@link-assistant/hive-mind 1.59.3 → 1.59.5

package/CHANGELOG.md

@@ -1,5 +1,209 @@
 # @link-assistant/hive-mind
 
+## 1.59.5
+
+### Patch Changes
+
+- bb24175: Fix `/merge` to correctly detect active CI runs on the default branch — issue
+  #1722.
+
+  The `/merge` command merged PR #1719 even though a CI/CD workflow run was
+  still in progress on `main`. The merge triggered a new run, which cancelled
+  the previous one. Verbose log:
+
+  ```
+  [VERBOSE] /merge: Checking for active CI runs on link-assistant/hive-mind branch main...
+  [VERBOSE] /merge: Error checking active runs on main: stdout maxBuffer length exceeded
+  [VERBOSE] /merge: No active CI runs on main branch. Ready to proceed.
+  ```
+
+  Two compounding root causes in
+  [`src/github-merge.lib.mjs`](./src/github-merge.lib.mjs)
+  `getActiveBranchRuns()` (and the parallel
+  [`src/github-merge-repo-actions.lib.mjs`](./src/github-merge-repo-actions.lib.mjs)
+  `getAllActiveRepoRuns()` introduced by issue #1503):
+  1. **No `maxBuffer` override on `gh api --paginate --slurp`.** Node's default
+     `child_process.exec` buffer is 1 MB; the unfiltered `actions/runs` response
+     on this repo's `main` was 12.7 MB, so `exec` rejected with
+     `stdout maxBuffer length exceeded`.
+  2. **Fetch errors became "no active runs".** The `catch` block returned
+     `hasActiveRuns: false`, which the caller (`waitForBranchCI`) interpreted as
+     "branch CI is idle, ready to merge". A transient fetch/buffer/parse error
+     was indistinguishable from genuine idleness.
+
+  Fix:
+  - **Server-side `?status=` filter**, looped over the active set
+    (`in_progress`, `queued`, `waiting`, `requested`, `pending`) with run-id
+    dedup. Response size scales with active-run count, not with historical-run
+    count — typically a few KB instead of 12+ MB.
+  - **Raise `exec` `maxBuffer` to `githubLimits.bufferMaxSize`** (10 MB, env
+    `HIVE_MIND_GITHUB_BUFFER_MAX_SIZE`) for all `gh` calls in
+    `github-merge.lib.mjs` and `github-merge-repo-actions.lib.mjs`. The existing
+    `githubLimits` infrastructure was already used in `github.batch.lib.mjs`;
+    this just wires it into the `/merge` paths.
+  - **Stop swallowing fetch errors as "idle".** Errors now propagate. The
+    surrounding `waitForBranchCI` / `waitForAllRepoActions` poll loops already
+    retry on the next tick; the timeout-final check has its own try/catch that
+    returns an explicit failure (instead of a false-positive "ready to merge").
+
+  Tests:
+  [`tests/test-active-branch-runs-buffer-1722.mjs`](./tests/test-active-branch-runs-buffer-1722.mjs)
+  shadows `gh` on `PATH` with a Node script that scripts active-run responses,
+  and asserts: (a) every call uses `?status=`, (b) duplicate runs across
+  statuses are deduplicated, (c) >1 MB responses are handled cleanly, (d)
+  `gh` failures throw rather than report idle, (e) `waitForBranchCI` keeps
+  polling on errors, (f) idle branches still resolve as ready,
+  (g) `getAllActiveRepoRuns` parity.
+
+  Documentation:
+  [`docs/case-studies/issue-1722/`](./docs/case-studies/issue-1722/README.md)
+  contains the timeline (with downloaded bot log, cancelled-run logs, run
+  metadata), facts, per-symptom root-cause analysis, and solution plan.
+  [`experiments/issue-1722-buffer-overflow.mjs`](./experiments/issue-1722-buffer-overflow.mjs)
+  is a minimal reproduction. No upstream report required — the fix lives
+  entirely in this repo.
+
+- 1a92ca1: Fix flaky CI `test-suites` job caused by `use-m`'s no-retry global npm install
+  — issue #1724.
+
+  CI run [25109962685](https://github.com/link-assistant/hive-mind/actions/runs/25109962685/job/73581228475)
+  on `main` failed in the `test-suites` job at the third test file
+  (`tests/test-active-branch-runs-buffer-1722.mjs`) with:
+
+  ```
+  Error: Failed to install command-stream@latest globally.
+    [cause]: Error: Command failed: npm install -g command-stream-v-latest@npm:command-stream@latest
+    npm error code ENOTEMPTY
+    npm error path /opt/hostedtoolcache/node/24.14.1/x64/lib/node_modules/command-stream-v-latest/js/src/commands
+  ```
+
+  Root cause: `src/github.lib.mjs` and `src/playwright-mcp.lib.mjs` call
+  `await use('command-stream')` at module top level (via `use-m`). Every test
+  file that transitively imports either module re-runs
+  `npm install -g command-stream-v-latest@npm:command-stream@latest`. `use-m`'s
+  `ensurePackageInstalled` issues a single `npm install -g` with no retry, and
+  npm intermittently fails with `ENOTEMPTY: directory not empty, rmdir` on
+  GitHub-hosted Ubuntu runners (a long-standing npm rmdir race against itself
+  when the previous global install left files behind).
+
+  Fix:
+  - New
+    [`scripts/preinstall-use-m-packages.mjs`](./scripts/preinstall-use-m-packages.mjs)
+    pre-installs every package the codebase loads through `use-m @latest`
+    (`command-stream`, `getenv`, `links-notation`, `@dotenvx/dotenvx`,
+    `telegraf`, `zx`, `yargs`) using the same alias scheme `use-m` does
+    (`<pkg-without-@-or-/>-v-latest`), with exponential-backoff retry on the
+    flake symptoms (`ENOTEMPTY` / `EBUSY` / `EPERM` / `ECONNRESET` / `ETIMEDOUT`
+    / `EAI_AGAIN` / `429` / `503`). After this step, `use-m`'s
+    `installedVersion === latestVersion` early-return path skips the install at
+    test time, so test imports never touch `npm install -g` again.
+  - The script also satisfies the case-study "verbose mode for next iteration"
+    requirement via `PREINSTALL_USE_M_VERBOSE=1` (or `RUNNER_DEBUG=1`), which
+    logs each attempt's command, stdout, stderr, and backoff delay, and
+    recognizes "package present on disk after a flake" as recovered success.
+  - Wires `node scripts/preinstall-use-m-packages.mjs` into the `test-suites`
+    and `test-execution` jobs in
+    [`.github/workflows/release.yml`](./.github/workflows/release.yml) right
+    after `npm install`, before any step that runs test files or `solve.mjs`.
+
+  Tests:
+  [`tests/test-preinstall-use-m-packages-1724.mjs`](./tests/test-preinstall-use-m-packages-1724.mjs)
+  covers the alias scheme, retryable-error matcher, exponential backoff, and
+  the four `installWithRetry` paths (first-success, retry-then-succeed,
+  non-retryable-abort, recovered-from-disk) deterministically (no real npm
+  calls). Marked `@hive-mind-test-suite default` so it runs in the same job
+  that previously flaked.
+
+  Documentation:
+  [`docs/case-studies/issue-1724/`](./docs/case-studies/issue-1724/README.md)
+  contains the timeline, verbatim error, downloaded failed-run logs, the
+  no-retry snippet from the live `use-m` source
+  (`logs/use-m-source.js`), the comparison with both pipeline templates
+  (JS/Rust — neither template uses `use-m @latest` at module load yet, so the
+  flake is hive-mind-specific until they do), and the implementation plan.
+
+## 1.59.4
+
+### Patch Changes
+
+- b2e0d12: Fix `/terminal_watch` uploading the full session log file when the watch
+  completes — addresses issue
+  [#1720](https://github.com/link-assistant/hive-mind/issues/1720).
+
+  Before this fix, `/terminal_watch` finished by calling
+  `bot.telegram.sendDocument(chatId, ...)` to attach the `<uuid>.log` file. That
+  had two unwanted effects:
+  - It duplicated work that the dedicated `/log` command already does.
+  - The bare `bot.telegram.sendDocument(chatId, ...)` call did not carry
+    `message_thread_id`, so in forum-enabled supergroups the document landed in
+    the **General** topic instead of the topic where `/terminal_watch` was
+    invoked, and it was not threaded as a reply.
+
+  `/terminal_watch` now only updates the live "✅ Terminal watch complete"
+  message at the end of the session. To download the log, use
+  `/log <uuid>` — it correctly replies in the originating topic via
+  `ctx.replyWithDocument`, which Telegraf annotates with `message_thread_id`
+  automatically.
+
+  A new regression test (`tests/test-issue-1720-terminal-watch-no-log.mjs`)
+  guards both behaviours, and `tests/test-issue-467-terminal-watch.mjs` was
+  updated to assert that no document is uploaded by the watcher.
+
+- 5c87a38: Fix `hive` to (a) stop forwarding `false` for solve options whose `type` is
+  `'string'` but whose `default` is `false`, and (b) exit non-zero when any
+  worker fails — issue #1718.
+
+  Previously, when a user ran `/hive` against several issues, every spawned
+  `solve` worker crashed with:
+
+  ```
+  Invalid --working-session-live-progress value: "false". Expected "comment" or "pr".
+  ```
+
+  …and `hive` itself still exited with code `0`, so the Telegram bot rendered a
+  green "Work session finished successfully" envelope even though zero PRs had
+  been created.
+
+  Two independent root causes:
+  1. **Auto-forwarder leaked `false` as a string.** In
+     [`src/hive.mjs`](./src/hive.mjs), the auto-forward block read:
+
+     ```js
+     } else if ((def.type === 'string' || def.type === 'number') && value !== undefined) {
+       args.push(`--${optionName}`, String(value));
+     }
+     ```
+
+     For `working-session-live-progress`, `solve.config.lib.mjs` declares
+     `type: 'string', default: false`. yargs preserves the boolean `false`
+     verbatim, so hive forwarded `--working-session-live-progress false`,
+     which `solve` rejects. The fix adds `&& value !== false` to the
+     predicate. Other `type:'string'` options whose `default` is `false`
+     are now also protected by a single defense-in-depth check.
+
+  2. **No non-zero exit on worker failures.** After `monitorWithSentry()`
+     resolved, hive returned without consulting `issueQueue.getStats()`. The
+     fix queries `finalStats = issueQueue.getStats()` and calls
+     `safeExit(1, …)` when `finalStats.failed > 0`, mirroring the exit
+     semantics solve already uses. Wrappers like `start-command`, the Telegram
+     bot, and CI now correctly observe the failure.
+
+  `--isolation screen` (R3 of the issue) was already wired through correctly;
+  no change required there. The verbose forwarder dump
+  (`📋 Command: ${solveCommand} ${args.join(' ')}`) — which is what allowed us
+  to diagnose this run in the first place — is preserved.
+
+  Tests: [`tests/test-issue-1718-hive-passthrough-false.mjs`](./tests/test-issue-1718-hive-passthrough-false.mjs)
+  locks the option shape, asserts both fixes are present in `src/hive.mjs`,
+  replays the forwarder logic on synthetic argv, and adds a defense-in-depth
+  sweep that no `type:'string'` / `default:false` option ever produces
+  `--<flag> false`.
+
+  Documentation: [`docs/case-studies/issue-1718/`](./docs/case-studies/issue-1718/README.md)
+  contains the timeline reconstructed from the user's `screen` log, the
+  distilled facts, the per-symptom root-cause analysis, the solution plan, and
+  notes confirming no upstream report (yargs / start-command) is required.
+
 ## 1.59.3
 
 ### Patch Changes

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@link-assistant/hive-mind",
-  "version": "1.59.3",
+  "version": "1.59.5",
   "description": "AI-powered issue solver and hive mind for collaborative problem solving",
   "main": "src/hive.mjs",
   "type": "module",

package/src/github-merge-repo-actions.lib.mjs

@@ -11,7 +11,14 @@
 
 import { promisify } from 'util';
 import { exec as execCallback } from 'child_process';
-const exec = promisify(execCallback);
+import { githubLimits } from './config.lib.mjs';
+const execRaw = promisify(execCallback);
+// Issue #1722: raise exec maxBuffer above Node's 1 MB default for paginated gh
+// API responses (workflow runs can easily exceed that on busy repos).
+const exec = (cmd, opts = {}) => execRaw(cmd, { maxBuffer: githubLimits.bufferMaxSize, ...opts });
+
+// Statuses we treat as "not yet finished".
+const ACTIVE_RUN_STATUSES = ['in_progress', 'queued', 'waiting', 'requested', 'pending'];
 
 /**
  * Get ALL active workflow runs across the entire repository (no branch filter).
@@ -21,20 +28,34 @@ const exec = promisify(execCallback);
  * @returns {Promise<{runs: Array, hasActiveRuns: boolean, count: number}>}
  */
 export async function getAllActiveRepoRuns(owner, repo, verbose = false) {
-  try {
-    const { stdout } = await exec(`gh api "repos/${owner}/${repo}/actions/runs?per_page=100" --paginate --slurp`);
-    const runs = JSON.parse(stdout.trim() || '[]')
-      .flatMap(page => page.workflow_runs || [])
-      .filter(run => ['in_progress', 'queued', 'waiting', 'requested', 'pending'].includes(run.status))
-      .map(run => ({ id: run.id, name: run.name, status: run.status, head_branch: run.head_branch, head_sha: run.head_sha?.slice(0, 7) }));
-    if (verbose && runs.length > 0) {
-      console.log(`[VERBOSE] repo-actions: ${runs.length} active run(s) in ${owner}/${repo}`);
-      for (const r of runs) console.log(`[VERBOSE] repo-actions:   ${r.name} (${r.status}) on ${r.head_branch}`);
+  // Issue #1722: filter on the server side per status to avoid pulling the full
+  // history of workflow runs (which can exceed exec maxBuffer). Also: do not
+  // swallow errors as "no active runs" — bubble them up so callers can retry
+  // instead of merging on top of a still-running CI run.
+  const seen = new Set();
+  const runs = [];
+  for (const status of ACTIVE_RUN_STATUSES) {
+    const { stdout } = await exec(`gh api "repos/${owner}/${repo}/actions/runs?status=${status}&per_page=100" --paginate --slurp`);
+    const pages = JSON.parse(stdout.trim() || '[]');
+    for (const page of pages) {
+      for (const run of page.workflow_runs || []) {
+        if (seen.has(run.id)) continue;
+        seen.add(run.id);
+        runs.push({
+          id: run.id,
+          name: run.name,
+          status: run.status,
+          head_branch: run.head_branch,
+          head_sha: run.head_sha?.slice(0, 7),
+        });
+      }
     }
-    return { runs, hasActiveRuns: runs.length > 0, count: runs.length };
-  } catch {
-    return { runs: [], hasActiveRuns: false, count: 0 };
   }
+  if (verbose && runs.length > 0) {
+    console.log(`[VERBOSE] repo-actions: ${runs.length} active run(s) in ${owner}/${repo}`);
+    for (const r of runs) console.log(`[VERBOSE] repo-actions:   ${r.name} (${r.status}) on ${r.head_branch}`);
+  }
+  return { runs, hasActiveRuns: runs.length > 0, count: runs.length };
 }
 
 /**
@@ -52,7 +73,16 @@ export async function waitForAllRepoActions(owner, repo, options = {}, verbose =
   let peakRunCount = 0;
 
   while (Date.now() - startTime < timeout) {
-    const active = await getAllActiveRepoRuns(owner, repo, verbose);
+    let active;
+    try {
+      active = await getAllActiveRepoRuns(owner, repo, verbose);
+    } catch (error) {
+      // Issue #1722: do not silently treat fetch errors as "no active runs".
+      // Log and retry on the next poll instead.
+      console.error(`[ERROR] repo-actions: Error checking repo CI: ${error.message}`);
+      await new Promise(resolve => setTimeout(resolve, pollInterval));
+      continue;
+    }
     if (onStatusUpdate) {
       try {
         await onStatusUpdate({ ...active, elapsedMs: Date.now() - startTime });
@@ -66,7 +96,15 @@ export async function waitForAllRepoActions(owner, repo, options = {}, verbose =
     peakRunCount = Math.max(peakRunCount, active.count);
     await new Promise(resolve => setTimeout(resolve, pollInterval));
   }
-  const finalRuns = await getAllActiveRepoRuns(owner, repo, verbose);
+  // Issue #1722: if the timeout-final check throws, surface that as an error
+  // rather than reporting "no remaining runs".
+  let finalRuns;
+  try {
+    finalRuns = await getAllActiveRepoRuns(owner, repo, verbose);
+  } catch (error) {
+    console.error(`[ERROR] repo-actions: Final CI check failed after timeout: ${error.message}`);
+    return { success: false, waitedForRuns: true, timedOut: true, remainingRuns: [] };
+  }
   return { success: false, waitedForRuns: true, timedOut: true, remainingRuns: finalRuns.runs };
 }
 

package/src/github-merge.lib.mjs

@@ -14,9 +14,17 @@
 import { promisify } from 'util';
 import { exec as execCallback } from 'child_process';
 
-const exec = promisify(execCallback);
+const execRaw = promisify(execCallback);
 
 import { parseGitHubUrl } from './github.lib.mjs';
+import { githubLimits } from './config.lib.mjs';
+
+// Issue #1722: gh api `--paginate --slurp` responses for repos with many
+// historical workflow runs can easily exceed Node's default 1 MB exec buffer
+// (observed 12.7 MB on this repo's main branch). Default to the configured
+// githubLimits.bufferMaxSize (10 MB; HIVE_MIND_GITHUB_BUFFER_MAX_SIZE) for all
+// gh calls in this file.
+const exec = (cmd, opts = {}) => execRaw(cmd, { maxBuffer: githubLimits.bufferMaxSize, ...opts });
 
 // Issue #1413: Import ready tag sync, timeline, and label constant from separate module
 // to keep this file under the 1500 line limit
@@ -674,9 +682,20 @@ export function parseRepositoryUrl(url) {
   };
 }
 
+/**
+ * Statuses we treat as "still running" / "not yet finished".
+ * Issue #1722: be exhaustive — GitHub uses several non-completed statuses.
+ */
+const ACTIVE_RUN_STATUSES = ['in_progress', 'queued', 'waiting', 'requested', 'pending'];
+
 /**
  * Get active workflow runs on a specific branch
  * Issue #1307: Used to check if there are any in-progress or queued runs on the target branch
+ * Issue #1722: Filter on the server side per status, otherwise the unfiltered
+ * `--paginate --slurp` response can overflow exec maxBuffer on busy repos
+ * (observed 12.7 MB on link-assistant/hive-mind main). Also: errors are now
+ * surfaced rather than swallowed as `hasActiveRuns: false`, which previously
+ * caused /merge to merge on top of a still-running CI run.
  * @param {string} owner - Repository owner
  * @param {string} repo - Repository name
  * @param {string} branch - Branch name (default: main)
@@ -684,36 +703,38 @@ export function parseRepositoryUrl(url) {
  * @returns {Promise<{runs: Array<Object>, hasActiveRuns: boolean, count: number}>}
  */
 export async function getActiveBranchRuns(owner, repo, branch = 'main', verbose = false) {
-  try {
-    // Query for in_progress and queued runs on the specified branch
-    const { stdout } = await exec(`gh api "repos/${owner}/${repo}/actions/runs?branch=${branch}&per_page=100" --paginate --slurp`);
-    const runs = JSON.parse(stdout.trim() || '[]')
-      .flatMap(page => page.workflow_runs || [])
-      .filter(run => run.status === 'in_progress' || run.status === 'queued')
-      .map(run => ({ id: run.id, name: run.name, status: run.status, created_at: run.created_at, html_url: run.html_url }));
-
-    if (verbose) {
-      console.log(`[VERBOSE] /merge: Found ${runs.length} active runs on ${owner}/${repo} branch ${branch}`);
-      for (const run of runs) {
-        console.log(`[VERBOSE] /merge:   - Run #${run.id}: ${run.name} (${run.status})`);
+  const seen = new Set();
+  const runs = [];
+  for (const status of ACTIVE_RUN_STATUSES) {
+    const { stdout } = await exec(`gh api "repos/${owner}/${repo}/actions/runs?branch=${branch}&status=${status}&per_page=100" --paginate --slurp`);
+    const pages = JSON.parse(stdout.trim() || '[]');
+    for (const page of pages) {
+      for (const run of page.workflow_runs || []) {
+        if (seen.has(run.id)) continue;
+        seen.add(run.id);
+        runs.push({
+          id: run.id,
+          name: run.name,
+          status: run.status,
+          created_at: run.created_at,
+          html_url: run.html_url,
+        });
       }
     }
+  }
 
-    return {
-      runs,
-      hasActiveRuns: runs.length > 0,
-      count: runs.length,
-    };
-  } catch (error) {
-    if (verbose) {
-      console.log(`[VERBOSE] /merge: Error checking active runs on ${branch}: ${error.message}`);
+  if (verbose) {
+    console.log(`[VERBOSE] /merge: Found ${runs.length} active runs on ${owner}/${repo} branch ${branch}`);
+    for (const run of runs) {
+      console.log(`[VERBOSE] /merge:   - Run #${run.id}: ${run.name} (${run.status})`);
     }
-    return {
-      runs: [],
-      hasActiveRuns: false,
-      count: 0,
-    };
   }
+
+  return {
+    runs,
+    hasActiveRuns: runs.length > 0,
+    count: runs.length,
+  };
 }
 
 /**
@@ -788,7 +809,20 @@ export async function waitForBranchCI(owner, repo, branch = 'main', options = {}
   }
 
   // Timeout reached
-  const finalCheck = await getActiveBranchRuns(owner, repo, branch, verbose);
+  // Issue #1722: if the final check throws, do NOT silently report "ready".
+  // Treat it the same as still-active (force a timeout failure), so /merge
+  // waits/retries instead of merging on top of a still-running CI run.
+  let finalCheck;
+  try {
+    finalCheck = await getActiveBranchRuns(owner, repo, branch, verbose);
+  } catch (error) {
+    return {
+      success: false,
+      waitedForRuns: true,
+      completedRuns: totalWaitedRuns,
+      error: `Timeout reached and final CI check failed on ${branch}: ${error.message}`,
+    };
+  }
   if (finalCheck.hasActiveRuns) {
     return {
       success: false,

package/src/hive.mjs

@@ -798,8 +798,8 @@ if (isRunningDirectly) {
                 for (const entry of value) {
                   args.push(`--${optionName}`, String(entry));
                 }
-              } else if ((def.type === 'string' || def.type === 'number') && value !== undefined) {
-                args.push(`--${optionName}`, String(value));
+              } else if ((def.type === 'string' || def.type === 'number') && value !== undefined && value !== false) {
+                args.push(`--${optionName}`, String(value)); // Issue #1718: skip false (some string options have default:false)
               }
             }
             // Log the actual command being executed so users can investigate/reproduce
@@ -1483,6 +1483,9 @@ if (isRunningDirectly) {
       await log(`   📁 Full log file: ${absoluteLogPath}`, { level: 'error' });
       await safeExit(1, 'Error occurred');
     }
+
+    const finalStats = issueQueue.getStats(); // Issue #1718: surface worker failures via exit code
+    if (finalStats.failed > 0) await safeExit(1, `${finalStats.failed} task(s) failed (completed: ${finalStats.completed})`);
   } catch (fatalError) {
     // Handle fatal errors during initialization or execution
     console.error('\n❌ Fatal error occurred during hive initialization or execution');

package/src/telegram-terminal-watch-command.lib.mjs

@@ -6,15 +6,12 @@
  */
 
 import fs from 'fs/promises';
-import path from 'path';
-import { constants as fsConstants } from 'fs';
 import { extractSessionIdFromText, decideLogDestination, resolveLogPath } from './telegram-log-command.lib.mjs';
 
 const DEFAULT_WIDTH = 120;
 const DEFAULT_HEIGHT = 25;
 const DEFAULT_INTERVAL_MS = 2500;
 const DEFAULT_MAX_CHARS = 3400;
-const TELEGRAM_DOCUMENT_MAX_BYTES = 50 * 1024 * 1024;
 const GITHUB_URL_RE = /https:\/\/github\.com\/[^\s"'`<>]+/i;
 const activeWatches = new Map();
 
@@ -135,23 +132,6 @@ export function formatTerminalWatchMessage({ sessionId, statusResult = null, log
   return lines.join('\n');
 }
 
-async function fileExists(filePath) {
-  try {
-    await fs.access(filePath, fsConstants.R_OK);
-    return true;
-  } catch {
-    return false;
-  }
-}
-
-async function fileSize(filePath) {
-  try {
-    return (await fs.stat(filePath)).size;
-  } catch {
-    return null;
-  }
-}
-
 async function readLogFile(logPath) {
   try {
     return await fs.readFile(logPath, 'utf8');
@@ -182,16 +162,6 @@ export async function resolveTerminalWatchRepository({ sessionInfo = null, statu
   }
 }
 
-async function sendLogDocument({ bot, chatId, logPath, sessionId, statusResult }) {
-  if (!(await fileExists(logPath))) return;
-  const size = await fileSize(logPath);
-  if (size !== null && size > TELEGRAM_DOCUMENT_MAX_BYTES) {
-    await bot.telegram.sendMessage(chatId, `⚠️ Full log for \`${sessionId}\` is ${(size / (1024 * 1024)).toFixed(1)} MB, above Telegram's 50 MB upload limit.`, { parse_mode: 'Markdown' });
-    return;
-  }
-  await bot.telegram.sendDocument(chatId, { source: logPath, filename: path.basename(logPath) }, { caption: `📄 Full log for session \`${sessionId}\`${statusResult?.status ? `\nStatus: \`${statusResult.status}\`` : ''}`, parse_mode: 'Markdown' });
-}
-
 async function querySessionStatusWithRetry(querySessionStatus, sessionId, verbose, attempts = 3) {
   for (let attempt = 1; attempt <= attempts; attempt++) {
     const statusResult = await querySessionStatus(sessionId, verbose);
@@ -201,7 +171,9 @@ async function querySessionStatusWithRetry(querySessionStatus, sessionId, verbos
   return null;
 }
 
-export function watchTerminalLogSession({ bot, chatId, messageId, sessionId, logPath, querySessionStatus, isTerminalSessionStatus, options = {}, repoDescription = null, verbose = false, attachLogOnComplete = true }) {
+// Note: /terminal_watch never uploads the full session log itself (issue #1720).
+// Use /log <uuid> if you want the log file delivered as a document.
+export function watchTerminalLogSession({ bot, chatId, messageId, sessionId, logPath, querySessionStatus, isTerminalSessionStatus, options = {}, repoDescription = null, verbose = false }) {
   const key = `${chatId}:${messageId}:${sessionId}`;
   activeWatches.get(key)?.stop();
 
@@ -225,7 +197,6 @@ export function watchTerminalLogSession({ bot, chatId, messageId, sessionId, log
       if (completed) {
         stopped = true;
         activeWatches.delete(key);
-        if (attachLogOnComplete) await sendLogDocument({ bot, chatId, logPath, sessionId, statusResult });
         return;
       }
     } catch (error) {
@@ -409,6 +380,5 @@ export const __INTERNAL_FOR_TESTS__ = {
   DEFAULT_HEIGHT,
   DEFAULT_INTERVAL_MS,
   DEFAULT_MAX_CHARS,
-  TELEGRAM_DOCUMENT_MAX_BYTES,
   GITHUB_URL_RE,
 };