settings.md was reading and writing .planning/config.json directly while
gsd-tools config-get/config-set route to .planning/workstreams/<slug>/config.json
when GSD_WORKSTREAM is active, causing silent write-read drift (Closes#2282).
- config.cjs: add cmdConfigPath() — emits the planningDir-resolved config path as
plain text (always raw, no JSON wrapping) so shell substitution works correctly
- gsd-tools.cjs: wire config-path subcommand
- settings.md: resolve GSD_CONFIG_PATH via config-path in ensure_and_load_config;
replace hardcoded cat .planning/config.json and Write to .planning/config.json
with $GSD_CONFIG_PATH throughout
- phase.cjs: fix renameDecimalPhases to preserve zero-padded prefix (06.3 → 06.2
not 6.2) — pre-existing test failure on main
- tests/config.test.cjs: add config-path command tests (#2282)
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(handoffs): include project identity in all Next Up blocks
Adds project_code and project_title to withProjectRoot() and updates
all 30 Next Up headings across 18 workflow files to include
[PROJECT_CODE] PROJECT_TITLE suffix for multi-project clarity.
Closes#1948
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(review): add withProjectRoot tests and fix placeholder syntax (#1951)
Address code review feedback:
- Add 4 tests for project_code/project_title injection in withProjectRoot()
- Fix inconsistent placeholder syntax in continuation-format.md canonical
template (bare-brace → shell-dollar to match variant examples)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(phase): preserve zero-padded prefix in renameDecimalPhases
Captures the zero-padded prefix (e.g. "06" from "06.3-slug") with
(0*${baseInt}) so renamed directories keep their original format
(06.2-slug) instead of stripping padding (6.2-slug).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Higgins <brandonscotthiggins@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extends /gsd-progress with opt-in --forensic mode that appends a
6-check integrity audit after the standard routing report. Default
behavior is byte-for-byte unchanged — the audit only runs when
--forensic is explicitly passed.
Checks: (1) STATE vs artifact consistency, (2) orphaned handoff files,
(3) deferred scope drift, (4) memory-flagged pending work, (5) blocking
operational todos, (6) uncommitted source code. Emits CLEAN or
N INTEGRITY ISSUE(S) FOUND verdict with concrete next actions.
Closes#2189
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds --all to /gsd-discuss-phase so users can skip the AskUserQuestion
area-selection step and jump straight into discussing all gray areas
interactively. Unlike --auto, --all does NOT auto-advance to plan-phase —
it only eliminates the selection friction while keeping full interactive
control over each discussion.
Closes#2188
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
After #1518 redefined --full as all three granular flags combined, passing
--discuss --research --validate individually bypassed $FULL_MODE and showed
a "DISCUSS + RESEARCH + VALIDATE" banner instead of "FULL".
Fix: add a normalization step in flag parsing — if all three granular flags
are set, promote to $FULL_MODE=true. Remove the now-unreachable banner case.
Closes#2181
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* docs: sync ARCHITECTURE.md command count to 74
commands/gsd/ has 74 .md files; the two count references in
ARCHITECTURE.md still said 73. Fixes the command-count-sync
regression test.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: normalize Windows paths in update scope detection (#2232)
On Windows with Git Bash, `pwd` returns POSIX-style /c/Users/... paths
while execution_context carries Windows-style C:/Users/... paths. The
string equality check for LOCAL vs GLOBAL install scope never matched,
so every local install on Windows was misdetected as GLOBAL and the
wrong (global) install was updated.
Fix: normalize both paths to POSIX drive-letter form before comparing,
using portable POSIX shell (case+printf+tr, no GNU extensions).
Closes#2232
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(commands): add gsd:inbox command for GitHub issue/PR triage
inbox.md was created but not committed, causing the command count
to read 73 in git while ARCHITECTURE.md correctly stated 74.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
When prompt files contain shell metacharacters (\$VAR, backticks,
\$(...)), passing them as -p "\$(cat file)" causes the shell to expand
those sequences before the CLI tool ever receives the text. This
silently corrupts prompts built from user-authored PLAN.md content.
Replace all -p "\$(cat /tmp/gsd-review-prompt-{phase}.md)" patterns
with cat file | cli -p - so the prompt bytes are passed verbatim via
stdin. Affected CLIs: gemini, claude, codex, qwen. OpenCode and cursor
already used the pipe-to-stdin pattern.
Closes#2200
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
autonomous.md reported 11,748 tokens (over the Claude Code Read tool's 10K
limit), causing it to be read in 150-line chunks and generating a warning
on every /gsd-autonomous invocation.
Extract the 280-line smart_discuss step into a new reference file
(get-shit-done/references/autonomous-smart-discuss.md) and replace the
step body with a lean stub that directs the agent to read the reference.
This follows the established planner decomposition pattern.
autonomous.md: 38,750 → 29,411 chars (~7,350 tokens, well under 10K limit)
Fixes#2196
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Prevents infinite config-get loops on Kimi K2.5 and other models that
re-execute bash tool calls when they encounter config-get subshell patterns.
Values are now bundled into the init plan-phase JSON so step 15 of
plan-phase.md can read them directly without separate shell calls.
Closes#2192
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Three gaps in the orchestrator file-protection block (#1756, #2040):
1. quick.md never received the pre-merge deletion guard added to
execute-phase.md in #2040. Added the same DELETIONS check: if the
worktree branch deletes any tracked .planning/ files, block the merge
with a clear message rather than silently losing those files.
2. Both workflows deleted STATE_BACKUP and ROADMAP_BACKUP on merge
conflict — destroying the recovery files at exactly the moment they
were needed. Changed conflict handler to: preserve both backup paths,
print restore instructions, and break (halt) instead of continue
(silently advancing to the next worktree).
3. Neither workflow used --no-ff. Without it a fast-forward merge
produces no merge commit, so HEAD~1 in the resurrection check points
to the worktree's parent rather than main's pre-merge HEAD. Added
--no-ff to both git merge calls so HEAD~1 is always reliable.
Closes#2208
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
PR #2038 added detect-custom-files to gsd-tools.cjs and the backup_custom_files
step to update.md, but commit 7bfb11b6 is not an ancestor of v1.36.0: main was
rebuilt after the merge, orphaning the change. Users on 1.36.0 running /gsd-update
silently lose any locally-authored files inside GSD-managed directories.
Root cause: git merge-base 7bfb11b6 HEAD returns aa3e9cf (Cline runtime, PR #2032),
117 commits before the release tag. The "merged" GitHub state reflects the PR merge
event, not reachability from the default branch.
Fix: re-apply the three changes from 7bfb11b6 onto current main:
- Add detect-custom-files subcommand to gsd-tools.cjs (walk managed dirs, compare
against gsd-file-manifest.json keys via path.relative(), return JSON list)
- Add 'detect-custom-files' to SKIP_ROOT_RESOLUTION set
- Restore backup_custom_files step in update.md before run_update
- Restore tests/update-custom-backup.test.cjs (7 tests, all passing)
Closes#2229Closes#1997
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
When USER-PROFILE.md signals a non-technical product owner (learning_style: guided,
jargon in frustration_triggers, or high-level explanation_depth), discuss-phase now
reframes gray area labels and advisor_research rationale paragraphs in product-outcome
language. Same technical decisions, translated framing so product owners can participate
meaningfully without needing implementation vocabulary.
Closes#2125
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Projects with more than 5 phases had active UAT sessions silently
dropped from the verify-work listing. Only the first 5 *-UAT.md files
were shown, causing /gsd-verify-work to report incomplete results.
Remove the | head -5 pipe so all UAT files are listed regardless of
phase count.
Closes#2171
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
When a user manually installs a dev branch where VERSION > npm latest,
gsd-check-update detects hooks as "stale" and the statusline showed
the red "⚠ stale hooks — run /gsd-update" message. Running /gsd-update
would incorrectly downgrade the dev install to the npm release.
Fix: detect dev install (cache.installed > cache.latest) in the
statusline and show an amber "⚠ dev install — re-run installer to sync
hooks" message instead, with /gsd-update reserved for normal upgrades.
Also expand the update.md workflow's installed > latest branch to
explain the situation and give the correct remediation command
(node bin/install.js --global --claude, not /gsd-update).
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(sdk): add typed query foundation and gsd-sdk query (Phase 1)
Add sdk/src/query registry and handlers with tests, GSDQueryError, CLI query wiring, and supporting type/tool-scoping hooks. Update CHANGELOG. Vitest 4 constructor mock fixes in milestone-runner tests.
Made-with: Cursor
* fix(2137): skip worktree isolation when .gitmodules detected
When a project contains git submodules, worktree isolation cannot
correctly handle submodule commits — three separate gaps exist in
worktree setup, executor commit protocol, and merge-back. Rather
than patch each gap individually, detect .gitmodules at phase start
and fall back to sequential execution, which handles submodules
transparently (Option B).
Affected workflows: execute-phase.md, quick.md
---------
Co-authored-by: David Sienkowski <dave@sienkowski.com>
Replace `git show HEAD:.planning/STATE.md` with `cp .planning/STATE.md`
in the worktree merge-back protection logic of execute-phase.md and
quick.md. The git show approach exits 128 when STATE.md has uncommitted
changes or is not yet in HEAD's committed tree, leaving an empty backup
and causing the post-merge restore guard to silently skip — zeroing or
staling the file. Using cp reads the actual working-tree file (including
orchestrator updates that haven't been committed yet), which is exactly
what "main always wins" should protect.
* test(2134): add failing test for code-review SUMMARY.md YAML parser section reset
Demonstrates bug #2134: the section-reset regex in the inline node parser
in get-shit-done/workflows/code-review.md uses \s+ (requires leading whitespace),
so top-level YAML keys at column 0 (decisions:, metrics:, tags:) never reset
inSection, causing their list items to be mis-classified as key_files.modified
entries.
RED test asserts that the buggy parser contaminates the file list with decision
strings. GREEN test and additional tests verify correct behaviour with the fix.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(2134): fix YAML parser section reset to handle top-level keys (\s* not \s+)
The inline node parser in compute_file_scope (Tier 2) used \s+ in the
section-reset regex, requiring leading whitespace. Top-level YAML keys at
column 0 (decisions:, metrics:, tags:) never matched, so inSection was never
cleared and their list items were mis-classified as key_files.modified entries.
Fix: change \s+ to \s* in both the reset check and its dash-guard companion so
any key at any indentation level (including column 0) resets inSection.
Before: /^\s+\w+:/.test(line) && !/^\s+-/.test(line)
After: /^\s*\w+:/.test(line) && !/^\s*-/.test(line)
Closes#2134
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(workflow): add opt-in TDD pipeline mode (workflow.tdd_mode)
Add workflow.tdd_mode config key (default: false) that enables
red-green-refactor as a first-class phase execution mode. When
enabled, the planner aggressively applies type: tdd to eligible
tasks and the executor enforces RED/GREEN/REFACTOR gate sequence
with fail-fast on unexpected GREEN before RED. An end-of-phase
collaborative review checkpoint verifies gate compliance.
Closes#1871
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(test): allowlist plan-phase.md in prompt injection scan
plan-phase.md exceeds 50K chars after TDD mode integration.
This is legitimate orchestration complexity, not prompt stuffing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* ci: trigger CI run
* ci: trigger CI run
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Add a new pattern mapper agent that analyzes the codebase for existing
patterns before planning, producing PATTERNS.md with per-file analog
assignments and code excerpts. Integrated into plan-phase workflow as
Step 7.8 (between research and planning), controlled by the
workflow.pattern_mapper config key (default: true).
Changes:
- New agent: agents/gsd-pattern-mapper.md
- New config key: workflow.pattern_mapper in VALID_CONFIG_KEYS and CONFIG_DEFAULTS
- init plan-phase: patterns_path field in JSON output
- plan-phase.md: Step 7.8 spawns pattern mapper, PATTERNS_PATH in planner files_to_read
- gsd-plan-checker.md: Dimension 12 (Pattern Compliance)
- model-profiles.cjs: gsd-pattern-mapper profile entry
- Tests: tests/pattern-mapper.test.cjs (5 tests)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace minutes-based task sizing with context-window percentage sizing.
Add planner_authority_limits section prohibiting difficulty-based scope
decisions. Expand decision coverage matrix to multi-source audit covering
GOAL, REQ, RESEARCH, and CONTEXT artifacts. Add Source Audit gap handling
to plan-phase orchestrator (step 9c). Update plan-checker to detect
time/complexity language in scope reduction scans. Add 374 CI regression
tests preventing prohibited language from leaking back into artifacts.
Closes#2091Closes#2092
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
When CONTEXT_WINDOW < 200000, executor and planner agent prompts strip
extended examples and anti-pattern lists into reference files for
on-demand @ loading, reducing static overhead by ~40% while preserving
behavioral correctness for standard (200K-500K) and enriched (500K+) tiers.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add /gsd:extract-learnings command and backing workflow that extracts
decisions, lessons, patterns, and surprises from completed phase artifacts
into a structured LEARNINGS.md file with YAML frontmatter metadata.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add optional cross-AI delegation step that lets execute-phase delegate
plans to external AI runtimes via stdin-based prompt delivery. Activated
by --cross-ai flag, plan frontmatter cross_ai: true, or config key
workflow.cross_ai_execution. Adds 3 config keys, template defaults,
and 18 tests.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds workflow.code_review_command config key that allows solo devs to
plug external AI review tools into the ship flow. When configured, the
ship workflow generates a diff, builds a review prompt with stats and
phase context, pipes it to the command via stdin, and parses JSON output
with verdict/confidence/issues. Handles timeout (120s) and failures
gracefully by falling through to the existing manual review flow.
Closes#1876
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add plan bounce feature that allows plans to be refined through an external
script between plan-checker approval and requirements coverage gate. Activated
via --bounce flag or workflow.plan_bounce config. Includes backup/restore
safety (pre-bounce.md), YAML frontmatter validation, and checker re-run on
bounced plans.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add CURSOR_SESSION_ID env var detection in review.md so Cursor skips
itself as a reviewer (matching the CLAUDE_CODE_ENTRYPOINT pattern).
Add Qwen Review and Cursor Review sections to the REVIEWS.md template.
Update ja-JP and ko-KR FEATURES.md to include --opencode, --qwen, and
--cursor flags in the /gsd-review command signature.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Removes the .next-call-count counter file guard (which fired on clean usage and missed
real incomplete work) and replaces it with a scan of all prior phases for plans without
summaries, unoverridden VERIFICATION.md failures, and phases with CONTEXT.md but no plans.
When gaps are found, shows a structured report with Continue/Stop/Force options; the
Continue path writes a formal 999.x backlog entry and commits it before routing. Clean
projects route silently with no interruption.
Closes#2089
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Display examples showing 'cd $TARGET_PATH' and 'cd $WORKSPACE_PATH/repo1'
were unquoted, causing path splitting when project paths contain spaces
(e.g. Windows paths like C:\Users\First Last\...).
Quote all path variable references in user-facing guidance blocks so
the examples shown to users are safe to copy-paste directly.
The actual bash execution blocks (git worktree add, rm -rf, etc.) were
already correctly quoted — this fixes only the display examples.
Fixes#2088
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
When a user selects "Other" in AskUserQuestion with no text body, the
answer_validation block was treating the empty result as a generic empty
response and retrying the question — causing 2-3 cascading question rounds
instead of pausing for freeform user input as intended by the Other handling
on line 795.
Add an explicit exception in answer_validation: "Other" + empty text signals
freeform intent, not a missing answer. The workflow must output one prompt line
and stop rather than retry or generate more questions.
Fixes#2085
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Two correctness bugs from @trek-e review:
1. Grep pattern `^<task` only matched unindented task tags, missing
indented tasks in PLAN.md templates that use indentation. Fixed to
`^\s*<task[[:space:]>]` which matches at any indentation level and
avoids false positives on <tasks> or </task>.
2. Threshold=0 was documented to disable inline routing but the
condition `TASK_COUNT <= INLINE_THRESHOLD` evaluated 0<=0 as true,
routing empty plans inline even when the feature was disabled.
Fixed by guarding with `INLINE_THRESHOLD > 0`.
Added tests/inline-plan-threshold.test.cjs (8 tests) covering:
- config-set accepts the key and threshold=0
- VALID_CONFIG_KEYS and planning-config.md contain the entry
- Routing pattern matches indented tasks and rejects <tasks>/</task>
- Inline routing is guarded by INLINE_THRESHOLD > 0
Review feedback on #2061 from @trek-e.
Plans with 1-2 tasks now execute inline (Pattern C) instead of spawning
a subagent (Pattern A). This avoids ~14K token subagent spawn overhead
and preserves the orchestrator's prompt cache for small plans.
The threshold is configurable via workflow.inline_plan_threshold
(default: 2). Set to 0 to always spawn subagents. Plans above the
threshold continue to use checkpoint-based routing as before.
Closes#1979
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Per approved spec in #1969, the planner must include CONTEXT.md and
SUMMARY.md from any phases listed in the current phase's 'Depends on:'
field in ROADMAP.md, in addition to the 3 most recent completed phases.
This ensures explicit dependencies are always visible to the planner
regardless of recency — e.g., Phase 7 declaring 'Depends on: Phase 2'
always sees Phase 2's context, not just when Phase 2 is among the 3
most recent.
Review feedback on #2058 from @trek-e.
When CONTEXT_WINDOW >= 500000 (1M models), the planner loaded ALL prior
phase CONTEXT.md and SUMMARY.md files for cross-phase consistency. On
projects with 20+ phases, this consumed significant context budget with
diminishing returns — decisions from phase 2 are rarely relevant to
phase 22.
Limit to the 3 most recent completed phases, which provides enough
cross-phase context for consistency while keeping the planner's context
budget focused on the current phase's plans.
Closes#1969
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Closes#2070
Two-layer fix for the bug where executor agents in worktree isolation mode
could leave SUMMARY.md uncommitted, then have it silently destroyed by
`git worktree remove --force` during post-wave cleanup.
Layer 1 — Clarify executor instruction (execute-phase.md):
Added explicit REQUIRED note to the <parallel_execution> block making
clear that SUMMARY.md MUST be committed before the agent returns,
and that the git_commit_metadata step in execute-plan.md handles the
SUMMARY.md-only commit path automatically in worktree mode.
Layer 2 — Orchestrator safety net (execute-phase.md):
Before force-removing each worktree, check for any uncommitted SUMMARY.md
files. If found, commit them on the worktree branch and re-merge into the
main branch before removal. This prevents data loss even when an executor
skips the commit step due to misinterpreting the "do not modify
orchestrator files" instruction.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(review): add Qwen Code and Cursor CLI as peer reviewers (#1938, #1960)
Add qwen and cursor to the /gsd-review pipeline following the
established pattern from CodeRabbit and OpenCode integrations:
- CLI detection via command -v
- --qwen and --cursor flags
- Invocation blocks with empty-output fallback
- Install help URLs
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(review): correct qwen/cursor invocations and add doc surfaces (#1966)
Address review feedback from trek-e, kturk, and lawsontaylor:
- Use positional form for qwen (qwen "prompt") — -p flag is deprecated
upstream and will be removed in a future version
- Fix cursor invocation to use cursor agent -p --mode ask --trust
instead of cursor --prompt which launches the editor GUI
- Add --qwen and --cursor flags to COMMANDS.md, FEATURES.md, help.md,
commands/gsd/review.md, and localized docs (ja-JP, ko-KR)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The deviation rules and task commit protocol were duplicated between
gsd-executor.md (agent definition) and execute-plan.md (workflow).
The copies had diverged: the agent had scope boundary and fix attempt
limits the workflow lacked; the workflow had 3 extra commit types
(perf, docs, style) the agent lacked.
Consolidate gsd-executor.md as the single source of truth:
- Add missing commit types (perf, docs, style) to gsd-executor.md
- Replace execute-plan.md's ~90 lines of duplicated content with
concise references to the agent definition
Saves ~1,600 tokens per workflow spawn and eliminates maintenance
drift between the two copies.
Closes#1968
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* fix(workflow): offer recommendation instead of hard redirect when UI-SPEC.md missing
When plan-phase detects frontend indicators but no UI-SPEC.md, replace the
AskUserQuestion hard-exit block with an offer_next-style recommendation that
displays /gsd-ui-phase as the primary next step and /gsd-plan-phase --skip-ui
as the bypass option. Also registers --skip-ui as a parsed flag so it silently
bypasses the UI gate.
Closes#2011
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* ci: retrigger CI — resolve stale macOS check
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
AskUserQuestion is a Claude Code-only tool. When running GSD on OpenAI Codex,
Gemini CLI, or other non-Claude runtimes, the model renders the tool call as a
markdown code block instead of executing it, so the interactive TUI never
appears and the session stalls without collecting user input.
The workflow.text_mode / --text flag mechanism already handles this in 5 of
the 37 affected workflows. This commit adds the same TEXT_MODE fallback
instruction to all remaining 32 workflows so that, when text_mode is enabled,
every AskUserQuestion call is replaced with a plain-text numbered list that
any runtime can handle.
Fixes#2012
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
