feat: MistakeTracker + file-provenance ledger (v2.7.4)

Two native-inference hardening features from boocode_code_review_v2 §1 #12.

MistakeTracker: new pure mistake-tracker.ts tracks consecutive heterogeneous
tool failures (kinds surfaced per tool from tool-phase.ts). On 3 in a row the
turn loop soft-nudges (model-facing recovery guidance + mistake_recovery
sentinel + reset), then escalates to stopping the turn (cap-hit-style, Continue
affordance) on a re-trip. Complements doom-loop (identical repeats) + cap-hit.

File-provenance ledger: compaction.ts derives a deterministic ## Files Read list
from the head messages' read-tool calls and injects it into the rolling-summary
prompt so provenance survives compaction (no new table; read-only).

mistake_recovery sentinel: MessageMetadata arm (server + web) + MessageBubble
render branch. Built by 2 parallel agents. Server 545 tests passing (23 new);
build + web tsc clean. Native-inference only. Builds on v2.7.3.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -2,6 +2,14 @@
 
 All notable changes per release tag. Most recent on top, ordered by tag creation date (which matches the git history). Tag names follow `vMAJOR.MINOR.PATCH-slug` — the slug describes what shipped, so the tag name alone is enough to recall the batch.
 
+## v2.7.4-mistake-tracker-ledger — 2026-06-01
+
+Two native-inference hardening features from `boocode_code_review_v2.md` §1 #12 (cline, algorithm-reimplemented). **MistakeTracker:** complements the doom-loop guard (identical repeats) and cap-hit (budget) by catching a run of consecutive tool *failures*. A new pure `mistake-tracker.ts` tracks heterogeneous failure kinds (`zod_reject`/`tool_not_found`/`exec_error`/`api_error`/`permission_denied`, surfaced per tool from `tool-phase.ts`); after 3 consecutive failures the `turn.ts` loop does a **soft nudge** — injects model-facing recovery guidance into the next step + drops a `mistake_recovery` UI sentinel + resets — then **escalates** to stopping the turn (cap-hit-style, with a Continue affordance) if it re-trips without an intervening success, so heterogeneous failures can't burn the whole step budget. **File-provenance ledger:** `compaction.ts` now derives a deterministic, sorted `## Files Read` list from the head messages' read-tool calls (`view_file`/`grep`/`find_files`/`list_dir`) and injects it into the rolling-summary prompt so file provenance survives compaction (no new table; prompt-driven merge, read-only since BooChat has no write tools). The `mistake_recovery` sentinel adds an arm to `MessageMetadata` in both server + web type copies plus a `MessageBubble` render branch. Built by two parallel agents (backend + frontend sentinel) over disjoint apps; server 545 tests passing (23 new: 12 mistake-tracker + 11 compaction), build + web tsc clean. Native-inference only (external agents run their own loops). Builds on `v2.7.3-sampling-streamjson-tokens`; openspec `mistake-tracker-file-ledger`.
+
+## v2.7.3-sampling-streamjson-tokens — 2026-06-01
+
+Three small BooCode wins from `boocode_code_review_v2.md` §1 #11/#7/#8. **Sampling knobs:** per-agent `top_n_sigma` + the `dry_*` repetition family (`dry_multiplier`/`dry_base`/`dry_allowed_length`/`dry_penalty_last_n`) are now first-class Agent frontmatter fields, parsed in `agents.ts` and threaded into the llama-swap chat-completion body via `providerOptions.openaiCompatible` (the `@ai-sdk/openai-compatible` extra-body channel). This surfaced and fixed a **latent bug**: `top_k` (rejected by the AI-SDK provider as unsupported) and `min_p` (never passed to `streamText` at all) had been dead on the wire — no agent's `top_k`/`min_p` ever affected sampling; both now route through the same channel, so agents that set them will start using them. `--reasoning-budget` is documented in `data/AGENTS.md` (already works via `llama_extra_args`, permitted by the deny-list validator). **Live PTY stream-json:** qwen/claude PTY dispatch sliced stdout opaque; a new `stream-json-parser.ts` line-buffers the Claude-Code-compatible NDJSON and emits text/reasoning/tool frames live as they arrive (mirroring the ACP/opencode paths) + persists the structured parts, with a clean fallback to the old opaque slice when output isn't NDJSON (claude now runs `--output-format stream-json --verbose`). **Token UI:** the per-`(chat,agent)` `agent_sessions.input_tokens`/`output_tokens`/`cost` columns (accumulated since `v2.6.8` but dropped by the read route + wire type) now flow through and render condensed beside the AgentComposerBar session chip. Built by three parallel agents over disjoint subsystems; server 523 + coder 245 tests passing (incl. 11 new stream-json-parser + new agent-parse tests), all builds + web tsc clean. Builds on `v2.7.2-checkpoint-idor`; openspec `sampling-streamjson-tokens`. The qwen-vs-claude `usage` field names in #7 are best-guess pending a live smoke.
+
 ## v2.7.2-checkpoint-idor — 2026-06-01
 
 Closes two IDOR authorization holes in the `v2.7.1-write-edit-robustness` checkpoint routes, flagged by the automated push security review. The `GET /api/sessions/:id/checkpoints?chat_id=` list route scoped its `chat_id` branch by `chat_id` alone — any session's `chat_id` would read its checkpoints; it now joins through `chats` and gates on `chats.session_id` (authoritative; `checkpoints.session_id` is a nullable denormalized hint). The `restoreCheckpoint` scope guard was fail-open — `cp.session_id && cp.session_id !== sessionId` fell through whenever the checkpoint's denormalized `session_id` was null, allowing a cross-session restore (worktree reset + transcript trim) — it now resolves the owning session via the checkpoint's chat and denies on any missing-or-mismatched row. A DB-integration regression covers the exact null-`session_id` cross-session case. Real-world blast radius is small (BooCoder is single-user behind Authelia on loopback), but both are genuine authorization bugs. Coder suite 234 passing (7/7 checkpoint tests incl. the regression against live postgres+git), typecheck clean. Hotfix on `v2.7.1-write-edit-robustness`.

apps/coder/src/routes/agent-sessions.ts

@@ -16,6 +16,11 @@ export interface AgentSessionRow {
   status: string;
   has_session: boolean;
   last_active_at: string | null;
+  // v2.6.8 per-(chat,agent) running token/cost totals (sampling-streamjson-tokens
+  // #8). BIGINT columns arrive as strings over the wire; the frontend coerces.
+  input_tokens: number;
+  output_tokens: number;
+  cost: number;
 }
 
 export function registerAgentSessionRoutes(app: FastifyInstance, sql: Sql): void {
@@ -39,7 +44,10 @@ export function registerAgentSessionRoutes(app: FastifyInstance, sql: Sql): void
           a.agent AS agent,
           a.status AS status,
           (a.agent_session_id IS NOT NULL) AS has_session,
-          a.last_active_at AS last_active_at
+          a.last_active_at AS last_active_at,
+          a.input_tokens AS input_tokens,
+          a.output_tokens AS output_tokens,
+          a.cost AS cost
         FROM agent_sessions a
         JOIN chats c ON c.id = a.chat_id
         WHERE c.session_id = ${sessionId}

apps/coder/src/services/__tests__/stream-json-parser.test.ts

@@ -0,0 +1,189 @@
+import { describe, it, expect } from 'vitest';
+import {
+  makeStreamJsonParser,
+  makeStreamJsonState,
+  parseStreamJsonLine,
+  type AgentEventList,
+} from '../stream-json-parser.js';
+import type { AgentEvent } from '../agent-backend.js';
+import type { AcpToolSnapshot } from '../acp-tool-snapshot.js';
+
+// Helpers to JSON-encode the representative Claude-Code stream-json lines.
+const sys = (sessionId: string) =>
+  JSON.stringify({ type: 'system', subtype: 'init', session_id: sessionId, tools: ['read', 'edit'] });
+
+const streamEvent = (event: unknown) => JSON.stringify({ type: 'stream_event', event });
+
+const textDelta = (index: number, text: string) =>
+  streamEvent({ type: 'content_block_delta', index, delta: { type: 'text_delta', text } });
+
+const thinkingDelta = (index: number, thinking: string) =>
+  streamEvent({ type: 'content_block_delta', index, delta: { type: 'thinking_delta', thinking } });
+
+const toolStart = (index: number, id: string, name: string) =>
+  streamEvent({ type: 'content_block_start', index, content_block: { type: 'tool_use', id, name } });
+
+const inputJsonDelta = (index: number, partial: string) =>
+  streamEvent({ type: 'content_block_delta', index, delta: { type: 'input_json_delta', partial_json: partial } });
+
+const blockStop = (index: number) => streamEvent({ type: 'content_block_stop', index });
+
+const resultLine = (input: number, output: number, sessionId?: string) =>
+  JSON.stringify({ type: 'result', subtype: 'success', session_id: sessionId, usage: { input_tokens: input, output_tokens: output } });
+
+describe('parseStreamJsonLine (pure per-line mapping)', () => {
+  it('captures session_id from the system init line and emits no events', () => {
+    const state = makeStreamJsonState();
+    const events = parseStreamJsonLine(sys('sess-abc'), state);
+    expect(events).toEqual([]);
+    expect(state.sessionId).toBe('sess-abc');
+  });
+
+  it('maps a text_delta stream_event → a text event', () => {
+    const state = makeStreamJsonState();
+    expect(parseStreamJsonLine(textDelta(0, 'Hello'), state)).toEqual([{ type: 'text', text: 'Hello' }]);
+  });
+
+  it('maps a thinking_delta stream_event → a reasoning event', () => {
+    const state = makeStreamJsonState();
+    expect(parseStreamJsonLine(thinkingDelta(0, 'pondering'), state)).toEqual([
+      { type: 'reasoning', text: 'pondering' },
+    ]);
+  });
+
+  it('tolerates a garbage / non-JSON line (returns [], no throw)', () => {
+    const state = makeStreamJsonState();
+    expect(parseStreamJsonLine('not json at all {{{', state)).toEqual([]);
+    expect(parseStreamJsonLine('', state)).toEqual([]);
+    expect(parseStreamJsonLine('   ', state)).toEqual([]);
+    // A truncated/partial JSON object also yields [] rather than throwing.
+    expect(parseStreamJsonLine('{"type":"stream_event","eve', state)).toEqual([]);
+  });
+
+  it('ignores unknown top-level line types and the user (tool-result) line', () => {
+    const state = makeStreamJsonState();
+    expect(parseStreamJsonLine(JSON.stringify({ type: 'user', message: {} }), state)).toEqual([]);
+    expect(parseStreamJsonLine(JSON.stringify({ type: 'whatever' }), state)).toEqual([]);
+  });
+
+  it('assembles a tool call across input_json_delta chunks (split across lines)', () => {
+    const state = makeStreamJsonState();
+    // start → tool_call (running, empty args)
+    const start = parseStreamJsonLine(toolStart(1, 'toolu_1', 'edit_file'), state);
+    expect(start).toHaveLength(1);
+    expect(start[0]!.type).toBe('tool_call');
+    const startSnap = (start[0] as { type: 'tool_call'; toolCall: AcpToolSnapshot }).toolCall;
+    expect(startSnap.toolCallId).toBe('toolu_1');
+    expect(startSnap.title).toBe('edit_file');
+    expect(startSnap.status).toBe('in_progress');
+    expect(startSnap.rawInput).toEqual({});
+
+    // args streamed in fragments — no events until stop
+    expect(parseStreamJsonLine(inputJsonDelta(1, '{"path":"a'), state)).toEqual([]);
+    expect(parseStreamJsonLine(inputJsonDelta(1, '.ts","content":'), state)).toEqual([]);
+    expect(parseStreamJsonLine(inputJsonDelta(1, '"hi"}'), state)).toEqual([]);
+
+    // stop → tool_update with the parsed, fully-assembled input
+    const stop = parseStreamJsonLine(blockStop(1), state);
+    expect(stop).toHaveLength(1);
+    expect(stop[0]!.type).toBe('tool_update');
+    const stopSnap = (stop[0] as { type: 'tool_update'; toolCall: AcpToolSnapshot }).toolCall;
+    expect(stopSnap.toolCallId).toBe('toolu_1');
+    expect(stopSnap.status).toBe('completed');
+    expect(stopSnap.rawInput).toEqual({ path: 'a.ts', content: 'hi' });
+  });
+
+  it('falls back to {_raw} when accumulated tool args are not valid JSON', () => {
+    const state = makeStreamJsonState();
+    parseStreamJsonLine(toolStart(0, 'toolu_x', 'run'), state);
+    parseStreamJsonLine(inputJsonDelta(0, '{"broken'), state);
+    const stop = parseStreamJsonLine(blockStop(0), state);
+    const snap = (stop[0] as { type: 'tool_update'; toolCall: AcpToolSnapshot }).toolCall;
+    expect(snap.rawInput).toEqual({ _raw: '{"broken' });
+  });
+
+  it('captures usage from message_delta and result lines', () => {
+    const state = makeStreamJsonState();
+    parseStreamJsonLine(streamEvent({ type: 'message_delta', usage: { output_tokens: 42 } }), state);
+    expect(state.usage.outputTokens).toBe(42);
+    parseStreamJsonLine(resultLine(100, 250, 'sess-z'), state);
+    expect(state.usage.inputTokens).toBe(100);
+    expect(state.usage.outputTokens).toBe(250);
+    expect(state.sessionId).toBe('sess-z');
+  });
+
+  it('maps a terminal assistant message (fallback) → text + reasoning + tool events', () => {
+    const state = makeStreamJsonState();
+    const line = JSON.stringify({
+      type: 'assistant',
+      session_id: 'sess-asst',
+      message: {
+        content: [
+          { type: 'thinking', thinking: 'let me think' },
+          { type: 'text', text: 'Here is the answer' },
+          { type: 'tool_use', id: 'toolu_9', name: 'view_file', input: { path: 'x.ts' } },
+        ],
+        usage: { input_tokens: 5, output_tokens: 7 },
+      },
+    });
+    const events = parseStreamJsonLine(line, state);
+    expect(events).toEqual([
+      { type: 'reasoning', text: 'let me think' },
+      { type: 'text', text: 'Here is the answer' },
+      {
+        type: 'tool_update',
+        toolCall: { toolCallId: 'toolu_9', title: 'view_file', kind: null, status: 'completed', rawInput: { path: 'x.ts' } },
+      },
+    ]);
+    expect(state.usage).toEqual({ inputTokens: 5, outputTokens: 7 });
+    expect(state.sessionId).toBe('sess-asst');
+  });
+});
+
+describe('makeStreamJsonParser (stateful wrapper over a full turn)', () => {
+  it('streams a representative turn: init → text → thinking → tool → result', () => {
+    const parser = makeStreamJsonParser();
+    const all: AgentEvent[] = [];
+    const feed = (line: string): AgentEventList => {
+      const evs = parser.push(line);
+      all.push(...evs);
+      return evs;
+    };
+
+    feed(sys('sess-1'));
+    feed(textDelta(0, 'Reading '));
+    feed(textDelta(0, 'the file. '));
+    feed(thinkingDelta(0, 'I should edit it'));
+    feed(toolStart(1, 'toolu_a', 'edit_file'));
+    feed(inputJsonDelta(1, '{"path":'));
+    feed(inputJsonDelta(1, '"main.ts"}'));
+    feed(blockStop(1));
+    feed(textDelta(0, 'Done.'));
+    feed(resultLine(120, 80, 'sess-1'));
+
+    expect(all).toEqual([
+      { type: 'text', text: 'Reading ' },
+      { type: 'text', text: 'the file. ' },
+      { type: 'reasoning', text: 'I should edit it' },
+      {
+        type: 'tool_call',
+        toolCall: { toolCallId: 'toolu_a', title: 'edit_file', kind: null, status: 'in_progress', rawInput: {} },
+      },
+      {
+        type: 'tool_update',
+        toolCall: { toolCallId: 'toolu_a', title: 'edit_file', kind: null, status: 'completed', rawInput: { path: 'main.ts' } },
+      },
+      { type: 'text', text: 'Done.' },
+    ]);
+
+    expect(parser.usage()).toEqual({ inputTokens: 120, outputTokens: 80 });
+    expect(parser.sessionId()).toBe('sess-1');
+  });
+
+  it('a garbage line interleaved mid-turn does not derail subsequent parsing', () => {
+    const parser = makeStreamJsonParser();
+    expect(parser.push(textDelta(0, 'a'))).toEqual([{ type: 'text', text: 'a' }]);
+    expect(parser.push('>>> not json <<<')).toEqual([]);
+    expect(parser.push(textDelta(0, 'b'))).toEqual([{ type: 'text', text: 'b' }]);
+  });
+});

apps/coder/src/services/dispatcher.ts

@@ -410,6 +410,52 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
         outputSummary = result.output.slice(0, 500);
         await persistExternalAgentTurn(sql, assistantId, result.toolSnapshots, acpReasoning);
       } else {
+        // v#7 (stream-json): claude + qwen run with --output-format stream-json.
+        // Parse the NDJSON live in pty-dispatch and forward AgentEvents here so we
+        // publish the SAME live frames the warm-ACP / opencode paths emit (text,
+        // reasoning, tool) and persist structured parts. Accumulate for the final
+        // message content + persistence; fall back to the opaque stdout slice when
+        // nothing parsed (agent ran without the flag, or crashed before emitting).
+        const ptyTextChunks: string[] = [];
+        const ptyReasoningChunks: string[] = [];
+        const ptyToolSnaps = new Map<string, AcpToolSnapshot>();
+
+        const onPtyEvent = (e: AgentEvent): void => {
+          switch (e.type) {
+            case 'text':
+              ptyTextChunks.push(e.text);
+              broker.publishFrame(sessionId, {
+                type: 'delta',
+                message_id: assistantId,
+                chat_id: chatId,
+                content: e.text,
+              } as WsFrame);
+              break;
+            case 'reasoning':
+              ptyReasoningChunks.push(e.text);
+              broker.publishFrame(sessionId, {
+                type: 'reasoning_delta',
+                message_id: assistantId,
+                chat_id: chatId,
+                content: e.text,
+              } as WsFrame);
+              break;
+            case 'tool_call':
+            case 'tool_update':
+              ptyToolSnaps.set(e.toolCall.toolCallId, e.toolCall);
+              broker.publishFrame(sessionId, {
+                type: 'tool_call',
+                message_id: assistantId,
+                chat_id: chatId,
+                tool_call: snapshotToWireToolCall(e.toolCall),
+              } as WsFrame);
+              break;
+            case 'commands':
+              // stream-json carries no commands today; ignore if it ever does.
+              break;
+          }
+        };
+
         const result = await dispatchViaPty({
           agent,
           task: task.input,
@@ -420,17 +466,33 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
           thinkingOptionId: task.thinking_option_id ?? undefined,
           signal: ac.signal,
           log,
+          onEvent: onPtyEvent,
         });
-        assistantContent = (result.stdout || result.stderr || '(no output)').slice(0, 50_000);
-        outputSummary = (result.stdout || result.stderr).slice(0, 500);
 
-        if (assistantContent) {
+        if (result.streamed) {
-          broker.publishFrame(sessionId, {
+          assistantContent = ptyTextChunks.join('').slice(0, 50_000);
-            type: 'delta',
+          // stream-json text can be empty for a tool-only turn — surface stderr or a
-            message_id: assistantId,
+          // placeholder so the message row isn't blank.
-            chat_id: chatId,
+          if (!assistantContent) {
-            content: assistantContent,
+            assistantContent = (result.stderr || '(no text output)').slice(0, 50_000);
-          } as WsFrame);
+          }
+          outputSummary = (ptyTextChunks.join('') || result.stderr).slice(0, 500);
+          acpReasoning = ptyReasoningChunks.join('').slice(0, 200_000);
+          await persistExternalAgentTurn(sql, assistantId, [...ptyToolSnaps.values()], acpReasoning);
+        } else {
+          // Fallback: agent produced no parseable NDJSON (ran without the flag, or
+          // crashed). Preserve today's opaque stdout-slice + single delta behavior.
+          assistantContent = (result.stdout || result.stderr || '(no output)').slice(0, 50_000);
+          outputSummary = (result.stdout || result.stderr).slice(0, 500);
+
+          if (assistantContent) {
+            broker.publishFrame(sessionId, {
+              type: 'delta',
+              message_id: assistantId,
+              chat_id: chatId,
+              content: assistantContent,
+            } as WsFrame);
+          }
         }
       }
 

apps/coder/src/services/pty-dispatch.ts

@@ -1,13 +1,29 @@
 /**
  * PTY dispatch — runs external agents directly on the host.
+ *
+ * claude + qwen run with `--output-format stream-json` and emit Claude-Code's
+ * stream-json NDJSON on stdout. When an `onEvent` callback is supplied we
+ * line-buffer that stdout (split on `\n`, hold the partial tail) and feed complete
+ * lines to `makeStreamJsonParser` so deltas surface live as AgentEvents. The raw
+ * stdout is still accumulated + returned for back-compat (and the dispatcher's
+ * fallback when nothing parsed). See `stream-json-parser.ts`.
  */
 import type { FastifyBaseLogger } from 'fastify';
 import { spawn } from 'node:child_process';
+import type { AgentEvent } from './agent-backend.js';
+import { makeStreamJsonParser, type StreamJsonUsage } from './stream-json-parser.js';
 
 export interface DispatchResult {
   exitCode: number;
   stdout: string;
   stderr: string;
+  /** True iff at least one NDJSON AgentEvent was parsed from stdout (v#7). When
+   *  false the dispatcher falls back to slicing stdout as the assistant content. */
+  streamed: boolean;
+  /** Final usage parsed from the stream-json `result` / `message_delta`, if any. */
+  usage?: StreamJsonUsage;
+  /** Provider session id from the stream-json `system` init line, if any. */
+  agentSessionId?: string | null;
 }
 
 export interface PtyDispatchOpts {
@@ -20,6 +36,10 @@ export interface PtyDispatchOpts {
   installPath?: string;
   signal?: AbortSignal;
   log: FastifyBaseLogger;
+  /** Optional live event sink. When set, stdout is line-buffered + NDJSON-parsed
+   *  and each AgentEvent is forwarded here as it arrives. Absent → opaque (old)
+   *  behavior: stdout is accumulated and returned, no parsing. */
+  onEvent?: (e: AgentEvent) => void;
 }
 
 interface PtySpawnSpec {
@@ -40,7 +60,9 @@ function buildPtySpawnSpec(
 
   switch (agent) {
     case 'claude': {
-      const args = ['-p'];
+      // stream-json on -p requires --verbose (Claude Code rejects stream-json
+      // print mode without it). qwen needs no such flag.
+      const args = ['-p', '--output-format', 'stream-json', '--verbose'];
       if (model) args.push('--model', model);
       if (modeId) args.push('--permission-mode', modeId);
       if (thinkingOptionId) args.push('--effort', thinkingOptionId);
@@ -73,7 +95,7 @@ function buildPtySpawnSpec(
 }
 
 export async function dispatchViaPty(opts: PtyDispatchOpts): Promise<DispatchResult> {
-  const { agent, task, worktreePath, model, modeId, thinkingOptionId, installPath, signal, log } = opts;
+  const { agent, task, worktreePath, model, modeId, thinkingOptionId, installPath, signal, log, onEvent } = opts;
 
   const cmd = buildPtySpawnSpec(agent, task, model, modeId, thinkingOptionId, installPath);
   if (!cmd) {
@@ -81,6 +103,7 @@ export async function dispatchViaPty(opts: PtyDispatchOpts): Promise<DispatchRes
       exitCode: 1,
       stdout: '',
       stderr: `Agent '${agent}' is not yet supported for PTY dispatch.`,
+      streamed: false,
     };
   }
 
@@ -102,7 +125,32 @@ export async function dispatchViaPty(opts: PtyDispatchOpts): Promise<DispatchRes
     let stderr = '';
     let killed = false;
 
-    child.stdout!.on('data', (chunk: Buffer) => { stdout += chunk.toString(); });
+    // Live NDJSON parsing (only when a sink is supplied). Line-buffer: split on
+    // '\n', dispatch complete lines, hold the partial tail until the next chunk.
+    const parser = onEvent ? makeStreamJsonParser() : null;
+    let lineBuf = '';
+    let streamed = false;
+    const feedLine = (line: string): void => {
+      if (!parser || !onEvent) return;
+      for (const e of parser.push(line)) {
+        streamed = true;
+        onEvent(e);
+      }
+    };
+
+    child.stdout!.on('data', (chunk: Buffer) => {
+      const text = chunk.toString();
+      stdout += text;
+      if (!parser) return;
+      lineBuf += text;
+      let nl = lineBuf.indexOf('\n');
+      while (nl !== -1) {
+        const line = lineBuf.slice(0, nl);
+        lineBuf = lineBuf.slice(nl + 1);
+        feedLine(line);
+        nl = lineBuf.indexOf('\n');
+      }
+    });
     child.stderr!.on('data', (chunk: Buffer) => { stderr += chunk.toString(); });
 
     const cleanup = () => {
@@ -116,7 +164,7 @@ export async function dispatchViaPty(opts: PtyDispatchOpts): Promise<DispatchRes
     if (signal) {
       if (signal.aborted) {
         cleanup();
-        resolve({ exitCode: 130, stdout: '', stderr: 'Aborted before start' });
+        resolve({ exitCode: 130, stdout: '', stderr: 'Aborted before start', streamed: false });
         return;
       }
       signal.addEventListener('abort', cleanup, { once: true });
@@ -124,8 +172,18 @@ export async function dispatchViaPty(opts: PtyDispatchOpts): Promise<DispatchRes
 
     child.on('close', (code) => {
       if (signal) signal.removeEventListener('abort', cleanup);
-      log.info({ agent, exitCode: code }, 'pty-dispatch: completed');
+      // Flush any final line with no trailing newline.
-      resolve({ exitCode: code ?? 1, stdout, stderr });
+      if (lineBuf.trim()) feedLine(lineBuf);
+      lineBuf = '';
+      log.info({ agent, exitCode: code, streamed }, 'pty-dispatch: completed');
+      resolve({
+        exitCode: code ?? 1,
+        stdout,
+        stderr,
+        streamed,
+        usage: parser?.usage(),
+        agentSessionId: parser?.sessionId() ?? null,
+      });
     });
 
     child.on('error', (err) => {

apps/coder/src/services/stream-json-parser.ts

@@ -0,0 +1,296 @@
+/**
+ * Claude-Code-compatible stream-json NDJSON parser (feature #7,
+ * openspec `sampling-streamjson-tokens`).
+ *
+ * qwen (`--output-format stream-json`) and claude (`--output-format stream-json`)
+ * both emit Claude-Code's stream-json NDJSON on stdout: one JSON object per line.
+ * This module turns that stream into the same transport-agnostic `AgentEvent`s the
+ * ACP / opencode-server backends emit, so the PTY dispatch path can publish live
+ * broker frames + persist structured parts instead of slicing stdout opaque.
+ *
+ * Two surfaces:
+ *  - `parseStreamJsonLine(line, state)` — PURE per-line mapping (unit-testable).
+ *    `state` is the caller-owned accumulator (open tool blocks + usage/session_id).
+ *  - `makeStreamJsonParser()` — a thin stateful wrapper holding the state, with a
+ *    `push(line)` that returns the events for that line and getters for the final
+ *    `usage` / `sessionId`.
+ *
+ * Defensive by contract: a non-JSON / partial / garbage line yields `[]` and never
+ * throws. Tool args (`input_json_delta`) arrive fragmented across many lines; we
+ * accumulate the partial JSON string per content-block index and only surface the
+ * parsed `rawInput` once the block stops (or, as a fallback, off the terminal
+ * `assistant` message which carries the fully-assembled `tool_use` blocks).
+ *
+ * Schema (keyed on top-level `type`):
+ *  - `system`     — init: { session_id, tools, ... }
+ *  - `assistant`  — { message: { content: [ {type:'text'|'thinking'|'tool_use', ...} ], usage? } }
+ *  - `user`       — tool results (ignored — diffing the worktree captures effects)
+ *  - `result`     — final: { usage: { input_tokens, output_tokens }, session_id? }
+ *  - `stream_event` — { event: { type, index?, content_block?, delta?, usage? } }
+ *      event.type:
+ *        content_block_start  — { index, content_block: {type, id?, name?} }
+ *        content_block_delta  — { index, delta: {type, text?|thinking?|partial_json?} }
+ *        content_block_stop   — { index }
+ *        message_delta        — { usage: { output_tokens } }
+ *        message_start        — { message: { usage } }
+ */
+import type { AgentEvent } from './agent-backend.js';
+import type { AcpToolSnapshot } from './acp-tool-snapshot.js';
+
+/** Convenience alias for the per-line return value. */
+export type AgentEventList = AgentEvent[];
+
+export interface StreamJsonUsage {
+  inputTokens?: number;
+  outputTokens?: number;
+}
+
+/** Per-open-content-block accumulation for tool args assembled across deltas. */
+interface OpenToolBlock {
+  toolCallId: string;
+  name: string;
+  /** Concatenated `input_json_delta.partial_json` fragments. */
+  partialJson: string;
+}
+
+export interface StreamJsonState {
+  /** content-block index → open tool block (only `tool_use` blocks are tracked). */
+  toolBlocks: Map<number, OpenToolBlock>;
+  sessionId: string | null;
+  usage: StreamJsonUsage;
+}
+
+export function makeStreamJsonState(): StreamJsonState {
+  return { toolBlocks: new Map(), sessionId: null, usage: {} };
+}
+
+function asRecord(value: unknown): Record<string, unknown> | null {
+  if (value && typeof value === 'object' && !Array.isArray(value)) {
+    return value as Record<string, unknown>;
+  }
+  return null;
+}
+
+function asString(value: unknown): string | undefined {
+  return typeof value === 'string' ? value : undefined;
+}
+
+function asNumber(value: unknown): number | undefined {
+  return typeof value === 'number' && Number.isFinite(value) ? value : undefined;
+}
+
+/** Pull token counts out of an Anthropic-shape `usage` object, mutating state. */
+function captureUsage(usage: Record<string, unknown> | null, state: StreamJsonState): void {
+  if (!usage) return;
+  const input = asNumber(usage.input_tokens);
+  const output = asNumber(usage.output_tokens);
+  if (input !== undefined) state.usage.inputTokens = input;
+  // output_tokens is reported incrementally on message_delta; keep the latest.
+  if (output !== undefined) state.usage.outputTokens = output;
+}
+
+/** Parse the accumulated tool-arg JSON; tolerate an unparseable/partial body. */
+function parseToolInput(partialJson: string): unknown {
+  const trimmed = partialJson.trim();
+  if (!trimmed) return {};
+  try {
+    return JSON.parse(trimmed);
+  } catch {
+    return { _raw: partialJson };
+  }
+}
+
+function toolSnapshot(block: OpenToolBlock, rawInput: unknown, status: AcpToolSnapshot['status']): AcpToolSnapshot {
+  return {
+    toolCallId: block.toolCallId,
+    title: block.name,
+    kind: null,
+    status,
+    rawInput,
+  };
+}
+
+/**
+ * Map one stream-event sub-object (the `event` field of a `stream_event` line) to
+ * AgentEvents, mutating `state` for open tool blocks + usage.
+ */
+function handleStreamEvent(event: Record<string, unknown>, state: StreamJsonState): AgentEvent[] {
+  const eventType = asString(event.type);
+  if (!eventType) return [];
+
+  switch (eventType) {
+    case 'content_block_start': {
+      const index = asNumber(event.index);
+      const block = asRecord(event.content_block);
+      if (index === undefined || !block) return [];
+      if (asString(block.type) !== 'tool_use') return [];
+      const toolCallId = asString(block.id) ?? `tool_${index}`;
+      const name = asString(block.name) ?? 'tool';
+      const open: OpenToolBlock = { toolCallId, name, partialJson: '' };
+      state.toolBlocks.set(index, open);
+      // Surface the tool start immediately (running, no args yet) so the UI shows
+      // the call before the args finish streaming.
+      return [{ type: 'tool_call', toolCall: toolSnapshot(open, {}, 'in_progress') }];
+    }
+
+    case 'content_block_delta': {
+      const index = asNumber(event.index);
+      const delta = asRecord(event.delta);
+      if (delta === null) return [];
+      const deltaType = asString(delta.type);
+      if (deltaType === 'text_delta') {
+        const text = asString(delta.text);
+        return text ? [{ type: 'text', text }] : [];
+      }
+      if (deltaType === 'thinking_delta') {
+        const text = asString(delta.thinking);
+        return text ? [{ type: 'reasoning', text }] : [];
+      }
+      if (deltaType === 'input_json_delta') {
+        // Accumulate tool args; no event until the block stops.
+        const fragment = asString(delta.partial_json);
+        if (index !== undefined && fragment) {
+          const open = state.toolBlocks.get(index);
+          if (open) open.partialJson += fragment;
+        }
+        return [];
+      }
+      return [];
+    }
+
+    case 'content_block_stop': {
+      const index = asNumber(event.index);
+      if (index === undefined) return [];
+      const open = state.toolBlocks.get(index);
+      if (!open) return [];
+      state.toolBlocks.delete(index);
+      const rawInput = parseToolInput(open.partialJson);
+      return [{ type: 'tool_update', toolCall: toolSnapshot(open, rawInput, 'completed') }];
+    }
+
+    case 'message_start': {
+      const message = asRecord(event.message);
+      captureUsage(asRecord(message?.usage), state);
+      return [];
+    }
+
+    case 'message_delta': {
+      captureUsage(asRecord(event.usage), state);
+      return [];
+    }
+
+    default:
+      return [];
+  }
+}
+
+/**
+ * Map the terminal `assistant` message (post-hoc full message) to AgentEvents. Used
+ * as a fallback for transports that emit only the assembled `assistant` line and no
+ * incremental `stream_event`s. When stream_events already streamed a block, the
+ * caller dedups by toolCallId, so re-emitting the assembled tool_use is harmless.
+ */
+function handleAssistantMessage(message: Record<string, unknown>, state: StreamJsonState): AgentEvent[] {
+  captureUsage(asRecord(message.usage), state);
+  const content = message.content;
+  if (!Array.isArray(content)) return [];
+  const out: AgentEvent[] = [];
+  let toolIdx = 0;
+  for (const rawBlock of content) {
+    const block = asRecord(rawBlock);
+    if (!block) continue;
+    const blockType = asString(block.type);
+    if (blockType === 'text') {
+      const text = asString(block.text);
+      if (text) out.push({ type: 'text', text });
+    } else if (blockType === 'thinking') {
+      const text = asString(block.thinking);
+      if (text) out.push({ type: 'reasoning', text });
+    } else if (blockType === 'tool_use') {
+      const toolCallId = asString(block.id) ?? `tool_${toolIdx}`;
+      const name = asString(block.name) ?? 'tool';
+      const rawInput = 'input' in block ? block.input : {};
+      out.push({
+        type: 'tool_update',
+        toolCall: { toolCallId, title: name, kind: null, status: 'completed', rawInput },
+      });
+    }
+    toolIdx++;
+  }
+  return out;
+}
+
+/**
+ * Pure per-line mapping. `line` is a single complete NDJSON line (no trailing
+ * newline required; surrounding whitespace tolerated). Returns the AgentEvents the
+ * line produces and mutates `state` (open tool blocks, usage, session_id). A blank,
+ * non-JSON, or unrecognized line yields `[]` and never throws.
+ */
+export function parseStreamJsonLine(line: string, state: StreamJsonState): AgentEvent[] {
+  const trimmed = line.trim();
+  if (!trimmed) return [];
+
+  let obj: Record<string, unknown> | null;
+  try {
+    const parsed: unknown = JSON.parse(trimmed);
+    obj = asRecord(parsed);
+  } catch {
+    return [];
+  }
+  if (!obj) return [];
+
+  const type = asString(obj.type);
+  switch (type) {
+    case 'system': {
+      const sid = asString(obj.session_id);
+      if (sid) state.sessionId = sid;
+      return [];
+    }
+
+    case 'stream_event': {
+      const event = asRecord(obj.event);
+      return event ? handleStreamEvent(event, state) : [];
+    }
+
+    case 'assistant': {
+      const sid = asString(obj.session_id);
+      if (sid) state.sessionId = sid;
+      const message = asRecord(obj.message);
+      return message ? handleAssistantMessage(message, state) : [];
+    }
+
+    case 'result': {
+      const sid = asString(obj.session_id);
+      if (sid) state.sessionId = sid;
+      captureUsage(asRecord(obj.usage), state);
+      return [];
+    }
+
+    default:
+      // `user` (tool results) and any unknown line type — ignore.
+      return [];
+  }
+}
+
+export interface StreamJsonParser {
+  /** Feed one complete NDJSON line; returns its AgentEvents (never throws). */
+  push(line: string): AgentEvent[];
+  /** Final usage (input/output tokens) accumulated so far. */
+  usage(): StreamJsonUsage;
+  /** Provider session id from the init `system` line / `result`, if seen. */
+  sessionId(): string | null;
+}
+
+/**
+ * Stateful wrapper around `parseStreamJsonLine`. Holds per-tool-block accumulation
+ * + usage/session_id across the turn. Line-buffering (splitting stdout on `\n` and
+ * holding the partial tail) is the caller's job — see `pty-dispatch.ts`.
+ */
+export function makeStreamJsonParser(): StreamJsonParser {
+  const state = makeStreamJsonState();
+  return {
+    push: (line: string) => parseStreamJsonLine(line, state),
+    usage: () => ({ ...state.usage }),
+    sessionId: () => state.sessionId,
+  };
+}

apps/server/src/services/__tests__/agents.test.ts

@@ -1,4 +1,4 @@
-import { describe, it, expect } from 'vitest';
+import { describe, it, expect, vi, afterEach } from 'vitest';
 import { isAgentRegistryMarkdown, parseAgentsMd } from '../agents.js';
 
 describe('isAgentRegistryMarkdown', () => {
@@ -31,3 +31,87 @@ Start here
     expect(r.errors.length).toBeGreaterThan(0);
   });
 });
+
+// v2.6 sampling-streamjson-tokens (#11): per-agent llama.cpp sampler extensions.
+describe('parseAgentsMd: v2.6 sampling knobs', () => {
+  afterEach(() => {
+    vi.restoreAllMocks();
+  });
+
+  const withFrontmatter = (lines: string) => `# Agents
+
+## Sampler
+---
+temperature: 0.6
+${lines}
+tools: [view_file]
+description: test
+---
+You sample.
+`;
+
+  it('parses top_n_sigma and the dry_* family from frontmatter', () => {
+    const md = withFrontmatter(
+      [
+        'top_n_sigma: 1.5',
+        'dry_multiplier: 0.8',
+        'dry_base: 1.75',
+        'dry_allowed_length: 2',
+        'dry_penalty_last_n: -1',
+      ].join('\n'),
+    );
+    const { agents, errors } = parseAgentsMd(md);
+    expect(errors).toHaveLength(0);
+    expect(agents).toHaveLength(1);
+    const a = agents[0]!;
+    expect(a.top_n_sigma).toBe(1.5);
+    expect(a.dry_multiplier).toBe(0.8);
+    expect(a.dry_base).toBe(1.75);
+    expect(a.dry_allowed_length).toBe(2);
+    expect(a.dry_penalty_last_n).toBe(-1);
+  });
+
+  it('defaults the new sampler fields to null when omitted', () => {
+    const { agents } = parseAgentsMd(withFrontmatter('top_p: 0.95'));
+    const a = agents[0]!;
+    expect(a.top_n_sigma).toBeNull();
+    expect(a.dry_multiplier).toBeNull();
+    expect(a.dry_base).toBeNull();
+    expect(a.dry_allowed_length).toBeNull();
+    expect(a.dry_penalty_last_n).toBeNull();
+  });
+
+  it('warns (does not error) on out-of-range top_n_sigma / dry_* values', () => {
+    const warn = vi.spyOn(console, 'warn').mockImplementation(() => {});
+    const md = withFrontmatter(
+      [
+        'top_n_sigma: -1',
+        'dry_multiplier: -0.5',
+        'dry_base: -2',
+        'dry_allowed_length: -3',
+        'dry_penalty_last_n: -5',
+      ].join('\n'),
+    );
+    const { agents, errors } = parseAgentsMd(md);
+    expect(errors).toHaveLength(0);
+    expect(agents).toHaveLength(1);
+    // Mirrors top_k/min_p: out-of-range still stored, with a warning.
+    expect(warn).toHaveBeenCalled();
+    const warnings = warn.mock.calls.map((c) => String(c[0])).join('\n');
+    expect(warnings).toContain('top_n_sigma');
+    expect(warnings).toContain('dry_multiplier');
+    expect(warnings).toContain('dry_base');
+    expect(warnings).toContain('dry_allowed_length');
+    expect(warnings).toContain('dry_penalty_last_n');
+  });
+
+  it('errors on non-numeric / non-integer sampler values', () => {
+    const md = withFrontmatter(
+      ['top_n_sigma: high', 'dry_allowed_length: 2.5'].join('\n'),
+    );
+    const { errors } = parseAgentsMd(md);
+    const joined = errors.map((e) => e.reason).join('\n');
+    expect(joined).toContain('top_n_sigma must be a number');
+    expect(joined).toContain('dry_allowed_length must be an integer');
+  });
+});

apps/server/src/services/__tests__/compaction.test.ts

@@ -7,6 +7,8 @@ import {
   select,
   buildPrompt,
   buildHeadPayload,
+  deriveFilesRead,
+  buildFilesReadContext,
   type CompactionMessage,
 } from '../compaction.js';
 import { SUMMARY_TEMPLATE } from '../compaction-prompt.js';
@@ -321,3 +323,105 @@ describe('buildHeadPayload reasoning render', () => {
     expect(out[1]!.content).not.toContain('<reasoning>');
   });
 });
+
+// ---- buildHeadPayload sentinel stripping (#12) -------------------------------
+
+describe('buildHeadPayload strips all UI sentinels', () => {
+  it('drops cap_hit, doom_loop, and mistake_recovery system rows', () => {
+    const out = buildHeadPayload([
+      mkMsg('user', 'do the thing'),
+      mkMsg('system', 'budget reached', { metadata: { kind: 'cap_hit' } }),
+      mkMsg('system', 'looping', { metadata: { kind: 'doom_loop' } }),
+      mkMsg('system', 'repeated errors', { metadata: { kind: 'mistake_recovery' } }),
+      mkMsg('assistant', 'answer'),
+    ]);
+    // Only the user + assistant rows survive; all three sentinels stripped.
+    expect(out).toHaveLength(2);
+    expect(out[0]!.role).toBe('user');
+    expect(out[1]!.role).toBe('assistant');
+  });
+
+  it('keeps a non-sentinel system row (e.g. compact bridge) untouched', () => {
+    const out = buildHeadPayload([
+      mkMsg('system', 'legacy compact', { kind: 'compact', metadata: null }),
+      mkMsg('user', 'q'),
+    ]);
+    expect(out[0]!.role).toBe('system');
+    expect(out[0]!.content).toBe('legacy compact');
+  });
+});
+
+// ---- file-provenance ledger (#12, Part B) -----------------------------------
+
+describe('deriveFilesRead', () => {
+  it('returns [] when the head has no read-tool calls', () => {
+    expect(deriveFilesRead([mkMsg('user', 'hi'), mkMsg('assistant', 'hello')])).toEqual([]);
+  });
+
+  it('extracts the path arg from view_file / list_dir / grep / find_files', () => {
+    const head = [
+      mkMsg('assistant', '', {
+        tool_calls: [
+          { id: 'c1', name: 'view_file', args: { path: 'src/index.ts' } },
+          { id: 'c2', name: 'list_dir', args: { path: 'src' } },
+          { id: 'c3', name: 'grep', args: { pattern: 'TODO', path: 'apps' } },
+          { id: 'c4', name: 'find_files', args: { pattern: '**/*.ts', path: 'lib' } },
+        ],
+      }),
+    ];
+    expect(deriveFilesRead(head)).toEqual(['apps', 'lib', 'src', 'src/index.ts']);
+  });
+
+  it('dedupes and sorts paths across multiple assistant turns', () => {
+    const head = [
+      mkMsg('assistant', '', { tool_calls: [{ id: 'c1', name: 'view_file', args: { path: 'b.ts' } }] }),
+      mkMsg('assistant', '', { tool_calls: [{ id: 'c2', name: 'view_file', args: { path: 'a.ts' } }] }),
+      mkMsg('assistant', '', { tool_calls: [{ id: 'c3', name: 'view_file', args: { path: 'b.ts' } }] }),
+    ];
+    expect(deriveFilesRead(head)).toEqual(['a.ts', 'b.ts']);
+  });
+
+  it('ignores non-read tools and grep calls without a path arg', () => {
+    const head = [
+      mkMsg('assistant', '', {
+        tool_calls: [
+          { id: 'c1', name: 'web_search', args: { query: 'x' } },
+          { id: 'c2', name: 'grep', args: { pattern: 'foo' } }, // no path → root, skipped
+          { id: 'c3', name: 'view_file', args: { path: 'kept.ts' } },
+        ],
+      }),
+    ];
+    expect(deriveFilesRead(head)).toEqual(['kept.ts']);
+  });
+
+  it('ignores read-tool calls on non-assistant rows', () => {
+    const head = [
+      mkMsg('user', '', { tool_calls: [{ id: 'c1', name: 'view_file', args: { path: 'nope.ts' } }] }),
+    ];
+    expect(deriveFilesRead(head)).toEqual([]);
+  });
+});
+
+describe('buildFilesReadContext', () => {
+  it('returns null when nothing was read (no empty section injected)', () => {
+    expect(buildFilesReadContext([mkMsg('user', 'hi')])).toBeNull();
+  });
+
+  it('formats a ## Files Read block with sorted bullet paths', () => {
+    const head = [
+      mkMsg('assistant', '', {
+        tool_calls: [
+          { id: 'c1', name: 'view_file', args: { path: 'z.ts' } },
+          { id: 'c2', name: 'view_file', args: { path: 'a.ts' } },
+        ],
+      }),
+    ];
+    expect(buildFilesReadContext(head)).toBe('## Files Read\n- a.ts\n- z.ts');
+  });
+});
+
+describe('SUMMARY_TEMPLATE includes the Files Read section (#12)', () => {
+  it('declares a ## Files Read section the model must maintain', () => {
+    expect(SUMMARY_TEMPLATE).toContain('## Files Read');
+  });
+});

apps/server/src/services/__tests__/mistake-tracker.test.ts

@@ -0,0 +1,164 @@
+import { describe, it, expect } from 'vitest';
+import {
+  MISTAKE_THRESHOLD,
+  freshMistakeState,
+  recordStep,
+  detectMistakePattern,
+  MISTAKE_RECOVERY_NOTE,
+  type FailureKind,
+} from '../inference/mistake-tracker.js';
+
+// ---- helpers ----------------------------------------------------------------
+// Replays a sequence of outcomes against a fresh state, returning the final
+// state so assertions can read .run / .nudges. The caller mimics turn.ts: after
+// each recordStep we consult detectMistakePattern and, if it returns 'nudge',
+// bump nudges + reset run (the loop's nudge-handling side effect).
+
+function replay(
+  outcomes: (FailureKind | 'success')[],
+  { applyNudge = false }: { applyNudge?: boolean } = {},
+) {
+  const state = freshMistakeState();
+  const decisions: (ReturnType<typeof detectMistakePattern>)[] = [];
+  for (const o of outcomes) {
+    recordStep(state, o);
+    const decision = detectMistakePattern(state);
+    decisions.push(decision);
+    if (applyNudge && decision === 'nudge') {
+      // Mirror turn.ts's nudge side effect: bump the counter, reset the streak.
+      state.nudges += 1;
+      state.run = [];
+    }
+  }
+  return { state, decisions };
+}
+
+// ---- fresh state ------------------------------------------------------------
+
+describe('freshMistakeState', () => {
+  it('starts with an empty run and zero nudges', () => {
+    const s = freshMistakeState();
+    expect(s.run).toEqual([]);
+    expect(s.nudges).toBe(0);
+  });
+});
+
+// ---- below threshold --------------------------------------------------------
+
+describe('detectMistakePattern — below threshold', () => {
+  it('returns null on a fresh state', () => {
+    expect(detectMistakePattern(freshMistakeState())).toBeNull();
+  });
+
+  it('returns null after fewer than MISTAKE_THRESHOLD failures', () => {
+    const { decisions } = replay(['zod_reject', 'exec_error']);
+    expect(decisions).toEqual([null, null]);
+  });
+});
+
+// ---- success reset ----------------------------------------------------------
+
+describe('recordStep — success resets', () => {
+  it("'success' clears both the run streak and the nudge counter", () => {
+    const state = freshMistakeState();
+    recordStep(state, 'zod_reject');
+    recordStep(state, 'exec_error');
+    state.nudges = 2; // simulate prior nudges
+    recordStep(state, 'success');
+    expect(state.run).toEqual([]);
+    expect(state.nudges).toBe(0);
+  });
+
+  it('a success mid-streak prevents the threshold from tripping', () => {
+    // fail, fail, success, fail, fail → streak never reaches 3.
+    const { decisions } = replay([
+      'zod_reject',
+      'exec_error',
+      'success',
+      'tool_not_found',
+      'permission_denied',
+    ]);
+    expect(decisions.every((d) => d === null)).toBe(true);
+  });
+});
+
+// ---- 3-streak nudge ---------------------------------------------------------
+
+describe('detectMistakePattern — nudge on 3-streak', () => {
+  it("returns 'nudge' the first time the streak reaches MISTAKE_THRESHOLD", () => {
+    const { decisions } = replay(['zod_reject', 'exec_error', 'tool_not_found']);
+    expect(decisions).toEqual([null, null, 'nudge']);
+  });
+
+  it("fires 'nudge' for a streak of identical kinds too (kind-agnostic)", () => {
+    const { decisions } = replay(['exec_error', 'exec_error', 'exec_error']);
+    expect(decisions[2]).toBe('nudge');
+  });
+});
+
+// ---- re-trip escalate -------------------------------------------------------
+
+describe('detectMistakePattern — escalate on re-trip', () => {
+  it("escalates when the streak re-trips after a nudge with no intervening success", () => {
+    // 3 fails → nudge (run reset, nudges=1), then 3 more fails → escalate.
+    const { decisions } = replay(
+      [
+        'zod_reject',
+        'exec_error',
+        'tool_not_found',
+        'permission_denied',
+        'exec_error',
+        'zod_reject',
+      ],
+      { applyNudge: true },
+    );
+    expect(decisions[2]).toBe('nudge');
+    expect(decisions[5]).toBe('escalate');
+  });
+
+  it("does NOT escalate if a success lands between the nudge and the next streak", () => {
+    const { decisions } = replay(
+      [
+        'zod_reject',
+        'exec_error',
+        'tool_not_found', // nudge here
+        'success', // clears nudges back to 0
+        'exec_error',
+        'zod_reject',
+        'tool_not_found', // 3-streak again → nudge, NOT escalate
+      ],
+      { applyNudge: true },
+    );
+    expect(decisions[2]).toBe('nudge');
+    expect(decisions[6]).toBe('nudge');
+    expect(decisions).not.toContain('escalate');
+  });
+});
+
+// ---- mixed kinds ------------------------------------------------------------
+
+describe('detectMistakePattern — mixed failure kinds', () => {
+  it('counts a streak of all five distinct kinds toward the threshold', () => {
+    const { state, decisions } = replay([
+      'zod_reject',
+      'tool_not_found',
+      'exec_error',
+    ]);
+    expect(decisions[2]).toBe('nudge');
+    expect(state.run).toEqual(['zod_reject', 'tool_not_found', 'exec_error']);
+  });
+});
+
+// ---- contract ---------------------------------------------------------------
+
+describe('MISTAKE_THRESHOLD + MISTAKE_RECOVERY_NOTE', () => {
+  it('threshold is a positive integer (tests assume 3)', () => {
+    expect(MISTAKE_THRESHOLD).toBeGreaterThan(0);
+    expect(Number.isInteger(MISTAKE_THRESHOLD)).toBe(true);
+  });
+
+  it('recovery note is a non-empty model-facing string', () => {
+    expect(typeof MISTAKE_RECOVERY_NOTE).toBe('string');
+    expect(MISTAKE_RECOVERY_NOTE.length).toBeGreaterThan(0);
+  });
+});

apps/server/src/services/agents.ts

@@ -88,6 +88,12 @@ interface ParsedFrontmatter {
   top_k?: number;
   min_p?: number;
   presence_penalty?: number;
+  // v2.6 sampling-streamjson-tokens (#11): llama.cpp sampler extensions.
+  top_n_sigma?: number;
+  dry_multiplier?: number;
+  dry_base?: number;
+  dry_allowed_length?: number;
+  dry_penalty_last_n?: number;
   tools?: string[];
   description?: string;
   model?: string;
@@ -178,6 +184,63 @@ function parseFrontmatter(yaml: string): { data: ParsedFrontmatter; errors: stri
       } else {
         errors.push(`presence_penalty must be a number (got "${valueRaw}")`);
       }
+    } else if (key === 'top_n_sigma') {
+      // v2.6 #11: llama.cpp top-n-sigma sampler. Float ≥ 0 (typical 0-3).
+      // Mirrors top_p/min_p: store then warn on out-of-range (non-numeric
+      // hard-fails the block).
+      const n = Number(valueRaw);
+      if (Number.isFinite(n)) {
+        data.top_n_sigma = n;
+        if (n < 0) {
+          console.warn(`agents: top_n_sigma ${n} out of range (≥0), ignoring (falling back to default)`);
+        }
+      } else {
+        errors.push(`top_n_sigma must be a number (got "${valueRaw}")`);
+      }
+    } else if (key === 'dry_multiplier') {
+      // v2.6 #11: DRY repetition-penalty multiplier. Float ≥ 0 (0 disables DRY).
+      const n = Number(valueRaw);
+      if (Number.isFinite(n)) {
+        data.dry_multiplier = n;
+        if (n < 0) {
+          console.warn(`agents: dry_multiplier ${n} out of range (≥0), ignoring (falling back to default)`);
+        }
+      } else {
+        errors.push(`dry_multiplier must be a number (got "${valueRaw}")`);
+      }
+    } else if (key === 'dry_base') {
+      // v2.6 #11: DRY penalty growth base. Float ≥ 0.
+      const n = Number(valueRaw);
+      if (Number.isFinite(n)) {
+        data.dry_base = n;
+        if (n < 0) {
+          console.warn(`agents: dry_base ${n} out of range (≥0), ignoring (falling back to default)`);
+        }
+      } else {
+        errors.push(`dry_base must be a number (got "${valueRaw}")`);
+      }
+    } else if (key === 'dry_allowed_length') {
+      // v2.6 #11: DRY max sequence length not penalized. Integer ≥ 0.
+      const n = Number(valueRaw);
+      if (Number.isInteger(n)) {
+        data.dry_allowed_length = n;
+        if (n < 0) {
+          console.warn(`agents: dry_allowed_length ${n} out of range (≥0), ignoring (falling back to default)`);
+        }
+      } else {
+        errors.push(`dry_allowed_length must be an integer (got "${valueRaw}")`);
+      }
+    } else if (key === 'dry_penalty_last_n') {
+      // v2.6 #11: DRY lookback window. Integer ≥ -1 (-1 = whole context, 0 = off).
+      const n = Number(valueRaw);
+      if (Number.isInteger(n)) {
+        data.dry_penalty_last_n = n;
+        if (n < -1) {
+          console.warn(`agents: dry_penalty_last_n ${n} out of range (≥-1), ignoring (falling back to default)`);
+        }
+      } else {
+        errors.push(`dry_penalty_last_n must be an integer (got "${valueRaw}")`);
+      }
     } else if (key === 'tools') {
       if (valueRaw === '') {
         data.tools = [];
@@ -354,6 +417,11 @@ function parseAgentSection(section: RawSection): Omit<Agent, 'source'> {
     top_k: typeof fm.top_k === 'number' ? fm.top_k : null,
     min_p: typeof fm.min_p === 'number' ? fm.min_p : null,
     presence_penalty: typeof fm.presence_penalty === 'number' ? fm.presence_penalty : null,
+    top_n_sigma: typeof fm.top_n_sigma === 'number' ? fm.top_n_sigma : null,
+    dry_multiplier: typeof fm.dry_multiplier === 'number' ? fm.dry_multiplier : null,
+    dry_base: typeof fm.dry_base === 'number' ? fm.dry_base : null,
+    dry_allowed_length: typeof fm.dry_allowed_length === 'number' ? fm.dry_allowed_length : null,
+    dry_penalty_last_n: typeof fm.dry_penalty_last_n === 'number' ? fm.dry_penalty_last_n : null,
     tools: filteredTools,
     model: typeof fm.model === 'string' && fm.model.length > 0 ? fm.model : null,
     max_tool_calls: typeof fm.max_tool_calls === 'number' ? fm.max_tool_calls : null,

apps/server/src/services/compaction-prompt.ts

@@ -31,10 +31,16 @@ export const SUMMARY_TEMPLATE = `Output exactly the Markdown structure shown ins
 
 ## Relevant Files
 - [file or directory path: why it matters, or "(none)"]
+
+## Files Read
+- [file or directory path that has been read/searched this session, or "(none)"]
 </template>
 
 Rules:
 - Keep every section, even when empty.
 - Use terse bullets, not prose paragraphs.
 - Preserve exact file paths, commands, error strings, and identifiers when known.
+- For ## Files Read: this is a cumulative provenance ledger. MERGE the paths
+  listed in any "## Files Read" block provided below with those already in the
+  previous summary — never drop a previously-recorded path. Sort and dedupe.
 - Do not mention the summary process or that context was compacted.`;

apps/server/src/services/compaction.ts

@@ -181,6 +181,54 @@ export function select(
   };
 }
 
+// === file-provenance ledger (#12, Part B) ===
+
+// Read tools whose path/target arg names a file or directory that was read.
+// BooChat (apps/server) is read-only — there are no write tools, so the ledger
+// only ever has a "Files Read" side (apps/coder can add "Modified" later).
+const READ_TOOL_ARG: Record<string, string> = {
+  view_file: 'path',
+  list_dir: 'path',
+  grep: 'path',
+  find_files: 'path',
+};
+
+// Derive a deterministic, deduped, sorted list of file/dir paths read by the
+// HEAD messages being summarized. Pure — scans assistant tool_calls only; the
+// boundary (which messages are "head") is decided by select() at the call site.
+// We derive at compaction time rather than via a live accumulator because
+// TurnArgs resets per turn and would miss reads on non-compacting turns; the
+// head messages are the authoritative record of what was read in the window
+// being summarized. The result propagates forward as summary text across
+// compactions (the LLM merges it into ## Files Read), so a path read long ago
+// survives even after its originating messages are compacted out.
+export function deriveFilesRead(head: CompactionMessage[]): string[] {
+  const paths = new Set<string>();
+  for (const m of head) {
+    if (m.role !== 'assistant') continue;
+    if (!m.tool_calls) continue;
+    for (const tc of m.tool_calls) {
+      const argName = READ_TOOL_ARG[tc.name];
+      if (!argName) continue;
+      const raw = (tc.args as Record<string, unknown> | null)?.[argName];
+      if (typeof raw === 'string' && raw.trim().length > 0) {
+        paths.add(raw.trim());
+      }
+    }
+  }
+  return [...paths].sort();
+}
+
+// Format the derived paths as a deterministic ## Files Read block for injection
+// into buildPrompt's context array. Returns null when nothing was read (so we
+// don't inject an empty section). The summarizer merges this into the rolling
+// summary's ## Files Read section per the SUMMARY_TEMPLATE instructions.
+export function buildFilesReadContext(head: CompactionMessage[]): string | null {
+  const paths = deriveFilesRead(head);
+  if (paths.length === 0) return null;
+  return ['## Files Read', ...paths.map((p) => `- ${p}`)].join('\n');
+}
+
 // === prompt assembly ===
 
 // Build the final user message that asks the model to (re)produce the
@@ -220,15 +268,26 @@ export interface OpenAiMessage {
   tool_call_id?: string;
 }
 
-function isCapHitSentinel(m: CompactionMessage): boolean {
+// #12: mirror inference/sentinels.ts:isAnySentinel over the CompactionMessage
-  return m.role === 'system' && m.metadata != null && m.metadata.kind === 'cap_hit';
+// shape (which carries metadata as { kind?: string } | null, not the full
+// Message type isAnySentinel expects). All UI-only sentinels are stripped from
+// the head payload — they never go to the summarizer LLM. Keep the kind list in
+// sync with isAnySentinel in sentinels.ts.
+const SENTINEL_KINDS = new Set(['cap_hit', 'doom_loop', 'mistake_recovery']);
+function isAnySentinel(m: CompactionMessage): boolean {
+  return (
+    m.role === 'system' &&
+    m.metadata != null &&
+    typeof m.metadata.kind === 'string' &&
+    SENTINEL_KINDS.has(m.metadata.kind)
+  );
 }
 
 // v1.13.6: exported for unit-test access (reasoning render coverage).
 export function buildHeadPayload(head: CompactionMessage[]): OpenAiMessage[] {
   const out: OpenAiMessage[] = [];
   for (const m of head) {
-    if (isCapHitSentinel(m)) continue;
+    if (isAnySentinel(m)) continue;
     if (m.role === 'assistant' && (m.status === 'streaming' || m.status === 'cancelled')) continue;
     if (m.kind === 'compact') {
       // Legacy compact row — pass through as system context. The new
@@ -417,7 +476,14 @@ export async function process(input: ProcessInput): Promise<void> {
   // user message carrying buildPrompt(previousSummary, []). No system prompt
   // — matches opencode (`system: []`); the template + anchor are sufficient.
   const headPayload = buildHeadPayload(sel.head);
-  const finalUser: OpenAiMessage = { role: 'user', content: buildPrompt(previousSummary, []) };
+  // #12 Part B: derive the file-provenance ledger from the head's read-tool
+  // calls and inject it as a deterministic ## Files Read context block so the
+  // summarizer merges it into the rolling summary. Empty → no injection.
+  const filesReadCtx = buildFilesReadContext(sel.head);
+  const finalUser: OpenAiMessage = {
+    role: 'user',
+    content: buildPrompt(previousSummary, filesReadCtx ? [filesReadCtx] : []),
+  };
   const payload = [...headPayload, finalUser];
 
   log.info(

apps/server/src/services/inference/index.ts

@@ -19,6 +19,14 @@ export type {
 } from './turn.js';
 export type { ToolPhaseResult } from './tool-phase.js';
 export { detectDoomLoop, DOOM_LOOP_THRESHOLD } from './sentinels.js';
+export {
+  detectMistakePattern,
+  freshMistakeState,
+  recordStep,
+  MISTAKE_THRESHOLD,
+  MISTAKE_RECOVERY_NOTE,
+} from './mistake-tracker.js';
+export type { FailureKind, MistakeState } from './mistake-tracker.js';
 export { buildMessagesPayload } from './payload.js';
 export { generateToolUseSummary } from './tool-summaries.js';
 export type { ToolInfo } from './tool-summaries.js';

apps/server/src/services/inference/mistake-tracker.ts

@@ -0,0 +1,69 @@
+// v#12 MistakeTracker: heterogeneous-failure recovery. Complements the
+// doom-loop guard (sentinels.ts:detectDoomLoop, which only catches *identical*
+// repeats) by catching a run of consecutive tool FAILURES the model isn't
+// recovering from — even when each failure is a *different* error. Algorithm
+// reimplemented from cline's mistake-counting pattern (NOT vendored).
+//
+// Pure module — mirrors sentinels.ts:detectDoomLoop. No DB, no I/O. The state
+// lives loop-local in TurnArgs (reset per runInference, like recentToolCalls).
+
+// The failure taxonomy already distinguished in tool-phase.ts:executeToolCall.
+// 'api_error' is reserved for upstream-model failures surfaced as tool outcomes
+// (no current emit site on apps/server, but the union mirrors the design doc
+// so a future caller can record it without a type change).
+export type FailureKind =
+  | 'zod_reject'
+  | 'tool_not_found'
+  | 'exec_error'
+  | 'api_error'
+  | 'permission_denied';
+
+// Smallest streak that doesn't false-positive on a model that retries once
+// after a transient error. Matches DOOM_LOOP_THRESHOLD's rationale.
+export const MISTAKE_THRESHOLD = 3;
+
+export interface MistakeState {
+  // The current consecutive-failure streak (any successful tool step clears it).
+  run: FailureKind[];
+  // How many recovery nudges have fired without an intervening success. Used to
+  // escalate (stop the turn) on the second trip rather than nudging forever.
+  nudges: number;
+}
+
+export function freshMistakeState(): MistakeState {
+  return { run: [], nudges: 0 };
+}
+
+// Record one tool step's outcome. A 'success' clears BOTH the streak and the
+// nudge counter (the model recovered). A FailureKind pushes onto the streak.
+export function recordStep(
+  state: MistakeState,
+  outcome: FailureKind | 'success',
+): void {
+  if (outcome === 'success') {
+    state.run = [];
+    state.nudges = 0;
+    return;
+  }
+  state.run.push(outcome);
+}
+
+// Decide whether to intervene given the current streak. When the streak has
+// reached MISTAKE_THRESHOLD: 'nudge' the first time (no nudge fired yet),
+// 'escalate' if it trips again while a nudge is already outstanding (no
+// intervening success cleared `nudges`). Below threshold → null.
+//
+// Pure — the caller is responsible for mutating `nudges`/`run` after acting on
+// the decision (mirrors how turn.ts consumes detectDoomLoop's result).
+export function detectMistakePattern(
+  state: MistakeState,
+): 'nudge' | 'escalate' | null {
+  if (state.run.length < MISTAKE_THRESHOLD) return null;
+  return state.nudges === 0 ? 'nudge' : 'escalate';
+}
+
+// Model-facing guidance injected (transiently, for the next step only) when a
+// nudge fires. Short + declarative for the same reliability reason as the
+// cap-hit / doom-loop notes.
+export const MISTAKE_RECOVERY_NOTE =
+  "You've hit several different errors in a row. Stop retrying variations — re-read the tool schemas, verify file paths and arguments exist before calling, and try a fundamentally different approach.";

apps/server/src/services/inference/sentinel-summaries.ts

@@ -86,7 +86,7 @@ export async function runCapHitSummary(
       ctx,
       session.model,
       messages,
-      { tools: null, temperature: agent?.temperature, top_p: agent?.top_p ?? undefined, top_k: agent?.top_k ?? undefined, min_p: agent?.min_p ?? undefined, presence_penalty: agent?.presence_penalty ?? undefined },
+      { tools: null, temperature: agent?.temperature, top_p: agent?.top_p ?? undefined, top_k: agent?.top_k ?? undefined, min_p: agent?.min_p ?? undefined, presence_penalty: agent?.presence_penalty ?? undefined, top_n_sigma: agent?.top_n_sigma ?? undefined, dry_multiplier: agent?.dry_multiplier ?? undefined, dry_base: agent?.dry_base ?? undefined, dry_allowed_length: agent?.dry_allowed_length ?? undefined, dry_penalty_last_n: agent?.dry_penalty_last_n ?? undefined },
       (delta) => {
         accumulated += delta;
         ctx.publish(sessionId, {
@@ -346,7 +346,7 @@ export async function runDoomLoopSummary(
       ctx,
       session.model,
       messages,
-      { tools: null, temperature: agent?.temperature, top_p: agent?.top_p ?? undefined, top_k: agent?.top_k ?? undefined, min_p: agent?.min_p ?? undefined, presence_penalty: agent?.presence_penalty ?? undefined },
+      { tools: null, temperature: agent?.temperature, top_p: agent?.top_p ?? undefined, top_k: agent?.top_k ?? undefined, min_p: agent?.min_p ?? undefined, presence_penalty: agent?.presence_penalty ?? undefined, top_n_sigma: agent?.top_n_sigma ?? undefined, dry_multiplier: agent?.dry_multiplier ?? undefined, dry_base: agent?.dry_base ?? undefined, dry_allowed_length: agent?.dry_allowed_length ?? undefined, dry_penalty_last_n: agent?.dry_penalty_last_n ?? undefined },
       (delta) => {
         accumulated += delta;
         ctx.publish(sessionId, {
@@ -545,7 +545,7 @@ export async function runStepCapSummary(
       ctx,
       session.model,
       messages,
-      { tools: null, temperature: agent?.temperature, top_p: agent?.top_p ?? undefined, top_k: agent?.top_k ?? undefined, min_p: agent?.min_p ?? undefined, presence_penalty: agent?.presence_penalty ?? undefined },
+      { tools: null, temperature: agent?.temperature, top_p: agent?.top_p ?? undefined, top_k: agent?.top_k ?? undefined, min_p: agent?.min_p ?? undefined, presence_penalty: agent?.presence_penalty ?? undefined, top_n_sigma: agent?.top_n_sigma ?? undefined, dry_multiplier: agent?.dry_multiplier ?? undefined, dry_base: agent?.dry_base ?? undefined, dry_allowed_length: agent?.dry_allowed_length ?? undefined, dry_penalty_last_n: agent?.dry_penalty_last_n ?? undefined },
       (delta) => {
         accumulated += delta;
         ctx.publish(sessionId, {
@@ -717,3 +717,57 @@ async function insertDoomLoopSentinel(
     metadata,
   });
 }
+
+// #12 MistakeTracker: heterogeneous-failure recovery sentinel. Mirrors
+// insertDoomLoopSentinel structurally — a role='system', status='complete' row
+// firing the standard message_started → delta → message_complete frame
+// sequence. Two variants distinguished by `escalated`:
+//   - escalated:false → a nudge fired; recovery guidance was injected into the
+//     model's next step and the loop continued. can_continue is true (the turn
+//     is still live).
+//   - escalated:true  → the nudge didn't break the failure run; the turn was
+//     stopped (cap-hit-style). can_continue is true so the UI can still offer a
+//     Continue affordance — a fresh user turn resets the tracker.
+export async function insertMistakeRecoverySentinel(
+  ctx: InferenceContext,
+  sessionId: string,
+  chatId: string,
+  opts: { failureKinds: string[]; count: number; escalated: boolean; canContinue: boolean },
+): Promise<void> {
+  const metadata: MessageMetadata = {
+    kind: 'mistake_recovery',
+    failure_kinds: opts.failureKinds,
+    count: opts.count,
+    escalated: opts.escalated,
+    can_continue: opts.canContinue,
+  };
+  const content = opts.escalated
+    ? `Repeated different errors persisted after a recovery nudge (${opts.count} in a row). Stopping the tool-call loop.`
+    : `Hit ${opts.count} different errors in a row. Injected recovery guidance and continuing.`;
+
+  const [row] = await ctx.sql<{ id: string }[]>`
+    INSERT INTO messages (session_id, chat_id, role, content, status, created_at, metadata)
+    VALUES (${sessionId}, ${chatId}, 'system', ${content}, 'complete', clock_timestamp(), ${ctx.sql.json(metadata as never)})
+    RETURNING id
+  `;
+
+  // Standard frame sequence — same as cap-hit / doom-loop sentinels.
+  ctx.publish(sessionId, {
+    type: 'message_started',
+    message_id: row!.id,
+    chat_id: chatId,
+    role: 'system',
+  });
+  ctx.publish(sessionId, {
+    type: 'delta',
+    message_id: row!.id,
+    chat_id: chatId,
+    content,
+  });
+  ctx.publish(sessionId, {
+    type: 'message_complete',
+    message_id: row!.id,
+    chat_id: chatId,
+    metadata,
+  });
+}

apps/server/src/services/inference/sentinels.ts

@@ -48,6 +48,18 @@ export function isDoomLoopSentinel(m: Message): boolean {
   );
 }
 
-export function isAnySentinel(m: Message): boolean {
+// #12: mistake-recovery sentinel. Same UI-only semantics as cap-hit /
-  return isCapHitSentinel(m) || isDoomLoopSentinel(m);
+// doom-loop — never sent to the LLM (filtered via the isAnySentinel check
+// below, which buildMessagesPayload + buildHeadPayload both consult).
+export function isMistakeRecoverySentinel(m: Message): boolean {
+  return (
+    m.role === 'system' &&
+    m.metadata !== null &&
+    typeof m.metadata === 'object' &&
+    (m.metadata as { kind?: unknown }).kind === 'mistake_recovery'
+  );
+}
+
+export function isAnySentinel(m: Message): boolean {
+  return isCapHitSentinel(m) || isDoomLoopSentinel(m) || isMistakeRecoverySentinel(m);
 }

apps/server/src/services/inference/stream-phase.ts

@@ -33,6 +33,39 @@ interface StreamOptions {
   top_k?: number | null;
   min_p?: number | null;
   presence_penalty?: number | null;
+  // v2.6 sampling-streamjson-tokens (#11): llama.cpp sampler extensions. These
+  // are NOT standard AI-SDK streamText options and are NOT serialized by the
+  // openai-compatible provider's standardized-settings path (topK is even
+  // explicitly dropped with an "unsupported feature: topK" warning). They reach
+  // llama-server only via providerOptions.openaiCompatible (see buildSamplerProviderOptions).
+  top_n_sigma?: number | null;
+  dry_multiplier?: number | null;
+  dry_base?: number | null;
+  dry_allowed_length?: number | null;
+  dry_penalty_last_n?: number | null;
+}
+
+// v2.6 #11: build the providerOptions.openaiCompatible extraBody object for the
+// llama.cpp sampler extensions. @ai-sdk/openai-compatible (2.0.47) merges every
+// non-reserved key under providerOptions.openaiCompatible straight into the
+// chat-completion request body (see its getArgs: the Object.fromEntries spread
+// filtered against openaiCompatibleLanguageModelChatOptions.shape). This is the
+// ONLY working passthrough for these params:
+//   - top_k / min_p were latently dropped before this: top_k was passed as the
+//     AI-SDK `topK` setting which the openai-compatible provider rejects as
+//     unsupported; min_p was never passed to streamText at all.
+//   - top_n_sigma + the dry_* family have no AI-SDK equivalent.
+// Keys use llama-server's snake_case body names so they land verbatim.
+function buildSamplerProviderOptions(opts: StreamOptions): Record<string, number> | undefined {
+  const body: Record<string, number> = {};
+  if (typeof opts.top_k === 'number') body.top_k = opts.top_k;
+  if (typeof opts.min_p === 'number') body.min_p = opts.min_p;
+  if (typeof opts.top_n_sigma === 'number') body.top_n_sigma = opts.top_n_sigma;
+  if (typeof opts.dry_multiplier === 'number') body.dry_multiplier = opts.dry_multiplier;
+  if (typeof opts.dry_base === 'number') body.dry_base = opts.dry_base;
+  if (typeof opts.dry_allowed_length === 'number') body.dry_allowed_length = opts.dry_allowed_length;
+  if (typeof opts.dry_penalty_last_n === 'number') body.dry_penalty_last_n = opts.dry_penalty_last_n;
+  return Object.keys(body).length > 0 ? body : undefined;
 }
 
 // v1.13.1-A: convert BooCode's OpenAI-shaped history into AI SDK
@@ -195,6 +228,14 @@ export async function streamCompletion(
     return toolCall;
   };
 
+  // v2.6 #11: llama.cpp sampler extensions (top_k, min_p, top_n_sigma, dry_*)
+  // ride providerOptions.openaiCompatible — they are NOT standardized streamText
+  // settings. NB: top_k used to be passed below as the AI-SDK `topK` setting;
+  // the openai-compatible provider dropped it with an "unsupported feature: topK"
+  // warning and min_p was never wired at all, so both were dead on the wire
+  // before this. They now go through the same extraBody path as the new params.
+  const samplerBody = buildSamplerProviderOptions(opts);
+
   const result = streamText({
     model: upstreamModel(ctx.config, model, agent ?? null),
     messages: aiMessages,
@@ -203,8 +244,8 @@ export async function streamCompletion(
       : {}),
     ...(typeof opts.temperature === 'number' ? { temperature: opts.temperature } : {}),
     ...(typeof opts.top_p === 'number' ? { topP: opts.top_p } : {}),
-    ...(typeof opts.top_k === 'number' ? { topK: opts.top_k } : {}),
     ...(typeof opts.presence_penalty === 'number' ? { presencePenalty: opts.presence_penalty } : {}),
+    ...(samplerBody ? { providerOptions: { openaiCompatible: samplerBody } } : {}),
     abortSignal: signal,
   });
 
@@ -398,6 +439,12 @@ export async function executeStreamPhase(
   const effectiveTopK = agent?.top_k ?? undefined;
   const effectiveMinP = agent?.min_p ?? undefined;
   const effectivePresencePenalty = agent?.presence_penalty ?? undefined;
+  // v2.6 #11: llama.cpp sampler extensions, threaded the same way as top_k/min_p.
+  const effectiveTopNSigma = agent?.top_n_sigma ?? undefined;
+  const effectiveDryMultiplier = agent?.dry_multiplier ?? undefined;
+  const effectiveDryBase = agent?.dry_base ?? undefined;
+  const effectiveDryAllowedLength = agent?.dry_allowed_length ?? undefined;
+  const effectiveDryPenaltyLastN = agent?.dry_penalty_last_n ?? undefined;
 
   // v1.12.2: ctx_max lookup is cached after the first hit per model, so this
   // is a Map probe in steady state. We capture nCtx once at the top of the
@@ -435,7 +482,19 @@ export async function executeStreamPhase(
       ctx,
       session.model,
       messages,
-      { tools: effectiveTools, temperature: effectiveTemperature, top_p: effectiveTopP, top_k: effectiveTopK, min_p: effectiveMinP, presence_penalty: effectivePresencePenalty },
+      {
+        tools: effectiveTools,
+        temperature: effectiveTemperature,
+        top_p: effectiveTopP,
+        top_k: effectiveTopK,
+        min_p: effectiveMinP,
+        presence_penalty: effectivePresencePenalty,
+        top_n_sigma: effectiveTopNSigma,
+        dry_multiplier: effectiveDryMultiplier,
+        dry_base: effectiveDryBase,
+        dry_allowed_length: effectiveDryAllowedLength,
+        dry_penalty_last_n: effectiveDryPenaltyLastN,
+      },
       (delta) => {
         state.accumulated += delta;
         ctx.publish(sessionId, {

apps/server/src/services/inference/tool-phase.ts

@@ -17,6 +17,7 @@ import { formatUnknownToolError } from './tool-suggestions.js';
 // prompted about paths we couldn't grant anyway (e.g. /etc/passwd).
 import { resolveGrantRoot } from '../grant_resolver.js';
 import { stripToolMarkup } from './tool-call-parser.js';
+import type { FailureKind } from './mistake-tracker.js';
 import type {
   InferenceContext,
   StreamResult,
@@ -33,13 +34,18 @@ async function executeToolCall(
   toolCall: ToolCall,
   extraRoots: readonly string[],
   toolCtx?: ToolExecCtx,
-): Promise<{ output: unknown; truncated: boolean; error?: string }> {
+): Promise<{ output: unknown; truncated: boolean; error?: string; outcome: FailureKind | 'success' }> {
+  // v#12 MistakeTracker: every return path carries an `outcome` so the turn
+  // loop can detect a run of heterogeneous failures. The failure taxonomy
+  // mirrors mistake-tracker.ts:FailureKind. Does NOT alter the existing
+  // output/truncated/error shape — outcome is purely additive.
   const tool = TOOLS_BY_NAME[toolCall.name];
   if (!tool) {
     return {
       output: null,
       truncated: false,
       error: formatUnknownToolError(toolCall.name, Object.keys(TOOLS_BY_NAME)),
+      outcome: 'tool_not_found',
     };
   }
   const parsed = tool.inputSchema.safeParse(toolCall.args);
@@ -64,6 +70,7 @@ async function executeToolCall(
       output: null,
       truncated: false,
       error: `tool '${toolCall.name}' rejected — ${hint}`,
+      outcome: 'zod_reject',
     };
   }
   try {
@@ -72,15 +79,16 @@ async function executeToolCall(
       typeof output === 'object' && output !== null && 'truncated' in output
         ? Boolean((output as { truncated: unknown }).truncated)
         : false;
-    return { output, truncated };
+    return { output, truncated, outcome: 'success' };
   } catch (err) {
     if (err instanceof PathScopeError) {
-      return { output: null, truncated: false, error: err.message };
+      return { output: null, truncated: false, error: err.message, outcome: 'permission_denied' };
     }
     return {
       output: null,
       truncated: false,
       error: err instanceof Error ? err.message : String(err),
+      outcome: 'exec_error',
     };
   }
 }
@@ -93,6 +101,12 @@ export interface ToolPhaseResult {
   toolCallCount: number;
   toolCalls: ToolCall[];
   nextAssistantId: string | null;
+  // v#12 MistakeTracker: one outcome per executed tool call, in no particular
+  // order (filled inside the Promise.all callbacks). The turn loop folds these
+  // into TurnArgs.mistakeTracker via recordStep. Pause/auto-grant control-flow
+  // tools record 'success' (they aren't model mistakes); the genuine error
+  // paths record their FailureKind.
+  outcomes: (FailureKind | 'success')[];
 }
 
 export async function executeToolPhase(
@@ -187,6 +201,10 @@ export async function executeToolPhase(
   // for the synthesis input. Race-free under Promise.all because each
   // callback pushes its own captured value.
   const synthEntries: Array<{ tc: ToolCall; output: unknown; error?: string }> = [];
+  // v#12 MistakeTracker: collect each tool's outcome. Concurrent pushes under
+  // Promise.all are safe (each callback appends its own value; order is not
+  // significant to recordStep which folds them sequentially).
+  const outcomes: (FailureKind | 'success')[] = [];
   await Promise.all(
     toolCalls.map(async (tc) => {
       const [toolRow] = await ctx.sql<{ id: string }[]>`
@@ -197,6 +215,7 @@ export async function executeToolPhase(
       const toolMessageId = toolRow!.id;
       if (tc.name === 'ask_user_input') {
         pausingForUserInput = true;
+        outcomes.push('success');
         const sentinel = { tool_call_id: tc.id, output: null, truncated: false };
         // v1.13.20: parts-only. The answer-endpoint UPDATE later
         // (messages.ts) will delete and re-insert this part when the user
@@ -227,7 +246,10 @@ export async function executeToolPhase(
         );
         if (!resolution.ok) {
           // Auto-deny without pausing. The model sees the reason on its
-          // next turn and decides what to do.
+          // next turn and decides what to do. Counts as a permission_denied
+          // failure for the mistake tracker (the model asked for a path it
+          // can't have — a recoverable mistake it should learn from).
+          outcomes.push('permission_denied');
           const stored = {
             tool_call_id: tc.id,
             output: `denied: ${resolution.reason}`,
@@ -255,6 +277,7 @@ export async function executeToolPhase(
         // pause. The grant endpoint re-derives the root at decision time
         // (state may have changed in the meantime) so we don't stash it here.
         pausingForUserInput = true;
+        outcomes.push('success');
         const sentinel = { tool_call_id: tc.id, output: null, truncated: false };
         // v1.13.20: parts-only write.
         await insertParts(
@@ -267,6 +290,10 @@ export async function executeToolPhase(
         return;
       }
       if (agent && !matchToolGlob(tc.name, agent.tools)) {
+        // Agent-scope denial — the model called a tool outside its whitelist.
+        // permission_denied for the mistake tracker (the model should pick a
+        // tool it's actually allowed to use).
+        outcomes.push('permission_denied');
         const stored = {
           tool_call_id: tc.id,
           output: null,
@@ -295,6 +322,10 @@ export async function executeToolPhase(
         sql: ctx.sql,
         sessionId,
       });
+      // v#12 MistakeTracker: record the real execution outcome (success or a
+      // FailureKind). This is the primary signal for heterogeneous-failure
+      // detection.
+      outcomes.push(tres.outcome);
       if (SYNTHESIS_TOOLS.has(tc.name)) {
         synthEntries.push({ tc, output: tres.output, ...(tres.error ? { error: tres.error } : {}) });
       }
@@ -340,6 +371,7 @@ export async function executeToolPhase(
       toolCallCount: toolCalls.length,
       toolCalls,
       nextAssistantId: null,
+      outcomes,
     };
   }
 
@@ -378,6 +410,7 @@ export async function executeToolPhase(
         toolCallCount: toolCalls.length,
         toolCalls,
         nextAssistantId: null,
+        outcomes,
       };
     }
     // ran === false → synthesis failed (timeout / model error) → fall through
@@ -397,5 +430,6 @@ export async function executeToolPhase(
     toolCallCount: toolCalls.length,
     toolCalls,
     nextAssistantId: nextAssistant!.id,
+    outcomes,
   };
 }

apps/server/src/services/inference/turn.ts

@@ -22,6 +22,13 @@ import { resolveToolBudget } from './budget.js';
 import {
   detectDoomLoop,
 } from './sentinels.js';
+import {
+  detectMistakePattern,
+  freshMistakeState,
+  recordStep,
+  MISTAKE_RECOVERY_NOTE,
+  type MistakeState,
+} from './mistake-tracker.js';
 import {
   buildMessagesPayload,
   loadContext,
@@ -39,6 +46,7 @@ import {
   runCapHitSummary,
   runDoomLoopSummary,
   runStepCapSummary,
+  insertMistakeRecoverySentinel,
 } from './sentinel-summaries.js';
 
 // v1.14.0: hard ceiling on the number of stream-and-tool iterations per
@@ -144,6 +152,16 @@ export interface TurnArgs {
   // boundaries by runInference, same as toolsUsed. Doom-loop check at the
   // top of runAssistantTurn slices the last DOOM_LOOP_THRESHOLD entries.
   recentToolCalls: ToolCall[];
+  // v#12 MistakeTracker: heterogeneous-failure recovery state. Loop-local,
+  // reset per runInference (user-message boundary) like recentToolCalls. Folds
+  // tool-phase outcomes via recordStep each iteration; detectMistakePattern
+  // gates the nudge/escalate decision.
+  mistakeTracker: MistakeState;
+  // v#12: transient model-facing recovery note set when a nudge fires. Consumed
+  // (appended as a role:'system' message + cleared) on the NEXT payload build.
+  // Never persisted — mirrors how the cap-hit/doom-loop notes live only inside
+  // the summary call's messages array.
+  pendingRecoveryNote?: string;
   signal: AbortSignal | undefined;
 }
 
@@ -188,6 +206,12 @@ export async function runAssistantTurn(
   let toolsUsed = args.toolsUsed;
   let recentToolCalls = args.recentToolCalls;
   let assistantMessageId = args.assistantMessageId;
+  // v#12 MistakeTracker: the tracker state is carried on `args` (mutated in
+  // place by recordStep). pendingRecoveryNote is a loop-local because it is a
+  // single-step transient — set when a nudge fires, consumed (injected into the
+  // next payload) and cleared on the following iteration.
+  const mistakeTracker = args.mistakeTracker;
+  let pendingRecoveryNote: string | undefined = args.pendingRecoveryNote;
 
   while (stepNumber < effectiveCap) {
     // ---- doom-loop check (moved from top-of-function) ----
@@ -196,7 +220,7 @@ export async function runAssistantTurn(
       // Need fresh history for the summary.
       const loaded = await loadContext(ctx.sql, sessionId, chatId);
       if (loaded) {
-        const iterArgs: TurnArgs = { sessionId, chatId, assistantMessageId, toolsUsed, recentToolCalls, signal };
+        const iterArgs: TurnArgs = { sessionId, chatId, assistantMessageId, toolsUsed, recentToolCalls, mistakeTracker, signal };
         await runDoomLoopSummary(ctx, iterArgs, loaded.session, loaded.project, loaded.history, agent, loop);
       }
       break;
@@ -206,7 +230,7 @@ export async function runAssistantTurn(
     if (toolsUsed >= budget) {
       const loaded = await loadContext(ctx.sql, sessionId, chatId);
       if (loaded) {
-        const iterArgs: TurnArgs = { sessionId, chatId, assistantMessageId, toolsUsed, recentToolCalls, signal };
+        const iterArgs: TurnArgs = { sessionId, chatId, assistantMessageId, toolsUsed, recentToolCalls, mistakeTracker, signal };
         await runCapHitSummary(ctx, iterArgs, loaded.session, loaded.project, loaded.history, agent, budget);
       }
       break;
@@ -265,7 +289,16 @@ export async function runAssistantTurn(
       }
     }
 
-    const iterArgs: TurnArgs = { sessionId, chatId, assistantMessageId, toolsUsed, recentToolCalls, signal };
+    // v#12 MistakeTracker: if the prior iteration's nudge fired, append the
+    // transient recovery note to THIS payload (consumed exactly once, then
+    // cleared). Never persisted — same lifecycle as the cap-hit/doom-loop
+    // summary notes, which live only inside the in-memory messages array.
+    if (pendingRecoveryNote) {
+      messages.push({ role: 'system', content: pendingRecoveryNote });
+      pendingRecoveryNote = undefined;
+    }
+
+    const iterArgs: TurnArgs = { sessionId, chatId, assistantMessageId, toolsUsed, recentToolCalls, mistakeTracker, signal };
     const state: StreamPhaseState = { accumulated: '', startedAt: null };
     let result: StreamResult;
     try {
@@ -305,10 +338,78 @@ export async function runAssistantTurn(
     recentToolCalls = [...recentToolCalls, ...toolPhaseResult.toolCalls];
     stepNumber++;
 
+    // v#12 MistakeTracker: fold this iteration's tool outcomes into the
+    // tracker, in order. recordStep mutates `mistakeTracker` in place (it is
+    // the same object referenced by args). A 'success' clears the streak.
+    for (const o of toolPhaseResult.outcomes) {
+      recordStep(mistakeTracker, o);
+    }
+
     if (toolPhaseResult.action !== 'continue') {
-      // 'paused' (user input) or 'synthesis_done' — stop the loop.
+      // 'paused' (user input) or 'synthesis_done' — stop the loop. The turn is
+      // already ending, so neither a nudge nor an escalate would change the
+      // control flow; we skip the mistake decision here.
       break;
     }
+
+    // v#12 MistakeTracker: heterogeneous-failure decision. Only evaluated on
+    // the 'continue' path (the only case where the loop would otherwise
+    // proceed to another step). Complements the doom-loop check above, which
+    // only catches *identical* repeats.
+    const mistake = detectMistakePattern(mistakeTracker);
+    if (mistake === 'nudge') {
+      // Soft intervention: inject model-facing recovery guidance into the NEXT
+      // step's payload, drop a UI sentinel, bump nudges, reset the streak, and
+      // continue. The note is consumed (and cleared) at the top of the next
+      // iteration's payload build.
+      pendingRecoveryNote = MISTAKE_RECOVERY_NOTE;
+      const failureKinds = [...mistakeTracker.run];
+      await insertMistakeRecoverySentinel(ctx, sessionId, chatId, {
+        failureKinds,
+        count: failureKinds.length,
+        escalated: false,
+        canContinue: true,
+      });
+      mistakeTracker.nudges += 1;
+      mistakeTracker.run = [];
+      ctx.log.info(
+        { sessionId, chatId, step: stepNumber, nudges: mistakeTracker.nudges, failureKinds },
+        'mistake_recovery nudge',
+      );
+      assistantMessageId = toolPhaseResult.nextAssistantId!;
+      continue;
+    }
+    if (mistake === 'escalate') {
+      // The nudge didn't break the failure run — stop the turn (cap-hit-style)
+      // to avoid burning the whole step budget on heterogeneous failures. The
+      // next assistant row is still 'streaming'; finalize it as a short note so
+      // the slot doesn't dangle, then drop the escalate sentinel.
+      const failureKinds = [...mistakeTracker.run];
+      assistantMessageId = toolPhaseResult.nextAssistantId!;
+      await ctx.sql`
+        UPDATE messages
+        SET content = '', status = 'complete', finished_at = clock_timestamp()
+        WHERE id = ${assistantMessageId}
+      `;
+      ctx.publish(sessionId, {
+        type: 'message_complete',
+        message_id: assistantMessageId,
+        chat_id: chatId,
+      });
+      await insertMistakeRecoverySentinel(ctx, sessionId, chatId, {
+        failureKinds,
+        count: failureKinds.length,
+        escalated: true,
+        canContinue: true,
+      });
+      ctx.publishUser({ type: 'chat_status', chat_id: chatId, status: 'idle', at: new Date().toISOString() });
+      ctx.log.info(
+        { sessionId, chatId, step: stepNumber, failureKinds },
+        'mistake_recovery escalate — stopping turn',
+      );
+      break;
+    }
+
     // 'continue' — advance to next assistant message.
     assistantMessageId = toolPhaseResult.nextAssistantId!;
   }
@@ -320,7 +421,7 @@ export async function runAssistantTurn(
   if (stepNumber >= effectiveCap && effectiveCap < Infinity) {
     const loaded = await loadContext(ctx.sql, sessionId, chatId);
     if (loaded) {
-      const capArgs: TurnArgs = { sessionId, chatId, assistantMessageId, toolsUsed, recentToolCalls, signal };
+      const capArgs: TurnArgs = { sessionId, chatId, assistantMessageId, toolsUsed, recentToolCalls, mistakeTracker, signal };
       await runStepCapSummary(ctx, capArgs, loaded.session, loaded.project, loaded.history, agent, stepNumber, effectiveCap);
     }
   }
@@ -378,12 +479,16 @@ export async function runInference(
   // per-call budget.
   // v1.11.6: recentToolCalls also resets — doom-loop detection is scoped
   // to a single user-message turn, so a Continue starts with no history.
+  // v#12 MistakeTracker: fresh per user-message turn, like recentToolCalls.
+  // Tracks consecutive heterogeneous tool failures across the loop's
+  // stream-and-tool iterations within this turn.
   return runAssistantTurn(ctx, {
     sessionId,
     chatId,
     assistantMessageId,
     toolsUsed: 0,
     recentToolCalls: [],
+    mistakeTracker: freshMistakeState(),
     signal,
   });
 }

apps/server/src/types/api.ts

@@ -117,6 +117,15 @@ export interface Agent {
   top_k: number | null;  // null means omit from request body
   min_p: number | null;  // null means omit from request body
   presence_penalty: number | null;  // null means omit from request body
+  // v2.6 sampling-streamjson-tokens (#11): llama.cpp sampler extensions.
+  // null = omit from request body. top_n_sigma + the DRY repetition family
+  // help the doom-loop-prone local model. All travel via the same
+  // providerOptions.openaiCompatible extraBody channel as top_k/min_p.
+  top_n_sigma: number | null;
+  dry_multiplier: number | null;
+  dry_base: number | null;
+  dry_allowed_length: number | null;
+  dry_penalty_last_n: number | null;
   tools: string[];       // whitelist of tool names; empty = no tools allowed
   model: string | null;  // null means "session.model wins"
   source: AgentSource;
@@ -198,10 +207,15 @@ export type ErrorReason =
   | 'summary_after_cap_failed';
 
 // v1.8.2 / v1.11.6: shapes stored in messages.metadata. Discriminated on `kind`.
-//   cap_hit    — system sentinel emitted when tool budget is exhausted
+//   cap_hit          — system sentinel emitted when tool budget is exhausted
-//   doom_loop  — system sentinel emitted when the model called the same
+//   doom_loop        — system sentinel emitted when the model called the same
-//                tool with the same args DOOM_LOOP_THRESHOLD times in a row
+//                      tool with the same args DOOM_LOOP_THRESHOLD times in a row
-//   error      — attached to a failed assistant message so UI can show reason
+//   mistake_recovery — system sentinel emitted when a run of consecutive
+//                      *heterogeneous* tool failures is detected (#12). A nudge
+//                      (escalated:false) injects model-facing recovery guidance
+//                      and continues; an escalate (escalated:true) stops the
+//                      turn after the nudge failed to break the failure run.
+//   error            — attached to a failed assistant message so UI can show reason
 export type MessageMetadata =
   | {
       kind: 'cap_hit';
@@ -216,6 +230,14 @@ export type MessageMetadata =
       args: Record<string, unknown>;
       threshold: number;
     }
+  | {
+      // PINNED CONTRACT (#12) — mirrored byte-for-byte in apps/web/src/api/types.ts.
+      kind: 'mistake_recovery';
+      failure_kinds: string[];
+      count: number;
+      escalated: boolean;
+      can_continue?: boolean;
+    }
   | {
       kind: 'error';
       error_reason: ErrorReason;

apps/web/src/api/client.ts

@@ -34,6 +34,12 @@ export interface AgentSessionInfo {
   status: string;
   has_session: boolean;
   last_active_at: string | null;
+  // v2.6.8 per-(chat,agent) running token/cost totals (sampling-streamjson-tokens
+  // #8). input_tokens/output_tokens are BIGINT and may arrive as strings; cost is
+  // DOUBLE. AgentComposerBar coerces with Number(...) before rendering.
+  input_tokens: number;
+  output_tokens: number;
+  cost: number;
 }
 
 // write-edit-robustness #4: a pre-turn worktree snapshot anchored to an

apps/web/src/api/types.ts

@@ -155,6 +155,9 @@ export type ErrorReason =
 //                budget + agent name + whether Continue is still allowed.
 //   doom_loop  — sentinel emitted when the model called the same tool with
 //                the same arguments threshold times in a row.
+//   mistake_recovery — sentinel emitted when the model hit repeated *different*
+//                errors; non-escalated means recovery guidance was injected and
+//                the turn continues, escalated means the turn was stopped.
 //   error      — attached to a failed assistant message so the bubble can show
 //                a specific reason on reload (WS error frame is one-shot).
 export type MessageMetadata =
@@ -171,6 +174,13 @@ export type MessageMetadata =
       args: Record<string, unknown>;
       threshold: number;
     }
+  | {
+      kind: 'mistake_recovery';
+      failure_kinds: string[];
+      count: number;
+      escalated: boolean;
+      can_continue?: boolean;
+    }
   | {
       kind: 'error';
       error_reason: ErrorReason;

apps/web/src/components/AgentComposerBar.tsx

@@ -185,6 +185,14 @@ interface Props {
   hasPriorTurn?: boolean;
 }
 
+// Condensed token count: 950 → "950", 12_400 → "12.4K", 3_200_000 → "3.2M".
+// Sub-1000 stays exact; thousands/millions get one decimal, trailing .0 trimmed.
+function abbrevTokens(n: number): string {
+  if (!Number.isFinite(n) || n < 1000) return String(Math.max(0, Math.round(n)));
+  if (n < 1_000_000) return `${(n / 1000).toFixed(1).replace(/\.0$/, '')}K`;
+  return `${(n / 1_000_000).toFixed(1).replace(/\.0$/, '')}M`;
+}
+
 // Relative-time formatter for the resumed-chip title (e.g. "3m ago").
 function relativeTime(iso: string | null): string {
   if (!iso) return 'unknown';
@@ -353,6 +361,21 @@ export function AgentComposerBar({ projectPath, value, onChange, onProviderComma
           : { label: 'new session', title: `${value.provider} starts a fresh session this turn` }
       : null;
 
+  // sampling-streamjson-tokens #8: condensed per-(chat,agent) token/cost readout
+  // beside the session chip. Coerce — input/output are BIGINT (string over wire).
+  // Hidden when no session row or all totals are zero (e.g. native boocode, which
+  // holds no agent_sessions row, or a provider that hasn't run yet).
+  const usageReadout = (() => {
+    if (!sessionChip || !sessionRow) return null;
+    const inTok = Number(sessionRow.input_tokens) || 0;
+    const outTok = Number(sessionRow.output_tokens) || 0;
+    const cost = Number(sessionRow.cost) || 0;
+    if (inTok <= 0 && outTok <= 0 && cost <= 0) return null;
+    const parts = [`${abbrevTokens(inTok)} in`, `${abbrevTokens(outTok)} out`];
+    if (cost > 0) parts.push(`$${cost.toFixed(2)}`);
+    return parts.join(' · ');
+  })();
+
   return (
     <div className="flex flex-wrap items-center gap-1 px-2 py-1 border-b border-border bg-muted/20 shrink-0">
       <CompactPicker
@@ -374,6 +397,14 @@ export function AgentComposerBar({ projectPath, value, onChange, onProviderComma
           {sessionChip.label}
         </span>
       )}
+      {usageReadout && (
+        <span
+          className="text-[10px] text-muted-foreground tabular-nums whitespace-nowrap shrink-0"
+          title="Tokens in · out · cost for this agent session"
+        >
+          {usageReadout}
+        </span>
+      )}
       <CompactPicker
         label="Mode"
         value={value.modeId ?? ''}

apps/web/src/components/MessageBubble.tsx

@@ -1,6 +1,6 @@
 import { useEffect, useState } from 'react';
 import type { ReactNode } from 'react';
-import { ChevronDown, ChevronRight, Copy, RefreshCw, Check, Share2, RotateCw, GitFork, Trash2, Brain, History } from 'lucide-react';
+import { ChevronDown, ChevronRight, Copy, RefreshCw, Check, Share2, RotateCw, GitFork, Trash2, Brain, History, AlertCircle } from 'lucide-react';
 import { toast } from 'sonner';
 import type { Chat, ErrorReason, Message } from '@/api/types';
 import { api } from '@/api/client';
@@ -637,6 +637,76 @@ function ReasoningBlock({ text, streaming }: { text: string; streaming: boolean
   );
 }
 
+// feature #12: mistake-recovery sentinel. Inserted by the backend as a
+// role='system', metadata.kind='mistake_recovery' row when the model hit
+// repeated *different* errors (distinct from doom_loop, which is the same
+// call repeated). Visual treatment mirrors CapHitSentinel / DoomLoopSentinel
+// (amber card + alert icon). Non-escalated → recovery guidance was injected
+// and the turn continues. Escalated → the turn was stopped; if can_continue
+// is set, offer the same Continue affordance as the cap-hit sentinel.
+// Loose `!= null` guards per the CLAUDE.md coder-message note (coder rows pass
+// metadata as undefined, not null).
+function MistakeRecoverySentinel({ message }: { message: Message }) {
+  const meta = message.metadata;
+  const isMistakeRecovery =
+    meta != null && typeof meta === 'object' && meta.kind === 'mistake_recovery';
+  const failureKinds = isMistakeRecovery ? meta.failure_kinds : [];
+  const escalated = isMistakeRecovery ? meta.escalated : false;
+  const canContinue = isMistakeRecovery ? meta.can_continue === true : false;
+
+  const [continuing, setContinuing] = useState(false);
+
+  async function handleContinue() {
+    if (continuing || !canContinue) return;
+    setContinuing(true);
+    try {
+      await api.chats.continue(message.chat_id, message.id);
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : 'continue failed');
+    } finally {
+      setContinuing(false);
+    }
+  }
+
+  const kindsLabel =
+    Array.isArray(failureKinds) && failureKinds.length > 0
+      ? failureKinds.join(', ')
+      : null;
+
+  return (
+    <div className="rounded-md border border-amber-500/40 bg-amber-500/10 text-sm">
+      <div className="px-3 py-2 flex items-start gap-2">
+        <AlertCircle className="size-4 text-amber-500 shrink-0 mt-0.5" />
+        <div className="flex-1 min-w-0 space-y-1">
+          <div className="text-xs font-medium text-amber-700 dark:text-amber-300">
+            {escalated ? 'Repeated errors — turn stopped' : 'Recovering from repeated errors'}
+          </div>
+          <div className="text-xs text-muted-foreground">
+            {escalated
+              ? 'Repeated errors persisted — stopped the turn.'
+              : kindsLabel
+                ? `Hit repeated different errors (${kindsLabel}) — recovery guidance injected, continuing.`
+                : 'Hit repeated different errors — recovery guidance injected, continuing.'}
+          </div>
+          {escalated && canContinue && (
+            <div className="pt-1">
+              <Button
+                type="button"
+                size="sm"
+                variant="outline"
+                onClick={() => void handleContinue()}
+                disabled={continuing}
+              >
+                {continuing ? 'Continuing…' : 'Continue'}
+              </Button>
+            </div>
+          )}
+        </div>
+      </div>
+    </div>
+  );
+}
+
 export function MessageBubble({
   message,
   sessionChats,
@@ -681,6 +751,13 @@ export function MessageBubble({
     return <DoomLoopSentinel message={message} />;
   }
 
+  // feature #12: mistake-recovery sentinel. Non-escalated rows narrate that
+  // recovery guidance was injected mid-turn; escalated rows report the turn
+  // was stopped and (when can_continue) offer the cap-hit-style Continue.
+  if (message.role === 'system' && message.metadata?.kind === 'mistake_recovery') {
+    return <MistakeRecoverySentinel message={message} />;
+  }
+
   // v1.8.2: tool messages and assistant tool_calls are now rendered by
   // MessageList via ToolCallLine / ToolCallGroup. Tool-role messages reach
   // this point only if MessageList didn't consume them (shouldn't happen,

data/AGENTS.md

@@ -6,6 +6,10 @@ Operating rules for every agent in this registry. Full procedures live in the `c
 
 **Worktrees** — Isolate work in a worktree when it is parallel to in-progress work, risky/experimental, a hotfix interrupting other work, or splits into independent units — just create when clear, propose in one line when ambiguous, skip quick/small single-stream work. Branch from a stable base (default branch); worktrees persist (never auto-remove or auto-merge); they isolate code state, not runtime (ports/DBs/services still collide). Full heuristic: invoke `using-worktrees`.
 
+**Sampling knobs** — Each `## Name` frontmatter block accepts these per-agent sampler fields, threaded into the llama-swap chat-completion request: `temperature`, `top_p`, `top_k`, `min_p`, `presence_penalty`, and (v2.6) `top_n_sigma`, `dry_multiplier`, `dry_base`, `dry_allowed_length`, `dry_penalty_last_n`. The `top_n_sigma` + `dry_*` repetition family curb the doom-loop-prone local model. Omit a field to leave it at the server default. Example: `top_n_sigma: 1.0`, `dry_multiplier: 0.8`, `dry_base: 1.75`, `dry_allowed_length: 2`, `dry_penalty_last_n: -1` (-1 = whole context).
+
+**Reasoning budget** — To cap a reasoning model's thinking tokens, pass `--reasoning-budget` through `llama_extra_args` (already permitted by the deny-list validator; routes the agent to llama-sidecar). Example frontmatter line: `llama_extra_args: ["--reasoning-budget", "2048"]`. This is a sidecar process flag, not a chat-completion body param — distinct from the sampling knobs above.
+
 ## Code Reviewer
 ---
 temperature: 0.6

openspec/changes/mistake-tracker-file-ledger/proposal.md

@@ -0,0 +1,70 @@
+# MistakeTracker + file-provenance ledger (#12)
+
+**Status:** in progress (started 2026-06-01)
+**Source:** `boocode_code_review_v2.md` §1 #12, §5e (cline — algorithm-reimplemented, not vendored).
+
+Two native-inference (apps/server) hardening features. One cohesive backend change (they share
+`TurnArgs` + the tool-phase observation point) + a small frontend sentinel render.
+
+## Part A — MistakeTracker (heterogeneous-failure recovery)
+
+Complements the doom-loop guard (`sentinels.ts:detectDoomLoop`, which only catches *identical*
+repeats) by catching a run of consecutive tool **failures** the model isn't recovering from.
+
+- New pure `apps/server/src/services/inference/mistake-tracker.ts` (mirrors `detectDoomLoop`):
+  - `FailureKind = 'zod_reject' | 'tool_not_found' | 'exec_error' | 'api_error' | 'permission_denied'`
+    (all already distinguished in `tool-phase.ts:executeToolCall`).
+  - `MISTAKE_THRESHOLD = 3`.
+  - State `{ run: FailureKind[]; nudges: number }` — `run` is the current consecutive-failure streak,
+    reset on ANY successful tool step; `nudges` counts recovery injections not yet cleared by a success.
+  - `recordStep(state, outcome)` where outcome is a failure kind or `'success'`.
+  - `detectMistakePattern(state): 'nudge' | 'escalate' | null` — `run.length >= 3` → `'nudge'` the first
+    time (`nudges === 0`), `'escalate'` if it trips again while `nudges >= 1` (no intervening success).
+- Lives in `TurnArgs` (loop-local, reset per `runInference`, like `recentToolCalls`).
+- Integration in `turn.ts` loop: after each tool phase, `recordStep` per tool outcome; then
+  `detectMistakePattern`:
+  - `'nudge'` (decision: soft + escalate): append a transient **model-facing** recovery-guidance system
+    message to the NEXT turn's payload (re-read schemas, verify paths exist before acting, try a
+    different approach — not retry variations), insert a `mistake_recovery` UI sentinel
+    (`escalated:false`), bump `nudges`, reset `run`. Loop continues.
+  - `'escalate'`: stop the turn (break), insert a `mistake_recovery` sentinel (`escalated:true`,
+    `can_continue:true`, cap-hit-style), finalize. Prevents heterogeneous failures from burning the
+    whole step budget.
+
+## Part B — File-provenance ledger (Read-only)
+
+- Accumulate file paths read by `view_file`/`grep`/`find_files`/`list_dir` into `TurnArgs.filesRead:
+  Set<string>` (recorded at the tool-phase, like the failure outcomes).
+- On compaction (`compaction.ts:buildPrompt`), inject a deterministic, sorted `## Files Read` list into
+  the summary prompt context so the summarizer merges it into the rolling summary — **no new
+  table/column**; it propagates as summary text across compactions. `compaction-prompt.ts`'s
+  `SUMMARY_TEMPLATE` already has a `## Relevant Files` section to extend/merge with.
+- BooChat is **read-only** (no write tools on apps/server) → "Files Modified" is N/A here; only
+  "Files Read". (The apps/coder write side can add "Modified" later.)
+
+## Sentinel contract (pinned — backend + frontend must match)
+
+New sentinel kind on `MessageMetadata` in BOTH `apps/server/src/types/api.ts` AND
+`apps/web/src/api/types.ts`:
+```
+{ kind: 'mistake_recovery'; failure_kinds: string[]; count: number; escalated: boolean; can_continue?: boolean }
+```
+- `role='system'`, `status='complete'`, stripped from the LLM payload via `isAnySentinel` in
+  `payload.ts` (UI-only) and `compaction.ts:buildHeadPayload`.
+- Frontend render branch in `apps/web/src/components/MessageBubble.tsx`: `escalated:false` →
+  "Hit repeated different errors — recovery guidance injected, continuing." `escalated:true` →
+  "Repeated errors persisted — stopped the turn." (mirror the doom-loop/cap-hit branches).
+
+## Decisions (2026-06-01)
+- MistakeTracker intervention: **soft nudge + escalate**.
+- **UI sentinel** for recovery (`mistake_recovery`).
+
+## Files (backend, one agent) / (frontend, one agent)
+- Backend: `mistake-tracker.ts` (new), `turn.ts`, `tool-phase.ts`, `sentinels.ts`,
+  `sentinel-summaries.ts`, `payload.ts`, `compaction.ts`, `compaction-prompt.ts`, `types/api.ts` +
+  tests (`mistake-tracker.test.ts`, ledger/compaction assertions).
+- Frontend: `apps/web/src/api/types.ts` (MessageMetadata arm) + `MessageBubble.tsx` (render branch).
+  MUST NOT touch Sam's WIP web files.
+
+## Verify
+- `pnpm -C apps/server test`; `pnpm -C apps/server build`; `npx tsc -p apps/web/tsconfig.app.json --noEmit`

openspec/changes/sampling-streamjson-tokens/proposal.md

@@ -0,0 +1,45 @@
+# Small wins — sampling knobs + PTY stream-json + token UI
+
+**Status:** in progress (started 2026-06-01)
+**Source:** `boocode_code_review_v2.md` §1 #11 / #7 / #8 (config-adopt + qwen-code §5g + opencode §3 #4).
+
+Three independent BooCode improvements, disjoint subsystems (apps/server / apps/coder / apps/web).
+
+## #11 — New sampling knobs (apps/server)
+Per-agent `top_n_sigma` + the `dry_*` repetition family help the doom-loop-prone local model.
+Today the Agent type threads `temperature/top_p/top_k/min_p/presence_penalty` into the inference
+request (`stream-phase.ts:396–438`). Add `top_n_sigma`, `dry_multiplier`, `dry_base`,
+`dry_allowed_length`, `dry_penalty_last_n` as first-class Agent fields (`types/api.ts`), parse them in
+`agents.ts:parseFrontmatter` (same bounded per-field numeric pattern + out-of-range warn), and thread
+them into the request body **via the same mechanism `top_k`/`min_p` already use** (the agent must
+confirm whether that's an AI-SDK `providerOptions`/`extraBody` passthrough — these are llama.cpp
+extensions, not standard OpenAI fields — and ride it; surface it if `top_k`/`min_p` turn out to be
+silently dropped today). `--reasoning-budget` is a llama-server CLI flag already permitted by the
+deny-list validator, so it works via `llama_extra_args: ["--reasoning-budget","N"]` now — document it
+in `data/AGENTS.md`. apps/server only.
+
+## #7 — Live PTY stream-json NDJSON parsing (apps/coder)
+qwen/claude PTY dispatch slices stdout opaque (`dispatcher.ts` PTY path; qwen already runs
+`--output-format stream-json`). Add a parser for the Claude-Code-compatible NDJSON
+(`system`/`assistant`/`result`/`stream_event` → `content_block_delta` text/thinking/tool deltas +
+`usage` + `session_id`) that maps to the existing `AgentEvent` union (`agent-backend.ts`). **Live
+incremental** (decision 2026-06-01): line-buffer the PTY stdout `data` events, parse each complete
+NDJSON line as it arrives, and emit broker frames live (text/reasoning/tool) like the ACP/opencode
+paths — plus accumulate for `persistExternalAgentTurn`. claude gets `--output-format stream-json` too.
+One parser serves both (same schema). apps/coder only (`pty-dispatch.ts`, `dispatcher.ts`, new
+`stream-json-parser.ts` + test).
+
+## #8 — Surface opencode token usage (apps/coder route + apps/web)
+`agent_sessions.input_tokens/output_tokens/cost` are accumulated (v2.6.8) but the
+`GET /api/sessions/:id/agent-sessions` SELECT + the `AgentSessionInfo` type drop them. Add the 3
+columns to both, render condensed beside the existing session chip in `AgentComposerBar`
+(ChatThroughput styling: `tabular-nums`, muted, e.g. "12.4K in / 3.2K out / $0.25"). MUST NOT touch
+Sam's uncommitted WIP (`ChatTabBar`, `SessionLandingPage`, `Workspace`, `useWorkspacePanes`,
+`PaneHeaderActions`).
+
+## Decisions (2026-06-01)
+- #7 surfacing: **live incremental** streaming (not parse-at-end).
+
+## Verify
+- `pnpm -C apps/server test` (+ new agent-parse tests); `pnpm -C apps/coder test` (+ new parser tests)
+- `pnpm -C apps/server build && pnpm -C apps/coder build`; `npx tsc -p apps/web/tsconfig.app.json --noEmit`