From 1bbf3466875a8a274d6a528b096a236afae7bdc6 Mon Sep 17 00:00:00 2001 From: Florian BRUNIAUX Date: Tue, 17 Mar 2026 08:42:20 +0100 Subject: [PATCH] docs: update Claude Code releases to v2.1.77 - Opus 4.6 default max output raised to 64k tokens; upper bound 128k - Security fix: PreToolUse hooks bypass of enterprise deny rules - allowRead sandbox setting, /branch command, /copy N shorthand - Breaking: Agent tool resume param removed (use SendMessage instead) - Fixed auto-updater memory leak; fixed --resume history truncation Co-Authored-By: Claude Sonnet 4.6 --- CHANGELOG.md | 21 ++++++++++++++++ guide/core/claude-code-releases.md | 29 +++++++++++++++++++--- machine-readable/claude-code-releases.yaml | 16 ++++++++++-- 3 files changed, 61 insertions(+), 5 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index a688801..f2be315 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -6,8 +6,29 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/). ## [Unreleased] +### Documentation + +- **Claude Code Releases**: Updated tracking to v2.1.77 + - Opus 4.6 default max output raised to 64k tokens; upper bound for Opus 4.6 and Sonnet 4.6 raised to 128k tokens + - Security fix: `PreToolUse` hooks returning `"allow"` could bypass enterprise `deny` permission rules + - `allowRead` sandbox setting; `/branch` replaces `/fork`; `/copy N` for Nth-latest response + - Breaking: `Agent` tool `resume` parameter removed — use `SendMessage({to: agentId})` instead + - Fixed auto-updater GBs memory leak; fixed `--resume` truncating recent history + +### Changed + +- **External support section — clarified positioning** (`docs/for-cto.md`, `docs/for-tech-leads.md`, `docs/for-cio-ceo.md`): Brown Bag Lunch, talks, and speaker/panelist slots (1-3h) explicitly marked as free and done for networking/challenge purposes. Training/consulting missions framed as open-but-not-actively-sought with "contact for availability and potentially pricing" wording. Contact link updated to `florian.bruniaux.com` across all three files. + +### Fixed + +- **`/release` skill — 3 correctness gaps** (`.claude/commands/release.md` Step 4): (1) Quiz count command fixed: `grep -c ' - id:'` was returning per-file counts and taking only the last file's value — replaced with `grep -r ' - id:' quiz/questions/ | wc -l` to sum across the entire directory. (2) `llms-full.txt` now documents all 4 occurrences requiring update (Metadata block, "For Learning" URL text, "Template Library" section heading, repo tree comment) — previously only Metadata fields were listed, causing silent drift. (3) Verification gate added after Step 4: bash block prints all three llms files' key fields side-by-side against expected values before the commit, making any mismatch visible immediately. + ### Added +- **Packmind — Engineering Standards Distribution** (`guide/ecosystem/third-party-tools.md`, `guide/ultimate-guide.md`, `guide/ecosystem/mcp-servers-ecosystem.md`): Added Packmind (score 4/5, eval #076) as a new "Engineering Standards Distribution" section in third-party-tools. Tool distributes CLAUDE.md + slash commands + skills across repos and agents (Claude Code, Cursor, Copilot, Windsurf) from a single playbook, ships an MCP server, Apache-2.0 CLI self-hostable. Added cross-reference paragraph at end of ultimate-guide.md §3.5 (Team Configuration at Scale) linking the per-project `.claude/rules/` pattern to org-scale tooling. Added Packmind MCP server entry in mcp-servers-ecosystem.md Orchestration section. + +- **3 new AI roles** (`guide/roles/ai-roles.md`): Added MLOps Engineer (§14), AI Developer Advocate (§15), and AI Orchestration Engineer (§16) as full role profiles. Includes responsibilities, required skills, salary benchmarks, entry paths, and key distinctions from adjacent roles. Sections 14→17 renumbered accordingly. Removed "Orchestration engineer" from "What's Not a Role Yet" — job postings at Vista Equity, Zapier, Heidi Health, and Adobe confirm it's now a real title. Career Decision Matrix and Salary Benchmarks updated with all 3 roles. Based on Perplexity market research (March 2026). + - **Failure-triggered context drift pattern** (`guide/core/architecture.md` §Session Degradation Limits): New subsection documenting a distinct degradation mode from compaction drift — repeated tool failures accumulate error noise that dilutes the original intent without filling the context window. Pattern: re-inject core task instructions on every command failure via `PostToolUse` hook, not just after `/compact`. Source: Nick Tune (2026-03-01). Resource evaluation: `docs/resource-evaluations/2026-03-16-nick-tune-workflow-dsl-ddd.md` (score 3/5 — 1 of 3 patterns integrated). - **Identity re-injection after compaction** (`guide/ultimate-guide.md` §7.5 + `examples/hooks/bash/identity-reinjection.sh`): New hook pattern from Nick Tune (Feb 2026). Solves agent identity drift after context compaction in long sessions — `UserPromptSubmit` hook reads transcript, detects missing identity marker in last assistant message, re-injects `.claude/agent-identity.txt` as `additionalContext`. Configurable via `CLAUDE_IDENTITY_FILE` and `CLAUDE_IDENTITY_MARKER` env vars. `reference.yaml` updated with `identity_reinjection_hook` + `identity_reinjection_example` keys. diff --git a/guide/core/claude-code-releases.md b/guide/core/claude-code-releases.md index bb21e9b..d52cf08 100644 --- a/guide/core/claude-code-releases.md +++ b/guide/core/claude-code-releases.md @@ -10,13 +10,13 @@ tags: [reference, release] > **Full details**: [github.com/anthropics/claude-code/CHANGELOG.md](https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md) > **Machine-readable**: [claude-code-releases.yaml](../machine-readable/claude-code-releases.yaml) -**Latest**: v2.1.76 | **Updated**: 2026-03-14 +**Latest**: v2.1.77 | **Updated**: 2026-03-17 --- ## Quick Jump -- [2.1.x Series (January-March 2026)](#21x-series-january-march-2026) — Worktree isolation, background agents, ConfigChange hook, Fast mode Opus 4.6, 1M context, claude.ai MCP connectors, remote-control, auto-memory, /copy command, HTTP hooks, worktree config sharing, ultrathink re-introduced, InstructionsLoaded hook, 4 security fixes, Agent model override restored, 12x SDK token cost reduction, /context actionable suggestions, modelOverrides setting, 1M context Opus 4.6 default for Max/Team/Enterprise, MCP elicitation, PostCompact hook, /effort command +- [2.1.x Series (January-March 2026)](#21x-series-january-march-2026) — Worktree isolation, background agents, ConfigChange hook, Fast mode Opus 4.6, 1M context, claude.ai MCP connectors, remote-control, auto-memory, /copy command, HTTP hooks, worktree config sharing, ultrathink re-introduced, InstructionsLoaded hook, 4 security fixes, Agent model override restored, 12x SDK token cost reduction, /context actionable suggestions, modelOverrides setting, 1M context Opus 4.6 default for Max/Team/Enterprise, MCP elicitation, PostCompact hook, /effort command, Opus 4.6 64k/128k output tokens, allowRead sandbox setting, /branch command - [2.0.x Series (Nov 2025 - Jan 2026)](#20x-series-november-2025---january-2026) — Opus 4.5, Claude in Chrome, Background agents - [Breaking Changes Summary](#breaking-changes-summary) - [Milestone Features](#milestone-features) @@ -25,6 +25,28 @@ tags: [reference, release] ## 2.1.x Series (January-March 2026) +### v2.1.77 (2026-03-17) + +- **New**: ⭐ Opus 4.6 default maximum output tokens raised to 64k; upper bound for Opus 4.6 and Sonnet 4.6 raised to 128k tokens +- **New**: `allowRead` sandbox filesystem setting to re-allow read access within `denyRead` regions +- **New**: `/copy N` to copy the Nth-latest assistant response directly +- **New**: `/branch` command (replaces `/fork`; `/fork` still works as an alias) +- **New**: `SendMessage` now auto-resumes stopped agents in the background instead of returning an error +- **Fixed**: ⚠️ **Security** — `PreToolUse` hooks returning `"allow"` could bypass `deny` permission rules including enterprise managed settings +- **Fixed**: Auto-updater accumulating tens of gigabytes of memory when slash-command overlay repeatedly opened/closed, triggering overlapping binary downloads +- **Fixed**: `--resume` silently truncating recent conversation history due to a race between memory-extraction writes and the main transcript +- **Fixed**: "Always Allow" on compound bash commands (e.g. `cd src && npm test`) saving a single rule for the full string instead of per-subcommand, leading to dead rules and repeated permission prompts +- **Fixed**: Write tool silently converting line endings when overwriting CRLF files or creating files in CRLF directories +- **Fixed**: Cost and token usage not tracked when API falls back to non-streaming mode +- **Fixed**: `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS` not stripping beta tool-schema fields, causing proxy gateways to reject requests +- **Fixed**: Bash tool reporting errors for successful commands when system temp directory path contains spaces +- **Fixed**: Paste being lost when typing immediately after pasting; Ctrl+D in `/feedback` deleting forward instead of exiting +- **Fixed**: Various rendering fixes: ordered list numbers, CJK bleeding, background colors in tmux, hyperlinks opening twice in VS Code +- **Fixed**: Teammate panes not closing when leader exits; iTerm2 session crash when selecting text inside tmux over SSH +- **Breaking**: `Agent` tool no longer accepts a `resume` parameter — use `SendMessage({to: agentId})` to continue a previously spawned agent +- **VSCode**: Fixed gitignore patterns with commas silently excluding filetypes from `@`-mention file picker; improved scroll wheel responsiveness; improved plan preview tab titles +- **Performance**: Faster startup on macOS (~60ms) by reading keychain credentials in parallel with module loading; faster `--resume` on fork-heavy sessions (up to 45% faster, 100-150MB less peak memory) + ### v2.1.76 (2026-03-14) - **New**: ⭐ MCP elicitation support — MCP servers can now request structured input mid-task via an interactive dialog (form fields or browser URL) @@ -762,12 +784,13 @@ tags: [reference, release] | v2.0.58 | Managed settings prefer `C:\Program Files\ClaudeCode` | | v2.1.2 | Deprecated `C:\ProgramData\ClaudeCode` path | -### SDK +### SDK / Agent Tool | Version | Change | |---------|--------| | v2.0.25 | Removed legacy SDK entrypoint → `@anthropic-ai/claude-agent-sdk` | | v2.1.0 | Minimum zod peer dependency: `^4.0.0` | +| v2.1.77 | `Agent` tool no longer accepts `resume` parameter — use `SendMessage({to: agentId})` instead | ### API Ecosystem diff --git a/machine-readable/claude-code-releases.yaml b/machine-readable/claude-code-releases.yaml index df4792c..c939106 100644 --- a/machine-readable/claude-code-releases.yaml +++ b/machine-readable/claude-code-releases.yaml @@ -3,8 +3,8 @@ # Purpose: Track Claude Code product releases for documentation sync # Maintained: Manual updates when new releases are announced -latest: "2.1.76" -updated: "2026-03-14" +latest: "2.1.77" +updated: "2026-03-17" # ════════════════════════════════════════════════════════════════ # RELEASES (newest first, condensed highlights only) @@ -15,6 +15,18 @@ releases: # 2.1.x Series (January-March 2026) # ───────────────────────────────────────────────────────────── + - version: "2.1.77" + date: "2026-03-17" + highlights: + - "⭐ Opus 4.6 default max output tokens raised to 64k; upper bound for Opus 4.6 and Sonnet 4.6 raised to 128k tokens" + - "allowRead sandbox filesystem setting to re-allow read access within denyRead regions" + - "/copy N to copy Nth-latest assistant response; /branch replaces /fork (alias kept)" + - "Fixed PreToolUse hooks returning 'allow' could bypass enterprise deny permission rules — security fix" + - "Fixed auto-updater accumulating GBs of memory from overlapping binary downloads (slash-command overlay loop)" + - "Fixed --resume silently truncating recent history; Fixed Always Allow on compound bash commands saving dead rules" + breaking: + - "Agent tool no longer accepts resume parameter — use SendMessage({to: agentId}) to continue a previously spawned agent" + - version: "2.1.76" date: "2026-03-14" highlights: