micro-models-agent 0.41.2 → 0.43.0

package/dist/modules/execution/execution-plugin.js

@@ -1,272 +0,0 @@
-import { t } from "../../i18n/index";
-import { detectTestResults } from "../../tools/bash";
-import { forbiddenWindowsCommand } from "./windows-commands";
-import { platform } from "os";
-/**
- * Cooldown (in iterations) between stuck-recovery injections. Exported so the
- * ExecutionModule can seed its state's lastRecoveryIteration with the same
- * offset the plugin compares against.
- */
-export const STUCK_RECOVERY_COOLDOWN = 5;
-const MAX_PLAN_WARNINGS_BEFORE_BLOCK = 3;
-const FORCE_SKIP_THRESHOLD = 10;
-/**
- * The execution plugin hooks, extracted verbatim from ExecutionModule.getPlugin().
- * All shared mutable state (pendingMessages, forbiddenBashFailures, state,
- * tracker) is owned by the module and mutated by reference through `deps`.
- */
-export function createExecutionPlugin(deps) {
-    return {
-        name: "execution",
-        onBeforeThink: (ctx) => {
-            // Flush deferred <system-summary> messages first so they are visible
-            // to the next model call but never interleave tool messages.
-            if (ctx.contextManager && deps.pendingMessages.length > 0) {
-                for (const m of deps.pendingMessages.splice(0)) {
-                    ctx.contextManager.addMessage(m);
-                }
-            }
-            if (deps.trackerRef.current?.isComplete()) {
-                // All steps done/skipped — stop counting "no progress" against
-                // the last step and silence step-stuck warnings for a finished
-                // plan, BUT keep the tool-call history so a read-without-write
-                // loop is still caught (observed: plan complete + false-positive
-                // audit → the model re-read files for 50+ iterations with zero
-                // writes and no recovery hint).
-                deps.stuckDetector.resetStepProgress();
-                deps.state.consecutivePlanWarnings = 0;
-                deps.state.lastStepId = -1;
-                deps.state.stuckNotified = false;
-            }
-            else {
-                const step = deps.trackerRef.current?.getCurrentStep();
-                if (deps.trackerRef.current && step) {
-                    if (step.id !== deps.state.lastStepId) {
-                        deps.state.consecutivePlanWarnings = 0;
-                        deps.state.lastStepId = step.id;
-                        deps.state.stuckNotified = false;
-                    }
-                    deps.stuckDetector.setCurrentStep(step.id, step.description);
-                    deps.stuckDetector.recordIteration(step.id);
-                }
-                else {
-                    deps.stuckDetector.reset();
-                    deps.state.consecutivePlanWarnings = 0;
-                    deps.state.lastStepId = -1;
-                    deps.state.stuckNotified = false;
-                    const iter = typeof ctx.iteration === "number" ? ctx.iteration : 0;
-                    if (iter === 3 && !deps.trackerRef.current && ctx.contextManager) {
-                        ctx.contextManager.addMessage({
-                            role: "user",
-                            content: `<system-summary>You have made 3 tool calls without creating a plan. For any task that involves creating files, installing packages, or multiple steps — you MUST use plan create BEFORE continuing. Use the plan tool now with concrete steps (exact filenames, commands, deliverables). Do NOT make any more write/edit/bash calls until you have a plan.</system-summary>`,
-                        });
-                    }
-                    if (iter >= 6 && !deps.trackerRef.current && ctx.contextManager) {
-                        ctx.contextManager.addMessage({
-                            role: "user",
-                            content: `<system-summary>STOP. ${iter} iterations without a plan. You MUST call plan create RIGHT NOW. No more tool calls until you create a plan.</system-summary>`,
-                        });
-                    }
-                }
-            }
-            const stuckReason = deps.stuckDetector.getStuckReason();
-            if (stuckReason) {
-                // Log once per stuck episode instead of spamming every iteration.
-                const logIt = deps.stuckDetector.isStuck() ? !deps.state.stuckNotified : true;
-                if (logIt) {
-                    ctx.logger?.warn(stuckReason);
-                    ctx.sessionLog?.plan("stuck-warning", stuckReason, typeof ctx.iteration === "number" ? ctx.iteration : undefined);
-                    if (deps.stuckDetector.isStuck())
-                        deps.state.stuckNotified = true;
-                }
-            }
-            if (deps.stuckDetector.isStuck() ||
-                deps.stuckDetector.hasRepetitiveToolCalls() ||
-                deps.stuckDetector.hasReadOnlyLoop()) {
-                const currentIter = typeof ctx.iteration === "number" ? ctx.iteration : 0;
-                if (currentIter - deps.state.lastRecoveryIteration >= STUCK_RECOVERY_COOLDOWN) {
-                    const recovery = deps.stuckDetector.getRecoveryMessage();
-                    if (recovery && ctx.contextManager) {
-                        const lastError = deps.stuckDetector.getLastErrorOutput();
-                        const skillHint = lastError
-                            ? "\nIf you have relevant skills available, consider loading one with load_skill for expert guidance."
-                            : "";
-                        // Actionable hints based on actual error output
-                        const actionableHints = deps.stuckDetector.getActionableHints();
-                        const actionableHintStr = actionableHints.length > 0
-                            ? `\n${t("exec.hints", { hints: actionableHints.map((h) => `- ${h}`).join("\n") })}`
-                            : "";
-                        // Tool alternative suggestion
-                        const alternative = deps.stuckDetector.getToolAlternative();
-                        const altHint = alternative
-                            ? `\nTool "${deps.stuckDetector.getLastFailedTool()}" is failing. Try "${alternative}" instead.`
-                            : "";
-                        ctx.contextManager.addMessage({
-                            role: "user",
-                            content: `<system-summary>${recovery}${skillHint}${actionableHintStr}${altHint}</system-summary>`,
-                        });
-                    }
-                    const hints = deps.stuckDetector.getHints();
-                    if (hints.length > 0 && ctx.contextManager) {
-                        const hintMsg = t("exec.hints", {
-                            hints: hints.map((h) => `- ${h}`).join("\n"),
-                        });
-                        ctx.contextManager.addMessage({
-                            role: "user",
-                            content: `<system-summary>${hintMsg}</system-summary>`,
-                        });
-                    }
-                    deps.stuckDetector.recordEscalation();
-                    deps.state.lastRecoveryIteration = currentIter;
-                    if (deps.stuckDetector.shouldEscalate() && ctx.onMeta) {
-                        const escalation = t("exec.escalation", {
-                            stepId: String(deps.trackerRef.current?.getCurrentStep()?.id ?? "?"),
-                            description: deps.trackerRef.current?.getCurrentStep()?.description ?? "",
-                        });
-                        ctx.onMeta(escalation);
-                    }
-                    // Force skip after too many iterations on the same step
-                    if (deps.stuckDetector.getIterationsOnCurrentStep() >= FORCE_SKIP_THRESHOLD &&
-                        ctx.contextManager) {
-                        const step = deps.trackerRef.current?.getCurrentStep();
-                        ctx.contextManager.addMessage({
-                            role: "user",
-                            content: `<system-summary>STOP. Step ${step?.id ?? "?"} ("${step?.description ?? ""}") took ${deps.stuckDetector.getIterationsOnCurrentStep()} iterations with no progress. DO NOT continue this step. Immediately call: plan update step=${step?.id ?? "?"} status=done (if code works despite warnings) OR plan update step=${step?.id ?? "?"} status=skipped note="reason". Do NOT make any other tool calls before updating the plan.</system-summary>`,
-                        });
-                    }
-                }
-            }
-        },
-        onBeforeTool: (_ctx, call) => {
-            const warning = deps.checkPlanAlignment(call);
-            if (warning) {
-                deps.pendingMessages.push({
-                    role: "user",
-                    content: `<system-summary>${warning}</system-summary>`,
-                });
-                deps.state.consecutivePlanWarnings++;
-                if (deps.state.consecutivePlanWarnings >= MAX_PLAN_WARNINGS_BEFORE_BLOCK) {
-                    deps.pendingMessages.push({
-                        role: "user",
-                        content: `<system-summary>${t("exec.plan_blocked", { step: String(deps.trackerRef.current?.getCurrentStep()?.id ?? "?"), max: MAX_PLAN_WARNINGS_BEFORE_BLOCK })}</system-summary>`,
-                    });
-                    // Block with an explicit reason so the model sees WHY the call
-                    // was blocked (executor shows it via tool.blocked_reason).
-                    return t("exec.plan_blocked", {
-                        step: String(deps.trackerRef.current?.getCurrentStep()?.id ?? "?"),
-                        max: MAX_PLAN_WARNINGS_BEFORE_BLOCK,
-                    });
-                }
-            }
-            else {
-                deps.state.consecutivePlanWarnings = 0;
-            }
-            return true;
-        },
-        onToolCall: (ctx) => {
-            const toolName = ctx?.toolName;
-            const args = ctx?.args;
-            if (toolName && args) {
-                deps.stuckDetector.recordToolCall(toolName, args);
-            }
-        },
-        onAfterTool: (ctx, call, result) => {
-            // Remember every bash invocation (successful or not) so stuck
-            // detection can reason about e.g. repeated runs with empty output.
-            if (call.name === "bash") {
-                deps.stuckDetector.recordBashOutput(String(call.arguments?.command ?? ""), String(result.output ?? ""));
-            }
-            // A write/edit/bash that reports success but still carries a
-            // type/syntax error in its output is NOT a success — the model must
-            // fix the actual error, not keep rewriting the file blindly. Feed it
-            // to the stuck detector so the rewrite-loop hint fires.
-            const toolText = String(result.output ?? "");
-            if (/error TS\d+|\[Project typecheck failed\]|\[Syntax check failed\]/.test(toolText)) {
-                deps.stuckDetector.recordToolError(call.name, toolText.slice(0, 300));
-            }
-            if (!result.success) {
-                // Hard stop for forbidden Windows commands: after the second
-                // failure of the same forbidden command (grep/sed/ls/find/…),
-                // inject a STOP message so the session stops burning bash calls
-                // on commands cmd.exe does not have. Mirrors the bash tool's own
-                // win32-gated hard-stop (src/tools/bash.ts) — on macOS/Linux
-                // these commands exist and fail for legit reasons.
-                if (call.name === "bash" && platform() === "win32") {
-                    const cmd = String(call.arguments?.command ?? "");
-                    const forbidden = forbiddenWindowsCommand(cmd);
-                    if (forbidden) {
-                        const n = (deps.forbiddenBashFailures.get(forbidden) || 0) + 1;
-                        deps.forbiddenBashFailures.set(forbidden, n);
-                        if (n === 2) {
-                            deps.pendingMessages.push({
-                                role: "user",
-                                content: `<system-summary>${t("exec.forbidden_cmd", { cmd: forbidden })}</system-summary>`,
-                            });
-                        }
-                    }
-                }
-                deps.stuckDetector.recordToolError(call.name, result.output);
-                // Immediately queue an actionable hint on failure (don't wait for
-                // the stuck threshold). Flushed in onBeforeThink so it does not
-                // land between tool messages.
-                const actionableHints = deps.stuckDetector.getActionableHints();
-                const alternative = deps.stuckDetector.getToolAlternative();
-                if (actionableHints.length > 0 || alternative) {
-                    const parts = [...actionableHints];
-                    if (alternative) {
-                        parts.push(`Tool "${call.name}" crashed. Try "${alternative}" instead.`);
-                    }
-                    deps.pendingMessages.push({
-                        role: "user",
-                        content: `<system-summary>${t("exec.hints", { hints: parts.map((h) => `- ${h}`).join("\n") })}</system-summary>`,
-                    });
-                }
-            }
-            else {
-                deps.stuckDetector.recordToolSuccess();
-                // When a bash command runs code successfully, suggest marking the
-                // step done — unless the output shows failing tests, which must
-                // never be reported as a clean success.
-                if (call.name === "bash" && result.success) {
-                    const cmd = String(call.arguments?.command ?? "");
-                    const testRun = detectTestResults(String(result.output ?? ""));
-                    if (testRun && testRun.failed > 0) {
-                        deps.pendingMessages.push({
-                            role: "user",
-                            content: `<system-summary>${testRun.framework} reported ${testRun.failed} FAILING test(s) (${testRun.passed} passing). Do NOT mark the current step as done — fix the failing tests (read the failure output, correct the code) and re-run them until all pass.</system-summary>`,
-                        });
-                    }
-                    else if (testRun && testRun.failed === 0 && testRun.passed > 0) {
-                        deps.pendingMessages.push({
-                            role: "user",
-                            content: `<system-summary>${testRun.framework}: all ${testRun.passed} test(s) passed for "${cmd}". You may mark the current step as done via plan update step=N status=done.</system-summary>`,
-                        });
-                    }
-                    else if (/node|tsx|ts-node|python|npm\s+(start|test|run)/.test(cmd)) {
-                        deps.pendingMessages.push({
-                            role: "user",
-                            content: `<system-summary>The command "${cmd}" completed successfully. If this was testing your code, mark the current step as done via plan update step=N status=done.</system-summary>`,
-                        });
-                    }
-                }
-            }
-            if (result.success && (call.name === "write_file" || call.name === "edit_file")) {
-                const filePath = call.arguments?.path;
-                if (filePath) {
-                    deps.stuckDetector.recordFileRewrite(filePath);
-                    if (deps.stuckDetector.hasExcessiveRewrites()) {
-                        const file = deps.stuckDetector.getExcessiveRewriteFile();
-                        const count = deps.stuckDetector.getFileRewriteCount(file);
-                        if (ctx.onMeta) {
-                            ctx.onMeta(t("exec.file_rewrite_warning", {
-                                file: file,
-                                count: String(count),
-                            }));
-                        }
-                    }
-                }
-                deps.advancePlanIfStepComplete(ctx.contextManager, ctx.sessionLog);
-            }
-        },
-    };
-}