diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
index 1a1701b0..018f0384 100644
--- a/.claude-plugin/plugin.json
+++ b/.claude-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "name": "commitlore",
   "displayName": "CommitLore",
-  "version": "1.2.2",
+  "version": "1.2.3",
   "description": "Recorded decisions from git history, delivered to the agent before it edits. Constraints, alternatives already ruled out, and warnings left by whoever was here last.",
   "author": {
     "name": "MongLong0214",
diff --git a/.codex-plugin/plugin.json b/.codex-plugin/plugin.json
index 1a79f028..15cc39b1 100644
--- a/.codex-plugin/plugin.json
+++ b/.codex-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "commitlore",
-  "version": "1.2.2",
+  "version": "1.2.3",
   "description": "Decision memory from Git history, with verified capture for coding sessions.",
   "author": {
     "name": "MongLong0214",
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 51d244d1..8150c2b6 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,59 @@ Release notes for 1.0.0, 1.0.1 and 1.0.2 are on the
 [GitHub releases page](https://github.com/MongLong0214/commitlore/releases); they
 were not written here.
 
+## 1.2.3
+
+The capture prompt was the session, so on a long session there was no prompt.
+
+**`capture` embedded the whole transcript in its prompt (#873).** Measured by the
+reporter: a 67,981,436-byte transcript produced a 67,468,122-byte prompt — 99.3%
+of the file, and larger than any model can consume, so the pipeline could not be
+completed and no record was ever written. The pipeline itself was fine, which is
+how they proved it: `tail -n 400` on the same transcript gave 537,250 bytes and
+the same command in the same repository worked.
+
+**The failure was silent, which is why it reads as the reason capture is not
+used.** Prompt-only mode reports `outcome: "empty"`, `staged: false`, exit 0. An
+operator who tried capture once on a real session got a prompt they could not
+use and no statement that anything was wrong, and did not try again. In the
+repository where this was measured, `stale` reported 1 record in 1000 commits
+while roughly twenty commits in one recent session carried real decision context
+and produced none — with `{"mode":"auto","unattended":true}` already set, so
+permission was never the obstacle.
+
+**The prompt now carries the end of the transcript within a byte budget**, 256
+KiB by default and `COMMITLORE_TRANSCRIPT_BUDGET_BYTES` to change it. The end
+rather than the beginning: a decision is taken near the end of the session that
+implements it, and the diff being captured is that end.
+
+**Three things the bound deliberately does not do.** It does not renumber — the
+window keeps the line numbers it has in the whole transcript, because
+verification reads the whole transcript and a window renumbered from 1 would
+have every `L<start>-L<end>` locator name a different line of the file it is
+checked against. It does not reach the hash — `source_hashes` and every quote
+check are still over the whole transcript, so a quote from outside the window
+still verifies and a caller passing the session it actually had is never told
+the transcript was substituted. And it does not stay quiet — the prompt says
+which lines it shows and how many were left out, and `transcript_window` says
+the same to `capture --json` and to `commitlore_prepare_capture`. A bounded
+prompt that did not say so would be the old silence in a smaller package.
+
+**One line can outrun the budget by itself.** A JSONL transcript line can hold an
+entire tool result. That line is shown from its end rather than dropped, and the
+window marks it partial, because a window of no lines is worse than a window of
+one partial line. The byte slice never leaves a split codepoint at the front: a
+replacement character inside a quotable line is a character nobody can copy back.
+
+**Not claimed.** Where the useful boundary is. 256 KiB is comfortably readable by
+current models and carries far more than the ~537 KB slice measured to work; the
+reporter said they had not measured where a good record stops needing context,
+and neither has this. Also unchanged, and worth stating because automation is
+built on it: `outcome` is `staged`, `empty` or `rejected`, and all three exit 0
+— anything driving capture must read `--json`, because an exit-code check reads
+a rejected record as a success.
+
+**Not reviewed.** No cross-provider review was run on this change.
+
 ## 1.2.2
 
 One line of configuration, and the plugin's MCP server had never started for
diff --git a/README.ja.md b/README.ja.md
index b443698a..1ccf4a88 100644
--- a/README.ja.md
+++ b/README.ja.md
@@ -47,18 +47,18 @@
 </p>
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2
+curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3
 ```
 
 <details>
 <summary>先にインストーラーを読みたいですか？</summary>
 
 ```bash
-curl -fsSLO https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh
-sh install.sh v1.2.2
+curl -fsSLO https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh
+sh install.sh v1.2.3
 
 # あるいはスクリプトを使わずに。スクリプトが作るチェックアウトは自分でも作れます。
-git clone --depth 1 --branch v1.2.2 https://github.com/MongLong0214/commitlore
+git clone --depth 1 --branch v1.2.3 https://github.com/MongLong0214/commitlore
 node commitlore/dist/commitlore.mjs --version
 ```
 
@@ -107,13 +107,13 @@ CommitLore はその判断をコードのそばに残します。
 macOS と Linux:
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2
+curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3
 ```
 
 Windows:
 
 ```powershell
-& ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.ps1))) v1.2.2
+& ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.ps1))) v1.2.3
 ```
 
 Node.js 22.23.2+ と Git が必要です。スクリプトは何かを書き込む前に両方を確認します。
diff --git a/README.ko.md b/README.ko.md
index 396becd9..c5bb0b9f 100644
--- a/README.ko.md
+++ b/README.ko.md
@@ -47,18 +47,18 @@
 </p>
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2
+curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3
 ```
 
 <details>
 <summary>먼저 설치기를 읽어 보고 싶나요?</summary>
 
 ```bash
-curl -fsSLO https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh
-sh install.sh v1.2.2
+curl -fsSLO https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh
+sh install.sh v1.2.3
 
 # 또는 스크립트를 건너뜁니다. 스크립트가 만드는 체크아웃은 직접 만들 수 있습니다.
-git clone --depth 1 --branch v1.2.2 https://github.com/MongLong0214/commitlore
+git clone --depth 1 --branch v1.2.3 https://github.com/MongLong0214/commitlore
 node commitlore/dist/commitlore.mjs --version
 ```
 
@@ -107,13 +107,13 @@ CommitLore는 그 판단을 코드 곁에 보관합니다.
 macOS와 Linux:
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2
+curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3
 ```
 
 Windows:
 
 ```powershell
-& ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.ps1))) v1.2.2
+& ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.ps1))) v1.2.3
 ```
 
 Node.js 22.23.2+와 Git이 필요합니다. 스크립트는 무엇이든 쓰기 전에 둘을 확인합니다.
diff --git a/README.md b/README.md
index ce9331e3..9c0d0634 100644
--- a/README.md
+++ b/README.md
@@ -48,18 +48,18 @@
 </p>
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2
+curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3
 ```
 
 <details>
 <summary>Prefer to read the installer first?</summary>
 
 ```bash
-curl -fsSLO https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh
-sh install.sh v1.2.2
+curl -fsSLO https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh
+sh install.sh v1.2.3
 
 # Or skip the script: the checkout it makes is one you can make yourself.
-git clone --depth 1 --branch v1.2.2 https://github.com/MongLong0214/commitlore
+git clone --depth 1 --branch v1.2.3 https://github.com/MongLong0214/commitlore
 node commitlore/dist/commitlore.mjs --version
 ```
 
@@ -109,13 +109,13 @@ preserve, not for narrating every change.
 macOS and Linux:
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2
+curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3
 ```
 
 Windows:
 
 ```powershell
-& ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.ps1))) v1.2.2
+& ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.ps1))) v1.2.3
 ```
 
 Requires Node.js 22.23.2+ and Git. The script checks both before it writes anything.
diff --git a/README.zh-CN.md b/README.zh-CN.md
index 8448a86b..337ab275 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -47,18 +47,18 @@
 </p>
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2
+curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3
 ```
 
 <details>
 <summary>想先阅读安装器吗？</summary>
 
 ```bash
-curl -fsSLO https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh
-sh install.sh v1.2.2
+curl -fsSLO https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh
+sh install.sh v1.2.3
 
 # 或者跳过脚本：它创建的检出，你自己也能创建。
-git clone --depth 1 --branch v1.2.2 https://github.com/MongLong0214/commitlore
+git clone --depth 1 --branch v1.2.3 https://github.com/MongLong0214/commitlore
 node commitlore/dist/commitlore.mjs --version
 ```
 
@@ -105,13 +105,13 @@ CommitLore 把那份判断留在代码旁边。
 macOS 和 Linux：
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2
+curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3
 ```
 
 Windows：
 
 ```powershell
-& ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.ps1))) v1.2.2
+& ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.ps1))) v1.2.3
 ```
 
 需要 Node.js 22.23.2+ 和 Git。脚本会在写入任何内容前检查两者。
diff --git a/install.ps1 b/install.ps1
index 0e7bc0fd..055f8561 100644
--- a/install.ps1
+++ b/install.ps1
@@ -1,8 +1,8 @@
 <#
     Installs commitlore from source on Windows, for any agent that is not Claude Code.
 
-      irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.ps1 | iex
-      & ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.ps1))) v1.2.2
+      irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.ps1 | iex
+      & ([scriptblock]::Create((irm https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.ps1))) v1.2.3
 
     Claude Code users do not need this script. The repository is itself a plugin
     marketplace (ADR-0011), so two /plugin commands register the MCP server, the
diff --git a/install.sh b/install.sh
index c05ff382..f77ed7cb 100755
--- a/install.sh
+++ b/install.sh
@@ -1,8 +1,8 @@
 #!/bin/sh
 # Installs commitlore from source, for any agent that is not Claude Code.
 #
-#   curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh
-#   curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2
+#   curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh
+#   curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3
 #
 # **Claude Code users do not need this script.** The repository is itself a
 # plugin marketplace (ADR-0011), so two `/plugin` commands register the MCP
diff --git a/package-lock.json b/package-lock.json
index 24b1d386..282256ab 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -1,12 +1,12 @@
 {
   "name": "commitlore",
-  "version": "1.2.2",
+  "version": "1.2.3",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "commitlore",
-      "version": "1.2.2",
+      "version": "1.2.3",
       "license": "MIT",
       "dependencies": {
         "@modelcontextprotocol/sdk": "^1.30.0",
diff --git a/package.json b/package.json
index b2e96388..e3d0c2a1 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "commitlore",
-  "version": "1.2.2",
+  "version": "1.2.3",
   "description": "Git-native, lifecycle-aware decision memory for coding agents",
   "license": "MIT",
   "private": true,
diff --git a/server.json b/server.json
index f597fef4..503946f5 100644
--- a/server.json
+++ b/server.json
@@ -8,7 +8,7 @@
     "source": "github"
   },
   "websiteUrl": "https://github.com/MongLong0214/commitlore#readme",
-  "version": "1.2.2",
+  "version": "1.2.3",
   "_meta": {
     "io.modelcontextprotocol.registry/publisher-provided": {
       "registryFit": "Distribution is a tagged git checkout plus a Claude Code plugin marketplace (ADR-0011 registry-free git distribution, ADR-0026 no compiled executables and no uploaded release asset), so no official package type applies and this record relies on websiteUrl plus publisher metadata.",
@@ -18,8 +18,8 @@
           "/plugin marketplace add MongLong0214/commitlore",
           "/plugin install commitlore@commitlore"
         ],
-        "installer": "curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.2/install.sh | sh -s v1.2.2",
-        "release": "https://github.com/MongLong0214/commitlore/releases/tag/v1.2.2"
+        "installer": "curl -fsSL https://raw.githubusercontent.com/MongLong0214/commitlore/v1.2.3/install.sh | sh -s v1.2.3",
+        "release": "https://github.com/MongLong0214/commitlore/releases/tag/v1.2.3"
       },
       "runtime": {
         "transport": "stdio",
diff --git a/src/commands/capture.ts b/src/commands/capture.ts
index f7e8de75..90e70de9 100644
--- a/src/commands/capture.ts
+++ b/src/commands/capture.ts
@@ -39,7 +39,7 @@ import {
   configuredTrustedSignerFingerprints,
   configuredTrustedAuthors,
 } from '../core/trusted-authors.js';
-import { parseDraft } from '../core/harvest.js';
+import { parseDraft, type TranscriptWindow } from '../core/harvest.js';
 import { gcPending } from '../core/pending-gc.js';
 import type { GuardAdvisory } from '../core/pending.js';
 
@@ -80,6 +80,13 @@ export interface CaptureResult {
   nonce: string | null;
   staged: boolean;
   prompt?: string;
+  /**
+   * What of the transcript `prompt` carries, present whenever `prompt` is
+   * (#873). The prompt is bounded; verification is not — it reads the whole
+   * transcript. Without this a caller could not tell a slice from the session,
+   * and the prompt was previously the session, at whatever size that was.
+   */
+  transcript_window?: TranscriptWindow;
   guard_advisory?: GuardAdvisory | null;
   /**
    * Every reason a record was refused (#309). Both sources are included: the
@@ -238,6 +245,7 @@ const runCapturePipeline = (opts: {
       nonce: null,
       staged: false,
       prompt: prepareResult.prompt,
+      transcript_window: prepareResult.transcript_window,
       guard_advisory: prepareResult.guard_advisory,
     };
   }
@@ -383,7 +391,11 @@ export const register = (program: Command): void => {
     // when a subcommand was invoked, so `capture gc` — which needs no transcript
     // — would fail before its own action ran. The requirement is enforced in the
     // action below instead, where it applies only to the capture flow itself.
-    .option('--transcript <path>', 'path to the session transcript file')
+    .option(
+      '--transcript <path>',
+      'path to the session transcript file (the prompt carries its last 256 KiB; ' +
+        'COMMITLORE_TRANSCRIPT_BUDGET_BYTES changes that, and verification always reads all of it)',
+    )
     .option('--diff <path>', 'path to the diff file (defaults to the staged diff)')
     .option('--draft <path>', 'path to the draft JSON file (omit for prompt-only mode)')
     .option('--out <path>', 'write the pending nonce to a file')
diff --git a/src/core/capture-prepare.ts b/src/core/capture-prepare.ts
index da1f1f82..6045de63 100644
--- a/src/core/capture-prepare.ts
+++ b/src/core/capture-prepare.ts
@@ -10,7 +10,7 @@ import { createHash, randomBytes } from 'node:crypto';
 import { markCaptureError } from './capture-outcome.js';
 import { execGitOrThrow } from './git.js';
 import { guard, renderGuardMatch, type GuardResult } from './guard.js';
-import { buildHarvestPrompt } from './harvest.js';
+import { buildHarvestPromptWithWindow, type TranscriptWindow } from './harvest.js';
 import { policySourceLabel, resolvePolicy } from './capture-policy.js';
 import {
   createPending,
@@ -136,6 +136,13 @@ export interface PrepareResult {
   policy_identity_hash: string;
   source_hashes: { transcript: string; diff: string };
   prompt: string;
+  /**
+   * What of the transcript the prompt carries (#873). The prompt is bounded;
+   * `source_hashes.transcript` and every verification are not — they are over
+   * the whole transcript. A caller that assumed the prompt was the session had
+   * no way to tell, so this says it.
+   */
+  transcript_window: TranscriptWindow;
   guard_advisory: GuardAdvisory | null;
   /**
    * A named reason when a policy file exists but could not be used (T-1110).
@@ -158,6 +165,7 @@ interface PreparedValues {
   policy_identity_hash: string;
   source_hashes: { transcript: string; diff: string };
   prompt: string;
+  transcript_window: TranscriptWindow;
   guard_advisory: GuardAdvisory | null;
   policy_error: string | null;
 }
@@ -248,13 +256,16 @@ const prepareValues = (opts: {
           : { trustedSignerFingerprints: opts.trustedSignerFingerprints }),
       });
 
+  const harvest = buildHarvestPromptWithWindow({ transcript, diff });
+
   return {
     base_head: baseHead,
     staged_diff_hash: stagedDiffHash,
     staged_tree_oid: stagedTreeOid,
     policy_identity_hash: policy.identityHash,
     source_hashes: sourceHashes,
-    prompt: buildHarvestPrompt({ transcript, diff }),
+    prompt: harvest.prompt,
+    transcript_window: harvest.window,
     guard_advisory: advisory,
     policy_error: policy.error,
   };
@@ -297,6 +308,7 @@ export const prepareCaptureContext = (opts: PrepareCaptureOptions): PrepareResul
     policy_identity_hash: prepared.policy_identity_hash,
     source_hashes: prepared.source_hashes,
     prompt: prepared.prompt,
+    transcript_window: prepared.transcript_window,
     policy_error: prepared.policy_error,
     guard_advisory: prepared.guard_advisory,
   };
@@ -335,6 +347,7 @@ export const prepareCaptureContextReadOnly = (opts: PrepareCaptureOptions & {
     policy_identity_hash: prepared.policy_identity_hash,
     source_hashes: prepared.source_hashes,
     prompt: prepared.prompt,
+    transcript_window: prepared.transcript_window,
     policy_error: prepared.policy_error,
     guard_advisory: prepared.guard_advisory,
     pending,
diff --git a/src/core/harvest.ts b/src/core/harvest.ts
index dbc51c1a..b07552d6 100644
--- a/src/core/harvest.ts
+++ b/src/core/harvest.ts
@@ -332,12 +332,156 @@ const vocabularyBlock = (entries: VocabularyEntry[]): string[] => {
   return lines;
 };
 
-/** Line-numbered so a citation can name a range the verifier can find again. */
-const numberLines = (text: string): string => {
+/**
+ * Line-numbered so a citation can name a range the verifier can find again.
+ *
+ * `firstLine` is the number the first line of `text` carries in the whole
+ * transcript, which is 1 unless the prompt is showing a window of it. The
+ * numbers have to be the transcript's own: verification reads the whole
+ * transcript, and a locator renumbered to a window would name a different line
+ * of the file it is checked against (#873).
+ */
+const numberLines = (text: string, firstLine = 1): string => {
   const lines = text.split('\n');
   if (lines.length > 1 && lines[lines.length - 1] === '') lines.pop();
-  const width = String(lines.length).length;
-  return lines.map((line, index) => `${String(index + 1).padStart(width)} | ${line}`).join('\n');
+  const width = String(firstLine + lines.length - 1).length;
+  return lines
+    .map((line, index) => `${String(firstLine + index).padStart(width)} | ${line}`)
+    .join('\n');
+};
+
+/**
+ * How much transcript the prompt may carry.
+ *
+ * #873: the prompt embedded the transcript whole, so a 67,981,436-byte session
+ * produced a 67,468,122-byte prompt — larger than any model can read, which
+ * makes the capture pipeline unusable exactly on the long sessions that have
+ * the most to record. The reporter measured the pipeline working again at
+ * roughly half a megabyte and said, correctly, that they had not measured where
+ * the useful boundary is. Neither has this: 256 KiB is a bound chosen to be
+ * comfortably readable by current models while still carrying far more than the
+ * ~537,250-byte slice that was shown to work, not a claim about how much
+ * context a good record needs.
+ *
+ * It is the last bytes rather than the first: a decision is taken near the end
+ * of the session that implements it, and the diff being captured is that end.
+ */
+const DEFAULT_TRANSCRIPT_BUDGET_BYTES = 256 * 1024;
+
+/** An override, for an operator whose model reads more, or less, than this. */
+const transcriptBudgetBytes = (): number => {
+  const raw = Number(process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES']);
+  return Number.isFinite(raw) && raw > 0 ? Math.floor(raw) : DEFAULT_TRANSCRIPT_BUDGET_BYTES;
+};
+
+/**
+ * What the prompt actually showed, so a caller is never left to assume it was
+ * the session. Nothing here changes what verification reads — the whole
+ * transcript is still hashed and still searched for every quote.
+ */
+export interface TranscriptWindow {
+  /** 1-based number, in the whole transcript, of the window's first line. */
+  first_line: number;
+  last_line: number;
+  total_lines: number;
+  /** Bytes of the whole transcript, so the caller can see the size it did not get. */
+  total_bytes: number;
+  /** Bytes of the window as the prompt carries it. */
+  window_bytes: number;
+  /** False when the whole transcript fitted and the window is the session. */
+  truncated: boolean;
+  /**
+   * True when `first_line` is shown from its middle. One JSONL line can hold a
+   * whole tool result and outrun the budget by itself, and a window of no lines
+   * at all would be worse than a window of one partial one.
+   */
+  first_line_partial: boolean;
+}
+
+/** Byte length as the prompt will carry it, newline included. */
+const lineBytes = (line: string): number => Buffer.byteLength(line, 'utf8') + 1;
+
+/**
+ * The last `budget` bytes of one line, without a split codepoint at the front.
+ * Slicing a UTF-8 buffer mid-character yields U+FFFD, and a replacement
+ * character inside a quotable line is a character nobody can copy back.
+ */
+const tailBytes = (line: string, budget: number): string =>
+  Buffer.from(line, 'utf8').subarray(-budget).toString('utf8').replace(/^\uFFFD+/, '');
+
+/**
+ * The tail of `transcript` that fits in `budget` bytes, in whole lines where
+ * whole lines fit.
+ */
+export const windowTranscript = (
+  transcript: string,
+  budget: number = transcriptBudgetBytes(),
+): { text: string; window: TranscriptWindow } => {
+  const lines = transcript.split('\n');
+  if (lines.length > 1 && lines[lines.length - 1] === '') lines.pop();
+  const totalLines = lines.length;
+  const totalBytes = Buffer.byteLength(transcript, 'utf8');
+
+  let kept = 0;
+  let bytes = 0;
+  for (let index = totalLines - 1; index >= 0; index -= 1) {
+    const next = bytes + lineBytes(lines[index]!);
+    if (next > budget && kept > 0) break;
+    if (next > budget) break; // not even the last line fits whole
+    bytes = next;
+    kept += 1;
+  }
+
+  if (kept === 0) {
+    // One line longer than the whole budget. Showing its tail keeps the window
+    // anchored where the decision is, and the line number stays the file's.
+    const last = lines[totalLines - 1] ?? '';
+    const text = tailBytes(last, budget);
+    return {
+      text,
+      window: {
+        first_line: totalLines,
+        last_line: totalLines,
+        total_lines: totalLines,
+        total_bytes: totalBytes,
+        window_bytes: Buffer.byteLength(text, 'utf8'),
+        truncated: true,
+        first_line_partial: true,
+      },
+    };
+  }
+
+  const firstLine = totalLines - kept + 1;
+  const text = lines.slice(firstLine - 1).join('\n');
+  return {
+    text,
+    window: {
+      first_line: firstLine,
+      last_line: totalLines,
+      total_lines: totalLines,
+      total_bytes: totalBytes,
+      window_bytes: Buffer.byteLength(text, 'utf8'),
+      truncated: firstLine > 1,
+      first_line_partial: false,
+    },
+  };
+};
+
+/**
+ * The line the prompt carries when it is showing a window, so the reader is
+ * told rather than left to infer it from a transcript that starts mid-sentence.
+ */
+const windowNotice = (window: TranscriptWindow): string[] => {
+  if (!window.truncated) return [];
+  const omitted = window.first_line - 1;
+  return [
+    `(This is the end of the transcript: lines ${window.first_line}-${window.last_line} of ` +
+      `${window.total_lines}, ${omitted} earlier line(s) omitted to bound this prompt` +
+      `${window.first_line_partial ? `, and line ${window.first_line} is shown from its middle` : ''}. ` +
+      'The numbers below are the transcript\'s own, so a locator you write still names ' +
+      'the line in the whole file. Cite only what you can see here.)',
+    '',
+  ];
 };
 
 const outputBlock = (entries: VocabularyEntry[]): string[] => {
@@ -414,15 +558,19 @@ export const buildHarvestContract = (): string => {
 };
 
 /**
- * Builds the prompt contract handed to the user's agent session. Deterministic
- * by construction — no clock, no randomness, no model — so the same transcript
- * and diff always produce the same bytes.
+ * Builds the prompt contract handed to the user's agent session, and says what
+ * of the transcript it carries. Deterministic by construction — no clock, no
+ * randomness, no model — so the same transcript, diff and budget always produce
+ * the same bytes.
  */
-export const buildHarvestPrompt = (input: HarvestInput): string => {
+export const buildHarvestPromptWithWindow = (
+  input: HarvestInput,
+): { prompt: string; window: TranscriptWindow } => {
   const entries = loadVocabulary().filter((entry) => entry.key !== 'Verified');
   const diff = input.diff.trim() === '' ? '(no diff)' : input.diff.replace(/\n+$/, '');
+  const { text, window } = windowTranscript(input.transcript);
 
-  return [
+  const prompt = [
     '# CommitLore harvest',
     '',
     'You are recording the decision context for a change that is about to be',
@@ -443,15 +591,22 @@ export const buildHarvestPrompt = (input: HarvestInput): string => {
     '',
     '## TRANSCRIPT',
     '',
-    numberLines(input.transcript),
+    ...windowNotice(window),
+    numberLines(text, window.first_line),
     '',
     '## DIFF',
     '',
     diff,
     '',
   ].join('\n');
+
+  return { prompt, window };
 };
 
+/** The prompt alone, for the callers that only emit it. */
+export const buildHarvestPrompt = (input: HarvestInput): string =>
+  buildHarvestPromptWithWindow(input).prompt;
+
 const RECORD_FIELDS = ['trailers', 'evidence'];
 const EVIDENCE_FIELDS = ['key', 'source', 'quote', 'locator'];
 const EVIDENCE_SOURCES: readonly EvidenceSource[] = ['transcript', 'diff'];
diff --git a/src/mcp/server.ts b/src/mcp/server.ts
index 5f2767f9..92292b34 100644
--- a/src/mcp/server.ts
+++ b/src/mcp/server.ts
@@ -379,7 +379,10 @@ const TOOLS: readonly Tool[] = [
     description:
       'Prepare a capture transaction: computes binding conditions (HEAD, staged diff, tree, ' +
       'policy hash), generates the prompt contract for the agent to use, and persists a ' +
-      'phase:"prepared" pending transaction. Returns the nonce needed for verify and stage.',
+      'phase:"prepared" pending transaction. Returns the nonce needed for verify and stage. ' +
+      'The prompt carries the end of the transcript rather than all of it; transcript_window ' +
+      'says which lines, numbered as the whole transcript numbers them. Verification still ' +
+      'reads the whole transcript, so quote only what the prompt shows you.',
     inputSchema: {
       type: 'object',
       properties: {
@@ -638,6 +641,11 @@ export const createServer = (opts: McpServerOptions = {}): Server => {
         policy_identity_hash: result.policy_identity_hash,
         source_hashes: result.source_hashes,
         prompt: result.prompt,
+        // What of the transcript that prompt carries (#873). It travels here
+        // for the same reason the two fields below do: MCP is the first-class
+        // surface for every agent but the plugin, and an agent handed a slice
+        // of its own session with no way to tell would cite the whole of it.
+        transcript_window: result.transcript_window,
         // MCP is the first-class surface for every agent other than the Claude
         // Code plugin, so both of these must travel here and not only to the
         // pending file and the CLI. `guard_advisory` is always present, never
diff --git a/test/capture-prompt-budget.test.ts b/test/capture-prompt-budget.test.ts
new file mode 100644
index 00000000..87c69d50
--- /dev/null
+++ b/test/capture-prompt-budget.test.ts
@@ -0,0 +1,230 @@
+/**
+ * #873: `capture` embedded the transcript in the prompt whole, so the prompt
+ * was the session. Measured by the reporter: a 67,981,436-byte transcript
+ * produced a 67,468,122-byte prompt — 99.3% of the file, and larger than any
+ * model can read. The pipeline itself was fine; the reporter confirmed that by
+ * slicing the transcript to `tail -n 400` (537,250 bytes) and watching the same
+ * command work.
+ *
+ * That made the failure worse than a size: it is silent. Prompt-only mode
+ * returns `outcome: "empty"`, `staged: false`, **exit 0**, so an operator who
+ * tried capture once on a real session got a prompt they could not use and no
+ * statement that anything was wrong. In the repository where it was measured,
+ * `stale` reported 1 record in 1000 commits with unattended capture already
+ * enabled — permission was never the obstacle.
+ *
+ * Three properties are load-bearing here, and each has a test below that fails
+ * when its half of the fix is reverted:
+ *
+ * 1. The prompt is bounded, whatever the transcript weighs.
+ * 2. The line numbers in that window are the *whole transcript's* numbers.
+ *    Verification reads the whole transcript, so a window renumbered from 1
+ *    would have every locator name a different line of the file it is checked
+ *    against.
+ * 3. The caller is told it received a window. The old prompt was the session,
+ *    so nothing downstream had any reason to ask; a bounded prompt that does
+ *    not say so is the same silence in a smaller package.
+ */
+
+import { execSync } from 'node:child_process';
+import { mkdtempSync, rmSync, writeFileSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+
+import { afterAll, afterEach, describe, expect, it } from 'vitest';
+
+import { prepareCaptureContext } from '../src/core/capture-prepare.js';
+import { verifyCaptureRecords } from '../src/core/capture-verify.js';
+import { buildHarvestPromptWithWindow, windowTranscript, type DraftRecord } from '../src/core/harvest.js';
+
+const scratch: string[] = [];
+afterAll(() => {
+  for (const dir of scratch) rmSync(dir, { recursive: true, force: true });
+});
+
+const makeRepo = (): string => {
+  const dir = mkdtempSync(join(tmpdir(), 'capture-prompt-budget-'));
+  scratch.push(dir);
+  execSync('git init --quiet --initial-branch=main', { cwd: dir });
+  execSync('git config user.name "Test"', { cwd: dir });
+  execSync('git config user.email "test@test.com"', { cwd: dir });
+  execSync('git config commit.gpgsign false', { cwd: dir });
+  writeFileSync(join(dir, 'a.txt'), 'hello\n');
+  execSync('git add a.txt', { cwd: dir });
+  execSync('git commit -m "init" --quiet', { cwd: dir });
+  writeFileSync(join(dir, 'a.txt'), 'hello\nworld\n');
+  execSync('git add a.txt', { cwd: dir });
+  return dir;
+};
+
+/** A transcript of `lines` numbered lines, each recognisable by its own number. */
+const transcriptOf = (lines: number, padTo = 0): string =>
+  Array.from({ length: lines }, (_, index) => {
+    const body = `line ${index + 1} of the session`;
+    return padTo > body.length ? body + ' '.repeat(padTo - body.length) : body;
+  }).join('\n');
+
+const DIFF = 'diff --git a/x b/x\n--- a/x\n+++ b/x\n@@ -1 +1 @@\n-old\n+new\n';
+
+describe('#873 windowTranscript bounds what the prompt can carry', () => {
+  it('is the whole transcript, unmarked, when the whole transcript fits', () => {
+    const transcript = transcriptOf(20);
+    const { text, window } = windowTranscript(transcript, 64 * 1024);
+
+    expect(text).toBe(transcript);
+    expect(window.truncated).toBe(false);
+    expect(window.first_line).toBe(1);
+    expect(window.last_line).toBe(20);
+    expect(window.total_lines).toBe(20);
+    expect(window.first_line_partial).toBe(false);
+  });
+
+  it('keeps the end of a transcript that does not fit, within the budget', () => {
+    const transcript = transcriptOf(5000);
+    const { text, window } = windowTranscript(transcript, 4096);
+
+    expect(Buffer.byteLength(text, 'utf8')).toBeLessThanOrEqual(4096);
+    expect(window.truncated).toBe(true);
+    expect(window.last_line).toBe(5000);
+    expect(window.total_lines).toBe(5000);
+    expect(window.first_line).toBeGreaterThan(1);
+    // The end, not the beginning: the decision is taken where the session ends.
+    expect(text.endsWith('line 5000 of the session')).toBe(true);
+    expect(text.startsWith(`line ${window.first_line} of the session`)).toBe(true);
+    expect(window.total_bytes).toBe(Buffer.byteLength(transcript, 'utf8'));
+    expect(window.window_bytes).toBe(Buffer.byteLength(text, 'utf8'));
+  });
+
+  it('shows the tail of one line that outruns the budget by itself', () => {
+    // A single JSONL line can hold an entire tool result. A window of no lines
+    // at all would be worse than a window of one partial line.
+    const transcript = `first\n${'x'.repeat(20_000)}`;
+    const { text, window } = windowTranscript(transcript, 1024);
+
+    expect(Buffer.byteLength(text, 'utf8')).toBeLessThanOrEqual(1024);
+    expect(window.first_line).toBe(2);
+    expect(window.last_line).toBe(2);
+    expect(window.first_line_partial).toBe(true);
+    expect(window.truncated).toBe(true);
+  });
+
+  it('never leaves a split codepoint at the front of a partial line', () => {
+    // Slicing a UTF-8 buffer mid-character yields U+FFFD, and a replacement
+    // character inside a quotable line is a character nobody can copy back.
+    const transcript = '한'.repeat(4000);
+    const { text } = windowTranscript(transcript, 1000);
+
+    expect(text.startsWith('�')).toBe(false);
+    expect(Buffer.byteLength(text, 'utf8')).toBeLessThanOrEqual(1000);
+  });
+
+  it('honours COMMITLORE_TRANSCRIPT_BUDGET_BYTES', () => {
+    const previous = process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'];
+    process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'] = '2048';
+    try {
+      const { window } = windowTranscript(transcriptOf(5000));
+      expect(window.window_bytes).toBeLessThanOrEqual(2048);
+      expect(window.truncated).toBe(true);
+    } finally {
+      if (previous === undefined) delete process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'];
+      else process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'] = previous;
+    }
+  });
+});
+
+describe('#873 the prompt numbers the window as the transcript numbers it', () => {
+  const budgeted = (lines: number): ReturnType<typeof buildHarvestPromptWithWindow> => {
+    const previous = process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'];
+    process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'] = '4096';
+    try {
+      return buildHarvestPromptWithWindow({ transcript: transcriptOf(lines), diff: DIFF });
+    } finally {
+      if (previous === undefined) delete process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'];
+      else process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'] = previous;
+    }
+  };
+
+  it('gives the window\'s first line its number in the whole transcript', () => {
+    const { prompt, window } = budgeted(5000);
+
+    // The exact property a renumbered window would break: the line the prompt
+    // labels N really is line N of the transcript verification will read.
+    expect(prompt).toContain(`${window.first_line} | line ${window.first_line} of the session`);
+    expect(prompt).toContain(`${window.last_line} | line ${window.last_line} of the session`);
+    expect(prompt).not.toContain('1 | line 1 of the session');
+  });
+
+  it('says it is a window, and how much was left out', () => {
+    const { prompt, window } = budgeted(5000);
+
+    expect(prompt).toContain(`lines ${window.first_line}-${window.last_line} of ${window.total_lines}`);
+    expect(prompt).toContain('earlier line(s) omitted to bound this prompt');
+  });
+
+  it('says nothing about a window when there is no window', () => {
+    const { prompt } = buildHarvestPromptWithWindow({ transcript: transcriptOf(5), diff: DIFF });
+
+    expect(prompt).not.toContain('omitted to bound this prompt');
+    expect(prompt).toContain('1 | line 1 of the session');
+  });
+});
+
+describe('#873 capture returns a prompt a model can read, and says what it is', () => {
+  const previous = process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'];
+  afterEach(() => {
+    if (previous === undefined) delete process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'];
+    else process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'] = previous;
+  });
+
+  it('does not return a prompt the size of the transcript', () => {
+    // The reporter's ratio, at a size a test can afford: the prompt used to be
+    // ~99.3% of the transcript, so it grew without bound with the session.
+    const cwd = makeRepo();
+    process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'] = '8192';
+    const transcript = transcriptOf(40_000, 200);
+
+    const result = prepareCaptureContext({ cwd, transcript });
+
+    expect(Buffer.byteLength(transcript, 'utf8')).toBeGreaterThan(4_000_000);
+    expect(Buffer.byteLength(result.prompt, 'utf8')).toBeLessThan(64 * 1024);
+    expect(result.transcript_window.truncated).toBe(true);
+    expect(result.transcript_window.total_lines).toBe(40_000);
+    expect(result.transcript_window.last_line).toBe(40_000);
+  });
+
+  it('hashes and verifies the whole transcript, not the window', () => {
+    // The bound is on the prompt alone. If it ever reached the hash, `verify`
+    // would be checking quotes against a slice — and a caller passing the
+    // session it actually had would be told the transcript was substituted.
+    const cwd = makeRepo();
+    process.env['COMMITLORE_TRANSCRIPT_BUDGET_BYTES'] = '4096';
+    const quote = 'we ruled out the queue worker because it loses ordering';
+    const transcript = `${quote}\n${transcriptOf(5000)}`;
+
+    const prepared = prepareCaptureContext({ cwd, transcript });
+    expect(prepared.transcript_window.first_line).toBeGreaterThan(1);
+
+    const draft: DraftRecord[] = [
+      {
+        trailers: [
+          { key: 'Ruled-out', value: 'queue worker | loses ordering' },
+          { key: 'Record-Id', value: 'r-window1' },
+        ],
+        evidence: [{ key: 'Ruled-out', source: 'transcript', quote, locator: 'L1-L1' }],
+      },
+    ];
+
+    const verified = verifyCaptureRecords({
+      nonce: prepared.nonce,
+      draft,
+      transcript,
+      diff: execSync('git diff --cached', { cwd, encoding: 'utf8' }),
+      cwd,
+    });
+
+    // The quote is on line 1, which the window does not show — and it verifies,
+    // because verification never looked at the window.
+    expect(verified.rejected).toHaveLength(0);
+    expect(verified.accepted).toHaveLength(1);
+  });
+});