micro-models-agent 2.37.0 → 2.38.0

package/CHANGELOG.md

@@ -4,6 +4,31 @@ All notable changes to Micro Models Agent (MMA) will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/), and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [2.38.0] - 2026-10-09
+
+Harness honesty for read-only work: the loop guard no longer treats "no mutation" as
+"no progress", every forced stop now yields a best-effort final answer instead of a silent
+kill, and the web chat shows why a turn stopped. Frontend seams landed so a future TUI can
+reuse the command registry and session snapshot, and `grep` on Windows was hardened.
+
+### Changed
+
+- **Host snapshot projection moved to `core`** (`src/core/host-state.ts`): `buildHostState`/`captureUsage`/`liveContextUsage`/`planEventFor`/`UsageSnapshot` lived in `src/modules/webui/host-state.ts`, yet `mma serve` already imported `planEventFor` from there - a structural subset of the web backend acting as a shared contract. `src/modules/webui/host-state.ts` is now a thin web wrapper keeping `WebIndexer`/`WebLsp` and the web-only `i18nStale` probe.
+- **`ReplContext` split into services and IO** (`src/cli/command-context.ts`): the old 20-member interface mixed frontend-agnostic services (config/agent/modules) with single-shot terminal IO (`rl: LineInput`, `output: OutputChannel`), so a TUI that redraws frames could not reuse the command registry. Now `CommandServices` (frontend-agnostic) + `CommandOutput` (2 required methods) + `CommandInput` (3 methods). `ReplContext` is an alias. `/wizard` and `/config reset --all` ask `ctx.input.withInput(...)`/`ctx.input.confirm(...)` instead of touching the terminal; `Spinner` takes only `writeOverlay`.
+- **Declarative slash-command catalog** (`src/core/commands/`): three descriptions drifted silently - the hand-synced `COMMAND_GROUPS` table (29 lines), the REPL registrations, and the web `HELP` list. All 29 commands are now data (`spec.ts`) plus a registry with per-surface availability (`catalog.ts`). Actions are deliberately not unified - terminal prints ANSI, the browser emits plain SSE lines, which is the `CommandOutput` adapter's job. `/help` builds from the catalog; `COMMAND_GROUPS` is deleted. `isAvailable` defaults to terminal + TUI so a future TUI gets the whole registry.
+- **Forced final answer instead of a silent kill** (`src/core/agent.ts`, `tool-batch.ts`): every harness terminal (duplicate budget, failed-batch honest-stop, `maxToolIterations` exhaustion) now makes one tool-less LLM call and accepts the text as the final answer (critic/audit/claim gates skipped, answer persisted to history, tokens/cost counted); `AgentResult.stopReason` carries the trigger. An empty forced answer is still a failure.
+
+### Added
+
+- **Anti-drift guards**: `tests/core/command-catalog.test.ts` asserts the catalog covers exactly the commands the REPL registers (both directions) and that web-marked commands match the real web dispatcher; `tests/cli/repl-help.test.ts` pins the user-visible `/help` output; `tests/cli/command-context.test.ts` locks the surface a new frontend must implement.
+
+### Fixed
+
+- **Read-only tasks are no longer hard-stopped** (`src/core/agent/tool-batch.ts`): the guard treated "no work mutation" as "no progress", so a genuinely read-only turn (analysis) always climbed the streak and was killed at 12 batches before the model could answer (`ses_mv0pirf5_hnmbxj`). Progress is now exact tool-call repetition (new `canonicalArgsKey`, `src/core/agent/canonical-args.ts`), so broad exploration with new files/patterns never trips it.
+- **Web shows why a turn stopped** (`webui/session/{run-agent,forward}.ts`, `assets/js/feed/turn.ts`): the harness reason (`AgentResult.stopReason` / `error`) was dropped on the web path, leaving a generic `turn failed`. It is now forwarded through SSE `turn_end` and rendered.
+- **`grep` with `|` on Windows** (`src/tools/grep-tool.ts`): an npm `.cmd`/`.ps1` `rg` shim runs through `cmd.exe` under Bun, so alternation patterns were read as pipes and the search failed with no JS fallback. Shim/spawn failures (exit 255/9009, shell signature) are now treated as engine-unavailable and fall through to the JS walker.
+- **Stuck-detector read-only nudge removed** (`src/modules/execution/stuck-detector.ts`, `execution-plugin.ts`): the second "no mutation = stuck" heuristic (10 read-only calls → `exec.read_only_loop`, "No write/exec for N tool calls") fired on legitimate read-only analysis. Removed — real loops stay covered by repetition sampling, step iteration, failure streaks and the duplicate-tool budget. Keys `exec.read_only_loop`, `exec.read_only_loop_recovery`, `exec.hint_read_only` dropped (en+ru).
+
 ## [2.37.0] - 2026-10-09
 
 `mma serve` now describes its own API: a live OpenAPI 3.1 document plus a

package/README.md

@@ -35,7 +35,7 @@ Most agents (Devin, Cursor, Copilot) are glued to cloud models or pricey APIs
 
 - **6-state agent loop** (`INIT → THINK → ACT → OBSERVE → OUTPUT → ERROR`) — clean, predictable, debuggable
 - **Context budget management** — per-model budgets with sliding-window compaction that preserves facts, decisions, and errors, keeps a `[Tool ledger]` of what was already read, and calibrates its token estimate against the provider's real `prompt_tokens`
-- **Honest-stop guards** — failed tool batches, read-only loops and retry storms stop the loop with a truthful status instead of letting a small model grind (honest stop, read-only budget, `retry_loop_stopped`)
+- **Honest-stop guards** — failed tool batches and duplicate-tool loops stop the loop with a truthful status instead of letting a small model grind (honest stop, duplicate-tool budget, forced tool-less final answer)
 - **Hallucination detection** — fact checks (file paths), consistency checks (decision rollbacks), confidence checks, plus claim verification: "created file X" / "output says Y" / "the price is Z" are checked against real tool output
 - **Execution guarantees** — LLM-driven auto-planning (no keyword heuristics), stuck detection, off-track warnings, automatic plan advancement
 

package/dist/assets/js/feed/turn.js

@@ -1,5 +1,5 @@
-import{msgs as A,el as i,icon as w,announce as a}from"../dom.js";import{t as f,pluralCount as _}from"../i18n.js";import{streamStateFor as C,tabs as k,session as g}from"../store.js";import{FILE_TOOLS as x,OPENABLE_TOOLS as F,LINK_TOOLS as B,categoryOf as y,labelOf as I,argText as U,formatMs as S,verbosity as b,firstOutputLine as L}from"../format.js";import{copyText as q}from"../../md.js";import{imageCard as P}from"./image-card.js";import{activeFeed as j,liveFeed as D,feedForEvent as h,clearBusy as O,clearEmpty as T,appendTo as d,insertAfter as N,appendActionsRow as $,flushRun as K,scrollToBottom as z}from"./core.js";import{finishMd as E,finishThink as M}from"./stream.js";export function onToolStart(e){const n=h(e.sessionId);if(O(n),!n)return;T(n),E(n),M(n);const o=C(n);o.pendingArgs.set(e.toolCallId,e.args||{});const s=i("div","act "+y(e.name));s.appendChild(i("span","g","◇")),s.appendChild(i("span","lab",I(e.name)));const l=U(e.args);l&&(s.appendChild(i("span","sep","·")),s.appendChild(i("span","args",l)));const t=i("span","ok","");s.appendChild(t),s.classList.add("running"),d(n,s),o.toolNodes.set(e.toolCallId,{status:t,line:s}),y(e.name)!=="recon"&&(o.runReconOnly=!1),o.runCount++}function R(e,n){if(typeof n!="string"||!n)return;const o=i("div","toolout"),s=i("div","code");if(s.textContent=n,o.appendChild(s),n.length>600||n.split(`
-`).length>12){o.classList.add("collapsed");const u="▸ "+_("web.tool.output",n.length),r=i("span","more",u);r.addEventListener("click",()=>{const c=o.classList.toggle("collapsed");r.textContent=c?u:"▾ "+f("web.tool.collapse")}),o.appendChild(r)}const t=e&&e.parentNode?null:D();e&&e.parentNode?e.parentNode.insertBefore(o,e.nextSibling):d(t,o)}function G(e){if(typeof e!="string"||!e)return null;try{const n=new URL(e,location.href);if(n.protocol==="http:"||n.protocol==="https:"||n.protocol==="file:")return n.href}catch{return null}return null}function H(e){const n=String(e??"").replace(/\\/g,"/");try{const o=g&&g.host&&g.host.workspace?String(g.host.workspace).replace(/\\/g,"/").replace(/\/+$/,""):"";if(o&&(n===o||n.indexOf(o+"/")===0)){const s=n===o?"":n.slice(o.length+1);if(s)return s}}catch{}return n}function m(e,n,o,s){const l=h(o);if(!l)return null;if(x.indexOf(e)!==-1){const t=n.path||n.file||n.dir;if(!t)return null;const u=H(t),r=i("div","filerow "+y(e));if(r.appendChild(i("span","g","◇")),r.appendChild(i("span","lab",I(e))),r.appendChild(i("span","sep"," · ")),F.indexOf(e)!==-1){const p=i("a","path",u||String(t));p.href="/api/file?path="+encodeURIComponent(String(t)),p.target="_blank",p.rel="noopener noreferrer",p.title=f("web.card.open_title")+" · "+String(t),r.appendChild(p)}else{const p=i("span","tail",u||String(t));p.title=String(t),r.appendChild(p)}const c=i("button","copy-btn");return c.type="button",c.title=f("web.card.copy_title"),c.setAttribute("aria-label",f("web.card.copy_title")),c.appendChild(w("copy")),c.addEventListener("click",()=>{q(String(t)).then(p=>{if(c.replaceChildren(w(p?"check":"copy")),!p){const v=r.querySelector(".st");v&&(v.textContent=" ⚠")}window.setTimeout(()=>c.replaceChildren(w("copy")),1400)})}),r.appendChild(c),r.appendChild(i("span","st",s===!1?" ✗":s===!0?" ✓":"")),s===!1&&(r.dataset.fail="1"),/\.drawio(\.xml)?$/i.test(String(t))&&b()==="verbose"&&r.appendChild(i("span","detail","↳ "+f("web.card.drawio_hint"))),d(l,r),r}if(B.indexOf(e)!==-1&&n.url){const t=i("div","linkline");t.appendChild(document.createTextNode("↳ "));const u=G(n.url);if(u){const r=i("a","",n.url);r.href=u,r.target="_blank",r.rel="noopener noreferrer",t.appendChild(r);try{const c=new URL(u).hostname;c&&t.appendChild(i("span","dom"," · "+c))}catch{}}else t.appendChild(i("span","",n.url));return d(l,t),t}return null}function J(e){if(!e.success||!Array.isArray(e.images)||!e.images.length)return;const n=h(e.sessionId);if(n)for(const o of e.images){const s=P(String(e.name||""),String(o));s&&d(n,s)}}export function onToolEnd(e){Q(e),J(e)}function Q(e){const n=h(e.sessionId),o=n?C(n):null,s=o?o.toolNodes.get(e.toolCallId):void 0,l=s?s.line:null;if(s&&o){s.line.classList.remove("running");const r=S(e.durationMs);s.status.textContent=e.success?" ✓"+(r?" "+r:""):" ✗"+(r?" "+r:""),e.success||(s.status.classList.add("fail"),o.runFails++,o.runReconOnly=!1),o.toolNodes.delete(e.toolCallId)}const t=(o?o.pendingArgs.get(e.toolCallId):void 0)||{};o&&o.pendingArgs.delete(e.toolCallId);const u=t.path||t.file||t.dir;if(x.indexOf(e.name)!==-1&&u){if(l&&l.parentNode&&l.remove(),b()==="quiet"){if(e.success){const c=m(e.name,t,e.sessionId,e.success);c&&o&&o.runNodes.push(c),o&&(o.runDuration+=typeof e.durationMs=="number"?e.durationMs:0)}else{const c=m(e.name,t,e.sessionId,e.success),p=L(e.output);p&&N(c||l,i("div","detail","↳ "+p)),c&&o&&o.runNodes.push(c)}return}const r=m(e.name,t,e.sessionId,e.success);R(r||l,e.output);return}if(b()==="quiet"){if(e.success){l&&o&&o.runNodes.push(l);const r=m(e.name,t,e.sessionId,e.success);r&&o&&o.runNodes.push(r),o&&(o.runDuration+=typeof e.durationMs=="number"?e.durationMs:0)}else{const r=L(e.output);r&&N(l,i("div","detail","↳ "+r))}return}e.success&&m(e.name,t,e.sessionId,e.success),R(l,e.output)}export function onTurnEnd(e){g.busy=!1;const n=h(e.sessionId);O(n),M(n),E(n),K(n);const o=n?C(n):null;o&&o.toolNodes.forEach(t=>{t&&t.line&&n&&t.line.parentNode===n&&t.line.classList.remove("running")});const s=k.tabs.find(t=>t.id===e.sessionId);if(s&&(s.planCard=null),!n||!o){resetTurnText(n);return}if(e.success===!1)e.interrupted?(d(n,i("div","warnline","⚠ "+f("web.turn.stopped"))),a(f("web.turn.stopped"))):(d(n,i("div","errline","✗ "+f("web.turn.failed"))),a(f("web.turn.failed")));else{const t=[];typeof e.toolCalls=="number"&&e.toolCalls>0&&t.push(f("web.turn.tools",{count:e.toolCalls})),typeof e.durationMs=="number"&&t.push(S(e.durationMs)),t.length&&d(n,i("div","metric","⟡ "+t.join(" · "))),a(f("web.turn.end",{count:e.toolCalls||0}))}const l=o.turnRoots.map(t=>t.innerText||t.textContent||"").join(`
+import{msgs as _,el as i,icon as w,announce as a}from"../dom.js";import{t as d,pluralCount as A}from"../i18n.js";import{streamStateFor as C,tabs as k,session as g}from"../store.js";import{FILE_TOOLS as x,OPENABLE_TOOLS as F,LINK_TOOLS as B,categoryOf as y,labelOf as I,argText as U,formatMs as S,verbosity as b,firstOutputLine as L}from"../format.js";import{copyText as q}from"../../md.js";import{imageCard as P}from"./image-card.js";import{activeFeed as j,liveFeed as D,feedForEvent as h,clearBusy as O,clearEmpty as T,appendTo as f,insertAfter as N,appendActionsRow as $,flushRun as K,scrollToBottom as z}from"./core.js";import{finishMd as E,finishThink as R}from"./stream.js";export function onToolStart(e){const n=h(e.sessionId);if(O(n),!n)return;T(n),E(n),R(n);const o=C(n);o.pendingArgs.set(e.toolCallId,e.args||{});const r=i("div","act "+y(e.name));r.appendChild(i("span","g","◇")),r.appendChild(i("span","lab",I(e.name)));const l=U(e.args);l&&(r.appendChild(i("span","sep","·")),r.appendChild(i("span","args",l)));const t=i("span","ok","");r.appendChild(t),r.classList.add("running"),f(n,r),o.toolNodes.set(e.toolCallId,{status:t,line:r}),y(e.name)!=="recon"&&(o.runReconOnly=!1),o.runCount++}function M(e,n){if(typeof n!="string"||!n)return;const o=i("div","toolout"),r=i("div","code");if(r.textContent=n,o.appendChild(r),n.length>600||n.split(`
+`).length>12){o.classList.add("collapsed");const p="▸ "+A("web.tool.output",n.length),s=i("span","more",p);s.addEventListener("click",()=>{const c=o.classList.toggle("collapsed");s.textContent=c?p:"▾ "+d("web.tool.collapse")}),o.appendChild(s)}const t=e&&e.parentNode?null:D();e&&e.parentNode?e.parentNode.insertBefore(o,e.nextSibling):f(t,o)}function G(e){if(typeof e!="string"||!e)return null;try{const n=new URL(e,location.href);if(n.protocol==="http:"||n.protocol==="https:"||n.protocol==="file:")return n.href}catch{return null}return null}function H(e){const n=String(e??"").replace(/\\/g,"/");try{const o=g&&g.host&&g.host.workspace?String(g.host.workspace).replace(/\\/g,"/").replace(/\/+$/,""):"";if(o&&(n===o||n.indexOf(o+"/")===0)){const r=n===o?"":n.slice(o.length+1);if(r)return r}}catch{}return n}function m(e,n,o,r){const l=h(o);if(!l)return null;if(x.indexOf(e)!==-1){const t=n.path||n.file||n.dir;if(!t)return null;const p=H(t),s=i("div","filerow "+y(e));if(s.appendChild(i("span","g","◇")),s.appendChild(i("span","lab",I(e))),s.appendChild(i("span","sep"," · ")),F.indexOf(e)!==-1){const u=i("a","path",p||String(t));u.href="/api/file?path="+encodeURIComponent(String(t)),u.target="_blank",u.rel="noopener noreferrer",u.title=d("web.card.open_title")+" · "+String(t),s.appendChild(u)}else{const u=i("span","tail",p||String(t));u.title=String(t),s.appendChild(u)}const c=i("button","copy-btn");return c.type="button",c.title=d("web.card.copy_title"),c.setAttribute("aria-label",d("web.card.copy_title")),c.appendChild(w("copy")),c.addEventListener("click",()=>{q(String(t)).then(u=>{if(c.replaceChildren(w(u?"check":"copy")),!u){const v=s.querySelector(".st");v&&(v.textContent=" ⚠")}window.setTimeout(()=>c.replaceChildren(w("copy")),1400)})}),s.appendChild(c),s.appendChild(i("span","st",r===!1?" ✗":r===!0?" ✓":"")),r===!1&&(s.dataset.fail="1"),/\.drawio(\.xml)?$/i.test(String(t))&&b()==="verbose"&&s.appendChild(i("span","detail","↳ "+d("web.card.drawio_hint"))),f(l,s),s}if(B.indexOf(e)!==-1&&n.url){const t=i("div","linkline");t.appendChild(document.createTextNode("↳ "));const p=G(n.url);if(p){const s=i("a","",n.url);s.href=p,s.target="_blank",s.rel="noopener noreferrer",t.appendChild(s);try{const c=new URL(p).hostname;c&&t.appendChild(i("span","dom"," · "+c))}catch{}}else t.appendChild(i("span","",n.url));return f(l,t),t}return null}function J(e){if(!e.success||!Array.isArray(e.images)||!e.images.length)return;const n=h(e.sessionId);if(n)for(const o of e.images){const r=P(String(e.name||""),String(o));r&&f(n,r)}}export function onToolEnd(e){Q(e),J(e)}function Q(e){const n=h(e.sessionId),o=n?C(n):null,r=o?o.toolNodes.get(e.toolCallId):void 0,l=r?r.line:null;if(r&&o){r.line.classList.remove("running");const s=S(e.durationMs);r.status.textContent=e.success?" ✓"+(s?" "+s:""):" ✗"+(s?" "+s:""),e.success||(r.status.classList.add("fail"),o.runFails++,o.runReconOnly=!1),o.toolNodes.delete(e.toolCallId)}const t=(o?o.pendingArgs.get(e.toolCallId):void 0)||{};o&&o.pendingArgs.delete(e.toolCallId);const p=t.path||t.file||t.dir;if(x.indexOf(e.name)!==-1&&p){if(l&&l.parentNode&&l.remove(),b()==="quiet"){if(e.success){const c=m(e.name,t,e.sessionId,e.success);c&&o&&o.runNodes.push(c),o&&(o.runDuration+=typeof e.durationMs=="number"?e.durationMs:0)}else{const c=m(e.name,t,e.sessionId,e.success),u=L(e.output);u&&N(c||l,i("div","detail","↳ "+u)),c&&o&&o.runNodes.push(c)}return}const s=m(e.name,t,e.sessionId,e.success);M(s||l,e.output);return}if(b()==="quiet"){if(e.success){l&&o&&o.runNodes.push(l);const s=m(e.name,t,e.sessionId,e.success);s&&o&&o.runNodes.push(s),o&&(o.runDuration+=typeof e.durationMs=="number"?e.durationMs:0)}else{const s=L(e.output);s&&N(l,i("div","detail","↳ "+s))}return}e.success&&m(e.name,t,e.sessionId,e.success),M(l,e.output)}export function onTurnEnd(e){g.busy=!1;const n=h(e.sessionId);O(n),R(n),E(n),K(n);const o=n?C(n):null;o&&o.toolNodes.forEach(t=>{t&&t.line&&n&&t.line.parentNode===n&&t.line.classList.remove("running")});const r=k.tabs.find(t=>t.id===e.sessionId);if(r&&(r.planCard=null),!n||!o){resetTurnText(n);return}if(e.success===!1){const t=e.error||(e.interrupted?d("web.turn.stopped"):d("web.turn.failed")),p=e.interrupted?"warnline":"errline",s=e.interrupted?"⚠ ":"✗ ";f(n,i("div",p,s+t)),a(t)}else{if(e.stopReason){const p=d("web.turn.stop_"+e.stopReason);f(n,i("div","warnline","⚠ "+p)),a(p)}const t=[];typeof e.toolCalls=="number"&&e.toolCalls>0&&t.push(d("web.turn.tools",{count:e.toolCalls})),typeof e.durationMs=="number"&&t.push(S(e.durationMs)),t.length&&f(n,i("div","metric","⟡ "+t.join(" · "))),a(d("web.turn.end",{count:e.toolCalls||0}))}const l=o.turnRoots.map(t=>t.innerText||t.textContent||"").join(`
 
-`);$(n,l,o.turnMd),resetTurnText(n)}export function resetTurnText(e){if(!e)return;const n=C(e);n.turnMd="",n.turnRoots=[]}export function onErrorEvent(e){if(!e||!e.error)return;const n=h(e.sessionId);n&&d(n,i("div","errline","✗ "+e.error)),a(String(e.error))}export function onCommandOutput(e,n){const o=h(n)||A;T(o),String(e).split(`
-`).forEach(s=>d(o,i("div","cmdout",s))),z()}export function onLog(e){if(!e||typeof e.text!="string"||!e.text||e.diagnostic&&b()!=="verbose")return;const n=h(e.sessionId)||j();if(!n)return;const o=e.level,s=o==="error"?"errline":o==="warn"?"warnline":"detail",l=o==="error"?"✗ ":o==="warn"?"⚠ ":"↳ ",t=i("div",s);t.textContent=l+e.text,d(n,t),(o==="error"||o==="warn")&&a(e.text)}
+`);$(n,l,o.turnMd),resetTurnText(n)}export function resetTurnText(e){if(!e)return;const n=C(e);n.turnMd="",n.turnRoots=[]}export function onErrorEvent(e){if(!e||!e.error)return;const n=h(e.sessionId);n&&f(n,i("div","errline","✗ "+e.error)),a(String(e.error))}export function onCommandOutput(e,n){const o=h(n)||_;T(o),String(e).split(`
+`).forEach(r=>f(o,i("div","cmdout",r))),z()}export function onLog(e){if(!e||typeof e.text!="string"||!e.text||e.diagnostic&&b()!=="verbose")return;const n=h(e.sessionId)||j();if(!n)return;const o=e.level,r=o==="error"?"errline":o==="warn"?"warnline":"detail",l=o==="error"?"✗ ":o==="warn"?"⚠ ":"↳ ",t=i("div",r);t.textContent=l+e.text,f(n,t),(o==="error"||o==="warn")&&a(e.text)}