# The `rigorquant` agent preset, declared (docs/adr/0002-declared-preset-on-dsh-0.1.7.md,
# Decision 25; carried forward by docs/adr/0003-rigorquant-on-dsh-0.2.0.md,
# Decision 26). DSH ignores directory presets; a preset exists
# because some bundle declares a `@deepseek-ai/dsh-agent-preset` row. This file
# is that row, in the shape of web-app's `presets/standard.patch.yml`, and
# package.json lists it after `cordis.patch.yml` in `dsh.bundle.patch`.
#
# The child list is 0.5.0's `agent.cordis.yml`, minus two rows:
# - the child `skill-filesystem` row: the host `skill-filesystem-rigorquant`
#   row in cordis.patch.yml already serves the skill to every agent;
# - the jacobian MCP row: on a declared preset a single child cannot be
#   patched, so the escalation lane mounts at runtime through `rq_escalate`.
#
# Child `baseUrl` is the PROFILE root (`file://$DSH_HOME/profiles/<profile>/`),
# not this package. No child may resolve a path from it.
#
# The package requires DSH `>=0.2.0-rc.2 <0.2.1` (package.json
# `peerDependencies`, restated declaratively under `engines.dsh`); the CLI
# skips a bundle whose range the running core does not satisfy.
- insert:
    # An override by id (a patch keyed `preset-rigorquant` in the profile, or
    # an edit saved from the Web preset editor) replaces this row's WHOLE
    # `config`: id, name, description, order and every child. There is no
    # per-child patch; an override must restate the full list.
    - id: preset-rigorquant
      name: '@deepseek-ai/dsh-agent-preset'
      config:
        id: rigorquant
        name: RigorQuant
        description: >-
          Session-scoped unattended empirical/computational mathematics research
          framework (econ/finance/portfolio/simulation): context-isolated
          multi-agent exploration in the style of Jin's Crouzeix run and Tao's
          collaborative projects, dual-track ground-truth derivation,
          adversarial counterexample-only audit, a four-part pre-implementation
          check battery, fixed-seed LLN conventions, and a jacobian/Lean
          escalation lane for proof-critical claims.
        order: 5
        plugins:
          # ── identity ────────────────────────────────────────────────

          - id: persona
            name: '@deepseek-ai/dsh-persona'
            config:
              # DSH >= 0.1.3-alpha.2 splits the persona into a required `prefix`
              # (rendered before first-party guidance) and an optional `suffix`
              # (rendered after it). The single `text` key this preset used through
              # 0.1.2 no longer exists and its absence rejects the whole mount.
              prefix: >-
                You are RigorQuant, the root orchestrator of an unattended mathematical
                research framework for empirical and computational work: economics,
                finance, portfolio construction and optimization, simulation, and
                computational econ/finance. You are powered by the {{model}} model.

                Your mission: carry a hard quantitative problem from statement to a method
                whose mathematical validity is established on simplified/special cases
                BEFORE numerical implementation, then implement it and proceed to the next
                stage. The operating pattern follows Shanmu Jin's Crouzeix-conjecture run
                (divergent multi-agent search, adversarial audit, counterexample-only
                elimination, persistence) and Terence Tao's collaborative proof projects
                (blueprint-style decomposition, automated checkers as the truth gate).

                NON-NEGOTIABLE CONSTRAINTS:

                1. A producer cannot certify its own output. Ground-truth derivations are
                   produced independently, never by the agent that authored the method
                   under test.
                2. A route is eliminated ONLY by a concrete failing case (a
                   counterexample), never by style, vibes, or "this feels wrong".
                3. Every stochastic run records its seed. Partial progress does not count
                   unless it resolves the sub-problem exactly.
                4. Do not handwave a load-bearing unproven claim — settle it before
                   implementing.

                ISOLATION, stated precisely: context isolation between you and your
                teammates is enforced by the harness (a fresh teammate never sees your
                conversation) — that is the only wall. Everything else is composed per
                teammate by the RigorQuant team plugin: role by name (a teammate named
                `<role>-<n>` gets that role's persona and model), tool budget by scope
                (blind roles lose web and skills, web-denied roles lose web), topology
                by guard (a teammate may message only you, cannot list the roster or
                the board, and touches only its own task; web-denied roles are refused
                network commands in bash). It is still not a network wall: filesystem
                scope and producer≠checker stay procedural, backed by the adversary.

                TEAM, stated precisely: the Team policy creates teammates only on an
                explicit request; a RigorQuant study is that explicit request, so
                create the study's teammates without asking again. Spawn only while
                your context carries the line "RigorQuant team guard: armed". If that
                line is absent, do not spawn: the team plugin is not composing this
                session, so report the missing plugin and stop.

                UNATTENDED, stated precisely: the framework runs unattended WITHIN one
                live session. Crossing a session boundary disarms the goal, and one human
                turn ("continue") is required to rearm it: on that turn, before any
                other study work, call `get_goal`, then `update_goal` with action
                `resume` and its exact `goal_id` and `revision`. Rounds do not continue
                autonomously across restarts; checkpoint state and resume explicitly.

                ASKING, stated precisely: `ask_user_question` runs in the harness's
                timed mode. A batch you ask returns `pending` when the user has not
                answered within the wait: pending is neither an answer nor permission.
                Keep working on steps that do not depend on it, never re-ask the same
                question, and never read pending as consent. The reply arrives later as
                a user message carrying JSON with `kind: "answer_to_pending_question"`
                and the original `callId`; match it to that open decision, record it in
                the journal, then proceed on what it gates. Pass `timeout: -1` only when
                no further step is possible without the answer — an install or an
                authorization only the user can give.

                When a tool result is spilled (a bounded preview plus a locator), read
                the full result back by its locator; never re-run the tool to see it.

                When a RigorQuant study begins, load the `rigorquant` skill and follow
                its operating procedure (workspace layout, registry schema, tolerances,
                escalation triggers, anti-patterns, lifecycle).
              suffix: Your working directory is {{cwd}}.

          - id: agent-instructions
            name: '@deepseek-ai/dsh-agent-instructions'
            config:
              maxBytes: 65536

          # ── shell ───────────────────────────────────────────────────

          - 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'

          # ── filesystem ──────────────────────────────────────────────

          - id: tool-fs
            name: '@deepseek-ai/dsh-tool-fs'

          - id: tool-fs-search
            name: '@deepseek-ai/dsh-tool-fs-search'
            config:
              sampleOverCapGlobResults: false

          # ── background jobs ─────────────────────────────────────────

          - id: tool-jobs
            name: '@deepseek-ai/dsh-tool-jobs'

          # ── skills ──────────────────────────────────────────────────

          - id: tool-skill
            name: '@deepseek-ai/dsh-tool-skill'

          # ── goals ───────────────────────────────────────────────────

          # The human-facing `/goal` command. The web-app bundle disables it at the host
          # plane, so every preset must re-mount it; without this row an unattended run
          # has no way to arm the goal whose round driver is the framework's loop.
          - id: command-goal
            name: '@deepseek-ai/dsh-command-goal'

          - id: tool-goal
            name: '@deepseek-ai/dsh-tool-goal'

          # ── plan mode ───────────────────────────────────────────────

          - id: planning
            name: cordis:group
            group: true
            isolate:
              planMode: true
            config:
              - id: plan-mode
                name: '@deepseek-ai/dsh-plan-mode'
                config:
                  section: |
                        You are in plan mode. Stay in plan mode until exit_plan_mode succeeds or the user switches the session mode. Imperative language to implement changes means plan the implementation, not execute it. A user's conversational agreement — including an answer confirming something you asked — approves nothing and does not end plan mode; fold the confirmed decision into the plan and submit it through exit_plan_mode.

                        Explore first. Use non-mutating reads, searches, static analysis, and checks to ground the plan in the actual repository. Do not edit or write files, change configuration, run formatters or code generation that rewrites tracked files, commit, or otherwise carry out the plan. Prefer existing functions and patterns over new machinery.

                        The tool catalog stays the same across modes for request-cache stability. These plan-mode rules override any later tool description or guidance that suggests using mutation tools; those tools remain listed to keep the tool catalog unchanged. Do not use todo_write to track this planning phase: it tracks implementation after an approved plan, while the plan itself belongs in exit_plan_mode.

                        Resolve discoverable facts by inspection. Use ask_user_question only for user-owned choices or material ambiguity that inspection cannot answer. Do not ask the user where code lives or how current behavior works when you can find out.

                        Make the plan decision-complete: state the goal and success criteria; group implementation changes by subsystem; identify public API, schema, and data-flow changes; cover edge cases, failure modes, tests, acceptance criteria, and explicit assumptions. Keep it concise enough to review but detailed enough that another engineer can implement it without making design decisions.

                        When ready, call exit_plan_mode with the complete plan markdown, starting with a # title. Make exit_plan_mode the only and final tool call in that assistant response: it presents the plan for approval, and implementation begins only in a later step after approval. Do not paste the final plan as a plain reply or ask "should I proceed?" through prose or ask_user_question. If review rejects it, incorporate the feedback and present again. If the review channel is unavailable or aborted, stay in plan mode and ask the user to switch modes manually; do not proceed with implementation.

          # ── compaction ──────────────────────────────────────────────

          - id: compaction
            name: cordis:group
            group: true
            isolate:
              compaction: true
              toolResultPruner: true
            config:
              - id: compaction-basic
                name: '@deepseek-ai/dsh-compaction-basic'
                config:
                  # Compact at 60% of the routed model's context window instead of the
                  # default 80%: earlier compaction shrinks the cached prefix that is
                  # re-sent on every step, cutting per-step input (and cache-read)
                  # tokens on long runs.
                  thresholdRatio: 0.6

              - id: command-compact
                name: '@deepseek-ai/dsh-command-compact'

              - id: tool-result-pruner
                name: '@deepseek-ai/dsh-compaction-tool-result-pruner'
                config:
                  # Shipped at 8192/4096/1024; rigorquant writes heavy outputs to files
                  # and re-reads them, so retaining 4 KiB of every tool result is enough
                  # context per step while halving the surface growth on long runs.
                  thresholdChars: 4096
                  headChars: 2048
                  tailChars: 1024

          # ── delegation and workflows ────────────────────────────────

          # Team-only (docs/architecture.md, Decision 24;
          # docs/adr/0001-rigorquant-on-agent-teams.md). Teammates are created by the
          # Agent Teams bundles' `spawn_teammate`, not by any row here: the seven
          # per-role delegation rows, the subagent control and list rows, and the fork
          # row are gone. The Team tools register `send_message`, `list_agents` and
          # `interrupt_agent` under the legacy names, so the classic control rows would
          # have mounted a second tool per name. Each teammate's persona, tool budget
          # and hub-and-spoke guard are applied from its NAME (`<role>-<n>`) by the
          # bundle's `dsh-rigorquant/team` row, and its model by the router — none of
          # that lives in this composition any more. No other delegation tool is
          # mounted, so depth-one delegation holds by construction: the rows left in
          # this group are all disabled.
          - id: delegation
            name: cordis:group
            group: true
            isolate:
              workflowEngine: true
            config:
              # Named external-agent instances (DSH >= 0.1.1-rc.1): install the
              # provider bundle once per profile (`dsh plugin --profile <p> add
              # @deepseek-ai/dsh-subagent-codex` / `...-claude-code`) and configure its
              # named instances there (providerName / permissionMode / env /
              # disposeGraceMs). Permission modes: Codex `never` (default) |
              # `approve-for-me` | `dangerously-bypass-approvals-and-sandbox`; Claude
              # Code `dontAsk` (default) | `acceptEdits` | `auto` | `plan` |
              # `bypassPermissions`. The escalation lane prefers plan/never modes and
              # reserves bypass modes for the approval-gated escalation flow
              # (references/escalation.md). Enabling a row whose bundle is not installed
              # fails children at spawn, not the preset mount, so rows stay disabled
              # here; flip on per study.
              - id: tool-subagent-codex
                name: '@deepseek-ai/dsh-tool-subagent'
                disabled: true
                config:
                  provider: codex
                  toolName: subagent_codex
                  backgroundMode: one-shot
                  maxDepth: provider-managed

              - id: tool-subagent-claude-code
                name: '@deepseek-ai/dsh-tool-subagent'
                disabled: true
                config:
                  provider: claude-code
                  toolName: subagent_claude_code
                  backgroundMode: one-shot
                  maxDepth: provider-managed

              # The workflow engine and its model-facing tool are DISABLED
              # (feature/subagent-context): workflow `agent()` children carry no role
              # tag, and the engine's start path expresses neither a persona nor a
              # per-child toolFilter, so every workflow child ran as an untagged agent
              # wearing the root persona with the full catalog (74 of 162 children in
              # the 0.4.x study logs; ~23k-token headers). The framework's fan-out is
              # the goal-round driver plus named teammates; a deployment
              # that wants workflows should use another preset, not this one. While
              # these rows are disabled no deny list may name `workflow`.
              #
              # The engine row follows the 0.1.6 rename: `workflow-worker-thread` no
              # longer exists as a package. A disabled row is never imported, so the
              # stale name mounted nothing either way — but it read as a typo and the
              # harness probe reported it UNRESOLVED. Enabled it would be fatal:
              # discovery marks a preset naming a missing package "Failed to load" and
              # the picker hides it.
              - id: workflow-ptc
                name: '@deepseek-ai/dsh-workflow-ptc'
                disabled: true
                config:
                  provider: spawn

              - id: tool-workflow
                name: '@deepseek-ai/dsh-tool-workflow'
                disabled: true

              # Ralph is DISABLED for the same reason as workflow: ralph rounds spawn
              # untagged children with no persona and no expressible per-child scope,
              # and the framework's iteration loop is the goal-round driver. While this
              # row is disabled no deny list may name `ralph`.
              - id: tool-ralph
                name: '@deepseek-ai/dsh-tool-ralph'
                disabled: true
                config:
                  subagentProvider: spawn
                  maxRounds: 64

          # ── remaining model-facing rows ─────────────────────────────

          # DSH 0.2.0 added `timed` mode to this tool (its Config grew
          # `mode`/`timeout`). RigorQuant takes it: a study runs unattended, so
          # a question the user has not answered within the wait returns
          # `pending` instead of parking the round, and the answer arrives
          # later as an `answer_to_pending_question` user message. The timeout is
          # stated rather than inherited so the framework's choice is pinned;
          # the orchestrator passes `timeout: -1` per call only when no further
          # step is possible without the answer (see the persona and Step 1's
          # study-root resolution).
          - id: tool-ask-user
            name: '@deepseek-ai/dsh-tool-ask-user'
            config:
              mode: timed
              timeout: 300

          - 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

          # Declaring the study's deliverables. NO bundle mounts this tool — only a
          # preset does — so without this row a RigorQuant session has no `present` at
          # all and the 0.1.5 deliverables flow (right-Sidebar preview, open in default
          # app, reveal in file manager) has nothing to render. The skill's deliverables
          # procedure ends with a `present` call for exactly this reason.
          - id: present
            name: '@deepseek-ai/dsh-tool-present'