feat(recovery): failure-ticket routing + review-gate RECOVER + return-to-sender

Retry-agency invariant: a stage may only retry a gate it has the capability to change. A write-less stage failing a build/contract/static gate (e.g. freestyle final_verification with allowedTools=[shell]) can never fix it, so retrying in place is futile until the budget drains (Vikunja #41). Instead: open a FailureTicketOpenedEvent (category + requiredCapability derived deterministically from the gate id, no LLM) and route to a recovery stage that holds the capability, bounded by a small per-stage route budget. - Slice 2: SalvageDecision.RECOVER (ternary CONTINUE/RECOVER/FAIL) lets the review-gate salvage judge hand off to recovery. Shared routeToRecovery() unifies the deterministic agency guard and the judge on one destination; decideGateExhaustion now returns StepResult?. - Return-to-sender: recovery's exit is dynamic (recoveryReturnMove) — it goes back to the exact ticket-origin stage to re-run its gate, so a write-less gate anywhere in the graph is handled without skipping intervening stages. The synthesized recovery->terminal edge is now only a no-ticket fallback. - Freestyle wiring: ExecutionPlanCompiler injectRecovery flag (Main passes true) synthesizes a write-capable recovery stage; ticket evidence reaches it via buildRecoveryTicketEntry. - Read-only tools always present: StageConfig.effectiveAllowedTools adds file_read/list_dir to any tool-granting stage (fixes the verifier reaching for `shell ls -R` to inspect the tree). Tests: RecoveryRoutingTest (deterministic gate, no static return edge — proves dynamic return) + GateRetryBudgetExhaustionTest RECOVER case + two compiler tests. All green; detekt clean.
2026-07-07 12:05:26 +04:00
parent 597a26d551
commit 879672a47d
13 changed files with 672 additions and 23 deletions
@@ -8,6 +8,7 @@ import com.correx.core.events.events.RepoKnowledgeHit
 import com.correx.core.context.model.ContextEntry
 import com.correx.core.context.model.ContextLayer
 import com.correx.core.context.model.EntryRole
+import com.correx.core.events.events.FailureTicketOpenedEvent
 import com.correx.core.events.events.RefinementIterationEvent
 import com.correx.core.events.events.RetryAttemptedEvent
 import com.correx.core.events.events.StoredEvent
@@ -38,6 +39,33 @@ fun buildRetryFeedbackEntry(events: List<StoredEvent>, stageId: StageId): Contex
    )
 }

+/**
+ * Feeds the open failure ticket into the recovery stage's context. The recovery stage was entered
+ * by kernel routing (not a normal edge) precisely because an upstream write-less stage failed a gate
+ * it lacked the capability to fix; without the ticket the recovery stage has no idea what to repair.
+ * Surfaces the failing stage, the gate, and the captured gate output (evidence, recorded at
+ * ticket-open per invariant #9) so recovery can act on recorded fact, not a re-observation.
+ */
+fun buildRecoveryTicketEntry(events: List<StoredEvent>, stageId: StageId): ContextEntry? {
+    val ticket = events
+        .mapNotNull { it.payload as? FailureTicketOpenedEvent }
+        .lastOrNull { it.routeTo == stageId } ?: return null
+    val content = "## Recovery ticket\n" +
+        "Stage '${ticket.stageId.value}' failed the '${ticket.gate}' gate (${ticket.category}) and " +
+        "lacks the '${ticket.requiredCapability}' capability to fix it, so the failure was routed to " +
+        "you. Repair the underlying cause, then control returns to verification.\n" +
+        "### Gate output\n${ticket.evidence}"
+    return ContextEntry(
+        id = ContextEntryId(UUID.randomUUID().toString()),
+        layer = ContextLayer.L1,
+        content = content,
+        sourceType = "recoveryTicket",
+        sourceId = stageId.value,
+        tokenEstimate = content.length / 4,
+        role = EntryRole.SYSTEM,
+    )
+}
+
 fun criticArtifactIds(
    events: List<StoredEvent>,
    graph: WorkflowGraph,
@@ -7,6 +7,7 @@ import com.correx.core.events.events.RefinementIterationEvent
 import com.correx.core.events.events.RetryAttemptedEvent
 import com.correx.core.events.events.RetrySalvageDecidedEvent
 import com.correx.core.events.events.SalvageDecision
+import com.correx.core.events.events.FailureTicketOpenedEvent
 import com.correx.core.events.events.StoredEvent
 import com.correx.core.events.events.TransitionExecutedEvent
 import com.correx.core.events.events.WorkflowCompletedEvent
@@ -94,6 +95,12 @@ class DefaultOrchestrationReducer : OrchestrationReducer {
            refinementIterations = state.refinementIterations + (p.cycleKey to p.iteration),
        )

+        // Charge the failing stage's recovery-route budget. routeAttempt is the 1-based count the
+        // orchestrator computed, so fold it directly (idempotent under replay).
+        is FailureTicketOpenedEvent -> state.copy(
+            recoveryRoutes = state.recoveryRoutes + (p.stageId.value to p.routeAttempt),
+        )
+
        else -> state
    }
 }
@@ -15,6 +15,7 @@ import com.correx.core.events.events.ArtifactContentStoredEvent
 import com.correx.core.events.events.ClarificationAnswer
 import com.correx.core.events.events.ClarificationAnsweredEvent
 import com.correx.core.events.events.ArtifactValidatedEvent
+import com.correx.core.events.events.FailureTicketOpenedEvent
 import com.correx.core.events.events.OrchestrationResumedEvent
 import com.correx.core.events.events.RefinementIterationEvent
 import com.correx.core.events.events.RetrySalvageDecidedEvent
@@ -29,6 +30,7 @@ import com.correx.core.events.types.ClarificationRequestId
 import com.correx.core.events.types.ArtifactLifecyclePhase
 import com.correx.core.events.types.SessionId
 import com.correx.core.events.types.StageId
+import com.correx.core.events.types.TransitionId
 import com.correx.core.kernel.execution.WorkflowResult
 import com.correx.core.kernel.orchestration.subagent.InSessionSubagentRunner
 import com.correx.core.kernel.orchestration.subagent.SubagentRunRequest
@@ -54,6 +56,25 @@ private val log = LoggerFactory.getLogger(DefaultSessionOrchestrator::class.java
 // Fallback refinement-loop cap when a stage does not declare maxRetries.
 private const val DEFAULT_MAX_REFINEMENT = 3

+// Own, small budget for the retry-agency recovery route: how many times a single stage may be
+// routed to a recovery stage before the run fails terminally. Deliberately smaller than the
+// per-gate retry budget — recovery is a purpose-built repair-then-reverify loop, not open-ended.
+private const val RECOVERY_ROUTE_BUDGET = 2
+
+// Deterministic gate id → capability the failing stage must hold to change the failure condition.
+// A gate not listed here is not eligible for recovery routing (retries in place as before).
+private val GATE_REQUIRED_CAPABILITY: Map<String, String> = mapOf(
+    "execution" to "file_write",
+    "contract" to "file_write",
+    "static_analysis" to "file_write",
+)
+
+// Deterministic failure category from the gate id (no LLM — keeps routing off the untrusted path).
+private fun ticketCategory(gate: String): String = when (gate) {
+    "plan_compile" -> "planning"
+    else -> "implementation"
+}
+
 class DefaultSessionOrchestrator(
    private val repositories: OrchestratorRepositories,
    engines: OrchestratorEngines,
@@ -138,6 +159,12 @@ class DefaultSessionOrchestrator(
        val resolved = resolveTransition(
            enriched.graph, enriched.sessionId, enriched.currentStageId, stageArtifacts, artifactContent,
        )
+        // Recovery return-to-sender: a recovery stage is entered by kernel routing from whichever
+        // write-less stage failed a gate, so its exit must go BACK to that exact stage (re-run its
+        // gate) — not follow a static edge that would skip intervening stages. Derived from the latest
+        // failure ticket routed here (replay-deterministic). Overrides the resolved edge; an operator
+        // preempt below still wins.
+        val baseDecision = recoveryReturnMove(enriched) ?: resolved
        // Preemptive redirect override (freestyle graph re-routing): if the operator confirmed a
        // jump (a recorded, unconsumed PreemptRedirectEvent), it overrides the resolver's edge —
        // unless the target is unknown or its needs aren't satisfied, in which case it is blocked
@@ -161,9 +188,9 @@ class DefaultSessionOrchestrator(
                    "[Orchestrator] redirect blocked session={} to={} reason={}",
                    enriched.sessionId.value, outcome.event.toStageId.value, outcome.event.reason,
                )
-                resolved
+                baseDecision
            }
-            PreemptRedirect.Outcome.None -> resolved
+            PreemptRedirect.Outcome.None -> baseDecision
        }
        log.debug(
            "[Orchestrator] transition session={} stage={} decision={}",
@@ -229,6 +256,114 @@ class DefaultSessionOrchestrator(
    }


+    /**
+     * Retry-agency guard (ticket + recovery model). When [failure]'s gate requires a capability the
+     * failing stage does not hold (e.g. a build/contract gate needs `file_write` on a write-less
+     * verifier), retrying the stage in place can never change the outcome. Instead open a
+     * [FailureTicketOpenedEvent] and route to the graph's recovery stage — which does hold the
+     * capability — bounded by the stage's own [RECOVERY_ROUTE_BUDGET].
+     *
+     * Returns a [StepResult] when it took over the failure (routed, or budget-exhausted terminal),
+     * or null to fall through to the normal per-gate retry path. Null in every backward-compatible
+     * case: gate not capability-gated, stage already has the capability, or no recovery stage exists.
+     */
+    private suspend fun maybeRouteToRecovery(
+        ctx: EnrichedExecutionContext,
+        stageId: StageId,
+        failure: StageExecutionResult.Failure,
+        state: OrchestrationState,
+    ): StepResult? {
+        val requiredCapability = GATE_REQUIRED_CAPABILITY[failure.gate] ?: return null
+        val stageTools = ctx.graph.stages[stageId]?.allowedTools ?: emptySet()
+        if (requiredCapability in stageTools) return null // stage has agency — retry in place is valid
+        return routeToRecovery(ctx, stageId, failure.gate, requiredCapability, failure.reason, state)
+    }
+
+    /**
+     * Open a [FailureTicketOpenedEvent] and route [stageId]'s failure to the graph's recovery stage,
+     * bounded by [RECOVERY_ROUTE_BUDGET]. Shared by the deterministic agency guard
+     * ([maybeRouteToRecovery]) and the review-gate salvage judge ([decideGateExhaustion] on
+     * [SalvageDecision.RECOVER]) so both routing paths reach the same recovery destination.
+     *
+     * Returns null when the graph declares no recovery stage (caller falls back to its legacy path);
+     * a terminal [StepResult] when the route budget is spent; otherwise the recovery stage's result.
+     */
+    private suspend fun routeToRecovery(
+        ctx: EnrichedExecutionContext,
+        stageId: StageId,
+        gate: String,
+        requiredCapability: String,
+        reason: String,
+        state: OrchestrationState,
+    ): StepResult? {
+        val recoveryId = findRecoveryStage(ctx.graph, stageId) ?: return null // no recovery stage: legacy path
+        val used = state.recoveryRoutes[stageId.value] ?: 0
+        if (used >= RECOVERY_ROUTE_BUDGET) {
+            return StepResult.Terminal(
+                failWorkflow(
+                    ctx.sessionId,
+                    stageId,
+                    "recovery route budget exhausted for stage ${stageId.value} (gate=$gate): $reason",
+                    retryExhausted = true,
+                ),
+            )
+        }
+        emit(
+            ctx.sessionId,
+            FailureTicketOpenedEvent(
+                sessionId = ctx.sessionId,
+                stageId = stageId,
+                gate = gate,
+                category = ticketCategory(gate),
+                requiredCapability = requiredCapability,
+                routeTo = recoveryId,
+                evidence = reason,
+                routeAttempt = used + 1,
+            ),
+        )
+        log.info(
+            "[Orchestrator] failure ticket session={} stage={} gate={} lacks={} -> route to recovery={} ({}/{})",
+            ctx.sessionId.value, stageId.value, gate, requiredCapability,
+            recoveryId.value, used + 1, RECOVERY_ROUTE_BUDGET,
+        )
+        val advancedTo = advanceStage(
+            ctx.sessionId,
+            stageId,
+            TransitionDecision.Move(transitionId = TransitionId("recovery-route"), to = recoveryId),
+        )
+        return enterStage(ctx.copy(currentStageId = advancedTo), recoveryId)
+    }
+
+    /**
+     * The graph's recovery stage: a stage marked `metadata["role"] == "recovery"` (or, as a
+     * convenience, one whose id is literally "recovery"), other than [failingStageId] itself.
+     * Null when the graph declares none — recovery routing is opt-in per workflow.
+     */
+    private fun findRecoveryStage(graph: WorkflowGraph, failingStageId: StageId): StageId? =
+        graph.stages.entries.firstOrNull { (id, cfg) ->
+            id != failingStageId && (cfg.metadata["role"] == "recovery" || id.value == "recovery")
+        }?.key
+
+    /**
+     * When the current stage is a recovery stage that has finished, route it back to the stage the
+     * failure came from — the origin of the latest [FailureTicketOpenedEvent] routed here — so the
+     * gate that triggered the route is re-run. This makes recovery correct for arbitrary graph
+     * topology (a write-less gate anywhere routes here and returns to exactly its own stage) rather
+     * than relying on a static recovery→terminal edge that would skip intervening stages. Returns
+     * null when the current stage is not a recovery stage or no ticket routed here (then the
+     * resolver's normal edge — the compiler's fallback recovery→terminal — is used).
+     */
+    private fun recoveryReturnMove(ctx: EnrichedExecutionContext): TransitionDecision.Move? {
+        val cfg = ctx.graph.stages[ctx.currentStageId] ?: return null
+        val isRecovery = cfg.metadata["role"] == "recovery" || ctx.currentStageId.value == "recovery"
+        if (!isRecovery) return null
+        val origin = repositories.eventStore.read(ctx.sessionId)
+            .mapNotNull { it.payload as? FailureTicketOpenedEvent }
+            .lastOrNull { it.routeTo == ctx.currentStageId }
+            ?.stageId ?: return null
+        return TransitionDecision.Move(transitionId = TransitionId("recovery-return"), to = origin)
+    }
+
    /** Returns the cached validated artifact content for the given session + artifact id, or null if absent. */
    fun validatedArtifactContent(sessionId: SessionId, artifactId: ArtifactId): String? =
        artifactContentCache["${sessionId.value}:${artifactId.value}"]
@@ -450,6 +585,10 @@ class DefaultSessionOrchestrator(
                        return StepResult.Continue(ctx.copy(stageCount = ctx.stageCount + 1))
                    }
                    val refreshedState = orchestrationRepository.getState(ctx.sessionId)
+                    // Retry-agency invariant: a gate this stage lacks the capability to fix must not
+                    // be retried in place (futile). Route to a recovery stage that holds the
+                    // capability, if one exists and the route budget remains.
+                    maybeRouteToRecovery(ctx, stageId, result, refreshedState)?.let { return it }
                    val gateDecision = retryCoordinator.decide(
                        sessionId = ctx.sessionId,
                        stageId = stageId,
@@ -461,10 +600,9 @@ class DefaultSessionOrchestrator(
                    when (gateDecision) {
                        RetryDecision.Retry -> Unit // retry — loop and re-execute
                        RetryDecision.Exhausted -> {
-                            val terminal = decideGateExhaustion(
-                                ctx.sessionId, stageId, result.gate, result.reason, refreshedState,
-                            )
-                            if (terminal != null) return StepResult.Terminal(terminal)
+                            decideGateExhaustion(
+                                ctx, stageId, result.gate, result.reason, refreshedState,
+                            )?.let { return it }
                            // review-gate salvage said CONTINUE — budget was reset by the reducer;
                            // loop and re-execute the stage with a fresh budget.
                        }
@@ -482,17 +620,19 @@ class DefaultSessionOrchestrator(
     * FAIL is terminal. A gate that already spent its one salvage reset (state.gateSalvageUsed) fails
     * immediately on a second exhaustion, regardless of what the judge would say.
     *
-     * Returns the terminal [WorkflowResult] to fail with, or null to retry (loop and re-execute).
+     * Returns a terminal [StepResult] to fail/route with, the recovery stage's result on RECOVER,
+     * or null to retry in place (loop and re-execute).
     */
    private suspend fun decideGateExhaustion(
-        sessionId: SessionId,
+        ctx: EnrichedExecutionContext,
        stageId: StageId,
        gate: String,
        reason: String,
        state: OrchestrationState,
-    ): WorkflowResult? {
+    ): StepResult? {
+        val sessionId = ctx.sessionId
        if (gate != "review" || state.gateSalvageUsed.contains(gate)) {
-            return failWorkflow(sessionId, stageId, reason, retryExhausted = true)
+            return StepResult.Terminal(failWorkflow(sessionId, stageId, reason, retryExhausted = true))
        }
        // No judge wired: degrade safely with a deterministic allow-one-reset-then-fail policy —
        // the gateSalvageUsed check above already ensures this fires at most once per gate.
@@ -504,7 +644,12 @@ class DefaultSessionOrchestrator(
        emit(sessionId, RetrySalvageDecidedEvent(sessionId, stageId, gate, judgment.decision, judgment.rationale))
        return when (judgment.decision) {
            SalvageDecision.CONTINUE -> null
-            SalvageDecision.FAIL -> failWorkflow(sessionId, stageId, reason, retryExhausted = true)
+            SalvageDecision.FAIL -> StepResult.Terminal(failWorkflow(sessionId, stageId, reason, retryExhausted = true))
+            // The judge chose recovery: route to the recovery stage (file_write is the capability it
+            // provides). Degrade to terminal if the graph declares no recovery stage.
+            SalvageDecision.RECOVER ->
+                routeToRecovery(ctx, stageId, gate, "file_write", reason, state)
+                    ?: StepResult.Terminal(failWorkflow(sessionId, stageId, reason, retryExhausted = true))
        }
    }

@@ -505,6 +505,8 @@ abstract class SessionOrchestrator(
        val agentInstructionsEntries = emptyList<ContextEntry>()
        val retryFeedbackEntries = buildRetryFeedbackEntry(sessionEvents, stageId)
            ?.let { listOf(it) } ?: emptyList()
+        val recoveryTicketEntries = buildRecoveryTicketEntry(sessionEvents, stageId)
+            ?.let { listOf(it) } ?: emptyList()
        val vocabularyEntries = artifactKindRegistry
            ?.takeIf { stageConfig.metadata["injectArtifactKinds"] == "true" }
            ?.let { listOf(buildArtifactKindVocabularyEntry(it.list())) } ?: emptyList()
@@ -532,7 +534,7 @@ abstract class SessionOrchestrator(
            systemPrompt + profileEntries + projectProfileEntries + agentInstructionsEntries +
                journalEntries + repoMapEntries + claimedTaskEntries +
                needsEntries + schemaEntries + vocabularyEntries + promptEntries + steeringEntries +
-                clarificationEntries + retryFeedbackEntries,
+                clarificationEntries + retryFeedbackEntries + recoveryTicketEntries,
        )
        val contextPack = runCatching {
            contextPackBuilder.build(
@@ -1021,7 +1023,7 @@ abstract class SessionOrchestrator(
                ),
            )
            val deniedByStage = toolCall.function.name != STAGE_COMPLETE_TOOL &&
-                toolCall.function.name !in stageConfig.allowedTools
+                toolCall.function.name !in stageConfig.effectiveAllowedTools
            val deniedByReadOnly = !deniedByStage && isReadOnlyMode(sessionId) &&
                (tool?.requiredCapabilities?.contains(ToolCapability.FILE_WRITE) == true)
            if (deniedByStage || deniedByReadOnly) {
@@ -2563,7 +2565,7 @@ abstract class SessionOrchestrator(
            tools = if (!withTools) {
                emptyList()
            } else {
-                stageConfig.allowedTools
+                stageConfig.effectiveAllowedTools
                    .mapNotNull { effectives.registry?.resolve(it) }
                    .filter { tool ->
                        // ponytail: filter write tools while read-before-write block is active; restored once a read completes