Appendix E: Version Evolution Log

The core analysis in this book is based on Claude Code v2.1.88 (with full source map, enabling recovery of 4,756 source files). This appendix records key changes in subsequent versions and their impact on each chapter.

Navigation tip: Each change links to the corresponding chapter's version evolution section. Click the chapter number to jump.

Since Anthropic removed source map distribution starting from v2.1.89, the following analysis is based on bundle string signal comparison + v2.1.88 source code-assisted inference, with limited depth.

v2.1.88 -> v2.1.91

Overview: cli.js +115KB | Tengu events +39/-6 | Environment variables +8/-3 | Source Map removed

High-Impact Changes

Medium-Impact Changes

Low-Impact Changes

v2.1.91 New Features in Detail

The following three features did not exist at all in v2.1.88 source code and are new in v2.1.91. Analysis is based on v2.1.91 bundle reverse engineering.

1. Powerup Lessons — Interactive Feature Tutorial System

Events: tengu_powerup_lesson_opened, tengu_powerup_lesson_completed

v2.1.88 status: Did not exist. No powerup or lesson-related code in restored-src/src/.

v2.1.91 reverse engineering findings:

Powerup Lessons is a built-in interactive tutorial system containing 10 course modules that teach users how to use Claude Code's core features. The complete course registry extracted from the bundle:

Technical implementation (from bundle reverse engineering):

// Course opened event
logEvent("tengu_powerup_lesson_opened", {
  lesson_id: lesson.id,           // Course ID
  was_already_unlocked: unlocked.has(lesson.id),  // Already unlocked?
  unlocked_count: unlocked.size   // Total unlocked count
})

// Course completed event
logEvent("tengu_powerup_lesson_completed", {
  lesson_id: id,
  unlocked_count: newUnlocked.size,
  all_unlocked: newUnlocked.size === lessons.length  // All completed?
})

Unlock state is persisted to user configuration via powerupsUnlocked. Each course contains a title, tagline, rich text content (with terminal animation demos), and the UI uses check/circle markers for completion status, triggering an "easter egg" animation when all courses are completed.

Book relevance: The 10 course modules of Powerup Lessons cover nearly all core topics from Parts 2 through 6 of this book — from permission modes (ch16-17) to sub-agents (ch20) to MCP (ch22). It represents Anthropic's official prioritization of "which features users should master" and can serve as a reference for this book's "What You Can Do" sections.

2. Write Append Mode — File Append Writing

Event: tengu_write_append_used

v2.1.88 status: Did not exist. v2.1.88's Write tool only supported overwrite (complete replacement) mode.

v2.1.91 reverse engineering findings:

The Write tool's inputSchema gained a new mode parameter:

// v2.1.91 bundle reverse engineering
inputSchema: {
  file_path: string,
  content: string,
  mode: "overwrite" | "append"  // New in v2.1.91
}

mode parameter description (extracted from bundle):

Write mode. 'overwrite' (default) replaces the file. Use 'append' to add content to the end of an existing file instead of rewriting the full content — e.g. for logs, accumulating output, or adding entries to a list.

Feature Gate: Append mode is controlled by GrowthBook flag tengu_maple_forge_w8k. When the flag is off, the mode field is .omit()'d from the schema, making it invisible to the model.

// v2.1.91 bundle reverse engineering
function getWriteSchema() {
  return getFeatureValue("tengu_maple_forge_w8k", false)
    ? fullSchema()           // Includes mode parameter
    : fullSchema().omit({ mode: true })  // Hides mode parameter
}

Book relevance: Affects ch02 (tool system overview) and ch08 (tool prompts). In v2.1.88, the Write tool's prompt explicitly stated "This tool will overwrite the existing file" — v2.1.91's append mode changes this constraint, and the model can now choose to append rather than overwrite.

3. Message Rating — Message Rating Feedback

Event: tengu_message_rated

v2.1.88 status: Did not exist. v2.1.88 had tengu_feedback_survey_* series events (session-level feedback) but no message-level rating.

v2.1.91 reverse engineering findings:

Message Rating is a message-level user feedback mechanism that allows users to rate individual Claude responses. Implementation extracted from bundle reverse engineering:

// v2.1.91 bundle reverse engineering
function rateMessage(messageUuid, sentiment) {
  const wasAlreadyRated = ratings.get(messageUuid) === sentiment
  // Clicking the same rating again → clear (toggle behavior)
  if (wasAlreadyRated) {
    ratings.delete(messageUuid)
  } else {
    ratings.set(messageUuid, sentiment)
  }

  logEvent("tengu_message_rated", {
    message_uuid: messageUuid,  // Message unique ID
    sentiment: sentiment,       // Rating direction (e.g., thumbs_up/thumbs_down)
    cleared: wasAlreadyRated    // Was the rating cleared?
  })

  // Show thank-you notification after rating
  if (!wasAlreadyRated) {
    addNotification({
      key: "message-rated",
      text: "thanks for improving claude!",
      color: "success",
      priority: "immediate"
    })
  }
}

UI mechanics:

• Rating functionality is injected into the message list via React Context (MessageRatingProvider)
• Rating state is stored in memory as Map<messageUuid, sentiment>
• Supports toggle — clicking the same rating again clears it
• After rating, a green notification "thanks for improving claude!" appears

Book relevance: Related to ch29 (Observability Engineering). v2.1.88's feedback system was session-level (tengu_feedback_survey_*); v2.1.91 adds message-level rating, refining feedback granularity from "was the whole session good" to "was this specific response good." This provides Anthropic with more fine-grained training signals for RLHF (Reinforcement Learning from Human Feedback).

Experimental Codename Events

The following events with random codenames are A/B tests with undisclosed purposes:

v2.1.91 -> v2.1.92 (Incremental Changes)

Based on signal differences extracted between v2.1.91 and v2.1.92 bundles. Full comparison report available at docs/version-diffs/v2.1.88-vs-v2.1.92.md.

Overview

Key Additions

Key Removals

New Environment Variables

Design Trends

The v2.1.91 -> v2.1.92 increment is small but directionally clear:

1. Security strategy descends from application layer to system layer (tree-sitter -> seccomp)
2. Tool system expands from pure execution to advisory (AdvisorTool)
3. Configuration management moves from purely static to runtime-mutable (Stop Hook dynamic management)
4. Enterprise onboarding barrier continues to lower (Bedrock wizard)

Use scripts/cc-version-diff.sh to generate diff data; docs/anchor-points.md provides subsystem anchor point locations

v2.1.92 -> v2.1.100

Overview: cli.js +870KB (+6.9%) | Tengu events +45/-21 (net +24) | Env vars +8/-2 | New audio-capture vendor

High Impact Changes

Medium Impact Changes

Low Impact Changes

v2.1.100 New Features in Detail

The following features did not exist in v2.1.92 or only had rudimentary form, and are incremental additions in v2.1.92→v2.1.100.

1. Kairos Dream — Background Scheduled Memory Consolidation

Event: tengu_kairos_dream

v2.1.92 status: v2.1.92 already had auto_dream and manual /dream trigger, but no background cron scheduling.

v2.1.100 addition:

Kairos Dream is the third trigger mode for the Dream system — executing memory consolidation automatically via cron scheduling in the background, without waiting for users to start new sessions. Cron expression generation extracted from the bundle:

// v2.1.100 bundle reverse engineering
function P_A() {
  let q = Math.floor(Math.random() * 360);
  return `${q % 60} ${Math.floor(q / 60)} * * *`;
  // Random minute+hour offset, avoids multi-user simultaneous triggers
}

Combined with the auto_dream_skipped event's reason field ("sessions"/"lock"), Kairos Dream implements a complete background memory consolidation lifecycle.

Book relevance: ch24 updated with Dream system analysis (three-tier trigger matrix); ch29 observability chapter can reference auto_dream_skipped skip reason distribution as an observability design case study.

2. Bedrock/Vertex Model Upgrade Wizard

Events: 18 events (9 Bedrock + 9 Vertex), symmetric structure

v2.1.92 status: v2.1.92 only had Bedrock's setup_started/complete/cancelled (3 events).

v2.1.100 addition:

Complete model upgrade detection and automatic switching mechanism. Design highlights:

1. Unpinned model detection: Scans user configuration to find model tiers not explicitly pinned via environment variables
2. Accessibility probing: probeBedrockModel / probeVertexModel verify whether new models are available in the user's account
3. User confirmation: Upgrades don't auto-execute; require user accept/decline
4. Persistent decline: Declined upgrades are recorded in user settings, preventing repeated prompting
5. Default fallback: When default model is inaccessible, automatic fallback to same-tier alternative

The Vertex wizard (vertex_setup_started etc.) is new in v2.1.100; v2.1.92 had no interactive Vertex setup.

3. Autofix PR — Remote Auto-Fix

Events: tengu_autofix_pr_started, tengu_autofix_pr_result

v2.1.92 status: Did not exist. v2.1.92 had ultraplan and ultrareview, but no autofix-pr.

v2.1.100 addition:

Autofix PR is the fourth remote agent task type, listed alongside remote-agent, ultraplan, and ultrareview in the XAY remote task type registry. Workflow extracted from the bundle:

// v2.1.100 bundle reverse engineering
// Remote task type registry
XAY = ["remote-agent", "ultraplan", "ultrareview", "autofix-pr", "background-pr"];

// Autofix PR launch
d("tengu_autofix_pr_started", {});
let b = await kt({
  initialMessage: h,
  source: "autofix_pr",
  branchName: P,
  reuseOutcomeBranch: P,
  title: `Autofix PR: ${k}/${R}#${v} (${P})`
});

Autofix PR spawns a remote Claude Code session that monitors a specified Pull Request and automatically fixes issues (CI failures, code review feedback). Unlike Ultraplan (planning) and Ultrareview (reviewing), Autofix PR focuses on executing fixes.

Note background-pr also appears in the task type list, suggesting another background PR processing mode.

4. Team Onboarding — Team Usage Report

Events: tengu_team_onboarding_invoked, tengu_team_onboarding_generated, tengu_team_onboarding_discovery_shown

v2.1.92 status: Did not exist.

v2.1.100 addition:

Team onboarding report generator that collects user usage data (session count, slash command count, MCP server count) and generates a guided document from a template. Key parameters extracted from the bundle:

• windowDays: Analysis window (1-365 days)
• sessionCount, slashCommandCount, mcpServerCount: Usage statistic dimensions
• GUIDE_TEMPLATE, USAGE_DATA: Report template variables

The cedar_inlet experiment event controls team onboarding discovery display (discovery_shown), suggesting this is an A/B tested feature.

Experiment Codename Events

The following events with random codenames are A/B tests with undisclosed purposes:

Experiments present in v2.1.92 but removed in v2.1.100: amber_lantern, editafterwrite_qpl, lean_sub_pf, maple_forge_w, relpath_gh.

Design Trends

The v2.1.92→v2.1.100 evolution direction:

1. Memory system from passive to active (auto_dream → kairos_dream scheduled execution + observable skip reasons)
2. Cloud platforms from configuration to wizards (manual env vars → interactive setup wizards + automatic model upgrade detection)
3. IDE bridge architecture restructured (bridge_repl largely removed, 16 events cleared — transitioning to new communication mechanism)
4. Remote agent family expansion (ultraplan/ultrareview → + autofix-pr + background-pr)
5. Context optimization refinement (tool_result_dedup reduces duplicates + MAX_CONTEXT_TOKENS user-controllable)

Use scripts/cc-version-diff.sh to generate diff data; docs/anchor-points.md provides subsystem anchor point locations