patchwork-os 1.2.0-beta.2.canary.837 → 1.2.0-beta.2.canary.839

package/dist/recipes/yamlRunner.js

@@ -67,7 +67,7 @@ import { resolveLocalEndpoint, resolveLocalModel } from "./localSettings.js";
 import { defaultDeprecationWarn, normalizeRecipeForRuntime, } from "./migrations/index.js";
 import { costRouter } from "./pricing/costRouter.js";
 import { loadPriceTable, costUsd as priceCostUsd, } from "./pricing/priceTable.js";
-import { ReplayIncompleteError } from "./replayBoundary.js";
+import { ReplayIncompleteError, TestModeRefusalError, } from "./replayBoundary.js";
 import { resolveRecipePath } from "./resolveRecipePath.js";
 import { RunBudget } from "./runBudget.js";
 import { recordAttemptRun } from "./runLedgers.js";
@@ -961,6 +961,9 @@ export async function runYamlRecipe(recipe, deps = {}, seedContext = {}) {
     // agent step would. It returns `{ value, ok }`; `ok: false` signals a
     // failed / silent-fail / empty agent response — the caller stops the loop
     // and treats it as exhausted (we don't re-judge a non-result).
+    // The agent step currently executing, for replay refusals raised from the
+    // refine-loop / nested-agent helpers below (they do not take a step id).
+    let currentAgentStepId;
     const runAgentText = async (prompt, driver, model, mcpAccess, downshift, providerOptions, 
     // P0-5: carry the reviewed/judge step's opt-in tool sandbox into refine-loop
     // re-runs so a sandboxed step STAYS sandboxed across revisions/re-judges.
@@ -995,7 +998,7 @@ export async function runYamlRecipe(recipe, deps = {}, seedContext = {}) {
             ...(deps.agentDisallowedTools?.length && { enforceSandbox: true }),
             ...(providerOptions && { providerOptions }),
             ...(dataPolicy !== undefined && { boundary: { dataPolicy } }),
-        }, buildAgentExecutorDeps(stepDeps, deps, undefined, runTaskId));
+        }, buildAgentExecutorDeps(stepDeps, deps, undefined, runTaskId, currentAgentStepId));
         runBudget.reconcile(
         // Prefer the driver executeAgent actually resolved+ran; fall back to
         // the routed value only when servedBy is absent (non-executeAgent
@@ -1076,7 +1079,7 @@ export async function runYamlRecipe(recipe, deps = {}, seedContext = {}) {
             ...(input.dataPolicy !== undefined && {
                 boundary: { dataPolicy: input.dataPolicy },
             }),
-        }, buildAgentExecutorDeps(stepDeps, deps, undefined, runTaskId));
+        }, buildAgentExecutorDeps(stepDeps, deps, undefined, runTaskId, currentAgentStepId));
         runBudget.reconcile(agentReturn.servedBy?.driver ?? input.driver ?? "auto", agentReturn.usage, agentReturn.servedBy?.model, {
             inputChars: input.prompt.length,
             outputChars: agentReturn.text.length,
@@ -1317,6 +1320,7 @@ export async function runYamlRecipe(recipe, deps = {}, seedContext = {}) {
             // cap is set; never disturbs injected (unit-test) price tables.
             runBudget.refreshPrices();
             const stepIdForEmit = step.into ?? step.agent?.into ?? `step_${stepsRun}`;
+            currentAgentStepId = stepIdForEmit;
             const stepTs = Date.now();
             stepStartTs.set(stepIdForEmit, stepTs);
             emit("recipe_step_start", {
@@ -1714,6 +1718,12 @@ export async function runYamlRecipe(recipe, deps = {}, seedContext = {}) {
                     // capture no `output`, so under replay there is never a mock for
                     // one — refuse here rather than let the catch below be the first
                     // to hear about it from the seam.
+                    // A recorded agent failure replays as that failure (thrown so the
+                    // catch below records it exactly as the original halt did).
+                    const replayedAgentFailure = deps.mockedFailures?.get(stepId);
+                    if (replayedAgentFailure !== undefined) {
+                        throw new Error(replayedAgentFailure.error);
+                    }
                     if (deps.replayOnly === true) {
                         throw new ReplayIncompleteError([stepId], "agent step");
                     }
@@ -1764,7 +1774,7 @@ export async function runYamlRecipe(recipe, deps = {}, seedContext = {}) {
                         ...(agentCfg.data_policy !== undefined && {
                             boundary: { dataPolicy: agentCfg.data_policy },
                         }),
-                    }, buildAgentExecutorDeps(stepDeps, deps, undefined, runTaskId));
+                    }, buildAgentExecutorDeps(stepDeps, deps, undefined, runTaskId, stepId));
                     agentResult = agentReturn.text;
                     runBudget.reconcile(
                     // Prefer the driver executeAgent actually resolved+ran; the routed
@@ -2007,7 +2017,16 @@ export async function runYamlRecipe(recipe, deps = {}, seedContext = {}) {
             // below (driven by `result`), so a replay shows how the recipe's
             // wiring behaves against captured evidence — only the tool call
             // itself is skipped. See RunnerDeps.mockedOutputs's doc comment.
-            if (deps.mockedOutputs?.has(stepId)) {
+            const replayedFailure = deps.mockedFailures?.get(stepId);
+            if (replayedFailure !== undefined) {
+                // The original run recorded this step as FAILED with no output. That
+                // record is the evidence: fail the same way, dispatch nothing, and do
+                // not retry — there is nothing to retry against.
+                thrownError = replayedFailure.error;
+                thrownErrorCode = replayedFailure.errorCode;
+                result = null;
+            }
+            else if (deps.mockedOutputs?.has(stepId)) {
                 result = deps.mockedOutputs.get(stepId) ?? null;
                 // A replayed http.post response is interpreted exactly as a live one,
                 // including a response captured before httpOutcome existed: it reads
@@ -2535,6 +2554,17 @@ export async function runYamlRecipe(recipe, deps = {}, seedContext = {}) {
         })(),
     };
 }
+/**
+ * Tools `recipe test` may run without a fixture: reads whose only effect goes
+ * through a dep `runTest` stubs (`readFile`, `gitLogSince`,
+ * `gitStaleBranches`, `getDiagnostics`). Everything else needs a mock.
+ */
+const TEST_STUBBED_TOOLS = new Set([
+    "file.read",
+    "git.log_since",
+    "git.stale_branches",
+    "diagnostics.get",
+]);
 export async function executeStep(step, ctx, deps, approvalEvidence, approvalDispatch, 
 /** Rollbackability the authority decision was made on (INV-1). */
 rollbackAtDecision) {
@@ -2551,6 +2581,14 @@ rollbackAtDecision) {
     if (deps.replayOnly === true) {
         throw new ReplayIncompleteError([step.into ?? toolId], `tool ${toolId}`);
     }
+    // `recipe test` boundary: only a mocked tool, or a read tool served wholly
+    // by injected stub deps, may run. Checked before the registry lookup for the
+    // same reason as replay: a missing fixture must never become a live call.
+    if (deps.testOnly === true &&
+        !deps.mockConnectors?.[toolId] &&
+        !TEST_STUBBED_TOOLS.has(toolId)) {
+        throw new TestModeRefusalError([step.into ?? toolId], `tool ${toolId}`);
+    }
     // Check if tool is registered in the new registry
     if (hasTool(toolId)) {
         const tool = getTool(toolId);
@@ -2947,6 +2985,12 @@ function refuseLiveModel(which) {
         "(claudeFn / localFn / claudeCodeFn) in the runner deps. " +
         "If the live call is the point, set PATCHWORK_TEST_ALLOW_LIVE=1 for that test.");
 }
+/** An agent driver fn that refuses: `recipe test` never reaches a live model. */
+function refuseInTest(which) {
+    return async () => {
+        throw new TestModeRefusalError(["agent"], `agent driver ${which}`);
+    };
+}
 function resolveStepDeps(deps, scope) {
     const workdir = deps.workdir ?? process.cwd();
     // Defense-in-depth: even if a file.* tool somehow forgets to call
@@ -2988,18 +3032,27 @@ function resolveStepDeps(deps, scope) {
             })),
         fetchFn: deps.fetchFn ?? globalThis.fetch,
         claudeFn: deps.claudeFn ??
-            (liveModelGuardActive() && process.env.ANTHROPIC_API_KEY
-                ? () => refuseLiveModel("defaultClaudeFn (Anthropic API)")
-                : defaultClaudeFn),
+            (deps.testOnly === true
+                ? refuseInTest("claude (Anthropic API)")
+                : liveModelGuardActive() && process.env.ANTHROPIC_API_KEY
+                    ? () => refuseLiveModel("defaultClaudeFn (Anthropic API)")
+                    : defaultClaudeFn),
         claudeCodeFn: deps.claudeCodeFn ??
-            (liveModelGuardActive()
-                ? () => refuseLiveModel("defaultClaudeCodeFn (claude CLI subprocess)")
-                : defaultClaudeCodeFn),
+            (deps.testOnly === true
+                ? refuseInTest("claude-code (CLI subprocess)")
+                : liveModelGuardActive()
+                    ? () => refuseLiveModel("defaultClaudeCodeFn (claude CLI subprocess)")
+                    : defaultClaudeCodeFn),
         localFn: deps.localFn ??
-            (liveModelGuardActive()
-                ? () => refuseLiveModel("defaultLocalFn (LOCAL_ENDPOINT)")
-                : defaultLocalFn),
-        providerDriverFn: deps.providerDriverFn ?? makeProviderDriverFn(),
+            (deps.testOnly === true
+                ? refuseInTest("local (LOCAL_ENDPOINT)")
+                : liveModelGuardActive()
+                    ? () => refuseLiveModel("defaultLocalFn (LOCAL_ENDPOINT)")
+                    : defaultLocalFn),
+        providerDriverFn: deps.providerDriverFn ??
+            (deps.testOnly === true
+                ? refuseInTest("provider driver")
+                : makeProviderDriverFn()),
         mockConnectors: deps.mockConnectors ?? {},
         recordFixturesDir: deps.recordFixturesDir,
         allowWrites: deps.allowWrites ?? [],
@@ -3035,6 +3088,7 @@ function resolveStepDeps(deps, scope) {
         // (`executeStep`, `buildAgentExecutorDeps`) can refuse without the run
         // loop's help. Only ever set by the replay entrypoints.
         ...(deps.replayOnly === true && { replayOnly: true }),
+        ...(deps.testOnly === true && { testOnly: true }),
         // Ephemeral rollback — same disk-availability gating as writeEffectLedger
         // above (deliberately: both share the operator's --ledger-dir/--attempt
         // inputs). No in-memory fallback: rollback only makes sense as a
@@ -3121,12 +3175,17 @@ function buildAgentExecutorDeps(stepDeps, runnerDeps, claudeCodeFnOverride,
  * the `runTaskIdRef` cell on its own deps at dispatch time, by which point
  * the run has started; the flat caller passes its local const directly.
  */
-runTaskId) {
+runTaskId, 
+/**
+ * The step this dispatch belongs to, so a replay refusal names it. Falls
+ * back to "agent" only when a caller genuinely does not know.
+ */
+stepId) {
     // Replay boundary (replayBoundary.ts, layer 3) — the single place agent
     // executor deps are built for every dispatch site in this runner, so a
     // replay's StepDeps can never hand a driver to an agent step.
     if (stepDeps.replayOnly === true) {
-        throw new ReplayIncompleteError(["agent"], "agent step");
+        throw new ReplayIncompleteError([stepId ?? "agent"], "agent step");
     }
     const claudeCliFn = claudeCodeFnOverride ?? stepDeps.claudeCodeFn;
     return {
@@ -3820,7 +3879,7 @@ recipeName) {
         }, buildAgentExecutorDeps(stepDeps, runnerDeps, claudeCodeFnOverride, 
         // Read at DISPATCH time, not build time — by now `runChainedRecipe` has
         // computed and published the run's `taskId`.
-        runTaskIdRef.current));
+        runTaskIdRef.current, opts?.stepId));
     };
     // ---------------------------------------------------------------------
     // BEGIN A-PR2 EDIT BLOCK — `loadNestedRecipe` jail (dogfood F-04).
@@ -4004,6 +4063,7 @@ export async function dispatchRecipe(recipe, deps, seedContext = {}) {
             runLog: deps.chainedOptions?.runLog,
             activityLog: deps.chainedOptions?.activityLog,
             mockedOutputs: deps.chainedOptions?.mockedOutputs,
+            mockedFailures: deps.chainedOptions?.mockedFailures,
             ...((deps.replayOnly === true ||
                 deps.chainedOptions?.replayOnly === true) && { replayOnly: true }),
             taskIdPrefix: deps.chainedOptions?.taskIdPrefix,