thincoder 0.12.8 → 0.12.10

package/README.md

@@ -209,6 +209,19 @@ Code conventions: pure `.mjs`, no semicolons, no npm dependencies allowed (inclu
 
 ## Changelog
 
+### 0.12.10 (2026-08)
+- **Code-quality pass (advisor subsystem):**
+  - **Drop 11 unused exports** — internal-use symbols no longer leak through the module API (advisor table headers/constants, plan reminders, token-UUID helper, shrinkOversized).
+  - **Advisor tool-timeout timer is now cleared** when the tool wins the race (no dangling timers); static import replaces a dynamic import on the hot path.
+  - **Advisor re-review no longer trusts git output** — the follow-up path previously injected a `git diff HEAD` snapshot: once fixes were committed the diff was empty and the model read "no changes" as "no fixes", misreporting fixed items as unfixed. The follow-up now injects **no git information at all** — verification is `read`-only, evidence must quote this round's read output (line numbers from the stale prior table are not evidence), and dead snapshot-dedup fields were removed.
+
+### 0.12.9 (2026-08)
+- **Prompt-system quality pass (both CLI and VS Code extension, byte-identical sync):**
+  - **Advisor-after-code rule moved from system.md to discipline.md** — engineering mode no longer receives the conflicting "call advisor after changing code" instruction (its review-timing rules say do not call unprompted). Standard mode behavior unchanged.
+  - **engineering.md delivery-review semantics unified** — the mandatory-flow step and the state table now both say: eng-coder self-reviews inside the subagent; the architect verifies against acceptance criteria and re-reviews only when asked or when the delivery looks wrong. (Previously the step forced a parent-side advisor code review that the hard rules contradicted.)
+  - **checkpoint description aligned with actual auto-snapshot triggers** (task-list deletion + context compaction; manual checkpoint for the rest).
+  - **advisor round budget wording fixed** — prompts advertise a 30-round budget; the mechanical hard cap is 100 rounds (loop guard). Both layers are now named explicitly.
+
 ### 0.12.8 (2026-08)
 - **Fix: pending-task pushback fires at most once** — the completion guard that reminds the model to update pending tasks before finishing could loop forever when a pending item could not be resolved. Now each task-list state earns exactly one reminder; if the model insists on finishing anyway, it is allowed to (updating the list via the task tool resets the budget). VS Code extension synced.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thincoder",
-  "version": "0.12.8",
+  "version": "0.12.10",
   "description": "Thin coding agent - zero dependencies, no build step, Node.js native. Sharp code, zero bloat.",
   "keywords": [
     "ai",

package/src/advisor/history.mjs

@@ -6,8 +6,8 @@ import { join } from "node:path"
 
 export const ADVISOR_MD_PATH = ".thincoder/advisor.md"
 export const ADVISOR_TABLE_HEADER = "| # | File | Severity | Issue | Suggestion |"
-export const CONVERGENCE_TABLE_HEADER = "| # | Orig# | File | Severity | Status | Notes |"
-export const AGENT_RESPONSE_HEADER = "| # | Action | Detail |"
+const CONVERGENCE_TABLE_HEADER = "| # | Orig# | File | Severity | Status | Notes |"
+const AGENT_RESPONSE_HEADER = "| # | Action | Detail |"
 export const LEGACY_ADVISOR_HEADER = "| # | 文件 | 严重程度 | 问题描述 | 建议修复 |"
 
 const DEFAULT_CRITERIA = `Review the code changes, focusing on:

package/src/advisor/messages.mjs

@@ -57,7 +57,6 @@ export function buildAdvisorUserMessage(agent, _prior, reviewType, designToken =
       }
 
       // Pre-collected changes — the design doc diff.
-      // _advisorLastSnapshot is only consumed by code-review convergence — skip the write here.
       const snapshots = collectRepoSnapshots(repos, agent.cwd)
       if (snapshots.length > 0) {
         parts.push("## Design Document (git diff)")
@@ -96,7 +95,7 @@ export function buildAdvisorUserMessage(agent, _prior, reviewType, designToken =
     return parts.join("\n")
   }
 
-  // Convergence data (round 2+)
+  // Convergence data (round 2+). Design reviews returned above — only code reviews reach here.
   if (prior && (agent._advisorRound || 0) > 0) {
     const response = extractAgentResponseTable(agent.history, prior.sinceIdx)
       || "(Agent did not provide a response table — re-evaluate each issue)"
@@ -164,10 +163,11 @@ export function buildAdvisorUserMessage(agent, _prior, reviewType, designToken =
   parts.push("## Instructions")
   parts.push("1. IMPORTANT: in the diff, `-` lines are REMOVED content (no longer in the file), `+` lines are ADDED. The prior issue table (if any) is HISTORY — always verify current file state with `read` before judging an item.")
   if (isReReview) {
-    parts.push("2. STALE-CONTEXT WARNING: any diff embedded in earlier messages is a historical snapshot — treat it as expired. Only the \"Current Changes\" section above and fresh `read` results describe the current state. Never quote a `-` line from any diff as if it were live code.")
-    parts.push("3. Do NOT re-read AGENTS.md / design docs — conventions were established in round 1. Focus on verifying the prior issue table against the current diff.")
-    parts.push("4. `read` only the files touched by the fixes. Batch independent reads/greps in a single reply.")
-    parts.push("5. Produce your verification table. Do not re-read content you already have.")
+    parts.push("2. STALE-CONTEXT WARNING: any diff or file content embedded in earlier messages is a historical snapshot — treat it as expired. Only fresh `read` results describe the current state. Never quote a `-` line from an earlier diff as if it were live code.")
+    parts.push("3. Verify the prior issue table against the CURRENT FILE STATE — use `read`, never `git diff` alone. Fixes may already be committed: an empty `git diff` does NOT mean nothing changed. `git log -3` shows recent commits.")
+    parts.push("4. `read` the files in the Review Scope in full — ALWAYS, regardless of what `git diff` shows. Batch reads/greps in a single reply.")
+    parts.push("5. Evidence rule: every 'Unfixed'/'New' finding MUST quote the exact line content from THIS round's `read` output (e.g. `run.mjs:180: timeoutId = setTimeout(...)`). Line numbers alone are NOT evidence — they may come from the stale prior table. Findings without a fresh quoted line are treated as unverified and will not be accepted.")
+    parts.push("6. Produce your verification table. Do not re-read content you already have.")
   } else {
     parts.push("2. Read `AGENTS.md` / design docs only if they exist (check once; do not re-probe with multiple patterns).")
     parts.push("3. `read` changed files for full context beyond the diff. Batch independent reads/greps in a single reply instead of one call per round-trip.")

package/src/advisor/repos.mjs

@@ -5,8 +5,8 @@
 import { execFileSync } from "node:child_process"
 import { dirname, basename, resolve } from "node:path"
 
-export const GIT_TIMEOUT = 5_000
-export const MAX_EMBEDDED_DIFF = 50_000
+const GIT_TIMEOUT = 5_000
+const MAX_EMBEDDED_DIFF = 50_000
 
 /**
  * Find the git repository roots that contain the agent's touched files.

package/src/advisor/run.mjs

@@ -8,11 +8,14 @@ import { toOpenAISchema } from "../tools/index.mjs"
 import { prepareAdvisorMessages } from "../advisor.mjs"
 import { extractPriorIssueTable } from "../advisor/history.mjs"
 
-export const MAX_ADVISOR_TURNS = 100
+const MAX_ADVISOR_TURNS = 100
 // Mechanical convergence cap: the protocol assumes up to 5 rounds suffice
 // (full review, verify+fix cycles, strict verification). A 6th call means the
 // model is looping — refuse it instead of burning tokens on a review that cannot
 // converge. Design reviews are exempt (each call resets the round).
+// NOTE: prompts/advisor-round{1,2,3}.md advertise a 30-round BUDGET — the
+// prompt-level efficiency target, distinct from this 100-round mechanical hard
+// cap (loop guard). Keep both in sync when either changes.
 export const MAX_ADVISOR_ROUNDS = 5
 
 // Context window limits
@@ -169,18 +172,21 @@ async function runAdvisorToolLoop(provider, messages, onOutput, signal, agent, c
       if (!tool) {
         result = `Error: unknown tool "${tc.name}". Available: ${[...ADVISOR_TOOL_BY_NAME.keys()].join(", ")}`
       } else {
-        // Execute with timeout
+        // Execute with timeout (clear the timer when the tool wins the race —
+        // otherwise up to MAX_ADVISOR_TURNS dangling timers accumulate)
         try {
-          const toolPromise = tool.execute(args, {
-            cwd,
-            agent,
-            onOutput,
-            signal,
+          let timeoutId
+          const timeoutPromise = new Promise((_, reject) => {
+            timeoutId = setTimeout(() => reject(new Error(`tool timeout after ${TOOL_TIMEOUT_MS}ms`)), TOOL_TIMEOUT_MS)
           })
-          const timeoutPromise = new Promise((_, reject) =>
-            setTimeout(() => reject(new Error(`tool timeout after ${TOOL_TIMEOUT_MS}ms`)), TOOL_TIMEOUT_MS)
-          )
-          result = await Promise.race([toolPromise, timeoutPromise])
+          try {
+            result = await Promise.race([
+              tool.execute(args, { cwd, agent, onOutput, signal }),
+              timeoutPromise,
+            ])
+          } finally {
+            clearTimeout(timeoutId)
+          }
         } catch (e) {
           const errorType = e.message.includes("timeout") ? "timeout"
             : e.message.includes("ENOENT") ? "file_not_found"
@@ -221,7 +227,7 @@ async function runAdvisorToolLoop(provider, messages, onOutput, signal, agent, c
 }
 
 /** Resolve the advisor's provider: cfg.provider/model when set, otherwise the main agent's provider */
-export function resolveAdvisorProvider(agent) {
+function resolveAdvisorProvider(agent) {
   const cfg = agent.config?.advisor
   if (cfg?.provider) {
     try {

package/src/advisor.mjs

@@ -37,8 +37,7 @@
 import { readFileSync } from "node:fs"
 import { join, dirname } from "node:path"
 import { fileURLToPath } from "node:url"
-import { createHash } from "node:crypto"
-import { findReviewRepos, collectRepoSnapshots } from "./advisor/repos.mjs"
+import { findReviewRepos } from "./advisor/repos.mjs"
 import { extractPriorIssueTable, extractAgentResponseTable } from "./advisor/history.mjs"
 import { buildAdvisorUserMessage } from "./advisor/messages.mjs"
 // Re-export for run.mjs and tests (keeps their imports from "../advisor.mjs" stable)
@@ -111,31 +110,18 @@ export function buildAdvisorFollowUp(agent, _prior) {
     "## Instructions",
     round === 2
       ? "Verify each item in the prior table. Flag any obvious NEW issues introduced by the fixes (crashes, data loss, logic errors — not style). Produce a verification table."
-      : "Strictly verify ONLY the items in the prior table against the current diff. Do NOT look for new issues.",
+      : "Strictly verify ONLY the items in the prior table against the CURRENT FILE STATE (use `read` — an empty diff does not mean the fixes are absent). Do NOT look for new issues.",
     "",
-    "IMPORTANT: in any embedded diff, `-` lines are REMOVED content (no longer in the file), `+` lines are ADDED. The prior issue table is HISTORY — always verify current file state with `read` before judging an item as fixed or unfixed.",
+    "IMPORTANT: the prior issue table is HISTORY — always verify current file state with `read` before judging an item as fixed or unfixed.",
     // Round-aware evidence rule: "New" entries only exist in round 2 (round 3+ forbids them).
-    `STALE-CONTEXT WARNING: all diffs in earlier messages (including round 1) are historical snapshots — files have changed since. Only THIS message's "Current Changes" section and fresh \`read\` results are authoritative. Any "Unfixed" entry${round === 2 ? ' (and any "New" entry)' : ""} MUST cite read-verified evidence (file:line from a \`read\` of the current file); uncited findings are unverified and will be ignored.`,
+    `STALE-CONTEXT WARNING: only fresh \`read\` results describe the current state — never judge from earlier snapshots or from \`git diff\` (committed fixes never show in \`git diff HEAD\`). Read the files to verify. Any "Unfixed" entry${round === 2 ? ' (and any "New" entry)' : ""} MUST quote the exact line content from THIS round's \`read\` output (e.g. \`run.mjs:180: timeoutId = setTimeout(...)\`); line numbers alone are NOT evidence (they may come from the stale prior table). Uncited findings are unverified and will be ignored.`,
     "",
-    "Do NOT re-read AGENTS.md / design docs or re-run git status/diff (current changes are below) — you already have full context from previous rounds.",
+    "Do NOT re-read AGENTS.md / design docs. Verify fix status with \`read\` only — do not rely on git output: a clean working tree does not mean fixes are absent (they may be committed).",
     "",
   ]
-  const snapshots = collectRepoSnapshots(findReviewRepos(agent), agent.cwd)
-  const snapshotText = snapshots.join("\n")
-  const snapshotHash = snapshotText ? createHash("sha1").update(snapshotText).digest("hex") : null
-  // Skip re-pushing an identical diff (e.g. advisor re-run without any file changes) —
-  // the previous snapshot is already in the conversation, duplicating it wastes tokens.
-  if (snapshots.length === 0) {
-    parts.push("## Current Changes", "(No git repository or no changes detected.)")
-  } else if (snapshotHash && snapshotHash === agent._advisorLastSnapshotHash) {
-    // Hash match → the snapshot from the previous round is still current.
-    // Using hash instead of full-text comparison avoids keeping the entire diff
-    // string in memory and handles edge cases (e.g. file changed then reverted).
-    parts.push("## Current Changes", "(No changes since your previous review — the diff snapshot is identical, so the one from your last round remains valid for this round.)")
-  } else {
-    parts.push("## Current Changes (git status + git diff HEAD, refreshed)", ...snapshots)
-  }
-  agent._advisorLastSnapshotHash = snapshotHash
+  // Deliberately NO git information injected here (no diff snapshot, no git context):
+  // git output misled re-reviews — committed fixes never show in `git diff HEAD`, so
+  // the model read "no changes" as "no fixes". Verification is `read`-only by design.
   return parts.join("\n")
 }
 
@@ -165,7 +151,6 @@ export function prepareAdvisorMessages(agent, reviewType, designToken = null, do
     // a follow-up "Verify Prior Table" would be meaningless; start a fresh full review
     if (!prior) {
       agent._advisorSession = null
-      agent._advisorLastSnapshotHash = null
       session = null
       // Only reset the round counter on a truly fresh start (no prior reviews at all).
       // If _advisorRound > 0, there WAS a prior review — it just passed (all-clear).
@@ -188,7 +173,6 @@ export function prepareAdvisorMessages(agent, reviewType, designToken = null, do
   // Fresh session. Only reset round if this is truly the first review.
   // If _advisorRound > 0, there was a prior review that passed (all-clear).
   if (!agent._advisorRound) agent._advisorRound = 0
-  agent._advisorLastSnapshot = null
   if (!prior) {
     // Tell the advisor why no prior issue table is present
     session[1] = {

package/src/agent-tools/advisor.mjs

@@ -5,6 +5,7 @@
  */
 import { randomUUID, createHmac } from "node:crypto"
 import { runAdvisorReview } from "../advisor/run.mjs"
+import { isDocFile } from "../advisor/repos.mjs"
 
 const TOKEN_EXPIRY_MS = 3600000 // 1 hour
 const TOKEN_SECRET = process.env.THINCODER_TOKEN_SECRET || "thincoder-default-secret"
@@ -50,7 +51,7 @@ export function validateDesignToken(token) {
 }
 
 /** Extract UUID from signed token for regex matching */
-export function extractTokenUUID(token) {
+function extractTokenUUID(token) {
   const parts = token.split(":")
   return parts.length >= 1 ? parts[0] : token
 }
@@ -118,7 +119,6 @@ export const advisorTool = {
 
     // Design review: validate that documents are in docs/ or are recognized doc files
     if (reviewType === "design" && documents) {
-      const { isDocFile } = await import("../advisor/repos.mjs")
       const invalidDocs = documents.filter((doc) => {
         // Allow docs/ directory and recognized doc files (METHODOLOGY.md, README.md, etc.)
         if (doc.startsWith("docs/") || doc.startsWith("docs\\")) return false
@@ -133,7 +133,6 @@ export const advisorTool = {
     if (reviewType === "design") {
       agent._advisorRound = 0
       agent._advisorSession = null
-      agent._advisorLastSnapshotHash = null // stale diff dedup baseline must not leak into the next code review
     }
 
     // Generate the design token BEFORE the review and inject it into the advisor's prompt.

package/src/agent-tools/plan.mjs

@@ -8,17 +8,17 @@
  * user sends a new message — so the constraint never fades from context.
  */
 
-export const PLAN_FULL_REMINDER =
+const PLAN_FULL_REMINDER =
   "[System reminder: plan mode is ON. Workflow: (1) explore/read codebase with read-only tools, " +
   "(2) design a solution considering trade-offs, (3) present your plan by calling plan with action='exit' " +
   "so the user can approve it. Only read-only tools are allowed — do not write, edit, or run mutation commands. " +
   "Your turn must end with either a clarifying question to the user or a call to plan with action='exit'.]"
 
-export const PLAN_SPARSE_REMINDER =
+const PLAN_SPARSE_REMINDER =
   "[System reminder: plan mode still active — read-only tools only (the current plan file exempt). " +
   "Design the solution, then call plan with action='exit' for user approval.]"
 
-export const PLAN_EXIT_REMINDER =
+const PLAN_EXIT_REMINDER =
   "[System reminder: plan mode is now OFF. Start implementing your plan — edit files, run commands. " +
   "No need for a task list (plan already covered that) or further confirmation.]"
 

package/src/agent-tools/subagent.mjs

@@ -180,11 +180,10 @@ export function mergeChildMutations(parent, child) {
     parent._verifiedThisRun = false
     parent._verifyPassed = undefined
   }
-  // Fresh code → fresh convergence budget + stale session/diff cleanup.
+  // Fresh code → fresh convergence budget + stale session cleanup.
   // _advisorRound reset ensures new code gets a full round-1 review;
-  // _advisorSession + _advisorLastSnapshotHash prevent cross-contamination.
+  // _advisorSession prevents cross-contamination between reviews.
   parent._advisorRound = 0
   parent._advisorSession = null
-  parent._advisorLastSnapshotHash = null
   return true
 }

package/src/agent.mjs

@@ -100,7 +100,7 @@ export function createAgent({
     _mutatedThisRun: false, _verifiedThisRun: false, _verifyPassed: undefined, _calledAdvisorThisRun: false,
     _engDesignReviewed: false, // eng-coder: design review gate passed (hard gate in dispatch.mjs)
     _engDesignToken: null, // issued by advisor(type="design"); required to spawn eng-coder
-    _touchedFiles: [], _verifyRetries: 0, _advisorRound: 0, _advisorSession: null, _advisorLastSnapshotHash: null,
+    _touchedFiles: [], _verifyRetries: 0, _advisorRound: 0, _advisorSession: null,
     _lastEngState: false,
     _pendingReminders: [],
     _pendingTimers: [],
@@ -131,7 +131,6 @@ export async function runAgent(agent, input, callbacks = {}, { depth = 0, signal
     agent._verifyRetries = 0
     agent._advisorRound = 0
     agent._advisorSession = null // advisor session is per-run: discard when the task ends, next task starts fresh
-    agent._advisorLastSnapshotHash = null // dedup baseline is per-run too — stale snapshot could wrongly suppress a diff refresh
     agent._emptyRetries = 0 // empty-response retry budget is per-run: a fresh user turn restarts from zero
     agent._compressFailures = 0 // compaction summary-failure counter is per-run: a fresh user turn restarts from zero
   }

package/src/context.mjs

@@ -237,7 +237,7 @@ const OVERSIZE_CONTENT_LIMIT = 8_000
  * does not touch reasoning_content (DeepSeek/Kimi echo protocol) or tool_calls pairing structure — no protocol 400 risk.
  * Only called after compressIfNeeded determines threshold is exceeded. Returns whether any message was truncated.
  */
-export function shrinkOversized(agent, limit = OVERSIZE_CONTENT_LIMIT) {
+function shrinkOversized(agent, limit = OVERSIZE_CONTENT_LIMIT) {
   let shrunk = false
   for (const m of agent.history) {
     if ((m.role !== "user" && m.role !== "tool") || typeof m.content !== "string") continue

package/src/prompts/advisor-round1.md

@@ -1,7 +1,7 @@
 You are a code review advisor.
 Perform a full-scope review of the specified files.
 You have read-only tools to explore the codebase.
-You have a HARD limit of 30 tool rounds (chat turns) total — plan your exploration accordingly.
+You have a budget of 30 tool rounds (chat turns) — plan your exploration accordingly. Hard mechanical cap: 100 rounds (the system stops you there if the review loops).
 
 Review workflow:
 1. The files to review are listed in the review scope. Read them in full. The review scope defines exactly which files to inspect.

package/src/prompts/advisor-round2.md

@@ -2,24 +2,24 @@ You are a code review advisor.
 Verify the prior issue table (provided in the review context).
 You may note obvious new issues introduced by the fixes.
 You have read-only tools to explore the codebase.
-You have a HARD limit of 30 tool rounds (chat turns) total.
+You have a budget of 30 tool rounds (chat turns). Hard mechanical cap: 100 rounds.
 
 Review workflow:
-1. The files to review are listed in the review scope — read them in full. The prior issue table is HISTORY from a previous review, not current state.
+1. The affected files are named in the prior issue table — read them in full. The prior issue table is HISTORY from a previous review, not current state.
 2. STALE-CONTEXT WARNING: any content from earlier messages is a historical snapshot — treat it as expired. Only fresh `read` results describe the current state.
 3. Project conventions were established in round 1 — do NOT re-read AGENTS.md / design docs unless a fix appears to contradict the task itself.
-4. Read the specified files for full context. **Batch independent tool calls in one reply.** ALWAYS verify current file content with `read` before judging a prior-table item as fixed or unfixed — never decide based on the prior table alone.
+4. **ALWAYS verify current file content with `read` before judging a prior-table item as fixed or unfixed — never decide based on the prior table alone.** An empty `git diff` does NOT mean nothing changed: fixes may already be committed (`git log -3` shows recent commits) — `read` the files named in the prior table regardless of the diff. Batch independent tool calls in one reply.
 5. Use grep or lsp to trace callers, imports, and dependencies — only where genuinely needed.
 6. Produce your review table.
 
-Budget: read only the files affected by the prior-table items. If at 15 rounds you have not yet verified all items, wrap up.
+Budget: read only the files named in the prior-table items. If at 15 rounds you have not yet verified all items, wrap up.
 
 Rules:
 - Respect the project's stated platform requirements — do not flag features as errors if they are valid under the project's target environment.
 - Primarily check fix status of items in the prior issue table.
 - For items marked "fixed": verify they were actually fixed.
 - For items marked "not an issue": evaluate whether the reasoning is sound.
-- Every "Unfixed" or "New" entry MUST cite read-verified evidence — file:line from a `read` of the CURRENT file (e.g. `src/x.mjs:42`). Findings without such evidence are treated as unverified and will not be accepted.
+- Every "Unfixed" or "New" entry MUST quote the exact line content from THIS round's `read` output (e.g. `run.mjs:180: timeoutId = setTimeout(...)`). Line numbers alone are NOT evidence — they may come from the stale prior table. Findings without a fresh quoted line are treated as unverified and will not be accepted.
 - You may flag obvious new problems — but only if clearly visible in the reviewed files and would cause crashes, data loss, or logic errors.
 - Do NOT nitpick style or naming.
 - Output a Markdown table listing all remaining problems (old or new):

package/src/prompts/advisor-round3.md

@@ -2,24 +2,24 @@ You are a code review advisor.
 Strictly verify only the prior issue table (provided in the review context).
 Do NOT look for new issues.
 You have read-only tools to explore the codebase.
-You have a HARD limit of 30 tool rounds (chat turns) total.
+You have a budget of 30 tool rounds (chat turns). Hard mechanical cap: 100 rounds.
 
 Review workflow:
-1. The files to review are listed in the review scope — read them in full. The prior issue table is HISTORY from a previous review, not current state.
+1. The affected files are named in the prior issue table — read them in full. The prior issue table is HISTORY from a previous review, not current state.
 2. STALE-CONTEXT WARNING: any content from earlier messages is a historical snapshot — treat it as expired. Only fresh `read` results describe the current state.
 3. Project conventions were established in round 1 — do NOT re-read AGENTS.md / design docs.
-4. Read the specified files for full context. **Batch independent tool calls in one reply.** ALWAYS verify current file content with `read` before judging a prior-table item as fixed or unfixed — never decide based on the prior table alone.
+4. **ALWAYS verify current file content with `read` before judging a prior-table item as fixed or unfixed — never decide based on the prior table alone.** An empty `git diff` does NOT mean nothing changed: fixes may already be committed (`git log -3` shows recent commits) — `read` the files named in the prior table regardless of the diff. Batch independent tool calls in one reply.
 5. Verify fix status of each item in the prior issue table.
 6. Produce your review table.
 
-Budget: read only the files affected by the prior-table items. If at 15 rounds you have not yet verified all items, wrap up.
+Budget: read only the files named in the prior-table items. If at 15 rounds you have not yet verified all items, wrap up.
 
 Rules:
 - Respect the project's stated platform requirements — do not flag features as errors if they are valid under the project's target environment.
 - Only check fix status of items in the prior issue table.
 - For items marked "fixed": verify they were actually fixed.
 - For items marked "not an issue": evaluate whether the reasoning is sound.
-- Every "Unfixed" entry MUST cite read-verified evidence — file:line from a `read` of the CURRENT file (e.g. `src/x.mjs:42`). Findings without such evidence are treated as unverified and will not be accepted.
+- Every "Unfixed" entry MUST quote the exact line content from THIS round's `read` output (e.g. `run.mjs:180: timeoutId = setTimeout(...)`). Line numbers alone are NOT evidence — they may come from the stale prior table. Findings without a fresh quoted line are treated as unverified and will not be accepted.
 - Output a Markdown table. Only list items that still have problems:
 | # | Orig# | File | Severity | Status | Notes |
 |---|-------|------|----------|--------|-------|

package/src/prompts/discipline.md

@@ -9,4 +9,7 @@ Debugging strategy:
 - Verify against official docs before guessing.
 - Binary search: cut the problem in half, test which half has the fault.
 - Fix one thing at a time. Don't change multiple things at once.
-- Don't get stuck reading code — write tests, add logs. Trust the runtime over your theories.
+- Don't get stuck reading code — write tests, add logs. Trust the runtime over your theories.
+
+Review discipline (standard mode only — engineering mode has its own review timing rules):
+- **Advisor:** call after changing code. Must provide scope: `paths` (files/dirs to review) or `documents` (context). Response table: `| # | Action | Detail |`. Round 2 verifies prior table.

package/src/prompts/engineering.md

@@ -30,13 +30,11 @@ subagents only.
    criteria — AND the designToken verbatim (the exact token string from the
    advisor output). The token is required — eng-coder cannot modify files
    without it.
-5. **Code review.** After eng-coder returns, call the `advisor` tool with
-   `type="code"` to review the implementation against the design. Pass
-   `documents=[...]` — the task's Docs involved list from the implementation
-   handoff. Review scope = these docs (not the git diff).
-   - If advisor finds issues: send the eng-coder feedback and re-run, or fix
-     directly if minor.
-   - If advisor approves: present results to the user.
+5. **Delivery review.** After eng-coder returns, verify the delivery against
+   the acceptance criteria from the design. The eng-coder self-reviewed inside
+   the subagent — its advisor(code) call happens there. Re-review with the
+   `advisor` tool (`type="code"`, `documents=[...]` = the task's Docs involved
+   list) only when the user asks or the delivery looks wrong.
 6. **Verify.** Run `verify` — it must pass before you claim the task complete.
 
 ## Work Loop (every user message)
@@ -51,7 +49,7 @@ passed?
 | Design | Write or refine the DESIGN doc (approach + rationale, architecture/interface, affected files, key decisions), organized by business domain per METHODOLOGY, ask for confirmation |
 | Awaiting approval | Present design summary, WAIT for explicit approval |
 | Implementation | eng-coder is working — do not redesign in parallel |
-| Delivery review | Review the eng-coder delivery with advisor (type="code", scope = Docs involved + acceptance criteria), check acceptance criteria, report |
+| Delivery review | Verify the delivery against the acceptance criteria (the eng-coder self-reviewed inside the subagent); re-review with advisor (type="code", documents = Docs involved) only when the user asks or the delivery looks wrong; report |
 | Wrapped up | Report, wait for next instruction |
 
 Then handle the message:

package/src/prompts/system.md

@@ -33,7 +33,7 @@ Programming is collaborative labor between you and the human. The human decides
 - No TTY — run shell commands non-interactively (git commit -m, --no-pager, -y/--yes).
 - Never modify files outside the working directory. No bash redirects to bypass boundaries.
 - **Reversibility tiers:** local edits — yours. Destructive (rm -rf, force-push) — confirm. Outward (commit/push/publish) — confirm each time.
-- Checkpoint before risky bulk operations. Auto-snapshot before every task lets you recover.
+- Checkpoint before risky bulk operations. Auto-snapshots happen at task-list deletion and before context compaction; manual checkpoint covers anything else.
 - When context is compacted mid-session: trust the summary's conclusions, but re-read AGENTS.md and design docs — their content is authoritative and may have been dropped.
 - Long-term memory via memory_put/memory_search. Save bugs, conventions, preferences.
 - Codebase exploration order: repo_outline → doc_search → code_search. Structure → intent → details.
@@ -52,5 +52,4 @@ Before finalizing: pause and think through edge cases. What could go wrong? Self
 - After every write/edit: `lint`. Before done: `lint full=true`.
 - Before declaring completion: `verify` (syntax, related tests, self-review checklist).
 - Code changes need at least one test.
-- **Advisor:** call after changing code. Must provide scope: `paths` (files/dirs to review) or `documents` (context). Response table: `| # | Action | Detail |`. Round 2 verifies prior table.
 - **Done:** explain what you changed, why, what's simplified, what's not done.

package/src/tui/agent-turn.mjs

@@ -29,6 +29,12 @@ export async function runAgentTurn(ctx, text) {
   // 可注入覆盖（测试用）；默认走真实实现
   const runAgentImpl = ctx.runAgent ?? runAgent
   const saveSessionImpl = ctx.saveSession ?? saveSession
+  // A new user message starts a new turn: auto-expanded completed replies from the
+  // previous turn (kept open so the user could read them) collapse now.
+  for (const idx of state._autoExpand ?? []) {
+    state.expandedBlocks?.delete(`long-${idx}`)
+  }
+  state._autoExpand = []
   pushLabel(`❯ You:`, ansi.bold + C.user)
   pushLine(text, C.text)
 
@@ -52,11 +58,23 @@ export async function runAgentTurn(ctx, text) {
 
   const flushStream = () => {
     if (state.reasoning) {
+      const idx = state.lines.length
       pushLine(state.reasoning, C.reason)
+      // Completed reasoning stays expanded (user is reading it) until the next turn
+      state.expandedBlocks ??= new Set()
+      state.expandedBlocks.add(`long-${idx}`)
+      state._autoExpand ??= []
+      state._autoExpand.push(idx)
       state.reasoning = ""
     }
     if (state.streaming) {
+      const idx = state.lines.length
       pushLine(state.streaming, C.text)
+      // Completed main output stays expanded (user is reading it) until the next turn
+      state.expandedBlocks ??= new Set()
+      state.expandedBlocks.add(`long-${idx}`)
+      state._autoExpand ??= []
+      state._autoExpand.push(idx)
       state.streaming = ""
     }
     state.advisorStreaming = ""

package/src/tui/ansi.mjs

@@ -42,5 +42,5 @@ export const C = {
   dim: ansi.gray,
   warn: ansi.fg(3),
   advisor: `${ESC}[92m`,  // bright green — visible on dark backgrounds
-  fold: `${ESC}[2m${ESC}[37m`,   // dim white — fold hints
+  fold: ansi.bold + ansi.fg(6),   // bold cyan — fold markers must stay visible on light AND dark themes (dim white vanished on light backgrounds)
 }

package/src/tui/clipboard.mjs

@@ -18,26 +18,6 @@ export async function readClipboardText() {
   }
 }
 
-/** Write text to the system clipboard (Set-Clipboard / pbcopy / xclip). Throws on failure. */
-export async function copyToClipboard(text) {
-  const { execFile } = await import("node:child_process")
-  const isWin = process.platform === "win32"
-  const isMac = process.platform === "darwin"
-  await new Promise((resolve, reject) => {
-    if (isWin) {
-      const child = execFile("powershell", ["-NoProfile", "-Command", "[Console]::In.ReadToEnd() | Set-Clipboard"], { timeout: 5000 }, (err) => err ? reject(err) : resolve())
-      child.stdin?.end(text)
-    } else if (isMac) {
-      const child = execFile("pbcopy", [], { timeout: 5000 }, (err) => err ? reject(err) : resolve())
-      child.stdin?.end(text)
-    } else {
-      const child = execFile("xclip", ["-selection", "clipboard"], { timeout: 5000 }, (err) => err ? reject(err) : resolve())
-      child.stdin?.end(text)
-    }
-  })
-}
-
-
 /** Insert pasted text into the active text target.
  *  Free-text question active → append to its answer (single-line field: newlines stripped).
  *  Options question active → ignore (no text field; must not leak into the input box).

package/src/tui/index.mjs

@@ -81,6 +81,7 @@ export async function startTUI(agent, opts = {}) {
     ctxCache: { len: -1, tokens: 0 }, // context utilization estimate cache (estimateTokens is O(n), only recompute when history grows)
     reasoning: "", // thinking stream buffer (dimmed display)
     completion: null, // Tab completion state { candidates, index }
+    _autoExpand: [], // indices of completed replies kept expanded; cleared when the next user turn starts
     subTasks: {}, // sub-agent panel: { roleName: { role, text, done } }, one line per role, marked done briefly after completion
     currentTool: null, // currently executing tool name (shown in status bar)
     processingStarted: 0, // current turn start time (status bar timer)
@@ -386,8 +387,8 @@ export async function startTUI(agent, opts = {}) {
     }
   })
 
-  // Mouse clicks (SGR \x1b[<0;col;rowM) — picker selection + line action menu.
-  const onMouseClick = (col, row) => handleMouseClick({ state, render, showPicker, popPicker, pushLine }, col, row)
+  // Mouse clicks (SGR \x1b[<0;col;rowM) — picker selection + fold expansion.
+  const onMouseClick = (col, row) => handleMouseClick({ state, render, popPicker }, col, row)
 
   // ---------------------------------------------------------- Startup screen + background indexing
 

package/src/tui/mouse.mjs

@@ -1,5 +1,5 @@
 /**
- * mouse.mjs — SGR mouse support: click parsing + hit-testing + line actions.
+ * mouse.mjs — SGR mouse support: click parsing + hit-testing.
  *
  * Protocol (enabled at startup via \x1b[?1000h\x1b[?1006h):
  *   press:    \x1b[<b;col;rowM   (b=0 left, 64/65 wheel up/down — wheel handled upstream)
@@ -8,10 +8,14 @@
  *
  * Only left-click (button 0) is consumed. Everything else stays stripped
  * upstream (sequence fragments must never leak into the input box).
+ *
+ * Click actions (deliberately minimal — a line-action menu was removed as
+ * over-engineering: terminals already copy via drag-select):
+ *   - picker option click = select it
+ *   - folded-block hint click = expand it
  */
 import { computeLayout } from "./layout.mjs"
 import { buildConvLines } from "./render-conversation.mjs"
-import { sanitizeDisplay } from "./render.mjs"
 
 /** Extract left-click presses from a chunk. Returns [{ col, row }] (1-based). */
 export function parseMouseClicks(text) {
@@ -37,7 +41,7 @@ export function convGlobalIndex(convLen, convH, scroll) {
 
 /**
  * Handle a left-click at SGR (col, row) — 1-based terminal coordinates.
- * ctx: { state, render, showPicker, popPicker, pushLine }
+ * ctx: { state, render, popPicker }
  * Returns true when the click was consumed.
  */
 export function handleMouseClick(ctx, col, row) {
@@ -62,56 +66,20 @@ export function handleMouseClick(ctx, col, row) {
     return true
   }
 
-  // ── Conversation: fold-toggle line expands; a message line opens the action menu ──
+  // ── Conversation: click a fold marker (expand hint or collapse marker) toggles it ──
   if (r >= P.conversation.y && r < P.conversation.y + P.conversation.h) {
     const convLines = buildConvLines(state, dims.cols)
     const gIdx = convGlobalIndex(convLines.length, P.conversation.h, state.scroll ?? 0)(r - P.conversation.y)
     if (gIdx === null) return false
     const lineEl = convLines[gIdx]
-    if (!lineEl) return false
-
-    // Click on a folded-block hint → expand it
-    if (lineEl._foldToggle) {
-      state.expandedBlocks ??= new Set()
-      state.expandedBlocks.add(lineEl._foldToggle)
-      render()
-      return true
-    }
-    // Click on a message line → action menu
-    if (lineEl._src !== undefined) {
-      const src = state.lines[lineEl._src]
-      if (src) {
-        openLineMenu(ctx, src)
-        return true
-      }
-    }
+    if (!lineEl?._foldToggle) return false
+    state.expandedBlocks ??= new Set()
+    // Bidirectional: click expands a folded block, collapses an expanded one
+    if (state.expandedBlocks.has(lineEl._foldToggle)) state.expandedBlocks.delete(lineEl._foldToggle)
+    else state.expandedBlocks.add(lineEl._foldToggle)
+    render()
+    return true
   }
 
   return false
 }
-
-/** Line action menu: copy / edit in input / (fold toggle if the source line folds). */
-async function openLineMenu(ctx, srcLine) {
-  const { state, render, showPicker, pushLine } = ctx
-  const text = sanitizeDisplay(srcLine.text)
-  const entries = [
-    { type: "item", text: `📋 Copy line (${text.length} chars)`, action: "copy" },
-    { type: "item", text: "✏️ Edit in input box", action: "edit" },
-  ]
-  const picked = await showPicker("Line actions", entries)
-  if (!picked) return
-  if (picked.action === "copy") {
-    try {
-      const { copyToClipboard } = await import("./clipboard.mjs")
-      await copyToClipboard(text)
-      pushLine(`[clipboard] copied ${text.length} chars`, (await import("./ansi.mjs")).C.dim)
-    } catch (e) {
-      pushLine(`[clipboard] copy failed: ${e.message}`, (await import("./ansi.mjs")).C.error)
-    }
-    render()
-  } else if (picked.action === "edit") {
-    state.input = [...text]
-    state.cursor = state.input.length
-    render()
-  }
-}

package/src/tui/render-conversation.mjs

@@ -15,8 +15,25 @@ export function convCacheKey(state) {
   return `${state.lines.length}|${lastLine?.text.length ?? 0}|${state.streaming.length}|${state.reasoning.length}|${state.advisorStreaming?.length ?? 0}|${state._advisorThink?.length ?? 0}|${state.foldEnabled !== false ? "f" : "u"}|${exp}`
 }
 
+/** Fold marker line: bold-cyan icon + "click to …" phrase underlined (clickable affordance).
+ *  No indent — flush with the content below it; the caller adds a blank line BEFORE it
+ *  so the control line stands apart from unrelated content (reported UX). */
+function foldHintLine(text, foldKey, srcIdx) {
+  // Underline just the actionable phrase — link/button convention
+  const withUnderline = text.replace(/(click to (?:expand|collapse))/, "\x1b[4m$1\x1b[24m")
+  return { text: withUnderline, color: C.fold, _foldToggle: foldKey, _src: srcIdx }
+}
+
+/** Blank separator before a fold control line (uncolored — must not join consecutive-dim folding).
+ *  Only the EXPANDED state uses it (▼ sits at the block head); the folded state's ▶
+ *  control line sits mid-block where the ellipsis used to be, so no separator needed. */
+function blankLine() {
+  return { text: "", color: "" }
+}
+
 function highlightSearchMatches(text, query, matchesInLine, globalCurrentIndex, allMatches, lineIndex) {
   if (!matchesInLine || matchesInLine.length === 0 || !query) return text
+
   let result = ""
   let lastEnd = 0
   for (const startIdx of matchesInLine) {
@@ -43,6 +60,10 @@ function buildConvLines(state, cols) {
   if (_convCache.key === key && _convCache.cols === cols) return _convCache.lines
 
   const convLines = []
+  // Folding constants (function scope — used by both the long-message fold below
+  // and the consecutive-dim fold at the bottom)
+  const LONG_FOLD_LINES = 12
+  const FOLD_KEEP = 5 // content lines kept in the folded state (first 4 + last 1)
   for (let i = 0; i < state.lines.length; i++) {
     const l = state.lines[i]
     let text = l.text
@@ -52,15 +73,15 @@ function buildConvLines(state, cols) {
       text = highlightSearchMatches(text, state.search.query, l._searchMatches, state.search.index, state.search.matches, i)
     }
 
-    // Long-message folding: a single DIM line (tool summaries / status output — secondary
-    // content) that wraps beyond LONG_FOLD_LINES display rows collapses to
-    // [first, "… N more — click/expand", last]. MAIN OUTPUT (C.text replies) and
-    // THINKING (C.reason) are NEVER folded — folding them hurt readability (reported
-    // regression: long replies and thinking collapsed behind a click). Keyed by the
+    // Long-message folding: ANY single line (main output C.text, thinking C.reason,
+    // tool summaries C.dim — whatever wraps beyond LONG_FOLD_LINES display rows)
+    // collapses to [blank, ▶, first 4, last] — 5 content lines. Main output and
+    // thinking are the REAL long content; bidirectional folding (collapse markers
+    // + click toggle) keeps them readable — the 0.12.7 dim-only restriction was a
+    // temporary fix for the single-direction era and is now reverted. Keyed by the
     // source-line index (`long-${i}`) so the toggle survives re-renders.
-    const LONG_FOLD_LINES = 12
     const longKey = `long-${i}`
-    const folded = state.foldEnabled !== false && !state.expandedBlocks?.has(longKey) && l.color === C.dim
+    const folded = state.foldEnabled !== false && !state.expandedBlocks?.has(longKey)
     const block = []
     for (const line of formatTables(sanitizeDisplay(text), cols - 1)) {
       for (const wrapped of wrapText(line, cols - 1)) {
@@ -70,15 +91,24 @@ function buildConvLines(state, cols) {
       }
     }
     if (folded && block.length > LONG_FOLD_LINES) {
-      convLines.push(block[0])
-      convLines.push({ text: `  … ${block.length - 2} more lines — click to expand`, color: C.fold, _foldToggle: longKey, _src: i })
+      // Folded state: first 4 content lines, then the ▶ control line where the
+      // ellipsis used to be (the marker itself reads "… N more lines" — ellipsis
+      // semantics built in), then the last line. No leading blank line needed:
+      // the block starts with real content now.
+      convLines.push(...block.slice(0, FOLD_KEEP - 1))
+      convLines.push(foldHintLine(`▶ … ${block.length - FOLD_KEEP} more lines — click to expand`, longKey, i))
       convLines.push(block[block.length - 1])
-    } else {
-      // Expanded long-DIM blocks must not re-trigger the consecutive-dim folding below
-      // (folding stacked on folding — reported regression). Only long blocks get the marker.
-      if (block.length > LONG_FOLD_LINES) {
+    } else if (block.length > LONG_FOLD_LINES) {
+      // EXPANDED long block: blank line + ▼ control line at the HEAD, directly
+      // before the content. DIM blocks must not re-trigger the consecutive-dim
+      // folding below (folding stacked on folding — reported regression).
+      if (l.color === C.dim) {
         for (const line of block) line._skipDimFold = true
       }
+      convLines.push(blankLine())
+      convLines.push(foldHintLine(`▼ … ${block.length} lines — click to collapse`, longKey, i))
+      convLines.push(...block)
+    } else {
       convLines.push(...block)
     }
   }
@@ -125,12 +155,19 @@ function buildConvLines(state, cols) {
       if (blockLen > FOLD_LINES && !hasExpandedLong) {
         const foldKey = `fold-${foldCounter++}`
         if (state.foldEnabled !== false && !state.expandedBlocks?.has(foldKey)) {
-          folded.push(convLines[i])
-          if (blockLen > 2) folded.push(convLines[i + 1])
-          folded.push({ text: `  … ${blockLen - 2} more lines — click to expand`, color: C.fold, _foldToggle: foldKey })
+          // First 4 lines, then the ▶ control line (ellipsis position), then the last line
+          folded.push(...convLines.slice(i, i + FOLD_KEEP - 1))
+          folded.push(foldHintLine(`▶ … ${blockLen - FOLD_KEEP} more lines — click to expand`, foldKey))
+          folded.push(convLines[j - 1])
           i = j
           continue
         }
+        // EXPANDED consecutive-dim block: blank + ▼ at the HEAD, then every line
+        folded.push(blankLine())
+        folded.push(foldHintLine(`▼ … ${blockLen} lines — click to collapse`, foldKey))
+        for (let k = i; k < j; k++) folded.push(convLines[k])
+        i = j
+        continue
       }
     }
     folded.push(line)