micro-models-agent 0.41.2 → 0.43.0

package/dist/modules/execution/plan-tool.js

@@ -1,508 +0,0 @@
-import { t } from "../../i18n/index";
-import { PlanCreator } from "./planner";
-import { checkPlanCoverage } from "./plan-coverage";
-/**
- * The plan / todo / verify tool definitions for the execution module. The
- * handlers were extracted verbatim from ExecutionModule.getToolDefinitions() —
- * all state access goes through the `deps` closures (the module owns the
- * tracker and the plan store).
- */
-export function createPlanToolDefinitions(deps) {
-    return [
-        {
-            name: "plan",
-            alwaysOn: true,
-            description: 'Create, update, show, abort, list, switch, or re-plan multi-step plans.\n\nActions:\n- create: Start a new plan. Previous active plan is auto-preserved: incomplete → draft, complete → archive.\n- update: Mark step status (done/failed/skipped), or rebuild plan with new steps.\n- show: Print current plan checklist.\n- abort: Archive current plan and clear active slot.\n- list: Show all plans (active, drafts, archived) with progress.\n- switch: Make a different plan active (by plan id).\n- re-plan: Iterative replanning: keep completed steps, replace remaining with new steps.\n\nWrite CONCRETE steps with exact file paths and commands:\n- Specify WHICH files to create with exact paths (e.g. "create src/components/Header.tsx with navigation and logo")\n- Specify WHICH packages to install (e.g. "run npm install react react-dom")\n- Specify WHICH CLI commands to run with exact arguments\n- Each step should include at least one file extension (.ts, .tsx, .json, etc.) or a command verb (install, create, build, run, add, init)\n- Good: "Создать src/components/Header.tsx с навигацией и логотипом"\n- Bad: "Настройка проекта" (too vague — what exactly needs to be configured?)\n- For steps that REMOVE files, pass kinds: ["delete"] aligned with steps (the done-gate verifies those files are gone instead of demanding they exist).',
-            parameters: {
-                type: "object",
-                properties: {
-                    action: {
-                        type: "string",
-                        enum: ["create", "update", "show", "abort", "list", "switch", "re-plan"],
-                    },
-                    title: { type: "string" },
-                    steps: { type: "array", items: { type: "string" } },
-                    kinds: {
-                        type: "array",
-                        items: { type: "string", enum: ["create", "delete"] },
-                        description: 'Optional per-step intent, aligned 1:1 with steps. Pass "delete" for steps whose purpose is REMOVING files — the done-gate then requires those files to be GONE (not present). Defaults to "create".',
-                    },
-                    step: { type: "number" },
-                    status: { type: "string", enum: ["done", "failed", "skipped"] },
-                    note: { type: "string" },
-                    id: { type: "string", description: "Plan id (for switch action)" },
-                },
-                required: ["action"],
-            },
-            handler: async (_ctx, args) => {
-                const action = String(args.action);
-                if (action === "list") {
-                    const metas = deps.store.listAll();
-                    if (metas.length === 0) {
-                        return { success: true, output: t("plan.list_empty") };
-                    }
-                    const lines = metas.map((m) => {
-                        const icon = m.status === "active" ? "[*]" : m.status === "draft" ? "[ ]" : "[-]";
-                        const namePart = m.name ? ` (${m.name})` : "";
-                        return `${icon} ${m.id}${namePart} — ${m.title}  ${m.doneCount}/${m.stepCount}`;
-                    });
-                    return {
-                        success: true,
-                        output: `${t("plan.list_header")}\n${lines.join("\n")}`,
-                    };
-                }
-                if (action === "switch") {
-                    const planId = String(args.id || "");
-                    if (!planId) {
-                        return { success: false, output: t("plan.switch_no_id") };
-                    }
-                    const found = deps.store.find(planId);
-                    if (!found) {
-                        return {
-                            success: false,
-                            output: t("plan.not_found", { id: planId }),
-                        };
-                    }
-                    deps.preserveActive();
-                    if (found.status === "draft") {
-                        deps.store.removeDraft(found.plan.id);
-                    }
-                    else if (found.status === "archived") {
-                        deps.store.removeArchived(found.plan.id);
-                    }
-                    deps.setPlan(found.plan);
-                    const display = PlanCreator.toPromptBlock(found.plan, 0);
-                    return {
-                        success: true,
-                        output: t("plan.switched", {
-                            id: found.plan.id,
-                            title: found.plan.title,
-                        }),
-                        display,
-                    };
-                }
-                if (action === "re-plan") {
-                    if (!deps.trackerRef.current) {
-                        return { success: false, output: t("plan.no_active") };
-                    }
-                    const newSteps = Array.isArray(args.steps) ? args.steps.map(String) : [];
-                    if (newSteps.length === 0) {
-                        return { success: false, output: t("plan.replan_no_steps") };
-                    }
-                    const tracker = deps.trackerRef.current;
-                    const oldPlan = tracker.getPlan();
-                    const replanned = PlanCreator.replan(oldPlan, newSteps, args.title ? String(args.title) : undefined);
-                    const keptCount = replanned.steps.length - newSteps.length;
-                    deps.setPlan(replanned);
-                    const display = PlanCreator.toPromptBlock(replanned, keptCount);
-                    return {
-                        success: true,
-                        output: t("plan.replanned", {
-                            kept: String(keptCount),
-                            steps: String(newSteps.length),
-                        }),
-                        display,
-                    };
-                }
-                if (action === "create") {
-                    const title = String(args.title || "Task Plan");
-                    const steps = Array.isArray(args.steps) ? args.steps.map(String) : [];
-                    if (steps.length === 0) {
-                        return { success: false, output: t("plan.no_steps") };
-                    }
-                    const parsed = deps.parseKinds(args, steps);
-                    if (parsed.error) {
-                        return { success: false, output: parsed.error };
-                    }
-                    // Never silently discard an in-progress plan (observed:
-                    // ses_msvuao0h — right after compaction the model called
-                    // `plan create` again, dumping the 1/6 plan into drafts and
-                    // starting from scratch). An active plan WITH progress blocks
-                    // creation until the model consciously aborts it. A fresh plan
-                    // (all steps pending) is safe to replace — the model may have
-                    // just changed its mind about the approach.
-                    const existing = deps.trackerRef.current;
-                    if (existing) {
-                        const activePlan = existing.getPlan();
-                        const progressed = activePlan.steps.some((s) => s.status === "done" || s.status === "failed" || s.status === "skipped");
-                        if (progressed) {
-                            const done = activePlan.steps.filter((s) => s.status === "done").length;
-                            return {
-                                success: false,
-                                output: t("plan.active_in_progress", {
-                                    id: activePlan.id,
-                                    done: String(done),
-                                    total: String(activePlan.steps.length),
-                                    current: String(existing.getCurrentStepIndex() + 1),
-                                }),
-                            };
-                        }
-                    }
-                    deps.preserveActive();
-                    const plan = PlanCreator.createPlan(title, steps, deps.baseDir, parsed.kinds ?? []);
-                    deps.setPlan(plan);
-                    const display = PlanCreator.toPromptBlock(plan, 0);
-                    let output = t("plan.created", {
-                        title,
-                        steps: String(steps.length),
-                    });
-                    // The model sometimes passes an id expecting the plan to be
-                    // created under it (observed: id "plan_error_check_e07pb2" was
-                    // ignored, a new id appeared). Be explicit so it doesn't look
-                    // for a plan by the wrong id afterwards.
-                    if (args.id) {
-                        output += `\n${t("plan.id_ignored")}`;
-                    }
-                    if (existing) {
-                        output += `\n${t("plan.existing_fresh")}`;
-                    }
-                    let displayOut = display;
-                    const taskText = deps.getTaskText(_ctx);
-                    if (taskText) {
-                        const { missing } = checkPlanCoverage(taskText, steps);
-                        if (missing.length > 0) {
-                            const warn = t("plan.coverage_warning", {
-                                missing: missing.join(", "),
-                            });
-                            output += `\n⚠️ ${warn}`;
-                            displayOut = `${display}\n⚠️ ${warn}`;
-                        }
-                    }
-                    return {
-                        success: true,
-                        output,
-                        display: displayOut,
-                    };
-                }
-                if (action === "show") {
-                    const planId = args.id ? String(args.id) : "";
-                    if (planId && planId !== deps.trackerRef.current?.getPlan()?.id) {
-                        // Show a specific plan (active, draft, or archived) instead
-                        // of ignoring the id and always printing the active one
-                        // (observed: the model asked for an archived plan three
-                        // times and got the active plan back every time).
-                        const found = deps.store.find(planId);
-                        if (!found) {
-                            return {
-                                success: false,
-                                output: t("plan.not_found", { id: planId }),
-                            };
-                        }
-                        const display = PlanCreator.toPromptBlock(found.plan, 0);
-                        return {
-                            success: true,
-                            output: `${t("plan.show_header")}\n${display}`,
-                            display,
-                        };
-                    }
-                    if (!deps.trackerRef.current) {
-                        return { success: false, output: t("plan.no_active") };
-                    }
-                    const display = deps.trackerRef.current.toPromptBlock();
-                    return {
-                        success: true,
-                        output: `${t("plan.show_header")}\n${display}`,
-                        display,
-                    };
-                }
-                if (action === "update" && args.step && deps.trackerRef.current) {
-                    const tracker = deps.trackerRef.current;
-                    const stepId = Number(args.step);
-                    const target = tracker.getStep(stepId);
-                    if (!target) {
-                        return { success: false, output: t("plan.step_not_found") };
-                    }
-                    const status = String(args.status || "done");
-                    // Completed plan: re-marking a done step of a finished plan is the
-                    // loop anchor (observed: 3× "plan update step=4 done" on a 4/4 plan
-                    // while a NEW task waited). Archive + clear so the model can create a
-                    // fresh plan for the new task instead of spinning.
-                    if (tracker.isComplete()) {
-                        const plan = tracker.getPlan();
-                        deps.store.archivePlan(plan);
-                        deps.trackerRef.current = null;
-                        const done = plan.steps.filter((s) => s.status === "done").length;
-                        return {
-                            success: true,
-                            output: t("plan.completed_archived", {
-                                id: plan.id,
-                                done: String(done),
-                                total: String(plan.steps.length),
-                            }),
-                        };
-                    }
-                    // Idempotency: re-marking a done step as done usually means the
-                    // plan was already auto-advanced (all step files exist) and the
-                    // model is confirming its own work. Return SUCCESS so the loop
-                    // isn't confused — a small model misreads success:false as a
-                    // failed action and retries/spins. Genuine spinning (repeated
-                    // identical calls without new work) is still caught by the
-                    // stuck-detector's consecutive-identical-call check.
-                    if (status === "done" && target.status === "done") {
-                        const progress = tracker.getProgressString();
-                        const display = tracker.toPromptBlock();
-                        return {
-                            success: true,
-                            output: `${t("plan.step_already_done", {
-                                step: String(stepId),
-                            })}\n${progress}`,
-                            display,
-                        };
-                    }
-                    // Order enforcement: only the CURRENT step may be marked done.
-                    // Prevents jumping ahead (step 5 marked done before steps 1-4,
-                    // observed in the post-mortem session) and the follow-on audit
-                    // failures ("5/5 steps done, 5 files missing").
-                    if (status === "done") {
-                        const blockers = tracker
-                            .getPlan()
-                            .steps.filter((s) => s.id < stepId && s.status !== "done" && s.status !== "skipped");
-                        if (blockers.length > 0) {
-                            return {
-                                success: false,
-                                output: t("plan.order_blocked", {
-                                    step: String(stepId),
-                                    first: String(blockers[0].id),
-                                    desc: blockers[0].description,
-                                }),
-                            };
-                        }
-                    }
-                    // Deliverable gate: marking a step done must match its declared
-                    // intent. "create" (default) → named files must exist; "delete"
-                    // → named files must be GONE. Intent is model-declared (kind),
-                    // never keyword-inferred (rule #10).
-                    if (status === "done") {
-                        if (target.kind === "delete") {
-                            const leftovers = deps.stillExistingDeliverables(target);
-                            if (leftovers.length > 0) {
-                                return {
-                                    success: false,
-                                    output: t("plan.deliverables_remain", {
-                                        step: String(stepId),
-                                        files: leftovers.join(", "),
-                                    }),
-                                };
-                            }
-                        }
-                        else {
-                            const missing = deps.missingStepDeliverables(target);
-                            if (missing.length > 0) {
-                                return {
-                                    success: false,
-                                    output: t("plan.deliverables_missing", {
-                                        step: String(stepId),
-                                        files: missing.join(", "),
-                                    }),
-                                };
-                            }
-                        }
-                    }
-                    tracker.updateStepStatus(stepId, status);
-                    if (args.note)
-                        tracker.addNote(Number(args.step), String(args.note));
-                    tracker.syncCurrentStep();
-                    deps.store.saveActive(tracker.getPlan());
-                    // If this update completed the last step → auto-archive. This breaks
-                    // the loop where a finished plan stays active and anchors the model
-                    // to "continue step N, do NOT create a new plan".
-                    if (tracker.isComplete()) {
-                        const plan = tracker.getPlan();
-                        deps.store.archivePlan(plan);
-                        deps.trackerRef.current = null;
-                        const done = plan.steps.filter((s) => s.status === "done").length;
-                        return {
-                            success: true,
-                            output: `${t("plan.step_status", { step: String(args.step), status: String(args.status || "done") })}\n${t("plan.completed_archived", {
-                                id: plan.id,
-                                done: String(done),
-                                total: String(plan.steps.length),
-                            })}`,
-                        };
-                    }
-                    const progress = tracker.getProgressString();
-                    const display = tracker.toPromptBlock();
-                    _ctx.sessionLog?.plan("step-update", `${t("plan.step_status", { step: String(args.step), status: String(args.status || "done") })} | ${progress} | current: step ${tracker.getCurrentStepIndex() + 1}`);
-                    return {
-                        success: true,
-                        output: `${t("plan.step_status", { step: String(args.step), status: String(args.status || "done") })}\n${progress}`,
-                        display,
-                    };
-                }
-                if (action === "update" &&
-                    Array.isArray(args.steps) &&
-                    args.steps.length > 0 &&
-                    deps.trackerRef.current) {
-                    const tracker = deps.trackerRef.current;
-                    const title = String(args.title || tracker.getPlan().title);
-                    const steps = args.steps.map(String);
-                    const parsed = deps.parseKinds(args, steps);
-                    if (parsed.error) {
-                        return { success: false, output: parsed.error };
-                    }
-                    const plan = PlanCreator.createPlan(title, steps, deps.baseDir, parsed.kinds ?? []);
-                    deps.setPlan(plan);
-                    const display = PlanCreator.toPromptBlock(plan, 0);
-                    const output = t("plan.updated", {
-                        title,
-                        steps: String(steps.length),
-                    });
-                    return {
-                        success: true,
-                        output,
-                        display,
-                    };
-                }
-                if (action === "abort") {
-                    const tracker = deps.trackerRef.current;
-                    if (tracker) {
-                        deps.store.archivePlan(tracker.getPlan());
-                    }
-                    deps.trackerRef.current = null;
-                    deps.store.clearActive();
-                    return { success: true, output: t("plan.aborted") };
-                }
-                if (!deps.trackerRef.current) {
-                    return { success: false, output: t("plan.no_active") };
-                }
-                return {
-                    success: false,
-                    output: t("plan.unknown_action", { action }),
-                };
-            },
-        },
-        {
-            name: "todo",
-            alwaysOn: true,
-            description: "Manage sub-tasks within current plan step. Use to break down complex steps into smaller tasks.",
-            parameters: {
-                type: "object",
-                properties: {
-                    action: { type: "string", enum: ["add", "done", "list"] },
-                    items: { type: "array", items: { type: "string" } },
-                },
-                required: ["action"],
-            },
-            handler: async (_ctx, args) => {
-                if (!deps.trackerRef.current) {
-                    return { success: false, output: t("plan.no_active") };
-                }
-                const tracker = deps.trackerRef.current;
-                const action = String(args.action);
-                const currentStep = tracker.getCurrentStep();
-                if (!currentStep) {
-                    return { success: false, output: t("plan.step_not_found") };
-                }
-                if (action === "add" && Array.isArray(args.items)) {
-                    const items = args.items.map(String);
-                    const existing = currentStep.subtasks ?? [];
-                    let nextId = existing.reduce((m, s) => Math.max(m, s.id), 0) + 1;
-                    for (const item of items) {
-                        existing.push({ id: nextId++, text: item, done: false });
-                    }
-                    currentStep.subtasks = existing;
-                    deps.store.saveActive(tracker.getPlan());
-                    const display = tracker.toPromptBlock();
-                    return {
-                        success: true,
-                        output: t("todo.added", {
-                            count: String(items.length),
-                            items: items.join(", "),
-                        }),
-                        display,
-                    };
-                }
-                if (action === "done") {
-                    const items = Array.isArray(args.items) ? args.items.map(String) : [];
-                    if (items.length === 0) {
-                        return {
-                            success: false,
-                            output: t("todo.no_items"),
-                        };
-                    }
-                    const subs = currentStep.subtasks ?? [];
-                    let marked = 0;
-                    for (const item of items) {
-                        const sub = subs.find((s) => !s.done && s.text.toLowerCase() === item.toLowerCase());
-                        if (sub) {
-                            sub.done = true;
-                            marked++;
-                        }
-                    }
-                    if (marked === 0) {
-                        return {
-                            success: false,
-                            output: t("todo.subtask_not_found", {
-                                items: items.join(", "),
-                            }),
-                        };
-                    }
-                    currentStep.subtasks = subs;
-                    deps.store.saveActive(tracker.getPlan());
-                    const doneCount = subs.filter((s) => s.done).length;
-                    const display = tracker.toPromptBlock();
-                    // Marking a sub-task done must NOT auto-complete the plan
-                    // step — the step is only done when the model confirms the
-                    // whole step via plan update (observed: 3 × "todo done"
-                    // returned "Step 3: done" and marked the step done early).
-                    let output = t("todo.marked_done", {
-                        count: String(marked),
-                    });
-                    output += ` (${doneCount}/${subs.length})`;
-                    if (doneCount === subs.length) {
-                        output += `\nAll sub-tasks done — call plan update step=${currentStep.id} status=done to complete this step.`;
-                    }
-                    return { success: true, output, display };
-                }
-                if (action === "list") {
-                    const subs = currentStep.subtasks ?? [];
-                    const lines = subs.length
-                        ? subs.map((s) => `${s.done ? "[x]" : "[ ]"} ${s.text}`)
-                        : ["(no sub-tasks)"];
-                    return {
-                        success: true,
-                        output: `Step ${currentStep.id}: ${currentStep.description}\n${lines.join("\n")}`,
-                    };
-                }
-                return {
-                    success: false,
-                    output: t("todo.unknown_action", { action }),
-                };
-            },
-        },
-        {
-            name: "verify",
-            alwaysOn: true,
-            description: "Run verification for the current step. Checks files mentioned in the step description.",
-            parameters: {
-                type: "object",
-                properties: {
-                    step: { type: "number", description: "Step number to verify" },
-                },
-            },
-            handler: async (_ctx, args) => {
-                if (!deps.trackerRef.current)
-                    return { success: false, output: t("plan.no_active") };
-                const tracker = deps.trackerRef.current;
-                const step = args.step ? tracker.getStep(Number(args.step)) : tracker.getCurrentStep();
-                if (!step)
-                    return { success: false, output: t("plan.step_not_found") };
-                const result = await deps.verifier.verifyStep(step.description);
-                if (result.noFiles) {
-                    return {
-                        success: true,
-                        output: t("verify.no_files", { step: String(step.id) }),
-                    };
-                }
-                return {
-                    success: result.passed,
-                    output: result.passed
-                        ? t("verify.passed")
-                        : t("verify.failed", {
-                            details: result.failed.map((f) => f.message).join("; "),
-                        }),
-                };
-            },
-        },
-    ];
-}

package/dist/modules/execution/windows-commands.js

@@ -1,41 +0,0 @@
-/**
- * Unix commands that do not exist (or have incompatible semantics) in Windows
- * cmd.exe. The 9B model keeps reaching for grep/sed/ls/find via the bash tool;
- * each failure returns a hint, but the model retries.
- *
- * This is a session-level, module-side companion to the bash tool's own
- * per-command hard-stop (`UNIX_TO_WIN_HINTS` / `FAILING_FIRST_WORDS` /
- * `HARD_BLOCK_THRESHOLD` in src/tools/bash.ts). That mechanism injects
- * "STOP using X" INTO the tool output, which a 9B model tends to ignore; this
- * one, gated to win32 like bash.ts, queues a <system-summary> that lands
- * before the next model call. Keep the two command lists consistent.
- *
- * This is a command allow/deny list (like the security module's command
- * blacklist), NOT task classification (rule #10).
- */
-const FORBIDDEN_COMMANDS = new Set([
-    "grep",
-    "sed",
-    "ls",
-    "find",
-    "rm",
-    "touch",
-    "which",
-    "diff",
-    "cp",
-    "mv",
-]);
-function firstWord(command) {
-    const segs = command.trim().split(/[\s;|&]+/);
-    return (segs[0] || "").replace(/[^\w-]/g, "").toLowerCase();
-}
-/** The forbidden command name if `command` starts with one (handles `cd x && …`). */
-export function forbiddenWindowsCommand(command) {
-    const m = command.match(/^\s*cd\s+\S+\s*&&\s*(\S+)/);
-    if (m)
-        return forbiddenWindowsCommand(m[1]);
-    const word = firstWord(command);
-    if (word && FORBIDDEN_COMMANDS.has(word))
-        return word;
-    return null;
-}