macro-agent 0.1.7 → 0.1.10

package/docs/conflict-recovery.md

@@ -0,0 +1,472 @@
+# Conflict Recovery Design (Draft)
+
+Parallel design to `docs/workspace-interfaces.md`. Defines how macro-agent handles merge/rebase/cascade conflicts produced by `WorkspaceManager` operations.
+
+**Status**: draft for iteration. Not yet implemented.
+
+**Relationship to other docs:**
+- `git-cascade-integration-gaps.md` — narrative, established that conflict recovery is its own concern (§9 Open).
+- `workspace-interfaces.md` — defines `ConflictRecord`, `WorkspaceManager.resolveConflict()`. This doc defines how recovery is *driven*.
+
+---
+
+## 1. Problem statement
+
+Conflicts arise from four operations:
+
+- `mergeStream(source → target)` — merge conflict between branches
+- `syncWithParent(stream)` — rebase conflict when parent advanced
+- `rebaseOntoStream(stream → target)` — same
+- `cascadeRebase({ rootStreamId })` — one or more dependent streams fail to rebase
+
+git-cascade records each via `createConflict()`; the stream is marked `conflicted`. Today, macro-agent has no handling — strategies use raw `execSync('git rebase')` which fails hard, and the event is lost.
+
+The redesign needs a **pluggable recovery mechanism** that sits between detection (git-cascade's conflict model) and resolution (which might be git strategies, a spawned resolver agent, a human, or abandonment). The mechanism must work for all four sources.
+
+---
+
+## 2. When conflicts happen — lifecycle
+
+```
+Landing / sync / cascade operation
+    │
+    ├─ Success: MergeResult.success = true
+    │            stream.committed event emitted
+    │
+    └─ Conflict: ConflictRecord created in git-cascade
+                 stream marked 'conflicted'
+                 stream.conflicted event emitted
+                 MergeResult.success = false, conflictId returned
+                 │
+                 ├─ LandingStrategy / operation caller receives result
+                 │
+                 └─ workspaceManager.recoverConflict({ conflictId, strategyName? })
+                           │
+                           └─ Recovery strategy dispatches
+                                      │
+                                      ├─ Sync resolution → returns ConflictResolution immediately
+                                      └─ Async resolution → returns pending; events fire on completion
+```
+
+**Key separation**: LandingStrategy decides *how to land*; ConflictRecoveryStrategy decides *what to do when landing fails*. They're parallel concerns, not nested.
+
+---
+
+## 3. Core types
+
+```ts
+export interface ConflictContext {
+  conflictId: string;
+  streamId: StreamId;                  // the conflicted stream
+  sourceCommit?: string;
+  targetCommit?: string;
+  targetStreamId?: StreamId;           // where we were trying to land (if applicable)
+  paths: string[];                     // conflicting file paths
+  operation: "merge" | "sync" | "rebase" | "cascade";
+  landingAgentId?: AgentId;            // who was trying to land (may be terminated)
+  recoveryDepth: number;               // for bounded recursion
+  strategyConfig?: Record<string, unknown>;
+  workspaceManager: WorkspaceManager;
+  agentManager: AgentManager;
+  inboxAdapter: InboxAdapter;
+}
+
+export type ConflictResolution =
+  | { kind: "resolved"; resolutionCommit: string }
+  | { kind: "deferred"; reason: string }
+  | { kind: "abandoned"; streamId: StreamId; reason: string }
+  | { kind: "escalated"; escalatedTo: AgentId | "human" }
+  | { kind: "retry-after"; backoffMs: number; reason: string }
+  | { kind: "failed"; error: string };
+
+export type ConflictResolutionMode = "sync" | "async";
+
+export interface ConflictRecoveryStrategy {
+  readonly name: string;
+  readonly mode: ConflictResolutionMode;
+
+  canHandle?(ctx: ConflictContext): boolean;
+  recover(ctx: ConflictContext): Promise<ConflictResolution>;
+
+  initialize?(): Promise<void>;
+  close?(): Promise<void>;
+}
+```
+
+**Rationale:**
+- `recoveryDepth` prevents infinite recursion if a resolver creates new conflicts.
+- `mode` is declarative — callers can decide whether to await or fire-and-forget based on it.
+- `operation` tells strategies what produced the conflict (merge-specific resolution differs from rebase-specific).
+- `landingAgentId` may be null or already terminated; strategies can't rely on it being alive.
+
+---
+
+## 4. Built-in strategies
+
+Five registered by default. Teams pick via YAML `on_conflict:` directive.
+
+### 4.1 `auto-resolve` — sync, git-native strategies
+
+```ts
+// strategyConfig: { strategy: "ours" | "theirs" | "union" }
+// Runs raw git merge with -X <strategy>; commits the resolution.
+// Returns: { kind: "resolved", resolutionCommit } or { kind: "failed", ... }
+```
+
+Use cases: low-risk conflicts where one side is authoritative. Example: a reviewer branch that should always defer to coder's changes.
+
+**Constraint**: Only applicable to `operation: "merge"`. Rebase conflicts need different handling.
+
+### 4.2 `defer` — sync, no-op
+
+```ts
+// Creates no new state; conflict record persists, stream stays 'conflicted'.
+// Returns: { kind: "deferred", reason: "no strategy configured" }
+```
+
+Use cases: the team doesn't want automated recovery. Something else (human, external process) will deal with it.
+
+### 4.3 `spawn-resolver` — async, LLM-driven
+
+```ts
+// strategyConfig: {
+//   role: string;               // e.g., "resolver"
+//   max_concurrent?: number;    // default 2
+//   timeout_ms?: number;        // default 1800000 (30 min)
+//   prompt_template?: string;   // optional prompt override
+// }
+//
+// 1. Emit conflict.recovery.started event
+// 2. AgentManager.spawn({ role, options: { conflictId, streamId, paths } })
+// 3. Resolver's topology places it on the conflicted stream via
+//    workspace: new_stream, stream_lineage: track_existing_branch,
+//    track_branch: <conflicted_branch>.
+// 4. Resolver has capabilities: [workspace.commit, workspace.resolve, workspace.read]
+// 5. Resolver reads conflict markers, edits, commits, calls resolve_conflict MCP tool.
+// 6. resolve_conflict → workspaceManager.resolveConflict({ conflictId, resolutionCommit })
+//    → emits conflict.resolved event
+//    → stream status back to 'active'
+// 7. Original operation retried (if requested via config).
+//
+// Returns: { kind: "resolved", resolutionCommit } on success
+//          { kind: "escalated", escalatedTo: "human" } on timeout
+```
+
+Use cases: default for teams with LLM budget. Handles most real conflicts via LLM judgment.
+
+**Recursion bound**: if resolver's `resolve_conflict` call produces a new conflict, `recoveryDepth` increments. Default max = 3; beyond that, falls back to `escalate`.
+
+### 4.4 `abandon` — sync, give up
+
+```ts
+// Calls workspaceManager.abandonStream(streamId, { reason }).
+// If operation was 'merge', source stream is abandoned; target untouched.
+// If operation was 'sync' / 'rebase' / 'cascade', conflicted stream is abandoned
+//   (cascade: just the failing stream; siblings continue per cascade strategy).
+// Returns: { kind: "abandoned", streamId, reason }
+```
+
+Use cases: throwaway exploration streams; CI-driven teams where broken work is discarded.
+
+### 4.5 `escalate` — async, human-in-the-loop
+
+```ts
+// strategyConfig: { notify: AgentId[] | "team-root" | "external-channel" }
+//
+// 1. Create inbox message with high importance to configured recipients
+// 2. Pause stream via workspaceManager.pauseStream(streamId, "awaiting human")
+// 3. Emit conflict.recovery.escalated event
+// 4. Wait indefinitely for external resolve_conflict call
+// 5. On external resolution: resume stream, emit conflict.resolved
+//
+// Returns: { kind: "escalated", escalatedTo }
+```
+
+Use cases: high-stakes merges; teams with human reviewers; fallback from `spawn-resolver` timeout.
+
+---
+
+## 5. Integration with LandingStrategy
+
+Landing strategies detect conflicts but don't handle them directly. Contract:
+
+```ts
+// Inside a LandingStrategy.land(ctx):
+
+const mergeResult = await ctx.workspaceManager.mergeStream({ ... });
+if (!mergeResult.success && mergeResult.conflictId) {
+  // Do NOT call recoverConflict from inside the strategy.
+  // Return the conflict up; the caller (agent's done flow) handles it.
+  return {
+    success: false,
+    conflictId: mergeResult.conflictId,
+  };
+}
+```
+
+**Why not in-strategy recovery?** Three reasons:
+1. Strategies can be registered without recovery knowledge. Separation of concerns.
+2. Recovery policy is role-level (`on_conflict:` YAML), not strategy-level.
+3. Async recovery means strategy would have to block on recovery resolution — fragile.
+
+Instead, the agent's `done()` flow owns recovery dispatch:
+
+```ts
+// In AgentManagerV2's done handler (or land MCP tool handler):
+
+const landingResult = await workspaceManager.land({ ... });
+if (!landingResult.success && landingResult.conflictId) {
+  const roleConfig = topologyPolicy.getRoleConfig(agent.role);
+  const recoveryStrategy = roleConfig.on_conflict ?? teamDefault.conflict_recovery.default_strategy;
+
+  const resolution = await workspaceManager.recoverConflict({
+    conflictId: landingResult.conflictId,
+    strategyName: recoveryStrategy,
+    strategyConfig: roleConfig.conflict_recovery_config,
+    operation: "merge",
+    landingAgentId: agent.id,
+  });
+
+  // Handle resolution result — may be async; if so, agent can terminate
+  // and let recovery run in background.
+}
+```
+
+---
+
+## 6. `WorkspaceManager` additions
+
+Extend the interface from `workspace-interfaces.md` §5:
+
+```ts
+interface WorkspaceManager {
+  // ... existing methods ...
+
+  // Conflict recovery registry
+  registerConflictRecoveryStrategy(strategy: ConflictRecoveryStrategy): void;
+  unregisterConflictRecoveryStrategy(name: string): void;
+  getConflictRecoveryStrategy(name: string): ConflictRecoveryStrategy | null;
+
+  // Recovery dispatch
+  recoverConflict(opts: {
+    conflictId: string;
+    strategyName?: string;              // defaults to team default
+    strategyConfig?: Record<string, unknown>;
+    operation?: "merge" | "sync" | "rebase" | "cascade";
+    landingAgentId?: AgentId;
+  }): Promise<ConflictResolution>;
+
+  // Existing method, formalized
+  resolveConflict(opts: {
+    conflictId: string;
+    resolvedBy: Principal;
+    resolutionCommit?: string;
+  }): void;                              // Called by resolver agents via MCP tool.
+                                         // Emits conflict.resolved event.
+
+  // Listing / inspection
+  listConflicts(filter?: {
+    streamId?: StreamId;
+    resolved?: boolean;
+  }): ConflictRecord[];
+
+  getConflict(conflictId: string): ConflictRecord | null;
+}
+```
+
+New events (extend `WorkspaceEvent` union):
+
+```ts
+| { kind: "conflict.created"; conflictId: string; streamId: StreamId; operation: string }
+| { kind: "conflict.recovery.started"; conflictId: string; strategyName: string }
+| { kind: "conflict.recovery.escalated"; conflictId: string; escalatedTo: string }
+| { kind: "conflict.recovery.timed-out"; conflictId: string; strategyName: string }
+| { kind: "conflict.resolved"; conflictId: string; resolutionCommit?: string; resolvedBy: Principal }
+| { kind: "conflict.abandoned"; conflictId: string; reason: string }
+```
+
+---
+
+## 7. YAML configuration
+
+```yaml
+macro_agent:
+  workspace:
+    roles:
+      worker:
+        on_conflict: defer              # strategy name
+      coder:
+        on_conflict: spawn-resolver
+        conflict_recovery_config:
+          role: resolver
+          timeout_ms: 1200000
+      resolver:
+        workspace: new_stream
+        stream_lineage: track_existing_branch
+        # track_branch is set at spawn time from conflict context
+        landing: none                   # resolver doesn't land itself
+        capabilities:
+          - workspace.commit
+          - workspace.resolve
+          - workspace.read
+
+  # Team-level defaults
+  conflict_recovery:
+    default_strategy: spawn-resolver    # used when role has no on_conflict
+    default_config:
+      role: resolver
+      max_concurrent: 2
+      timeout_ms: 1800000
+    escalation_target: team-root        # for escalate strategy
+    max_recovery_depth: 3               # recursion bound
+```
+
+**Resolution order for strategy selection:**
+1. Role-level `on_conflict` (most specific)
+2. Team-level `conflict_recovery.default_strategy`
+3. Hardcoded fallback: `defer` (safest default)
+
+---
+
+## 8. MCP tool surface
+
+New capability: `workspace.resolve`. Registered tools:
+
+```ts
+// ── Capability: workspace.resolve ────────────────────────────────
+export const resolveConflictToolInput = z.object({
+  conflictId: z.string(),
+  resolutionCommit: z.string().optional(),  // inferred from current HEAD if omitted
+});
+// Handler: workspaceManager.resolveConflict({ conflictId, resolvedBy: agentId, resolutionCommit })
+//   → emits conflict.resolved event
+//   → returns { resolved: true, conflictId, resolutionCommit }
+
+export const listConflictsToolInput = z.object({
+  streamId: z.string().optional(),
+  resolvedOnly: z.boolean().default(false),
+});
+// Handler: workspaceManager.listConflicts({ ... })
+
+export const getConflictToolInput = z.object({
+  conflictId: z.string(),
+});
+// Handler: workspaceManager.getConflict(conflictId)
+```
+
+**Capability → tool mapping:**
+
+| Capability | Tools |
+|---|---|
+| `workspace.resolve` | `resolve_conflict`, `list_conflicts`, `get_conflict` |
+
+Resolver roles need `workspace.resolve` + `workspace.commit` + `workspace.read`.
+
+---
+
+## 9. End-to-end flow: `spawn-resolver`
+
+```
+[Coder agent] lands via queue-to-branch strategy
+  └─ queue-to-branch.land(ctx):
+       workspaceManager.addToMergeQueue({ streamId: coderStream, targetBranch })
+       returns { success: true, queuedAt: ... }
+
+[Integrator agent] drains queue
+  └─ integrator.processNextMergeRequest():
+       next = workspaceManager.getNextToMerge(targetBranch)
+       result = workspaceManager.mergeStream({ sourceStreamId: next.streamId, ... })
+       → CONFLICT: result.success = false, result.conflictId = "c-abc"
+       workspaceManager.markMergeQueueReady(next.id, { status: "conflicted" })
+
+[Integrator's landing flow sees conflict in result]
+  └─ done handler:
+       roleConfig.on_conflict = "spawn-resolver" (from YAML)
+       resolution = await workspaceManager.recoverConflict({
+         conflictId: "c-abc",
+         strategyName: "spawn-resolver",
+         strategyConfig: { role: "resolver", timeout_ms: 1200000 },
+         operation: "merge",
+         landingAgentId: integratorId,
+       })
+
+[SpawnResolverStrategy.recover(ctx)]
+  emit conflict.recovery.started
+  agentId = await agentManager.spawn({
+    role: "resolver",
+    parent: ctx.landingAgentId,
+    task: `Resolve conflict ${ctx.conflictId} on stream ${ctx.streamId}`,
+    options: {
+      conflictId: ctx.conflictId,
+      streamId: ctx.streamId,
+      paths: ctx.paths,
+    },
+  })
+  // Topology places resolver on conflicted stream via track_existing_branch
+  // Resolver wakes up, reads conflict markers in worktree, edits files,
+  // calls `commit` MCP tool, then `resolve_conflict` MCP tool.
+
+[Resolver agent]
+  calls commit({ message: "resolve: merge conflict in auth/middleware.ts" })
+    → commitChanges returns { commit: "def456", changeId: "Change-I..." }
+  calls resolve_conflict({ conflictId: "c-abc" })
+    → workspaceManager.resolveConflict({ conflictId, resolvedBy, resolutionCommit: "def456" })
+    → emits conflict.resolved
+    → stream status → 'active'
+
+[SpawnResolverStrategy] subscribed to conflict.resolved
+  returns { kind: "resolved", resolutionCommit: "def456" }
+
+[Integrator's done handler]
+  resolution.kind === "resolved"
+  → retry mergeStream(coderStream → targetBranch) — this time it succeeds
+  → mark queue entry merged
+  → done
+```
+
+**Timing**: spawn-resolver is async from the caller's POV but the caller awaits the returned Promise. For truly fire-and-forget, caller doesn't await — just subscribes to `conflict.resolved` event.
+
+---
+
+## 10. Edge cases
+
+**E1: Resolver creates a new conflict.**
+Resolver's `commit` succeeds; but when the retry of the original merge runs, it conflicts again (e.g., rebase moved the target). `recoveryDepth` increments; strategy re-enters. Capped at `max_recovery_depth`; beyond that, falls back to `escalate`.
+
+**E2: Multiple conflicts on the same stream.**
+git-cascade allows multiple `ConflictRecord` entries per stream. Recovery serializes — `recoverConflict` for conflict B waits if conflict A is in-progress. Implementation: per-stream recovery lock.
+
+**E3: Resolver crashes / times out.**
+`spawn-resolver` has a `timeout_ms`. On timeout, strategy returns `{ kind: "escalated", escalatedTo: "human" }` — escalation happens via inbox message. Original agent sees escalation in its resolution result.
+
+**E4: Landing agent already terminated when resolution completes.**
+`landingAgentId` is advisory, not required. The retry logic in the agent's done handler is gone; so who retries? Options:
+- (a) Recovery strategy doesn't trigger retry; caller subscribes to event and handles
+- (b) Recovery strategy itself triggers retry via a "retry" phase
+
+Proposal: **(a)**. The resolver's `resolve_conflict` tool does not retry the original operation. Downstream retry is a separate concern handled by the integrator/queue drainer, which should pick up the now-resolved stream on next iteration. Avoids complex state in strategies.
+
+**E5: Conflict on `cascadeRebase`.**
+Cascade produces N potential conflicts across dependents. `CascadeResult.failed[]` lists each with a conflictId. Caller iterates and calls `recoverConflict` per failed stream, typically in parallel. Cascade strategy `defer_conflicts` lets the cascade complete past failures; recovery kicks in per stream.
+
+**E6: Abandoning a stream mid-recovery.**
+If a human abandons a conflicted stream while recovery is in-progress, the resolver agent's subsequent `resolve_conflict` call fails with `stream_abandoned`. Resolver handles gracefully (reports failure, terminates). Recovery strategy emits `conflict.recovery.failed` with reason.
+
+---
+
+## 11. Open questions
+
+- **Retry ownership** (E4). Proposed (a): caller subscribes to `conflict.resolved` and retries if it still cares. Downside: lost retries if no one subscribes. Alternative: retry hook on the original LandingStrategy. TBD.
+- **Recovery for read-only operations.** `stream_status` tool returning stale data after a conflict — not a recovery concern, but affects UX. Documented as a non-goal.
+- **Cross-team conflicts.** If two teams' streams conflict during merge to shared branch, which team's recovery policy applies? Propose: conflict is attached to the stream being merged; the *owning team's* policy runs. For federated scenarios, escalates to human by default.
+- **Observability.** Recovery can take minutes; need a structured progress channel. Proposal: resolver agents emit `RECOVERY_PROGRESS` signals on a team channel; subscribers can watch. Optional, enabled via YAML.
+- **Cost control for LLM-driven recovery.** Runaway recursion + expensive LLM calls are a risk. Hard cap: `max_recovery_depth` + `max_concurrent` per strategy config. Tokens should be tracked separately.
+
+---
+
+## 12. What's not defined here
+
+- Implementation of resolver role prompt / skills.
+- Conflict resolution heuristics (e.g., how an LLM resolver approaches different conflict types) — belongs in resolver role definition, not the strategy infrastructure.
+- UI / dashboard for human-escalated conflicts.
+- Cross-process conflict coordination (multi-instance macro-agent sharing a repo).