feat: [ENG-2323] brv harness reset + refine commands

src/agent/core/interfaces/i-harness-store.ts

@@ -66,6 +66,13 @@
    */
   deleteOutcomes(projectId: string, commandType: string): Promise<number>
 
+  /**
+   * Remove the pin for a `(projectId, commandType)` pair.
+   * Idempotent — returns `true` if a pin existed and was removed,
+   * `false` if no pin was set.
+   */
+  deletePin(projectId: string, commandType: string): Promise<boolean>
+
   /**
    * Delete a single scenario by its `(projectId, commandType, scenarioId)` key.
    * Returns `true` when the scenario existed and was deleted; `false` on miss.
@@ -78,6 +85,23 @@
     scenarioId: string,
   ): Promise<boolean>
 
+  /**
+   * Delete every scenario for a `(projectId, commandType)` pair.
+   * Returns the number of records deleted. Used by `brv harness reset`.
+   */
+  deleteScenarios(projectId: string, commandType: string): Promise<number>
+
+  /**
+   * Delete a single version by its `(projectId, commandType, versionId)` key.
+   * Returns `true` when the version existed and was deleted; `false` on miss.
+   * Used by `brv harness reset` to clear all versions for a pair.
+   */
+  deleteVersion(
+    projectId: string,
+    commandType: string,
+    versionId: string,
+  ): Promise<boolean>
+
   /**
    * Return the most-recently-written version for a `(projectId, commandType)`
    * pair — ranked by the stored `version` number, not by `heuristic`. This
@@ -192,7 +216,7 @@
    * Persist a new harness version. Templates bootstrap as v1; refinements
    * write v2, v3, … each pointing to a parent.
    *
    * @throws {HarnessStoreError} with code `VERSION_CONFLICT` when a version
    *   with the same `id`, or the same `(projectId, commandType, version)`
    *   tuple, already exists.
    */

src/agent/infra/agent/cipher-agent.ts

@@ -13,6 +13,7 @@ import type {AgentState, ExecutionContext, ICipherAgent} from '../../core/interf
 import type {IHistoryStorage} from '../../core/interfaces/i-history-storage.js'
 import type {ITokenizer} from '../../core/interfaces/i-tokenizer.js'
 import type {FileSystemService} from '../file-system/file-system-service.js'
+import type {HarnessSynthesizer} from '../harness/harness-synthesizer.js'
 import type {MemoryManager} from '../memory/memory-manager.js'
 import type {ProcessService} from '../process/process-service.js'
 import type {SystemPromptManager} from '../system-prompt/system-prompt-manager.js'
@@ -513,6 +514,15 @@ export class CipherAgent extends BaseAgent implements ICipherAgent {
     }
   }
 
+  /**
+   * Public accessor for the harness synthesizer. Used by `brv harness refine`
+   * via the agent-process task dispatch. Returns `undefined` when harness is
+   * disabled (synthesizer was never instantiated).
+   */
+  public getHarnessSynthesizer(): HarnessSynthesizer | undefined {
+    return this.services?.harnessSynthesizer
+  }
+
   /**
    * Get an existing session or create a new one.
    */
@@ -572,7 +582,7 @@ export class CipherAgent extends BaseAgent implements ICipherAgent {
 
   protected override async initializeServices(): Promise<CipherAgentServices> {
     // Pass pre-created event bus to service initializer
-    return createCipherAgentServices(this.config, this._agentEventBus)
+    return createCipherAgentServices(this.config, this._agentEventBus, () => this.buildHttpConfig())
   }
 
   /**
@@ -660,6 +670,8 @@ export class CipherAgent extends BaseAgent implements ICipherAgent {
     this.rebindCurateTools(services, httpConfig, sessionLLMConfig)
   }
 
+  // === Protected Methods (implement abstract from BaseAgent) ===
+
   /**
    * Reset the agent to initial state.
    * Resets execution state only. To reset sessions, use resetSession(sessionId).
@@ -671,8 +683,6 @@ export class CipherAgent extends BaseAgent implements ICipherAgent {
     }
   }
 
-  // === Protected Methods (implement abstract from BaseAgent) ===
-
   /**
    * Reset a specific session's conversation history.
    * @param sessionId - The session ID to reset

src/agent/infra/agent/service-initializer.ts

@@ -135,9 +135,10 @@
  * @param agentEventBus - Pre-created event bus from agent constructor
  * @returns Initialized shared services
  */
 export async function createCipherAgentServices(
   config: ValidatedAgentConfig,
   agentEventBus: AgentEventBus,
+  httpConfigProvider?: () => ByteRoverHttpConfig,
 ): Promise<CipherAgentServices> {
   // 1. Logger (uses provided event bus )
   const logger = new EventBasedLogger(agentEventBus, 'CipherAgent')
@@ -333,6 +334,9 @@
         : config.providerApiKey,
       baseUrl: config.providerBaseUrl,
       headers: config.providerHeaders,
+      // byterover provider needs httpConfig for API routing (sessionKey,
+      // projectId, spaceId, teamId). Resolved lazily from CipherAgent.
+      httpConfig: httpConfigProvider ? {...httpConfigProvider()} : undefined,
       httpReferer: config.httpReferer,
       maxTokens: 4096,
       model: refinementModel,

src/agent/infra/harness/harness-store.ts

@@ -119,6 +119,14 @@ export class HarnessStore implements IHarnessStore {
 
   // ── scenarios ─────────────────────────────────────────────────────────────
 
+  async deletePin(projectId: string, commandType: string): Promise<boolean> {
+    const key = this.pinKey(projectId, commandType)
+    const exists = await this.keyStorage.get(key)
+    if (exists === undefined) return false
+    await this.keyStorage.delete(key)
+    return true
+  }
+
   async deleteScenario(
     projectId: string,
     commandType: string,
@@ -149,6 +157,51 @@ export class HarnessStore implements IHarnessStore {
 
   // ── versions ───────────────────────────────────────────────────────────────
 
+  async deleteScenarios(projectId: string, commandType: string): Promise<number> {
+    const keys: StorageKey[] = []
+    for (const projectType of ProjectTypeSchema.options) {
+      // eslint-disable-next-line no-await-in-loop
+      const entries = await this.keyStorage.listWithValues<EvaluationScenario>([
+        HARNESS_PREFIX,
+        SCENARIO_PREFIX,
+        projectType,
+        projectId,
+        commandType,
+      ])
+      for (const entry of entries) keys.push(entry.key)
+    }
+
+    if (keys.length === 0) return 0
+
+    const operations: BatchOperation[] = keys.map((key) => ({key, type: 'delete' as const}))
+    await this.keyStorage.batch(operations)
+    this.logger.debug('HarnessStore.deleteScenarios cleared partition', {
+      commandType,
+      deleted: keys.length,
+      projectId,
+    })
+
+    return keys.length
+  }
+
+  async deleteVersion(
+    projectId: string,
+    commandType: string,
+    versionId: string,
+  ): Promise<boolean> {
+    const key = this.versionKey(projectId, commandType, versionId)
+    const exists = await this.keyStorage.exists(key)
+    if (!exists) return false
+
+    await this.keyStorage.delete(key)
+    this.logger.debug('HarnessStore.deleteVersion removed entry', {
+      commandType,
+      projectId,
+      versionId,
+    })
+    return true
+  }
+
   async getLatest(projectId: string, commandType: string): Promise<HarnessVersion | undefined> {
     // Delegate to `listVersions` rather than re-deriving the "max version"
     // comparator — a future change to the sort key can't silently break

src/agent/infra/llm/agent-llm-service.ts

@@ -29,6 +29,7 @@ import type {CompactionService} from './context/compaction/compaction-service.js
 import type {ICompressionStrategy} from './context/compression/types.js'
 
 import {getErrorMessage} from '../../../server/utils/error-helpers.js'
+import {sanitizeProjectPath} from '../../../server/utils/path-utils.js'
 import {AgentStateMachine} from '../../core/domain/agent/agent-state-machine.js'
 import {AgentState, TerminationReason} from '../../core/domain/agent/agent-state.js'
 import {LlmGenerationError, LlmMaxIterationsError, LlmResponseParsingError} from '../../core/domain/errors/llm-error.js'
@@ -901,14 +902,11 @@ export class AgentLLMService implements ILLMService {
     }
 
     try {
-      // Slug/path gap workaround (known issue — see
-      // outcome-collection.test.ts:32): the recorder derives projectId
-      // from `environmentContext.workingDirectory`, so we use the same
-      // source here. `bootstrapIfNeeded` takes both `projectId` (for
-      // store key partitioning) and `workingDirectory` (for filesystem
-      // detection); at present they're the same value, aliased for
-      // readability at the call site.
-      const projectId = this.workingDirectory
+      // projectId is the sanitized working directory — FileKeyStorage
+      // rejects path separators in key segments, so the raw absolute
+      // path cannot be used as a store partition key. `workingDirectory`
+      // stays unsanitized for filesystem detection (template language).
+      const projectId = sanitizeProjectPath(this.workingDirectory)
       const {workingDirectory} = this
 
       await harnessBootstrap.bootstrapIfNeeded(projectId, commandType, workingDirectory)

src/agent/infra/sandbox/sandbox-service.ts

@@ -21,6 +21,7 @@ import type { HarnessOutcomeRecorder } from '../harness/harness-outcome-recorder
 import type { SessionManager } from '../session/session-manager.js'
 import type { ISearchKnowledgeService, ToolsSDK } from './tools-sdk.js'
 
+import { sanitizeProjectPath } from '../../../server/utils/path-utils.js'
 import { ProjectTypeSchema } from '../../core/domain/harness/types.js'
 import {HarnessEvaluatorError} from '../harness/harness-evaluator-errors.js'
 import { OpsCounter } from '../harness/ops-counter.js'
@@ -232,7 +233,7 @@ export class SandboxService implements ISandboxService {
             conversationTurn: config?.conversationTurn,
             executionTimeMs: result.executionTime,
             harnessVersionId: this.harnessVersionIdBySession.get(sessionId),
-            projectId: this.environmentContext.workingDirectory,
+            projectId: sanitizeProjectPath(this.environmentContext.workingDirectory),
             projectType: this.resolveProjectType(),
             result,
             sessionId,