loopforge 2.6.0 → 2.11.0

package/skills/perception/SKILL.md

@@ -30,6 +30,31 @@ Keep the returned `sessionId`, `roundId`, prompt, and level.
 3. Check actual evidence before claiming success.
 4. Call `loopforge_next` with `sessionId` and a structured `evaluation`.
 
+### Evaluation fields
+
+| Field | Rule |
+| --- | --- |
+| `success` | `true` only when the whole objective and every hard constraint are satisfied. |
+| `output_summary` | State what changed and what was verified this round. |
+| `constraint_violations` | Report real violations. Use constraint IDs (`c-XXXXXXXX`) for exact matching. Do not hide violations to force progress. |
+| `should_continue` | `false` only when the entire task is complete. |
+| `execution_evidence` | `files_changed`, `test_results`, `success_criteria_met`, `success_criteria_remaining`, `progress_estimate`. Use criterion IDs (`cr-XXXXXXXX`) in `success_criteria_met`/`_remaining`. |
+| `discovered_constraints` | New guardrails found this round. Will be assigned stable IDs by the compiler. |
+| `objective_refinement` | Refine the objective when repository evidence changes its meaning. |
+| `emerged_subtasks` | New sub-problems that surfaced. Use clear, distinct descriptions. |
+| `completed_subtasks` | Sub-goal IDs (`sg-XXXXXXXX`) or descriptions you finished this round. |
+| `blocked_subtasks` | Sub-goal IDs or descriptions that cannot proceed. |
+| `canceled_subtasks` | Sub-goal IDs or descriptions that are no longer needed. |
+| `retracted_constraints` | Constraints the agent now believes are wrong or irrelevant. |
+| `revised_success_criteria` | Old → new revisions for criteria discovered to be incorrect. |
+| `wrong_assumptions` | Assumptions disproved this round. |
+| `next_action` | What you plan to do in the next round. Be specific — the verification gate compares this against your actual output. |
+| `compression_checkpoint` | Set `true` and include `checkpoint_label` when you complete a major phase. |
+| `drift_clarification` | Required when you pivot from your declared `next_action`. Must include specific constraint/criterion/sub-goal IDs or file paths. Filler text without concrete references will be rejected. |
+| `prompt_requests` | `emphasize` (what to highlight next round), `expand` (sections to expand), `confusion_points` (what you're confused about). |
+| `stop_reason` | When `should_continue=false` and `success=false`: `"gave_up"`, `"blocked"`, or `"needs_human_input"`. |
+| `worker_results` | Record delegated subtask results and discoveries from sub-agents. |
+
 Minimal evaluation:
 
 ```json
@@ -37,26 +62,68 @@ Minimal evaluation:
   "success": false,
   "output_summary": "Implemented the parser and added 8 passing tests.",
   "constraint_violations": [],
-  "should_continue": true
+  "should_continue": true,
+  "execution_evidence": {
+    "files_changed": ["src/parser.ts", "tests/parser.test.ts"],
+    "test_results": { "passed": 8, "failed": 0, "skipped": 0 },
+    "success_criteria_met": [],
+    "success_criteria_remaining": ["All parser edge cases covered"],
+    "progress_estimate": 0.4
+  }
 }
 ```
 
-Important fields:
+## Using stable IDs
 
-| Field | Rule |
-| --- | --- |
-| `success` | Use `true` only when the whole objective and every hard constraint are satisfied. |
-| `output_summary` | State what changed and what was verified. |
-| `constraint_violations` | Report real violations. Do not hide them to force progress. |
-| `should_continue` | Use `false` only when the entire task is complete. |
-| `execution_evidence` | Report changed files, test counts, met and remaining criteria, and progress when known. |
-| `discovered_constraints` | Add newly discovered guardrails. |
-| `objective_refinement` | Refine the objective when repository evidence changes its meaning. |
-| `wrong_assumptions` | Record assumptions disproved this round. |
-| `worker_results` | Record delegated subtask results and discoveries. |
+The prompt renders stable IDs beside every constraint, criterion, and sub-goal:
+
+- **Constraints:** `[c-a3f2b1c0] No plaintext passwords`
+- **Criteria:** `[cr-b5a6c7d8] Unit test coverage ≥ 90%`
+- **Sub-goals:** `[sg-f1e2d3c4] Add error handling to login`
+
+**Prefer IDs over descriptions** in `constraint_violations`, `success_criteria_met`/`_remaining`, `completed_subtasks`, `blocked_subtasks`, `canceled_subtasks`, and `drift_clarification`. Exact ID matching eliminates ambiguity — Jaccard similarity is a fallback, not the primary path.
+
+## Backtrack — when the loop rewinds
 
-Do not embed an evaluation marker in ordinary output when the MCP client accepts
-the structured `evaluation` argument.
+When progress stalls for multiple rounds, LoopForge triggers a **backtrack**:
+the round counter rolls back to the last clean round, failed work is discarded,
+and a diagnosis is injected into the prompt.
+
+**When you receive a backtrack prompt:**
+
+1. **Restore your workspace first.** The prompt includes a "Workspace Restore
+   Required" section with exact `git checkout -- . && git clean -fd` commands.
+   Your working directory still contains changes from the failed rounds —
+   you MUST discard them before working.
+2. **Read the restored state.** The prompt reflects the clean round's vault
+   state. The state file shows the correct full picture.
+3. **Do NOT repeat the failed approach.** The diagnosis explains why the
+   previous path failed. Choose a different technique or decomposition.
+4. **Preserved discoveries** from skipped rounds are merged into the active
+   constraints — they remain valid even though the rounds were discarded.
+
+If you skip the workspace restore and work on dirty files, the verification
+gate will **reject** your next submission with a `backtrack_workspace_not_restored`
+flag.
+
+## Drift — when you change direction
+
+The verification gate compares your `next_action` from the previous round
+against your actual `output_summary` this round. A mismatch triggers an
+`intent_drift` flag.
+
+**If you intentionally pivoted**, set `drift_clarification` with:
+- A specific reason for the pivot.
+- At least one concrete reference: a constraint ID (`c-XXXXXXXX`),
+  criterion ID (`cr-XXXXXXXX`), sub-goal ID (`sg-XXXXXXXX`), or changed
+  file path.
+
+✅ Good: `"Pivoted because constraint c-a3f2b1c0 was already satisfied by src/auth.ts. The CSS fix was higher priority."`
+
+❌ Rejected: `"I changed direction because I felt this approach was better."`
+
+Three consecutive weak clarifications (no concrete references) will
+**terminate the loop**.
 
 ## Interpret the result
 
@@ -66,22 +133,27 @@ If `enforcementAction` is `reject`, redo the same logical round. The round ID
 stays stable and the attempt number increases. Follow the retry requirements;
 the rejected attempt has not been committed.
 
+If `enforcementAction` is `backtrack`, the loop has been rolled back. Restore
+your workspace (see Backtrack section above), then work from the restored
+prompt at the new round number.
+
 If `prompt` is null, inspect `stopReason`:
 
 | Reason | Meaning |
 | --- | --- |
-| `task_complete` | The complete objective was reported finished. |
-| `circuit_breaker` | Repeated failed rounds require a different approach or user input. |
+| `completed` | The complete objective was reported finished. |
+| `failed` | The agent gave up (`should_continue=false`, `success=false`). |
+| `blocked` | Work cannot proceed under current constraints. Human input may be needed. |
 | `max_rounds` | The safety limit was reached. Summarize remaining work. |
 | `stalled` | Evaluation or execution did not produce a usable next state. |
-| `enforcement_terminated` | Repeated invalid claims or stalled progress terminated the loop. |
+| `enforcement_terminated` | Repeated invalid claims, stalled progress, or weak clarifications terminated the loop. |
 | `paused` | The durable session remains available for resume. |
 
 ## Prompt levels
 
-- L0 is a same-round retry with the rejection reason and changed evidence.
-- L1 is the normal compact continuation view.
-- L2 is a full rehydration for the first round, a checkpoint, or a goal change.
+- **L0** — Same-round retry with the rejection reason and changed evidence. ~3K chars.
+- **L1** — Normal compact continuation. ~7K chars.
+- **L2** — Full rehydration: first round, checkpoint, goal change, or recovery. In L2 pointer mode (default), the prompt shows a structured dashboard; the full state lives in `.loopforge/state/<loopId>-state.md`. Read it before acting.
 
 These levels control state density only. LoopForge does not choose a reasoning
 technique for you.
@@ -96,10 +168,9 @@ technique for you.
 - Use `loopforge_health` to inspect goal alignment and drift.
 - Use `loopforge_stop` only for an intentional terminal stop.
 
-The optional `.loopforge/state/<loopId>-state.md` file is a readable derived
-view. The typed JSON session and round documents are the recovery truth. If the
-state file exists, it can help you review the full state, but do not assume it is
-enabled.
+The `.loopforge/state/<loopId>-state.md` file is the authoritative full-state
+view in L2 pointer mode. Read it at the start of each L2 round — the prompt
+dashboard is a summary, not the complete picture.
 
 ## Delegation
 
@@ -111,8 +182,12 @@ self-contained subtask and relevant hard constraints. Add the returned result to
 
 1. Execute the prompt instead of generating another prompt.
 2. Submit one honest structured evaluation after each attempt.
-3. Treat required command evidence and verification errors as authoritative.
-4. Keep one loop focused on one user objective.
-5. Do not advance a paused or stopped session through another process.
-6. Continue within the same Agent task when possible; LoopForge preserves state,
-   but it does not create a background Agent.
+3. Use stable IDs (`c-`/`cr-`/`sg-XXXXXXXX`) for exact matching whenever possible.
+4. When backtrack occurs, restore the workspace before working.
+5. When you pivot from your `next_action`, explain with concrete references in `drift_clarification`.
+6. Treat required command evidence and verification errors as authoritative.
+7. Keep one loop focused on one user objective.
+8. Read the state file at the start of L2 rounds — it holds the complete picture.
+9. Do not advance a paused or stopped session through another process.
+10. Continue within the same Agent task when possible; LoopForge preserves state,
+    but it does not create a background Agent.

package/dist/interop.d.ts

@@ -1,35 +0,0 @@
-/** Dependency-free ecosystem checkpoint boundary.
- *
- * **@experimental** — This module is a lightweight stub (57 lines as of v2.0.0).
- * The checkpoint interface is stable but the integration surface (sinks, schema
- * versioning, external consumer contracts) has not been validated in production.
- * Use for prototyping; expect the envelope shape to evolve before v2.1.
- *
- * LangGraph checkpointers, workflow engines, databases, and agent SDKs can
- * consume this neutral envelope without LoopForge importing their packages.
- */
-import type { McpSession } from "./mcp/session.js";
-import type { RoundTransactionSnapshot } from "./round-transaction.js";
-import { getPolicyMetrics } from "./policy-metrics.js";
-export declare const COGNITIVE_CHECKPOINT_SCHEMA_VERSION: 1;
-export interface CognitiveStateCheckpoint {
-    schemaVersion: typeof COGNITIVE_CHECKPOINT_SCHEMA_VERSION;
-    checkpointId: string;
-    loopId: string;
-    sessionId: string;
-    task: string;
-    status: "running" | "stopped" | "stalled" | "paused";
-    round: number;
-    maxRounds: number;
-    roundId?: string;
-    prompt?: string | null;
-    successTrajectory: boolean[];
-    roundSnapshot?: RoundTransactionSnapshot;
-    policyMetrics: ReturnType<typeof getPolicyMetrics>;
-    updatedAt: string;
-}
-export interface CognitiveCheckpointSink {
-    save(checkpoint: CognitiveStateCheckpoint): void | Promise<void>;
-}
-export declare function createCognitiveCheckpoint(session: McpSession, updatedAt?: string): CognitiveStateCheckpoint;
-//# sourceMappingURL=interop.d.ts.map

package/dist/interop.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"interop.d.ts","sourceRoot":"","sources":["../src/interop.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,kBAAkB,CAAC;AACnD,OAAO,KAAK,EAAE,wBAAwB,EAAE,MAAM,wBAAwB,CAAC;AACvE,OAAO,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AAEvD,eAAO,MAAM,mCAAmC,EAAG,CAAU,CAAC;AAE9D,MAAM,WAAW,wBAAwB;IACvC,aAAa,EAAE,OAAO,mCAAmC,CAAC;IAC1D,YAAY,EAAE,MAAM,CAAC;IACrB,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE,SAAS,GAAG,SAAS,GAAG,SAAS,GAAG,QAAQ,CAAC;IACrD,KAAK,EAAE,MAAM,CAAC;IACd,SAAS,EAAE,MAAM,CAAC;IAClB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACvB,iBAAiB,EAAE,OAAO,EAAE,CAAC;IAC7B,aAAa,CAAC,EAAE,wBAAwB,CAAC;IACzC,aAAa,EAAE,UAAU,CAAC,OAAO,gBAAgB,CAAC,CAAC;IACnD,SAAS,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,uBAAuB;IACtC,IAAI,CAAC,UAAU,EAAE,wBAAwB,GAAG,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CAClE;AAED,wBAAgB,yBAAyB,CACvC,OAAO,EAAE,UAAU,EACnB,SAAS,SAA2B,GACnC,wBAAwB,CAoB1B"}

package/dist/interop.js

@@ -1,34 +0,0 @@
-/** Dependency-free ecosystem checkpoint boundary.
- *
- * **@experimental** — This module is a lightweight stub (57 lines as of v2.0.0).
- * The checkpoint interface is stable but the integration surface (sinks, schema
- * versioning, external consumer contracts) has not been validated in production.
- * Use for prototyping; expect the envelope shape to evolve before v2.1.
- *
- * LangGraph checkpointers, workflow engines, databases, and agent SDKs can
- * consume this neutral envelope without LoopForge importing their packages.
- */
-import { getPolicyMetrics } from "./policy-metrics.js";
-export const COGNITIVE_CHECKPOINT_SCHEMA_VERSION = 1;
-export function createCognitiveCheckpoint(session, updatedAt = new Date().toISOString()) {
-    const roundId = session.roundSnapshot?.roundId;
-    return {
-        schemaVersion: COGNITIVE_CHECKPOINT_SCHEMA_VERSION,
-        checkpointId: roundId
-            ? `${roundId}:session`
-            : `loop:${session.loopId}:round:${session.currentRound}:session`,
-        loopId: session.loopId,
-        sessionId: session.sessionId,
-        task: session.task,
-        status: session.status,
-        round: session.currentRound,
-        maxRounds: session.maxRounds,
-        roundId,
-        prompt: session.currentPrompt,
-        successTrajectory: [...session.successTrajectory],
-        roundSnapshot: session.roundSnapshot,
-        policyMetrics: getPolicyMetrics(session.loopId),
-        updatedAt,
-    };
-}
-//# sourceMappingURL=interop.js.map

package/dist/interop.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"interop.js","sourceRoot":"","sources":["../src/interop.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAIH,OAAO,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AAEvD,MAAM,CAAC,MAAM,mCAAmC,GAAG,CAAU,CAAC;AAuB9D,MAAM,UAAU,yBAAyB,CACvC,OAAmB,EACnB,SAAS,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;IAEpC,MAAM,OAAO,GAAG,OAAO,CAAC,aAAa,EAAE,OAAO,CAAC;IAC/C,OAAO;QACL,aAAa,EAAE,mCAAmC;QAClD,YAAY,EAAE,OAAO;YACnB,CAAC,CAAC,GAAG,OAAO,UAAU;YACtB,CAAC,CAAC,QAAQ,OAAO,CAAC,MAAM,UAAU,OAAO,CAAC,YAAY,UAAU;QAClE,MAAM,EAAE,OAAO,CAAC,MAAM;QACtB,SAAS,EAAE,OAAO,CAAC,SAAS;QAC5B,IAAI,EAAE,OAAO,CAAC,IAAI;QAClB,MAAM,EAAE,OAAO,CAAC,MAAM;QACtB,KAAK,EAAE,OAAO,CAAC,YAAY;QAC3B,SAAS,EAAE,OAAO,CAAC,SAAS;QAC5B,OAAO;QACP,MAAM,EAAE,OAAO,CAAC,aAAa;QAC7B,iBAAiB,EAAE,CAAC,GAAG,OAAO,CAAC,iBAAiB,CAAC;QACjD,aAAa,EAAE,OAAO,CAAC,aAAa;QACpC,aAAa,EAAE,gBAAgB,CAAC,OAAO,CAAC,MAAM,CAAC;QAC/C,SAAS;KACV,CAAC;AACJ,CAAC"}

package/dist/run.d.ts

@@ -1,51 +0,0 @@
-/** Autonomous loop runner — Library API entry point (v2.6).
- *
- * Wires the engine, compiler, self-eval extraction, and feedback persistence
- * into a single call that the host controls end-to-end. The agent never gets
- * a chance to "skip" the loop — the host calls execute(), extracts the result,
- * and feeds it back to LoopForge in the same tick.
- */
-import { type SelfEvaluation } from "./protocol.js";
-import type { LoopStore } from "./loop-store.js";
-import type { ExternalContextProvider, LoopTerminalSink } from "./protocol.js";
-import type { PromptArtifact } from "./protocol.js";
-export interface RunInput {
-    /** The task description — what the agent should accomplish. */
-    task: string;
-    /** Hard constraints that must not be violated across all rounds. */
-    constraintsFromPlan?: string[];
-    /** Maximum rounds before auto-stop. Falls back to policy default when omitted. */
-    maxRounds?: number;
-    /** The agent executor. Receives the compiled prompt and an AbortSignal.
-     *  Must return the agent's raw output (text). The runner extracts the
-     *  structured self-evaluation from it. */
-    execute: (prompt: string, context: {
-        signal: AbortSignal;
-    }) => Promise<string>;
-    /** Optional external context injected before each compile. */
-    contextProvider?: ExternalContextProvider;
-    /** Optional terminal event sinks (telemetry, logging, etc.). */
-    terminalSinks?: LoopTerminalSink[];
-    /** Optional typed store — falls back to .loopforge/ when omitted. */
-    store?: LoopStore;
-    /** Optional abort signal for cancelling the entire loop. */
-    signal?: AbortSignal;
-}
-export interface RunResult {
-    /** Whether the loop completed successfully (should_continue=false + success=true). */
-    success: boolean;
-    /** Why the loop stopped. */
-    stopReason: string;
-    /** Number of rounds completed. */
-    roundsCompleted: number;
-    /** True/false for each committed round. */
-    successTrajectory: boolean[];
-    /** The loop's stable identity — can be used to replay later. */
-    loopId: string;
-    /** The last self-evaluation submitted by the agent. */
-    lastEvaluation?: SelfEvaluation;
-    /** Prompt artifacts produced across the loop (for audit). */
-    artifacts: PromptArtifact[];
-}
-export declare function run(input: RunInput): Promise<RunResult>;
-//# sourceMappingURL=run.d.ts.map

package/dist/run.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"run.d.ts","sourceRoot":"","sources":["../src/run.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAGH,OAAO,EAA8B,KAAK,cAAc,EAAE,MAAM,eAAe,CAAC;AAEhF,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AACjD,OAAO,KAAK,EAAE,uBAAuB,EAAqB,gBAAgB,EAAE,MAAM,eAAe,CAAC;AAElG,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,eAAe,CAAC;AAEpD,MAAM,WAAW,QAAQ;IACvB,+DAA+D;IAC/D,IAAI,EAAE,MAAM,CAAC;IACb,oEAAoE;IACpE,mBAAmB,CAAC,EAAE,MAAM,EAAE,CAAC;IAC/B,kFAAkF;IAClF,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;8CAE0C;IAC1C,OAAO,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE;QAAE,MAAM,EAAE,WAAW,CAAA;KAAE,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC;IAC/E,8DAA8D;IAC9D,eAAe,CAAC,EAAE,uBAAuB,CAAC;IAC1C,gEAAgE;IAChE,aAAa,CAAC,EAAE,gBAAgB,EAAE,CAAC;IACnC,qEAAqE;IACrE,KAAK,CAAC,EAAE,SAAS,CAAC;IAClB,4DAA4D;IAC5D,MAAM,CAAC,EAAE,WAAW,CAAC;CACtB;AAED,MAAM,WAAW,SAAS;IACxB,sFAAsF;IACtF,OAAO,EAAE,OAAO,CAAC;IACjB,4BAA4B;IAC5B,UAAU,EAAE,MAAM,CAAC;IACnB,kCAAkC;IAClC,eAAe,EAAE,MAAM,CAAC;IACxB,2CAA2C;IAC3C,iBAAiB,EAAE,OAAO,EAAE,CAAC;IAC7B,gEAAgE;IAChE,MAAM,EAAE,MAAM,CAAC;IACf,uDAAuD;IACvD,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC,6DAA6D;IAC7D,SAAS,EAAE,cAAc,EAAE,CAAC;CAC7B;AAkCD,wBAAsB,GAAG,CAAC,KAAK,EAAE,QAAQ,GAAG,OAAO,CAAC,SAAS,CAAC,CA6J7D"}

package/dist/run.js

@@ -1,200 +0,0 @@
-/** Autonomous loop runner — Library API entry point (v2.6).
- *
- * Wires the engine, compiler, self-eval extraction, and feedback persistence
- * into a single call that the host controls end-to-end. The agent never gets
- * a chance to "skip" the loop — the host calls execute(), extracts the result,
- * and feeds it back to LoopForge in the same tick.
- */
-import { LoopForgeEngine, extractSelfEvaluation } from "./engine.js";
-import { Mode } from "./protocol.js";
-import { FileLoopStore } from "./loop-store.js";
-import { logEvent } from "./observability.js";
-function makeRunResult(loopId, lastResponse, roundsCompleted, successTrajectory, lastEval, artifacts = []) {
-    if (!lastResponse?.response) {
-        return {
-            success: false,
-            stopReason: "executor_failure",
-            roundsCompleted,
-            successTrajectory,
-            loopId,
-            lastEvaluation: lastEval,
-            artifacts,
-        };
-    }
-    return {
-        success: lastEval?.success === true && lastEval?.should_continue === false,
-        stopReason: lastEval?.should_continue === false
-            ? (lastEval?.success ? "completed" : "failed")
-            : "max_rounds",
-        roundsCompleted,
-        successTrajectory,
-        loopId,
-        lastEvaluation: lastEval,
-        artifacts,
-    };
-}
-export async function run(input) {
-    const store = input.store ?? new FileLoopStore();
-    const engine = new LoopForgeEngine(store);
-    const loopId = `run-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 8)}`;
-    const maxRounds = input.maxRounds ?? 20;
-    const artifacts = [];
-    let round = 1;
-    let lastEval;
-    const successTrajectory = [];
-    const task = input.task.trim();
-    if (!task)
-        throw new Error("run: task must be non-empty");
-    while (round <= maxRounds) {
-        // Check for external abort
-        if (input.signal?.aborted) {
-            return makeRunResult(loopId, null, round - 1, successTrajectory, lastEval, artifacts);
-        }
-        // ── 1. Compile prompt ──────────────────────────────────────────────
-        const request = {
-            task,
-            mode: Mode.LOOP_COMPILE,
-            feedback: null,
-            skill_name: null,
-            task_id: null,
-            loop_id: loopId,
-            round,
-            max_rounds: maxRounds,
-            constraints_from_plan: input.constraintsFromPlan ?? [],
-            goal_id: loopId,
-            domain: "",
-        };
-        // External context injection
-        if (input.contextProvider) {
-            try {
-                const ctx = await input.contextProvider({
-                    loopId,
-                    round,
-                    task,
-                    domain: "",
-                    lastEvaluation: lastEval,
-                });
-                if (ctx?.trim())
-                    request.external_context = ctx.trim();
-            }
-            catch {
-                logEvent("context_provider_error", { loopId, round });
-            }
-        }
-        // Forward last round result
-        if (lastEval) {
-            request.last_round_result = {
-                round: round - 1,
-                success: lastEval.success,
-                output_summary: lastEval.output_summary,
-                constraint_violations: lastEval.constraint_violations,
-                manual_fixes_needed: "",
-                discovered_constraints: lastEval.discovered_constraints ?? [],
-                objective_refinement: lastEval.objective_refinement ?? "",
-                emerged_subtasks: lastEval.emerged_subtasks ?? [],
-                execution_evidence: lastEval.execution_evidence,
-                retracted_constraints: lastEval.retracted_constraints ?? [],
-                revised_success_criteria: lastEval.revised_success_criteria ?? [],
-                wrong_assumptions: lastEval.wrong_assumptions ?? [],
-                worker_results: lastEval.worker_results ?? [],
-                compression_checkpoint: lastEval.compression_checkpoint ?? false,
-                checkpoint_label: lastEval.checkpoint_label ?? "",
-                next_action: lastEval.next_action,
-                completed_subtasks: lastEval.completed_subtasks ?? [],
-                blocked_subtasks: lastEval.blocked_subtasks ?? [],
-                canceled_subtasks: lastEval.canceled_subtasks ?? [],
-            };
-        }
-        let response;
-        try {
-            response = engine.invokeLoopCompile(request);
-        }
-        catch (exc) {
-            logEvent("compile_error", { loopId, round, error: String(exc) });
-            return makeRunResult(loopId, null, round - 1, successTrajectory, lastEval, artifacts);
-        }
-        if (response.status !== "ok" || !response.response?.prompt) {
-            return makeRunResult(loopId, response, round - 1, successTrajectory, lastEval, artifacts);
-        }
-        const artifact = response.response.prompt_artifact;
-        if (artifact)
-            artifacts.push(artifact);
-        // ── 2. Execute agent ───────────────────────────────────────────────
-        let output;
-        try {
-            output = await input.execute(response.response.prompt, {
-                signal: input.signal ?? new AbortController().signal,
-            });
-        }
-        catch (exc) {
-            logEvent("execute_error", { loopId, round, error: String(exc) });
-            // Persist terminal state before returning
-            if (lastEval)
-                engine.autoFeedback(lastEval, loopId, round - 1, task);
-            void notifyTerminal(input.terminalSinks, {
-                loopId, task, success: false, stopReason: "executor_failure",
-                roundsCompleted: round - 1, successTrajectory, lastEvaluation: lastEval,
-            });
-            return {
-                success: false,
-                stopReason: "executor_failure",
-                roundsCompleted: round - 1,
-                successTrajectory,
-                loopId,
-                lastEvaluation: lastEval,
-                artifacts,
-            };
-        }
-        // ── 3. Extract self-evaluation ─────────────────────────────────────
-        const selfEval = extractSelfEvaluation(output);
-        if (!selfEval) {
-            logEvent("extraction_failed", { loopId, round });
-            return makeRunResult(loopId, response, round - 1, successTrajectory, lastEval, artifacts);
-        }
-        lastEval = selfEval;
-        successTrajectory.push(selfEval.success);
-        // ── 4. Persist feedback ────────────────────────────────────────────
-        engine.autoFeedback(selfEval, loopId, round, task);
-        // ── 5. Check stop ──────────────────────────────────────────────────
-        if (!selfEval.should_continue) {
-            const result = makeRunResult(loopId, response, round, successTrajectory, lastEval, artifacts);
-            void notifyTerminal(input.terminalSinks, {
-                loopId, task,
-                success: result.success,
-                stopReason: result.stopReason,
-                roundsCompleted: round,
-                successTrajectory,
-                lastEvaluation: lastEval,
-            });
-            return result;
-        }
-        round++;
-    }
-    const result = makeRunResult(loopId, null, round - 1, successTrajectory, lastEval, artifacts);
-    result.stopReason = "max_rounds";
-    void notifyTerminal(input.terminalSinks, {
-        loopId, task,
-        success: false,
-        stopReason: "max_rounds",
-        roundsCompleted: round - 1,
-        successTrajectory,
-        lastEvaluation: lastEval,
-    });
-    return result;
-}
-function notifyTerminal(sinks, event) {
-    if (!sinks?.length)
-        return;
-    for (const sink of sinks) {
-        try {
-            const pending = sink(event);
-            if (pending && typeof pending.then === "function") {
-                void pending.catch(() => undefined);
-            }
-        }
-        catch {
-            // Sink failures must never fail the loop.
-        }
-    }
-}
-//# sourceMappingURL=run.js.map

package/dist/run.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"run.js","sourceRoot":"","sources":["../src/run.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,EAAE,eAAe,EAAE,qBAAqB,EAAE,MAAM,aAAa,CAAC;AACrE,OAAO,EAAE,IAAI,EAA6C,MAAM,eAAe,CAAC;AAChF,OAAO,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AAGhD,OAAO,EAAE,QAAQ,EAAE,MAAM,oBAAoB,CAAC;AAyC9C,SAAS,aAAa,CACpB,MAAc,EACd,YAAoC,EACpC,eAAuB,EACvB,iBAA4B,EAC5B,QAAyB,EACzB,YAA8B,EAAE;IAEhC,IAAI,CAAC,YAAY,EAAE,QAAQ,EAAE,CAAC;QAC5B,OAAO;YACL,OAAO,EAAE,KAAK;YACd,UAAU,EAAE,kBAAkB;YAC9B,eAAe;YACf,iBAAiB;YACjB,MAAM;YACN,cAAc,EAAE,QAAQ;YACxB,SAAS;SACV,CAAC;IACJ,CAAC;IACD,OAAO;QACL,OAAO,EAAE,QAAQ,EAAE,OAAO,KAAK,IAAI,IAAI,QAAQ,EAAE,eAAe,KAAK,KAAK;QAC1E,UAAU,EAAE,QAAQ,EAAE,eAAe,KAAK,KAAK;YAC7C,CAAC,CAAC,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,QAAQ,CAAC;YAC9C,CAAC,CAAC,YAAY;QAChB,eAAe;QACf,iBAAiB;QACjB,MAAM;QACN,cAAc,EAAE,QAAQ;QACxB,SAAS;KACV,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,GAAG,CAAC,KAAe;IACvC,MAAM,KAAK,GAAG,KAAK,CAAC,KAAK,IAAI,IAAI,aAAa,EAAE,CAAC;IACjD,MAAM,MAAM,GAAG,IAAI,eAAe,CAAC,KAAK,CAAC,CAAC;IAC1C,MAAM,MAAM,GAAG,OAAO,IAAI,CAAC,GAAG,EAAE,CAAC,QAAQ,CAAC,EAAE,CAAC,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC;IAC1F,MAAM,SAAS,GAAG,KAAK,CAAC,SAAS,IAAI,EAAE,CAAC;IACxC,MAAM,SAAS,GAAqB,EAAE,CAAC;IAEvC,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,IAAI,QAAoC,CAAC;IACzC,MAAM,iBAAiB,GAAc,EAAE,CAAC;IAExC,MAAM,IAAI,GAAG,KAAK,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC;IAC/B,IAAI,CAAC,IAAI;QAAE,MAAM,IAAI,KAAK,CAAC,6BAA6B,CAAC,CAAC;IAE1D,OAAO,KAAK,IAAI,SAAS,EAAE,CAAC;QAC1B,2BAA2B;QAC3B,IAAI,KAAK,CAAC,MAAM,EAAE,OAAO,EAAE,CAAC;YAC1B,OAAO,aAAa,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG,CAAC,EAAE,iBAAiB,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC;QACxF,CAAC;QAED,sEAAsE;QACtE,MAAM,OAAO,GAA4B;YACvC,IAAI;YACJ,IAAI,EAAE,IAAI,CAAC,YAAY;YACvB,QAAQ,EAAE,IAAI;YACd,UAAU,EAAE,IAAI;YAChB,OAAO,EAAE,IAAI;YACb,OAAO,EAAE,MAAM;YACf,KAAK;YACL,UAAU,EAAE,SAAS;YACrB,qBAAqB,EAAE,KAAK,CAAC,mBAAmB,IAAI,EAAE;YACtD,OAAO,EAAE,MAAM;YACf,MAAM,EAAE,EAAE;SACX,CAAC;QAEF,6BAA6B;QAC7B,IAAI,KAAK,CAAC,eAAe,EAAE,CAAC;YAC1B,IAAI,CAAC;gBACH,MAAM,GAAG,GAAG,MAAM,KAAK,CAAC,eAAe,CAAC;oBACtC,MAAM;oBACN,KAAK;oBACL,IAAI;oBACJ,MAAM,EAAE,EAAE;oBACV,cAAc,EAAE,QAAQ;iBACzB,CAAC,CAAC;gBACH,IAAI,GAAG,EAAE,IAAI,EAAE;oBAAE,OAAO,CAAC,gBAAgB,GAAG,GAAG,CAAC,IAAI,EAAE,CAAC;YACzD,CAAC;YAAC,MAAM,CAAC;gBACP,QAAQ,CAAC,wBAAwB,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC,CAAC;YACxD,CAAC;QACH,CAAC;QAED,4BAA4B;QAC5B,IAAI,QAAQ,EAAE,CAAC;YACb,OAAO,CAAC,iBAAiB,GAAG;gBAC1B,KAAK,EAAE,KAAK,GAAG,CAAC;gBAChB,OAAO,EAAE,QAAQ,CAAC,OAAO;gBACzB,cAAc,EAAE,QAAQ,CAAC,cAAc;gBACvC,qBAAqB,EAAE,QAAQ,CAAC,qBAAqB;gBACrD,mBAAmB,EAAE,EAAE;gBACvB,sBAAsB,EAAE,QAAQ,CAAC,sBAAsB,IAAI,EAAE;gBAC7D,oBAAoB,EAAE,QAAQ,CAAC,oBAAoB,IAAI,EAAE;gBACzD,gBAAgB,EAAE,QAAQ,CAAC,gBAAgB,IAAI,EAAE;gBACjD,kBAAkB,EAAE,QAAQ,CAAC,kBAAkB;gBAC/C,qBAAqB,EAAE,QAAQ,CAAC,qBAAqB,IAAI,EAAE;gBAC3D,wBAAwB,EAAE,QAAQ,CAAC,wBAAwB,IAAI,EAAE;gBACjE,iBAAiB,EAAE,QAAQ,CAAC,iBAAiB,IAAI,EAAE;gBACnD,cAAc,EAAE,QAAQ,CAAC,cAAc,IAAI,EAAE;gBAC7C,sBAAsB,EAAE,QAAQ,CAAC,sBAAsB,IAAI,KAAK;gBAChE,gBAAgB,EAAE,QAAQ,CAAC,gBAAgB,IAAI,EAAE;gBACjD,WAAW,EAAE,QAAQ,CAAC,WAAW;gBACjC,kBAAkB,EAAE,QAAQ,CAAC,kBAAkB,IAAI,EAAE;gBACrD,gBAAgB,EAAE,QAAQ,CAAC,gBAAgB,IAAI,EAAE;gBACjD,iBAAiB,EAAE,QAAQ,CAAC,iBAAiB,IAAI,EAAE;aACpD,CAAC;QACJ,CAAC;QAED,IAAI,QAAyB,CAAC;QAC9B,IAAI,CAAC;YACH,QAAQ,GAAG,MAAM,CAAC,iBAAiB,CAAC,OAAmD,CAAC,CAAC;QAC3F,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,QAAQ,CAAC,eAAe,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACjE,OAAO,aAAa,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG,CAAC,EAAE,iBAAiB,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC;QACxF,CAAC;QAED,IAAI,QAAQ,CAAC,MAAM,KAAK,IAAI,IAAI,CAAC,QAAQ,CAAC,QAAQ,EAAE,MAAM,EAAE,CAAC;YAC3D,OAAO,aAAa,CAAC,MAAM,EAAE,QAAQ,EAAE,KAAK,GAAG,CAAC,EAAE,iBAAiB,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC;QAC5F,CAAC;QAED,MAAM,QAAQ,GAAG,QAAQ,CAAC,QAAQ,CAAC,eAAe,CAAC;QACnD,IAAI,QAAQ;YAAE,SAAS,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QAEvC,sEAAsE;QACtE,IAAI,MAAc,CAAC;QACnB,IAAI,CAAC;YACH,MAAM,GAAG,MAAM,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAC,MAAM,EAAE;gBACrD,MAAM,EAAE,KAAK,CAAC,MAAM,IAAI,IAAI,eAAe,EAAE,CAAC,MAAM;aACrD,CAAC,CAAC;QACL,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,QAAQ,CAAC,eAAe,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACjE,0CAA0C;YAC1C,IAAI,QAAQ;gBAAE,MAAM,CAAC,YAAY,CAAC,QAAQ,EAAE,MAAM,EAAE,KAAK,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;YACrE,KAAK,cAAc,CAAC,KAAK,CAAC,aAAa,EAAE;gBACvC,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,kBAAkB;gBAC5D,eAAe,EAAE,KAAK,GAAG,CAAC,EAAE,iBAAiB,EAAE,cAAc,EAAE,QAAQ;aACxE,CAAC,CAAC;YACH,OAAO;gBACL,OAAO,EAAE,KAAK;gBACd,UAAU,EAAE,kBAAkB;gBAC9B,eAAe,EAAE,KAAK,GAAG,CAAC;gBAC1B,iBAAiB;gBACjB,MAAM;gBACN,cAAc,EAAE,QAAQ;gBACxB,SAAS;aACV,CAAC;QACJ,CAAC;QAED,sEAAsE;QACtE,MAAM,QAAQ,GAAG,qBAAqB,CAAC,MAAM,CAAC,CAAC;QAC/C,IAAI,CAAC,QAAQ,EAAE,CAAC;YACd,QAAQ,CAAC,mBAAmB,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC,CAAC;YACjD,OAAO,aAAa,CAAC,MAAM,EAAE,QAAQ,EAAE,KAAK,GAAG,CAAC,EAAE,iBAAiB,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC;QAC5F,CAAC;QAED,QAAQ,GAAG,QAAQ,CAAC;QACpB,iBAAiB,CAAC,IAAI,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;QAEzC,sEAAsE;QACtE,MAAM,CAAC,YAAY,CAAC,QAAQ,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,CAAC,CAAC;QAEnD,sEAAsE;QACtE,IAAI,CAAC,QAAQ,CAAC,eAAe,EAAE,CAAC;YAC9B,MAAM,MAAM,GAAG,aAAa,CAAC,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,iBAAiB,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC;YAC9F,KAAK,cAAc,CAAC,KAAK,CAAC,aAAa,EAAE;gBACvC,MAAM,EAAE,IAAI;gBACZ,OAAO,EAAE,MAAM,CAAC,OAAO;gBACvB,UAAU,EAAE,MAAM,CAAC,UAA6C;gBAChE,eAAe,EAAE,KAAK;gBACtB,iBAAiB;gBACjB,cAAc,EAAE,QAAQ;aACzB,CAAC,CAAC;YACH,OAAO,MAAM,CAAC;QAChB,CAAC;QAED,KAAK,EAAE,CAAC;IACV,CAAC;IAED,MAAM,MAAM,GAAG,aAAa,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG,CAAC,EAAE,iBAAiB,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC;IAC9F,MAAM,CAAC,UAAU,GAAG,YAAY,CAAC;IACjC,KAAK,cAAc,CAAC,KAAK,CAAC,aAAa,EAAE;QACvC,MAAM,EAAE,IAAI;QACZ,OAAO,EAAE,KAAK;QACd,UAAU,EAAE,YAAY;QACxB,eAAe,EAAE,KAAK,GAAG,CAAC;QAC1B,iBAAiB;QACjB,cAAc,EAAE,QAAQ;KACzB,CAAC,CAAC;IACH,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,SAAS,cAAc,CACrB,KAAqC,EACrC,KAAwB;IAExB,IAAI,CAAC,KAAK,EAAE,MAAM;QAAE,OAAO;IAC3B,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACzB,IAAI,CAAC;YACH,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC;YAC5B,IAAI,OAAO,IAAI,OAAO,OAAO,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;gBAClD,KAAK,OAAO,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;YACtC,CAAC;QACH,CAAC;QAAC,MAAM,CAAC;YACP,0CAA0C;QAC5C,CAAC;IACH,CAAC;AACH,CAAC"}