mindweave 2.4.9 → 2.5.1

package/dist/dynamo/engine.js

@@ -27,7 +27,7 @@ import { basePrompt } from "./prompt.js";
 import { basename } from "node:path";
 import { randomUUID } from "node:crypto";
 import { promises as fsp } from "node:fs";
-import { resolvePath, rootLabel, rootsOf } from "../tools/paths.js";
+import { relativize, resolvePath, rootLabel, rootsOf } from "../tools/paths.js";
 import { renderRules, renderSkillCatalog, reloadGovernance, governanceStamp, rescope } from "../governor/index.js";
 import { forkSession, reloadProjectMemory } from "../memory/session.js";
 import { selectActiveFiles } from "../memory/workingSet.js";
@@ -58,11 +58,11 @@ const MAX_COMPACT_FAILURES = 3;
 export function staticSystemPrompt(projectContext, projectMemory, memoryDir, memoryIndex, governance, workspace, priorSessions = 0) {
     let prompt = basePrompt(commandShellLabel());
     if (workspace) {
-        prompt += `
-
-This session spans more than one root folder. Each file is addressed as \`label/path\`; search tools cover every root unless you pass a specific \`path\`. The roots are:
-<workspace>
-${workspace}
+        prompt += `
+
+This session spans more than one root folder. Each file is addressed as \`label/path\`; search tools cover every root unless you pass a specific \`path\`. The roots are:
+<workspace>
+${workspace}
 </workspace>`;
     }
     // NOTE: the user's standing rules are deliberately NOT rendered here. They live
@@ -73,41 +73,41 @@ ${workspace}
     // skills are a reference catalog), so they alone get the salience boost. Keeping
     // them out of the prefix also stops a mid-session `remember_rule` from busting it.
     if (governance.forbidden) {
-        prompt += `
-
-You are FORBIDDEN from modifying these paths — never write, edit, or run a command that changes them. The tools also enforce this and will refuse, but do not even try:
-<forbidden>
-${governance.forbidden}
+        prompt += `
+
+You are FORBIDDEN from modifying these paths — never write, edit, or run a command that changes them. The tools also enforce this and will refuse, but do not even try:
+<forbidden>
+${governance.forbidden}
 </forbidden>`;
     }
     if (governance.forbiddenCommands) {
-        prompt += `
-
-You are FORBIDDEN from running these commands (or any command that contains one) — run_command will refuse them and only the user can lift that. Do not attempt them or a workaround:
-<forbidden_commands>
-${governance.forbiddenCommands}
+        prompt += `
+
+You are FORBIDDEN from running these commands (or any command that contains one) — run_command will refuse them and only the user can lift that. Do not attempt them or a workaround:
+<forbidden_commands>
+${governance.forbiddenCommands}
 </forbidden_commands>`;
     }
     if (governance.skills) {
-        prompt += `
-
-You have project skills available — named procedures you can run. To run one, call use_skill with its name; its full steps are loaded then (you only see the summary here). Use one when its description fits the task:
-<available_skills>
-${governance.skills}
+        prompt += `
+
+You have project skills available — named procedures you can run. To run one, call use_skill with its name; its full steps are loaded then (you only see the summary here). Use one when its description fits the task:
+<available_skills>
+${governance.skills}
 </available_skills>`;
     }
     if (projectContext) {
-        prompt += `
-
-The following describes the project and machine you're working in, captured at the start of this session (a snapshot — use tools for anything current or deeper):
+        prompt += `
+
+The following describes the project and machine you're working in, captured at the start of this session (a snapshot — use tools for anything current or deeper):
 ${projectContext}`;
     }
     if (projectMemory) {
-        prompt += `
-
-The project provides this context in its MINDWEAVE.md — treat it as background facts about this codebase:
-<project_memory>
-${projectMemory}
+        prompt += `
+
+The project provides this context in its MINDWEAVE.md — treat it as background facts about this codebase:
+<project_memory>
+${projectMemory}
 </project_memory>`;
     }
     // Its own past work in this project. The COUNT goes in the prompt (so the model
@@ -117,16 +117,16 @@ ${projectMemory}
     // and a question about past work gets a real answer instead of a deflection.
     if (priorSessions > 0) {
         const s = priorSessions === 1 ? "" : "s";
-        prompt += `
-
+        prompt += `
+
 You have worked in this project before: ${priorSessions} earlier session${s} of yours are saved, and you can read them. When the user refers to earlier work — "last session", "what did we do", "the bug we fixed" — call \`sessions\` to list them, then \`sessions\` again with an id to read the one they mean, and answer from what you find. It is not in your tool list until you load it with find_tools. Do not say you cannot see your past sessions, and do not guess from the project files instead. \`/continue\` is for the user to RESUME a session; it is not a substitute for you looking. Never present another tool's saved conversations as your own.`;
     }
     if (memoryDir) {
-        prompt += `
-
-Your cross-session memory for this project lives in \`${memoryDir}\` (read or grep the topic files there for the full text of any entry). Its index:
-<memory_index>
-${memoryIndex || "(empty — nothing has been saved to memory yet)"}
+        prompt += `
+
+Your cross-session memory for this project lives in \`${memoryDir}\` (read or grep the topic files there for the full text of any entry). Its index:
+<memory_index>
+${memoryIndex || "(empty — nothing has been saved to memory yet)"}
 </memory_index>`;
     }
     // The deferred pool's index. Roughly forty tokens standing in for several hundred of
@@ -134,8 +134,8 @@ ${memoryIndex || "(empty — nothing has been saved to memory yet)"}
     // missing feature, and the model routes around a capability it actually has.
     const deferred = deferredToolsIndex();
     if (deferred) {
-        prompt += `
-
+        prompt += `
+
 ${deferred}`;
     }
     return prompt;
@@ -156,8 +156,17 @@ ${deferred}`;
 const ACTIVE_FILES_FOR_NOTES = 20;
 export function volatileContext(rules, planMode, sessionMemory, approvedPlan = "", 
 /** Notes for the folders being worked in right now (see memory/projectNotes.ts). */
-directoryNotes = []) {
+directoryNotes = [], 
+/** Where the previous turn left the shell, when this turn started back at the root. */
+cwdResetFrom = "") {
     const parts = [];
+    // Each turn starts at the project root, and a model that `cd`-ed into a subfolder last
+    // turn does not know that unless it is told. It was not, and a real session ran
+    // `cargo run` from the root expecting the subfolder: "could not find Cargo.toml".
+    if (cwdResetFrom) {
+        parts.push(`Commands run from the project root. The previous turn had moved into ${cwdResetFrom}, but each ` +
+            `turn starts back at the root: cd there again, or use paths from the root.`);
+    }
     // Standing rules FIRST in the volatile tail. They're rebuilt every turn here (not
     // in the cached prefix), so a long conversation can never bury them — and they sit
     // at the top of the freshest context the model reads before it acts. Binding by
@@ -442,6 +451,12 @@ function workspaceText(session) {
 function modelSeesImages(session) {
     return manifestForModel(session.modelConfig.model).acceptsImages?.(session.modelConfig.model) ?? false;
 }
+/** "JPG and PNG" from `["image/png", "image/jpeg"]` — the names a user knows the files by. */
+export function imageTypeNames(types) {
+    const names = types.map((t) => ({ "image/jpeg": "JPG", "image/png": "PNG", "image/gif": "GIF", "image/webp": "WebP" })[t] ?? t);
+    const sorted = [...new Set(names)].sort();
+    return sorted.length <= 1 ? (sorted[0] ?? "") : `${sorted.slice(0, -1).join(", ")} and ${sorted[sorted.length - 1]}`;
+}
 async function loadImagePayloads(session) {
     // Nothing to load for a model that cannot look at one. This is the /provider switch
     // case: a picture attached while a vision model was running stays in the transcript,
@@ -471,6 +486,11 @@ function buildRequest(session, tools, imagePayloads = new Map(),
 /** Notes for the folders in play, resolved by the caller (it has to read disk). */
 directoryNotes = []) {
     const canSeeImages = modelSeesImages(session);
+    // The formats it takes, when narrower than everything core attaches. A manifest fact,
+    // like vision itself; absent means every type.
+    const imageTypes = canSeeImages
+        ? manifestForModel(session.modelConfig.model).imageTypes?.(session.modelConfig.model)
+        : undefined;
     const messages = [];
     for (const e of session.transcript) {
         if (e.role === "user" || e.role === "summary") {
@@ -487,6 +507,7 @@ directoryNotes = []) {
                 const images = [];
                 const missing = [];
                 const unseen = [];
+                const wrongType = [];
                 for (const ref of refs) {
                     // Told, never silently dropped. A message that mentions a screenshot and
                     // carries nothing reads to the model as a picture it failed to notice; the
@@ -495,6 +516,12 @@ directoryNotes = []) {
                         unseen.push(basename(ref.path));
                         continue;
                     }
+                    // A format the provider would reject: held back the same way, so the model
+                    // can tell the user which formats work instead of the whole request failing.
+                    if (imageTypes && !imageTypes.includes(ref.mediaType)) {
+                        wrongType.push(basename(ref.path));
+                        continue;
+                    }
                     const data = imagePayloads.get(ref.path);
                     if (data)
                         images.push({ path: ref.path, mediaType: ref.mediaType, data });
@@ -505,10 +532,17 @@ directoryNotes = []) {
                     ...(unseen.length > 0
                         ? [`${unseen.join(", ")} was attached, but the model now running cannot see images`]
                         : []),
+                    ...(wrongType.length > 0 && imageTypes
+                        ? [
+                            `${wrongType.join(", ")} was attached but not sent: the model now running ` +
+                                `accepts only ${imageTypeNames(imageTypes)} images, so tell the user to ` +
+                                `convert it or switch models`,
+                        ]
+                        : []),
                     ...(missing.length > 0 ? [`${missing.join(", ")} could not be read from disk`] : []),
                 ];
-                const content = notes.length > 0 ? `${said}
-
+                const content = notes.length > 0 ? `${said}
+
 [${notes.join("; ")}]` : said;
                 messages.push({ role: "user", content, ...(images.length > 0 ? { images } : {}) });
                 continue;
@@ -543,7 +577,12 @@ directoryNotes = []) {
                 approvedAt: session.toolContext.activePlanApprovedAt ?? "",
                 mode: "lightning",
             })
-            : "", directoryNotes),
+            : "", directoryNotes, 
+        // Only while the turn is still sitting where the reset put it. Once a command moves,
+        // the command's own result says where it is now.
+        session.toolContext.cwdResetFrom && session.toolContext.cwd === session.cwd
+            ? relativize(session.toolContext, session.toolContext.cwdResetFrom)
+            : ""),
         tools,
         model: session.modelConfig,
     };
@@ -558,55 +597,83 @@ async function backgroundEventNotes(session) {
     if (!mgr)
         return [];
     const events = await mgr.drainEvents();
-    return events.map(({ info, kind, tail, wake }) => {
-        // It came up. This is the only positive event a server ever produces, and it is
-        // what lets the model actually deliver the "I'll tell you when it's running" it
-        // was told to say. Nothing has gone wrong, so there is nothing to fix.
-        if (kind === "ready") {
-            return (`[Background shell #${info.id} (\`${info.command}\`) is up and running.]\n` +
-                `Recent output:\n${tail || "(no output)"}\n\n` +
-                `Tell the user in one short line that it's running. Nothing is wrong — do not investigate, ` +
-                `do not restart it, and do not change any files because of this.`);
-        }
-        // The watchdog thinks this running shell is stuck. It has produced nothing for a
-        // while — either blocked on a prompt it will never answer, or silently wedged on a
-        // command that should have kept working. The point is to stop it sitting invisible
-        // until the timeout, and to hand the model the two moves that resolve it.
-        if (kind === "stalled") {
-            const why = info.stallReason === "prompt"
-                ? `It looks like it is waiting for interactive input (its last line reads as a prompt). ` +
-                    `Kill it with kill_shell #${info.id} and re-run non-interactively — pipe the answer in ` +
-                    `(e.g. \`echo y | …\`) or add a non-interactive flag like \`-y\`/\`--yes\`.`
-                : `It has produced no output for a long time and may be wedged. Read it with shells #${info.id} ` +
-                    `to judge, then either keep waiting if it is genuinely mid-work, or kill it with ` +
-                    `kill_shell #${info.id} and look into why it hangs.`;
-            return (`[Background shell #${info.id} (\`${info.command}\`) appears to be stuck.]\n` +
-                `Recent output:\n${tail || "(no output)"}\n\n${why}`);
-        }
-        const status = info.status === "killed"
-            ? info.stoppedBy === "user"
-                ? "was stopped by the user"
-                : "was killed"
-            : `finished with exit code ${info.exitCode}`;
-        // An ending that is NOT worth interrupting for still arrives, so the model knows the
-        // thing is down and can answer about it. It is explicitly not a task: this is the
-        // path a user closing their own app takes, and treating it as news is what made the
-        // agent reopen it.
-        if (!wake) {
-            return (`[Background shell #${info.id} (\`${info.command}\`) ${status}. It had already started up, so ` +
-                `this is the user stopping their own app, not a failure.]\n` +
-                `This is background information only. Do NOT mention it unless it is relevant, do NOT restart ` +
-                `it, and do NOT change any files because of it. If the user later asks about this app, you now ` +
-                `know it is stopped.`);
-        }
-        // For a server, only a failure to come up reaches here: a normal stop does not wake.
-        const guidance = info.notify === "on_failure"
-            ? "This is a server or app that never came up, so the user never saw it running. Tell them what happened and offer to fix it — but do not restart it repeatedly on your own."
-            : "If it failed, tell the user briefly what went wrong and propose a fix — don't change files unless they agree.";
-        return (`[Background shell #${info.id} (\`${info.command}\`) ${status}.]\n` +
+    return events.map(backgroundEventNote).filter((note) => note !== null);
+}
+/**
+ * The note for one background-shell event, or null when there is nothing to say (pure).
+ *
+ * An ending the AGENT caused says nothing: `kill_shell` already told it the shell
+ * stopped. The note that used to follow declared "this is the user stopping their own
+ * app", which blamed the user for the agent's own restart and gave the model a second,
+ * contradictory account of the same event.
+ */
+export function backgroundEventNote({ info, kind, tail, wake, }) {
+    // It came up. This is the only positive event a server ever produces, and it is
+    // what lets the model actually deliver the "I'll tell you when it's running" it
+    // was told to say. Nothing has gone wrong, so there is nothing to fix.
+    if (kind === "ready") {
+        return (`[Background shell #${info.id} (\`${info.command}\`) is up and running.]\n` +
             `Recent output:\n${tail || "(no output)"}\n\n` +
-            guidance);
-    });
+            `Tell the user in one short line that it's running. Nothing is wrong — do not investigate, ` +
+            `do not restart it, and do not change any files because of this. Running only means the ` +
+            `process started: do not describe what it shows or say a change is visible unless you ` +
+            `have actually looked.`);
+    }
+    // The watchdog thinks this running shell is stuck. It has produced nothing for a
+    // while — either blocked on a prompt it will never answer, or silently wedged on a
+    // command that should have kept working. The point is to stop it sitting invisible
+    // until the timeout, and to hand the model the two moves that resolve it.
+    if (kind === "stalled") {
+        const why = info.stallReason === "prompt"
+            ? `It looks like it is waiting for interactive input (its last line reads as a prompt). ` +
+                `Kill it with kill_shell #${info.id} and re-run non-interactively — pipe the answer in ` +
+                `(e.g. \`echo y | …\`) or add a non-interactive flag like \`-y\`/\`--yes\`.`
+            : `It has produced no output for a long time and may be wedged. Read it with shells #${info.id} ` +
+                `to judge, then either keep waiting if it is genuinely mid-work, or kill it with ` +
+                `kill_shell #${info.id} and look into why it hangs.`;
+        return (`[Background shell #${info.id} (\`${info.command}\`) appears to be stuck.]\n` +
+            `Recent output:\n${tail || "(no output)"}\n\n${why}`);
+    }
+    const status = info.status === "killed"
+        ? info.stoppedBy === "user"
+            ? "was stopped by the user"
+            : "was killed"
+        : `finished with exit code ${info.exitCode}`;
+    // An ending that is NOT worth interrupting for still arrives, so the model knows the
+    // thing is down and can answer about it. It is explicitly not a task: this is the
+    // path a user closing their own app takes, and treating it as news is what made the
+    // agent reopen it.
+    if (info.status === "killed" && info.stoppedBy === "agent")
+        return null;
+    // Mindweave stopped it itself, because its output ran past the size cap. That is a
+    // runaway, not someone closing an app, and the model is the one who can explain it.
+    if (info.status === "killed" && info.stoppedBy === "system") {
+        return (`[Background shell #${info.id} (\`${info.command}\`) was stopped by Mindweave because its output ` +
+            `passed the size limit.]\n` +
+            `Recent output:\n${tail || "(no output)"}\n\n` +
+            `Something in it was writing without end. Tell the user, and look at the output above before ` +
+            `running it again.`);
+    }
+    if (!wake) {
+        // Only a stop the user made through the app is known to be theirs. An app that exited
+        // on its own after coming up was most likely closed by them, which is worth saying as
+        // the likely reading rather than as a fact.
+        const who = info.status === "killed" && info.stoppedBy === "user"
+            ? "the user stopping their own app"
+            : "most likely the user closing their own app";
+        return (`[Background shell #${info.id} (\`${info.command}\`) ${status}. It had already started up, so ` +
+            `this is ${who}, not a failure.]\n` +
+            `This is background information only. Do NOT mention it unless it is relevant, do NOT restart ` +
+            `it, and do NOT change any files because of it. If the user later asks about this app, you now ` +
+            `know it is stopped.`);
+    }
+    // For a server, only a failure to come up reaches here: a normal stop does not wake.
+    const guidance = info.notify === "on_failure"
+        ? "This is a server or app that never came up, so the user never saw it running. If you started it to check work you are still doing, getting it running is part of that work: find out why it failed and fix it. Otherwise tell them what happened and offer to fix it. Either way, do not restart it again without changing something first."
+        : "If it failed, tell the user briefly what went wrong and propose a fix — don't change files unless they agree.";
+    return (`[Background shell #${info.id} (\`${info.command}\`) ${status}.]\n` +
+        `Recent output:\n${tail || "(no output)"}\n\n` +
+        guidance);
 }
 /**
  * Produce Mindweave's next reply for the latest user message already on
@@ -696,13 +763,13 @@ export async function respond(session, options = {}) {
 function implementFromScratch(priorPath, plan) {
     const path = priorPath;
     const where = path
-        ? `
-
+        ? `
+
 If you need something exact from the planning that produced this — a snippet, an ` +
             `error message, a path — the full conversation is at: ${path}`
         : "";
-    return `Implement the following plan:
-
+    return `Implement the following plan:
+
 ${plan}${where}`;
 }
 /**
@@ -803,12 +870,16 @@ async function respondTurn(session, options = {}) {
     // (its lifecycle + tagged tool calls) up this same stream instead of running dark.
     session.toolContext.emitEvent = options.onEvent;
     session.toolContext.abortSignal = options.signal;
+    // So a tool can answer "which model are you running" instead of guessing.
+    session.toolContext.modelConfig = session.modelConfig;
     // WORKING-DIRECTORY RESET. Each turn starts at the project root — the working
     // directory is already set to the correct project directory automatically. Within a
     // turn cd still persists (so a multi-step command sequence works), but it never
     // carries a stale `cd` into the next
     // turn — the bug where `cd src-tauri` run in two turns became `…/src-tauri/src-tauri`.
     // The primary root (session.cwd) is fixed; only toolContext.cwd moves.
+    const leftIn = session.toolContext.cwd;
+    session.toolContext.cwdResetFrom = leftIn && leftIn !== session.cwd ? leftIn : undefined;
     session.toolContext.cwd = session.cwd;
     // TASK-BOUNDARY SWEEP. If the previous turn finished a task (a todo list completed)
     // and this new message opens a DIFFERENT one (not a "continue"), close the finished
@@ -836,6 +907,10 @@ async function respondTurn(session, options = {}) {
     const limits = taskLimits();
     const startedAt = Date.now();
     const usages = [];
+    // When each of those calls returned. Recorded as it happens: stamping them when the
+    // turn is saved gave every call in a turn the same time, so the call log could not say
+    // how a long turn's time was spent (one real session: 145 calls, 4 distinct times).
+    const usageTimes = [];
     // Verification-gate bookkeeping for this turn: did the model change any file,
     // did it ever run a check, and have we already nudged once (one-shot).
     let mutatedThisTurn = false;
@@ -898,7 +973,7 @@ async function respondTurn(session, options = {}) {
         // of them — those are different problems with different fixes, and the totals look
         // identical for all of them. Six numbers per call, capped, so a long session cannot
         // grow the meta file without bound.
-        session.callLog = [...(session.callLog ?? []), ...usages.map((u) => toCallRecord(u, session.modelConfig.model))].slice(-CALL_LOG_LIMIT);
+        session.callLog = [...(session.callLog ?? []), ...usages.map((u, i) => toCallRecord(u, session.modelConfig.model, usageTimes[i]))].slice(-CALL_LOG_LIMIT);
     };
     try {
         const reply = await runTurn();
@@ -1070,6 +1145,7 @@ async function respondTurn(session, options = {}) {
             emitUsage(result, options);
             if (result.usage) {
                 usages.push(result.usage);
+                usageTimes.push(Date.now());
                 writeCacheLog(cacheCallLine({
                     call: usages.length,
                     gapMs: sinceLastCall,
@@ -1337,6 +1413,7 @@ async function respondTurn(session, options = {}) {
                     summary: result.summary,
                     isError: result.isError,
                     detail: result.detail,
+                    detailFull: result.detailFull,
                     detailKind: result.detailKind,
                     quiet: result.quiet,
                     fullContentOf: result.fullContentOf,
@@ -1360,6 +1437,7 @@ async function respondTurn(session, options = {}) {
                     summary: r.summary ?? r.call.name,
                     error: r.isError ?? false,
                     detail: r.detail,
+                    ...("detailFull" in r && r.detailFull ? { detailFull: r.detailFull } : {}),
                     ...(r.detailKind ? { detailKind: r.detailKind } : {}),
                     ...(r.quiet ? { quiet: true } : {}),
                     ...(r.displayKind ? { displayKind: r.displayKind } : {}),
@@ -1493,8 +1571,10 @@ async function respondTurn(session, options = {}) {
                 session.transcript.push({
                     role: "user",
                     content: canSee
-                        ? `Here ${shots.length === 1 ? "is the image" : "are the images"} just captured (${names}).`
-                        : `${names} was captured and saved, but this model cannot see images, so you are ` +
+                        ? // Not "just captured": view_image opens files that already existed (the user's own
+                            // screenshots), and saying they were captured tells the model it took them.
+                            `Here ${shots.length === 1 ? "is the image" : "are the images"} from the tool call above (${names}).`
+                        : `${names} is ready, but this model cannot see images, so you are ` +
                             `being told about it rather than shown it. Describe what you expected to verify ` +
                             `and ask the user what they see, or switch to a model with vision using /model.`,
                     synthetic: true,
@@ -1784,9 +1864,9 @@ const CALL_LOG_LIMIT = 200;
 /** One call's usage, flattened for the session file. Exported so the recording is
  *  testable on its own — persisting a hand-built record proves nothing about what the
  *  engine actually writes. */
-export function toCallRecord(u, model) {
+export function toCallRecord(u, model, at = Date.now()) {
     return {
-        at: Date.now(),
+        at,
         prompt: u.promptTokens,
         hit: u.cacheHitTokens,
         miss: u.cacheMissTokens,