feat(workflows): per-step model overrides, team presets & llm_decision dynamic orchestration

.gitignore

@@ -59,4 +59,6 @@ e2e-test-project/
 tmp/
 worktrees/
 
-plugins/
+plugins/
+# jscpd JSON report (generated, .md is the tracked stub)
+report/jscpd-report.json

.openfox/agents/dream-reviewer.agent.md

@@ -0,0 +1,19 @@
+---
+id: dream-reviewer
+name: Dream Reviewer
+description: Top-level workflow agent for the dream demo's review step. Reviews the change for correctness and style, then calls step_done().
+subagent: false
+color: '#a855f7'
+allowedTools:
+  - read_file
+  - run_command
+---
+
+You are a code reviewer running as a top-level workflow step.
+
+Your job:
+
+1. Review the completed change for correctness and style.
+2. Approve it or note requested changes, concisely.
+
+When you are finished, you MUST call `step_done()` to signal completion. Do not loop: one review pass + `step_done()` is enough.

.openfox/agents/dream-verifier.agent.md

@@ -0,0 +1,22 @@
+---
+id: dream-verifier
+name: Dream Verifier
+description: Top-level workflow agent for the dream demo's verify step. Runs tests and lint, then calls step_done() so the llm_decision transition can route.
+subagent: false
+color: '#22c55e'
+allowedTools:
+  - read_file
+  - run_command
+  - session_metadata
+  - web_fetch
+---
+
+You are a verifier running as a top-level workflow step.
+
+Your job:
+
+1. Run the project's tests and lint (if any exist).
+2. Read the changed files to confirm the work meets the request.
+3. Report the outcome concisely.
+
+When you are finished, you MUST call `step_done()` to signal completion. Do not loop: a single verification pass + `step_done()` is enough.

.openfox/workflows/dream.workflow.json

@@ -0,0 +1,147 @@
+{
+  "metadata": {
+    "id": "dream",
+    "name": "Dream — Dynamic Orchestration",
+    "description": "Demo workflow for the per-step / team / llm_decision dream: an LLM orchestrator decides which step runs next. Bind a team (Phase 2) or set per-step overrides (Phase 1) so each step runs on its own model; the verify step's resolved model is the one asked to route.",
+    "version": "1.0.0",
+    "color": "#8b5cf6"
+  },
+  "entryStep": "build",
+  "settings": {
+    "maxIterations": 10
+  },
+  "steps": [
+    {
+      "id": "build",
+      "name": "Build",
+      "type": "agent",
+      "phase": "build",
+      "agentId": "builder",
+      "prompt": "Implement the requested change. When you are done, call step_done().",
+      "transitions": [
+        {
+          "when": {
+            "type": "always"
+          },
+          "goto": "verify"
+        }
+      ]
+    },
+    {
+      "id": "verify",
+      "name": "Verify & Route",
+      "type": "agent",
+      "phase": "verification",
+      "agentId": "dream-verifier",
+      "prompt": "Run the tests and lint. Report the outcome via step_done().",
+      "transitions": [
+        {
+          "when": {
+            "type": "custom",
+            "handler": "llm_decision",
+            "config": {
+              "prompt": "Given the verification outcome, what should we do next?",
+              "candidates": [
+                {
+                  "goto": "build",
+                  "label": "Retry build",
+                  "description": "Tests or lint failed; rebuild after fixes"
+                },
+                {
+                  "goto": "review",
+                  "label": "Send to review",
+                  "description": "Everything passed; have it reviewed"
+                },
+                {
+                  "goto": "$done",
+                  "label": "Finish",
+                  "description": "Nothing left to do"
+                }
+              ],
+              "thisGoto": "build"
+            }
+          },
+          "goto": "build"
+        },
+        {
+          "when": {
+            "type": "custom",
+            "handler": "llm_decision",
+            "config": {
+              "prompt": "Given the verification outcome, what should we do next?",
+              "candidates": [
+                {
+                  "goto": "build",
+                  "label": "Retry build",
+                  "description": "Tests or lint failed; rebuild after fixes"
+                },
+                {
+                  "goto": "review",
+                  "label": "Send to review",
+                  "description": "Everything passed; have it reviewed"
+                },
+                {
+                  "goto": "$done",
+                  "label": "Finish",
+                  "description": "Nothing left to do"
+                }
+              ],
+              "thisGoto": "review"
+            }
+          },
+          "goto": "review"
+        },
+        {
+          "when": {
+            "type": "custom",
+            "handler": "llm_decision",
+            "config": {
+              "prompt": "Given the verification outcome, what should we do next?",
+              "candidates": [
+                {
+                  "goto": "build",
+                  "label": "Retry build",
+                  "description": "Tests or lint failed; rebuild after fixes"
+                },
+                {
+                  "goto": "review",
+                  "label": "Send to review",
+                  "description": "Everything passed; have it reviewed"
+                },
+                {
+                  "goto": "$done",
+                  "label": "Finish",
+                  "description": "Nothing left to do"
+                }
+              ],
+              "thisGoto": "$done"
+            }
+          },
+          "goto": "$done"
+        },
+        {
+          "when": {
+            "type": "always"
+          },
+          "goto": "$done"
+        }
+      ]
+    },
+    {
+      "id": "review",
+      "name": "Review",
+      "type": "agent",
+      "phase": "review",
+      "agentId": "dream-reviewer",
+      "prompt": "Review the change for correctness and style. When done, call step_done().",
+      "transitions": [
+        {
+          "when": {
+            "type": "always"
+          },
+          "goto": "$done"
+        }
+      ]
+    }
+  ]
+}

docs/DREAM-ORCHESTRATION.md

@@ -0,0 +1,96 @@
+# Dream — Per-step dynamic orchestration
+
+This describes the four-phase feature that lets you choose, per workflow step,
+which LLM orchestrates, which produces code, and which verifies — transparently
+once configured, with no cap on the number of steps or models.
+
+| Phase | What it adds                                                                                     | Where                                                      |
+| ----- | ------------------------------------------------------------------------------------------------ | ---------------------------------------------------------- |
+| 0     | Plugin API extension point for custom transition handlers                                        | `src/server/providers/plugins/registry.ts`                 |
+| 1     | Per-step model override (`workflowId:stepId`)                                                    | `src/server/agents/model-overrides.ts`                     |
+| 2     | Teams: a named bundle of `stepId -> { providerId, model, reasoningEffort? }` bound to a workflow | `src/server/agents/teams.ts`, `src/server/routes/teams.ts` |
+| 3     | Built-in `llm_decision` transition handler: an LLM picks the next step                           | `src/server/workflows/llm-decision-handler.ts`             |
+
+## How they compose
+
+Resolution precedence for a step's LLM client (in `resolveLLMClientForStep`):
+
+```
+explicit step override  >  team assignment  >  agent override  >  session model
+```
+
+When a step's transitions use `when: { type: 'custom', handler: 'llm_decision',
+config }`, the executor passes that step's **resolved** client to the handler.
+So "which LLM orchestrates this step" is exactly the model you assigned to that
+step — per step, via a team, or via an explicit override. One LLM call is made
+per `(workflow, step, outcome)` and shared across sibling transitions; only the
+transition whose `thisGoto` matches the LLM's choice fires. If the call fails or
+the response is unparseable, the handler returns false and a following `always`
+fallback wins — routing never blocks.
+
+## Configuring it
+
+### 1. Bind a team to a workflow (recommended for N steps / N models)
+
+```bash
+# Create a team that assigns a different model to each step.
+curl -X PUT localhost:3000/api/teams/dream-team \
+  -H 'content-type: application/json' \
+  -d '{
+    "name": "Dream team",
+    "assignments": {
+      "build":  { "providerId": "openai",    "model": "gpt-4o" },
+      "verify": { "providerId": "anthropic", "model": "claude-sonnet-5", "reasoningEffort": "high" },
+      "review": { "providerId": "openai",    "model": "gpt-4o-mini" }
+    }
+  }'
+
+# Bind the workflow to the team (every run resolves steps from it).
+curl -X PUT localhost:3000/api/teams/bindings/dream \
+  -H 'content-type: application/json' \
+  -d '{"teamId":"dream-team"}'
+```
+
+### 2. Or override a single step
+
+```bash
+curl -X PUT localhost:3000/api/workflows/dream/steps/verify/model \
+  -H 'content-type: application/json' \
+  -d '{ "providerId": "anthropic", "model": "claude-sonnet-5", "reasoningEffort": "high" }'
+```
+
+### 3. Use `llm_decision` in the workflow
+
+See `.openfox/workflows/dream.workflow.json` for a full example. The routing
+step declares one `custom` transition per candidate, each with the shared
+`candidates` list and its own `thisGoto`:
+
+```json
+{
+  "when": {
+    "type": "custom",
+    "handler": "llm_decision",
+    "config": {
+      "prompt": "Given the verification outcome, what should we do next?",
+      "candidates": [
+        { "goto": "build", "label": "Retry build", "description": "Tests failed" },
+        { "goto": "review", "label": "Send to review", "description": "Passed" },
+        { "goto": "$done", "label": "Finish", "description": "Nothing left" }
+      ],
+      "thisGoto": "build"
+    }
+  },
+  "goto": "build"
+}
+```
+
+Repeat for each candidate's `thisGoto`, then add an `always` fallback. The
+handler is registered as a built-in before plugins load; a plugin registering
+`llm_decision` overrides it.
+
+## Verification
+
+`src/server/workflows/phase4-integration.test.ts` proves the composition
+end-to-end: the `llm_decision` call is made on the step-resolved (team) client,
+the session client is not consulted, and routing follows the LLM's choice (or
+falls through to the `always` fallback on an unparseable response).

src/provider/index.ts

@@ -1,5 +1,6 @@
 import type { ModelConfig } from '../shared/types.js'
 import type { LLMCompletionRequest, LLMCompletionResponse, LLMStreamEvent } from '../server/llm/types.js'
+import type { TransitionHandler } from '../server/workflows/transition-handlers.js'
 
 // ============================================================================
 // Auth Types
@@ -105,6 +106,12 @@ export interface ProviderPluginRegistry {
   registerAuth(adapter: ProviderAuthAdapter): void
   registerTransport(adapter: ProviderTransportAdapter): void
   registerPreset(preset: ProviderPreset): void
+  /**
+   * Register a custom workflow transition handler. A workflow step declares a
+   * transition with `when: { type: 'custom', handler: '<handlerId>' }` and the
+   * executor invokes the registered handler to decide whether it fires.
+   */
+  registerTransitionHandler(handlerId: string, handler: TransitionHandler): void
   readonly runtime: ProviderPluginRuntime
 }
 
@@ -170,3 +177,4 @@ export type {
   LLMToolDefinition,
 } from '../server/llm/types.js'
 export type { ModelConfig, ToolCall } from '../shared/types.js'
+export type { TransitionHandler, TransitionHandlerContext } from '../server/workflows/transition-handlers.js'

src/server/agents/__fixtures__/avocat-du-diable.agent.md

@@ -0,0 +1,44 @@
+---
+id: avocat-du-diable
+name: Avocat du diable
+description: "Olgenius : cherche activement ce qui va casser. Couvre 10 axes, écrit OBJECTIONS.md. N'approuve jamais. Ne code pas."
+subagent: false
+color: '#ef4444'
+allowedTools:
+  - read_file
+  - write_file
+  - run_command
+  - session_metadata
+---
+
+Tu es **l'Avocat du diable** du système Olgenius. Ton mandat : **chercher activement ce qui va casser.** Tu n'es pas là pour approuver. Si tu ne trouves rien, tu l'écris explicitement et tu justifies pourquoi chaque axe est couvert : un accord silencieux est un échec de ta part. Tu ne codes pas.
+
+## Grille à couvrir, dans l'ordre
+
+1. Hypothèses non vérifiées du planificateur
+2. Architecture : couplage, choix irréversibles, sur-ingénierie autant que sous-ingénierie
+3. Modèle de données et migrations
+4. Sécurité : authentification, autorisation, secrets, injections, données personnelles
+5. Cas limites et modes de défaillance
+6. Testabilité : ce que le découpage rend difficile à tester
+7. Dépendances externes : maturité, licence, verrouillage
+8. Exploitation : configuration, journalisation, reprise sur erreur, rollback
+9. Périmètre : ce qui a été ajouté sans être demandé, ce qui a été oublié
+10. Ordonnancement : phases qui vont devoir être refaites à cause de l'ordre choisi
+
+## Format de chaque objection
+
+```
+### OBJ-001 · [BLOQUANT|MAJEUR|MINEUR] · phase 03
+Constat     : ...
+Conséquence : ce qui casse concrètement, et quand
+Proposition : la correction précise
+```
+
+- **BLOQUANT** : le plan produira un système faux, non sécurisé, ou irréparable sans reprise lourde.
+- **MAJEUR** : coût significatif si on ne corrige pas maintenant.
+- **MINEUR** : préférence, style, optimisation.
+
+## Interdits
+
+Approuver par politesse. Inventer un risque non documenté pour remplir la grille. Recopier le contenu du PRD/PLAN dans tes messages : tu donnes des chemins et des références de ligne. Modifier le code.