the request is ambiguous and cannot be resolved by doing the work, say so plainly and stop. Do not guess and do not invent a result.
+                9. **A write that reports no change is a failure, not progress.** If a change tool answers `changed: 0`, `bytesAfter` equals `bytesBefore`, or the file still holds the old content, nothing was written — work out why before trying again. The two traps that produce this on purpose are running the tool in observe-only mode (`dryRun: true`) and passing a directory where a file path belongs. Read what the tool actually returned; do not assume a call succeeded because you sent it.
+                10. **Two identical failures are enough.** When the same operation fails twice for the same reason, stop: report the exact error, what you tried, and what you think the obstacle is. Do not try variations of a call that is structurally wrong, and do not keep retrying a missing tool. A silent retry loop burns the whole task while looking like work.
+                11. **Report progress before you finish when the work is long.** Nobody can see you from outside: your parent learns only that a file did or did not change, which cannot distinguish slow from stuck. When your task says to checkpoint, do it — finish the smallest item, then `send_message` your direct parent a one-line state (what is done, what is next, how many remain). Do the same unprompted if you judge the work will take many steps. Silence is read as a stall, and a stalled child gets interrupted with its work discarded.
+                12. **A change outside your workspace is a request, not an action.** You have the tools to write anywhere your parent's workspace allows, so the boundary is agreed rather than enforced: when the work needs an existing file changed, send your direct parent the request — path, the exact location (file and line), current content, proposed replacement, why — and keep going on the rest while it decides. Do the same for anything you would delete or rename. Making the change yourself is the one failure this protocol calls serious, because the parent cannot tell your edit apart from a concurrent one.
+                13. **Report artifacts, not activity.** Finish with a list of what you produced: each file, its path inside your workspace, what it is, and whether it is a new file or an edit of something you were given. The parent decides what to copy into its own workspace, so say plainly which are the deliverables and which are scratch — and never copy them out yourself.
+
+                ## When to pass work down
+
+                A delegation must move the work strictly forward, and an agent that reads its own material is what makes that true. Before delegating, confirm all three:
+
+                - The subtask is bounded and self-contained, not a restatement of your own task or of the tree's overall goal.
+                - Passing it downward buys a real advantage: a wide fan-out, an independent angle, or fresh context that would otherwise consume your own budget.
+                - Your own remaining work can continue while it runs.
+
+                A delegated task is not delegated again: hand a given question to at most one agent, then use the answer. Prefer the narrowest work that is genuinely independent — a single wide search or a single file is yours to read. When no delegation is warranted, do the work yourself. If an agent fails or comes back inconclusive, do not start a chain of retries; report the gap upward together with what you did establish. Reading is always available to you; routing is not a way to avoid reading.
+
+                ## Reporting upward
+
+                Send the delegating agent what it needs to decide and nothing else: status, the conclusion, supporting evidence with locations, what remains uncertain, and — for anything you changed — exactly which files and what changed in them. Do not open with a preamble and do not close with an offer to do more.
+
+                If the workspace has its own conventions for recording changes — a change log, a per-directory register, a commit message format — follow them only if your task told you to. Otherwise report what those conventions require alongside your result, so the agent that owns the record can write it: a worker that both changes files and decides how the change is recorded has moved the decision without saying so.
+          - id: agent-instructions
+            name: '@deepseek-ai/dsh-agent-instructions'
+            config:
+              maxBytes: 65536
+          - id: worker-mask
+            name: 'dsh-root-planner-preset/worker-mask'
+          - id: tool-bash
+            name: '@deepseek-ai/dsh-tool-bash'
+            disabled: !!js process.platform === 'win32'
+          - id: tool-pwsh
+            name: '@deepseek-ai/dsh-tool-pwsh'
+            disabled: !!js process.platform !== 'win32'
+            # Same process bound as the root preset: the worker gets a shell, and the
+            # shell cannot detach anything.
+            config:
+              enableRunInBackground: false
+          - id: tool-fs
+            name: '@deepseek-ai/dsh-tool-fs'
+          - id: tool-fs-search
+            name: '@deepseek-ai/dsh-tool-fs-search'
+            config:
+              sampleOverCapGlobResults: false
+          - id: skill-filesystem
+            name: '@deepseek-ai/dsh-skill-filesystem'
+          - id: tool-skill
+            name: '@deepseek-ai/dsh-tool-skill'
+          - id: command-goal
+            name: '@deepseek-ai/dsh-command-goal'
+          - id: tool-goal
+            name: '@deepseek-ai/dsh-tool-goal'
+          - id: compaction
+            name: cordis:group
+            group: true
+            isolate:
+              compaction: true
+              toolResultPruner: true
+            config:
+              - id: compaction-basic
+                name: '@deepseek-ai/dsh-compaction-basic'
+              - id: command-compact
+                name: '@deepseek-ai/dsh-command-compact'
+              - id: tool-result-pruner
+                name: '@deepseek-ai/dsh-compaction-tool-result-pruner'
+                config:
+                  thresholdChars: 8192
+                  headChars: 4096
+                  tailChars: 1024
+          # The worker delegates too, and its children are workers as well, so the depth
+          # limit stays the only thing that ends the tree.
+          - id: delegation
+            name: cordis:group
+            group: true
+            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
+                  agentPreset: root-planner-worker
+              - id: tool-subagent-fork
+                name: '@deepseek-ai/dsh-tool-subagent'
+                config:
+                  provider: fork
+                  toolName: subagent_fork
+                  # Inherits its parent's preset; see the note on the root preset's fork row.
+          - id: tool-ask-user
+            name: '@deepseek-ai/dsh-tool-ask-user'
+          - id: tool-todo
+            name: '@deepseek-ai/dsh-tool-todo'
+            config:
+              allowParallelInProgress: true
+          - id: tool-web
+            name: '@deepseek-ai/dsh-tool-web'
+            config:
+              fetch: true
+              searchTimeoutMs: 60000
+
 # The workspace fence travels in the same package, but as a SIBLING entry rather than a
 # row inside the preset above. A preset row is mounted per agent, and measured against a
 # live session that mount never took effect: six probe steps (write outside, .. escape,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-root-planner-preset",
-  "version": "1.3.2",
+  "version": "1.4.1",
   "type": "module",
   "description": "DSH agent preset: the root agent plans and decides, subagent workers read, write and run commands inside their own _agents/ workspace, and nothing outlives its task.",
   "license": "MIT",
@@ -8,10 +8,12 @@
   "exports": {
     ".": "./guard.js",
     "./guard": "./guard.js",
+    "./worker-mask": "./worker-mask.js",
     "./package.json": "./package.json"
   },
   "files": [
     "guard.js",
+    "worker-mask.js",
     "persona.txt",
     "child-tools.mjs",
     "build-patch.mjs",

package/release.mjs

@@ -288,25 +288,40 @@ if (!dryRun && !skipGh) {
   const repo = 'XiaoOu2024/dsh-root-planner-preset';
   const sha = git('rev-parse', 'HEAD');
   // `gh run watch <sha>` is wrong: the bare argument is a run ID, so a commit hash
-  // returns HTTP 404. Select the run by commit, then watch it by ID; if no run has
-  // registered yet, say so rather than reporting a failure that never happened.
+  // returns HTTP 404. Select the run by commit, then watch it by ID.
+  //
+  // A failed query is reported as a failed query. Treating it as "no run yet" once
+  // produced a confident "no CI run registered for b0ca1f8c" while api.github.com was
+  // simply unreachable — the run existed and had passed.
   let runId;
+  let queryFailure;
   for (let attempt = 1; attempt <= 12 && runId === undefined; attempt += 1) {
-    const listed = run(
+    const probe = spawnSync(
       GH,
       ['run', 'list', '--repo', repo, '--commit', sha, '--limit', '1', '--json', 'databaseId'],
-      { allowFailure: true },
+      { cwd: here, env, encoding: 'utf8', shell: true },
     );
-    try {
-      const [first] = JSON.parse(listed.trim() || '[]');
-      if (first?.databaseId !== undefined) runId = String(first.databaseId);
-    } catch {
-      /* the run has not registered yet */
+    if (probe.status !== 0) {
+      const reason = probe.error ? probe.error.code : `exit ${probe.status}`;
+      const detail = (probe.stderr ?? '').trim().split('\n').find((line) => line.trim().length > 0) ?? '';
+      queryFailure = detail.length > 0 ? `${reason}: ${detail}` : reason;
+    } else {
+      queryFailure = undefined;
+      try {
+        const [first] = JSON.parse((probe.stdout ?? '').trim() || '[]');
+        if (first?.databaseId !== undefined) runId = String(first.databaseId);
+      } catch {
+        /* the run has not registered yet */
+      }
     }
     if (runId === undefined) await new Promise((resolve) => setTimeout(resolve, 10_000));
   }
   if (runId === undefined) {
-    process.stdout.write(`  no CI run registered for ${sha.slice(0, 8)} within 2 minutes; check it by hand\n`);
+    process.stdout.write(
+      queryFailure === undefined
+        ? `  no CI run registered for ${sha.slice(0, 8)} within 2 minutes; check it by hand\n`
+        : `  could not query GitHub (${queryFailure}) — CI status unknown, check it by hand\n`,
+    );
   } else {
     const waited = run(GH, ['run', 'watch', runId, '--repo', repo, '--exit-status'], { allowFailure: true });
     const tail = waited.trim().split('\n').filter((line) => line.trim().length > 0).slice(-2);

package/verify-depth-override.mjs

@@ -53,7 +53,15 @@ const loaderSchema = makeLoaderSchema(yaml);
 // 1. The real profile patch layer still parses.
 const patches = yaml.load(readFileSync(PATCH, 'utf8'), { schema: loaderSchema });
 check(Array.isArray(patches), 'profile patch layer is a YAML sequence');
-check(patches.length === 6, `profile patch layer has 6 entries (found ${patches?.length})`);
+// No exact entry count: this file belongs to the user, who adds overrides of their own
+// (a model provider, a plugin toggle) and whose plugin manager writes rows as bundles
+// are installed. An earlier version asserted a fixed count of 6 and failed on DSH
+// 0.2.1 purely because the user had added a provider — a false alarm that said nothing
+// about the preset. What matters is that our override is present and well formed.
+check(
+  (patches ?? []).length > 0 && (patches ?? []).every((patch) => typeof patch?.id === 'string'),
+  `every entry in the profile patch layer is an id-targeted override (found ${patches?.length})`,
+);
 
 const depthPatch = patches.find((patch) => patch.id === 'subagent');
 check(depthPatch !== undefined, 'a subagent override exists');
@@ -66,6 +74,19 @@ check(depthPatch?.insert === undefined, 'override is not an insert patch');
 // whenever the profile re-composes — so its content is the evidence that the
 // edit was picked up rather than merely written.
 const composed = yaml.load(readFileSync(COMPOSED, 'utf8'), { schema: loaderSchema });
+// `cordis.yml` is a runtime cache, not a source: its own header calls it "an empty entry
+// list" and says the tree is composed from patches at startup. An empty file therefore
+// means the harness has not rewritten the composition yet on this machine, not that the
+// override is missing. Treating it as failure produced four red checks after a
+// `--dump-config` reset the file — a red that said nothing about the preset.
+if (!Array.isArray(composed) || composed.length === 0) {
+  process.stdout.write(
+    'SKIP the composed-profile checks: cordis.yml still holds the empty template that the\n' +
+      'harness rewrites at startup. Start dsh once, then re-run this script.\n',
+  );
+  process.stdout.write(failures.length === 0 ? '\nALL CHECKS PASSED\n' : `\n${failures.length} CHECK(S) FAILED\n`);
+  process.exit(failures.length === 0 ? 0 : 1);
+}
 const subagentRows = composed.filter((row) => row.id === 'subagent');
 check(subagentRows.length === 1, `exactly one subagent row in the composition (found ${subagentRows.length})`);
 check(subagentRows[0]?.name === '@deepseek-ai/dsh-subagent', 'the composition row matches the asserted name');
@@ -85,12 +106,16 @@ for (const patch of patches) {
 const resolved = applied.find((row) => row.id === 'subagent');
 check(resolved?.config?.maxDepth === 2, 'after patching, subagent.maxDepth resolves to 2');
 
-// 4. Only these two rows change; nothing else in the layer is disturbed.
-const otherIds = patches.filter((patch) => patch.id !== 'subagent').map((patch) => patch.id);
-check(
-  otherIds.join(',') === 'ui-settings-general,session-log-deepseek,ui-chat,agent-default-model,locale',
-  'the five pre-existing overrides are untouched',
-);
+// 4. Our override does not disturb the overrides that were already there. Asserted as a
+// SUBSET, not as an exact list: the profile layer is the user's, and new entries appear
+// in it through normal use (a model provider, a plugin toggle written by `dsh plugin`).
+// An earlier version compared the joined id list and failed the moment the user added
+// one — reporting "the five pre-existing overrides are untouched" as false while nothing
+// had touched them.
+const PRE_EXISTING = ['ui-settings-general', 'session-log-deepseek', 'ui-chat', 'agent-default-model', 'locale'];
+const presentIds = new Set((patches ?? []).map((patch) => patch.id));
+const missing = PRE_EXISTING.filter((id) => !presentIds.has(id));
+check(missing.length === 0, `pre-existing overrides are still present (missing: ${missing.join(', ') || 'none'})`);
 
 process.stdout.write(failures.length === 0 ? '\nALL CHECKS PASSED\n' : `\n${failures.length} CHECK(S) FAILED\n`);
 process.exit(failures.length === 0 ? 0 : 1);

package/verify-patch.mjs

@@ -144,51 +144,109 @@ check(existsSync(join(here, 'guard.js')), 'guard.js ships inside the package');
 }
 const delegationRows = delegation?.config ?? [];
 
-// Loader row id -> the model-facing tool name that row publishes.
+// DSH 0.2.1 removed `toolFilter` from the delegation tool. It does not warn: zod strips
+// the key and the child then runs the platform's DEFAULT preset, which is how a live child
+// ended up with all 39 tools and the wrong persona. The curated child toolset therefore
+// moved to two places at once — a worker preset (which decides the mounted tool plugins)
+// and `worker-mask` (which subtracts the app-level tools a preset cannot omit).
+const workerRow = inserted.find((row) => row.id === 'preset-root-planner-worker');
+check(workerRow !== undefined, 'the patch declares a worker preset');
+check(workerRow?.config?.id === 'root-planner-worker', 'the worker preset id is root-planner-worker');
+check(workerRow?.config?.order === 6, 'the worker preset order is 6');
+const workerPlugins = workerRow?.config?.plugins ?? [];
+check(
+  workerPlugins.some((row) => row.id === 'worker-mask' && row.name === 'dsh-root-planner-preset/worker-mask'),
+  'the worker preset mounts worker-mask (the tool restriction)',
+);
+const workerDelegation = workerPlugins.find((row) => row.id === 'delegation');
+const workerDelegationRows = workerDelegation?.config ?? [];
+
+// Loader row id -> the model-facing tool name that row publishes, plus the provider whose
+// capability decides whether a preset may be selected for it.
 const childTools = [
-  { id: 'tool-subagent', toolName: 'subagent' },
-  { id: 'tool-subagent-fork', toolName: 'subagent_fork' },
-].map(({ id, toolName }) => {
+  { id: 'tool-subagent', toolName: 'subagent', provider: 'spawn' },
+  { id: 'tool-subagent-fork', toolName: 'subagent_fork', provider: 'fork' },
+].map(({ id, toolName, provider }) => {
   const row = delegationRows.find((entry) => entry.id === id);
   check(row !== undefined, `delegation declares ${id}`);
   check(row?.config?.toolName === toolName, `${id} publishes the tool name ${toolName}`);
-  return { id, toolName, allow: row?.config?.toolFilter?.allow };
+  check(row?.config?.provider === provider, `${id} uses the ${provider} provider`);
+  check(
+    row?.config?.toolFilter === undefined,
+    `${id} carries no toolFilter (DSH 0.2.1 removed it and strips it silently)`,
+  );
+  // Only the spawn provider supports independent preset selection. A grandchild spawned
+  // through it must land in the worker preset too, or the tree degrades at depth 2: the
+  // platform default persona and the full tool list come back one level down.
+  if (provider === 'spawn') {
+    check(row?.config?.agentPreset === 'root-planner-worker', `${id} spawns into the worker preset`);
+    const nested = workerDelegationRows.find((entry) => entry.id === id);
+    check(nested?.config?.agentPreset === 'root-planner-worker', `worker ${id} spawns into the worker preset too`);
+  }
+  return { id, toolName, provider };
 });
 
-// A subagent is a worker: it keeps the write and shell tools, so the mask must
-// only exclude what a mask can name and what would unbounded a child's process.
-for (const { id, allow } of childTools) {
-  check(Array.isArray(allow) && allow.length > 0, `${id} declares a non-empty allow list`);
-  check(allow?.includes('read'), `${id} mask keeps read`);
-  for (const worker of ['write', 'edit', 'ops_replace', 'pwsh']) {
-    check(allow?.includes(worker), `${id} mask keeps ${worker} (a subagent must be able to work)`);
-  }
-  for (const scoped of SCOPE_LOCAL_TOOLS) {
-    check(
-      !allow?.includes(scoped),
-      `${id} mask excludes the agent-scoped tool "${scoped}" (tools.restrict() would refuse it)`,
-    );
-  }
+// The fork provider declares `agentPreset: false` together with
+// `inheritsParentContext: true`, and its row throws at activation when a preset is
+// configured for it — which is what made the whole preset fail to load on DSH 0.2.1.
+// Assert the absence in both presets so a later edit cannot reintroduce it.
+for (const [label, rows] of [
+  ['the root preset', delegationRows],
+  ['the worker preset', workerDelegationRows],
+]) {
+  const fork = rows.find((entry) => entry.id === 'tool-subagent-fork');
+  check(
+    fork?.config?.agentPreset === undefined,
+    `${label} sets no agentPreset on the fork row (the fork provider rejects one)`,
+  );
 }
 
-// "No unbounded process": with the shell tools present, the property lives in the
-// preset's own rows rather than in the mask.
-const pwshRow = plugins.find((row) => row.id === 'tool-pwsh');
-check(pwshRow !== undefined, 'the preset declares tool-pwsh');
-check(
-  pwshRow?.config?.enableRunInBackground === FORBIDDEN_CAPABILITIES.pwshConfig.enableRunInBackground,
-  `tool-pwsh sets enableRunInBackground=${FORBIDDEN_CAPABILITIES.pwshConfig.enableRunInBackground} ` +
-    `(found ${JSON.stringify(pwshRow?.config?.enableRunInBackground)}) — without it a subagent could detach a server`,
-);
+// A subagent is a worker: the curated set keeps the write and shell tools.
+for (const worker of ['read', 'write', 'edit', 'ops_replace', 'pwsh']) {
+  check(CHILD_TOOLS.includes(worker), `the curated set keeps ${worker} (a subagent must be able to work)`);
+}
+for (const scoped of SCOPE_LOCAL_TOOLS) {
+  check(
+    !CHILD_TOOLS.includes(scoped),
+    `the curated set excludes the agent-scoped tool "${scoped}" (tools.restrict() would refuse it)`,
+  );
+}
+
+// "No unbounded process": with the shell tools present, the property lives in each
+// preset's own rows rather than in a mask — the worker needs it as much as the root.
+for (const [label, rows] of [
+  ['the root preset', plugins],
+  ['the worker preset', workerPlugins],
+]) {
+  const row = rows.find((entry) => entry.id === 'tool-pwsh');
+  check(row !== undefined, `${label} declares tool-pwsh`);
+  check(
+    row?.config?.enableRunInBackground === FORBIDDEN_CAPABILITIES.pwshConfig.enableRunInBackground,
+    `${label} sets enableRunInBackground=${FORBIDDEN_CAPABILITIES.pwshConfig.enableRunInBackground} ` +
+      `(found ${JSON.stringify(row?.config?.enableRunInBackground)}) — without it a subagent could detach a server`,
+  );
+}
 
 /**
  * The global tool inventory of the author's profile, captured from a runtime
  * error rather than hand-kept: `tools.restrict()` prints every known global tool
  * when it refuses a name. Replace this list with the equivalent line from your
  * own environment — it is what makes the check below meaningful.
+ *
+ * Provenance: transcribed on DSH 0.2.1-rc.2. The upgrade from 0.2.0-rc.2 added
+ * `working_directory`, the four `schedule_*` tools, `cordis_inspect_list`,
+ * `cordis_inspect_query` and `plugin_manager`; all are registered by profile-level tool
+ * packages, so they are global. None of them is in the child mask, which is the intended
+ * behaviour of an allow list — a new tool is invisible to children until it is named.
+ *
+ * Regenerate after a DSH upgrade: break the mask on purpose (add a bogus name), let the
+ * agent-creation error print the list, and transcribe it — `tools/extract-tool-inventory.mjs`
+ * reads exactly that line out of a session log.
  */
 const OBSERVED_GLOBAL_TOOLS = new Set([
   'ask_user_question',
+  'cordis_inspect_list',
+  'cordis_inspect_query',
   'create_goal',
   'edit',
   'exit_plan_mode',
@@ -204,10 +262,15 @@ const OBSERVED_GLOBAL_TOOLS = new Set([
   'ops_grep',
   'ops_inspect_local_file',
   'ops_replace',
+  'plugin_manager',
   'present',
   'pwsh',
   'read',
   'read_image',
+  'schedule_create',
+  'schedule_delete',
+  'schedule_list',
+  'schedule_update',
   'send_message',
   'skill',
   'subagent_fork',
@@ -216,30 +279,30 @@ const OBSERVED_GLOBAL_TOOLS = new Set([
   'web_fetch',
   'web_search',
   'workflow',
+  'working_directory',
   'write',
 ]);
 
 /**
- * Tools that are agent-scoped rather than global, so the mask can never name
- * them. `subagent` is the one this bundle ran into: with
- * `modelSelectionSettings: true` it is mounted inside each composed Agent's own
- * scope, and `tools.restrict()` accepts global names only.
+ * Tools that are agent-scoped rather than global, so `tools.restrict()` can never name
+ * them. `subagent` is the one this bundle ran into: with `modelSelectionSettings: true`
+ * it is mounted inside each composed Agent's own scope, and a restriction accepts global
+ * names only. The curated set is the source of truth now — the old per-call mask is gone.
  */
-for (const { id, allow } of childTools) {
-  const unknown = (allow ?? []).filter((name) => !OBSERVED_GLOBAL_TOOLS.has(name));
-  check(
-    unknown.length === 0,
-    `${id} mask names only tools the runtime reports as global` +
-      (unknown.length ? ` (not global: ${unknown.join(', ')})` : ''),
-  );
-  const missing = CHILD_TOOLS.filter((name) => !(allow ?? []).includes(name));
-  const extra = (allow ?? []).filter((name) => !CHILD_TOOLS.includes(name));
+{
+  const unknown = CHILD_TOOLS.filter((name) => !OBSERVED_GLOBAL_TOOLS.has(name));
+  const scoped = new Set(SCOPE_LOCAL_TOOLS);
+  const unexpected = unknown.filter((name) => !scoped.has(name));
   check(
-    missing.length === 0 && extra.length === 0,
-    `${id} mask matches child-tools.mjs (${CHILD_TOOLS.length} entries)` +
-      (missing.length ? ` — missing: ${missing.join(', ')}` : '') +
-      (extra.length ? ` — unexpected: ${extra.join(', ')}` : ''),
+    unexpected.length === 0,
+    `the curated set names only tools the runtime reports as global` +
+      (unexpected.length ? ` (not global: ${unexpected.join(', ')})` : ''),
   );
+  // `subagent` reaches a worker through the preset's own delegation row rather than the
+  // restriction, so it may legitimately sit in the curated set while being scope-local.
+  for (const name of unknown) {
+    check(scoped.has(name), `the only scope-local name in the curated set is declared as such (${name})`);
+  }
 }
 
 process.stdout.write(failures.length === 0 ? '\nALL CHECKS PASSED\n' : `\n${failures.length} CHECK(S) FAILED\n`);

package/worker-mask.js

@@ -0,0 +1,30 @@
+// Restricts a Root Planner worker to the curated child toolset.
+//
+// DSH 0.2.1 removed `toolFilter` from `dsh-tool-subagent` — its error says to "configure
+// persona and tool plugins in an Agent Preset and select it with agentPreset (spawn
+// only)". A preset decides which tool plugins are mounted for that agent, but not which
+// globally mounted tools it may call: several newer tools (tool-cordis, plugin-manager,
+// working-directory, schedule) are registered at the app level and would reach every
+// agent regardless of the preset's rows. `tools.restrict()` is the mechanism that still
+// subtracts them, and this module is that call, mounted as a row inside the worker
+// preset so it applies to the worker's own agent scope.
+//
+// The allow list lives in `child-tools.mjs` — the same file the previous mask used, so
+// the curated set is still declared once.
+
+import { CHILD_TOOLS } from './child-tools.mjs'
+
+export const name = 'root-planner-worker-mask'
+
+export const inject = ['tools']
+
+export function apply(ctx) {
+  // `tools.restrict()` requires a scoped context: it refuses to run from a global one
+  // because that would mask every agent in the process. A preset row is mounted into the
+  // agent's own scope, which is exactly the scope this needs.
+  ctx.effect(() => {
+    const tools = typeof ctx.get === 'function' ? ctx.get('tools') : ctx.tools
+    if (tools === undefined) return () => {}
+    return tools.restrict({ allow: CHILD_TOOLS })
+  })
+}