Fixes#3168
The Claude Code subagent dispatcher tool is named `Agent` (with `subagent_type`
parameter). The `Task*` namespace (TaskCreate, TaskList, TaskGet, TaskUpdate,
TaskOutput, TaskStop) is the separate task-tracker. GSD's commands, workflows,
and agents were partially migrated and still referenced `- Task` / `Task(` in
55 files, causing orchestrators to silently fall back to inline execution when
no `Task` tool appeared on their tool surface.
Changes:
- `commands/gsd/*.md` allowed-tools: replaced `- Task` with `- Agent` in 24
files; removed duplicate `- Task` from autonomous.md (already had `- Agent`)
- `get-shit-done/workflows/*.md`: replaced dispatcher `Task(` → `Agent(` in
29 workflow files (~133 call sites); TaskCreate/List/Get/Update/Output/Stop
left untouched
- `agents/gsd-debug-session-manager.md`: replaced `Task` → `Agent` in tools
frontmatter (the only remaining agent with the wrong name)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(workflows): add atomic Write+commit ordering directive for SUMMARY.md
Adds explicit prompt-ordering language to executor spawn prompts and
plan-execution steps so agents commit SUMMARY.md before emitting any
concluding narrative. Mitigates the truncation-between-Write-and-commit
failure mode that has made the #2070 rescue net load-bearing.
Refs #2806
* fix(workflows): condense REQUIRED ORDER blocks to fit XL budget
The two REQUIRED ORDER directives added in bd1956df pushed
execute-phase.md to 1712 lines, exceeding the 1700-line XL budget.
Collapse each 6-line block into a single line that preserves the
semantic intent (Write SUMMARY.md → commit → narration; no text
between Write and commit; #2070 rescue is not primary defense).
File is now exactly 1700 lines; workflow-size-budget test passes.
* fix(execute-plan): move self-check before commit to preserve atomic Write+commit (#2939)
* fix(workflows): assert HEAD on per-agent branch before worktree commits
Worktree-mode setup could leave HEAD attached to a protected branch (master),
causing agent commits to land there. The previous response was a destructive
self-recovery via 'git update-ref refs/heads/master <sha>', which silently
rewinds the protected branch and destroys concurrent commits in multi-active
scenarios (parallel agents, user committing while agent runs).
- Reorder <worktree_branch_check> in execute-phase.md and quick.md to assert
HEAD via 'git symbolic-ref' BEFORE any 'git reset --hard'. HALT with a
blocker if HEAD is on main/master/develop/trunk/release/* or detached.
- Add a per-commit HEAD assertion (step 0) to gsd-executor.md
<task_commit_protocol>; HEAD attachment can drift after 'git checkout <sha>'.
- Forbid 'git update-ref refs/heads/<protected>' in
<destructive_git_prohibition>; surface the blocker rather than self-heal.
- Remove '--no-verify' as the worktree-mode default in execute-phase.md,
execute-plan.md, quick.md, and references/git-integration.md. Hooks now
run on every executor commit; opt out only via workflow.worktree_skip_hooks.
- Add regression test that parses the worktree_branch_check blocks structurally
and asserts the symbolic-ref check precedes the reset --hard, no workflow
performs update-ref on a protected ref, and --no-verify is no longer the
default in any parallel-execution prompt.
* fix(#2924): address CodeRabbit review findings on worktree HEAD PR
- Add positive worktree-agent-* allow-list to <task_commit_protocol> step 0
in gsd-executor.md and to <worktree_branch_check> in execute-phase.md and
quick.md. The deny-list (main|master|develop|trunk|release/*) silently
allowed feature/* and other arbitrary branches outside the agent namespace.
- Register workflow.worktree_skip_hooks in both config schemas
(sdk/src/query/config-schema.ts and get-shit-done/bin/lib/config-schema.cjs)
and document it in docs/CONFIGURATION.md so config-set accepts it.
- Fix stash lifecycle in execute-phase.md post-wave hook validation: stash
under a named ref and pop after the hook run; warn on pop failure.
- Pre-dispatch PLAN.md commit in quick.md: gate on git diff --cached --quiet
for idempotency and exit 1 with a clear error on commit failure (both the
--no-verify and the normal branches) — no more swallowing real errors.
- Test fixes (tests/bug-2924-worktree-head-attachment.test.cjs):
- Parse the protected-branch alternation structurally and require
main, master, develop, trunk, release/.* (release/* was previously
skipped by the \\b...\\b regex).
- Use fs.readdirSync(dir, { recursive: true }) so workflows in nested
subdirectories are also asserted against the update-ref ban.
- Add allow-list assertions for execute-phase.md, quick.md, and
gsd-executor.md to lock in the new positive namespace check.
* test(#2924): assert sub-section end marker exists before slicing
* test(#2924): use section boundary instead of fixed window for parallel-agents slice
* fix(#2767): pass paths via --files to gsd-sdk query commit + lint guard
Workflows, agents, commands, and references passed file paths positionally
to `gsd-sdk query commit`, which silently appended them to the commit
subject and triggered the `.planning/` wholesale-stage fallback in
sdk/src/query/commit.ts:136. Regression of #733/#798.
Inserted `--files` before the path list at every site (81 invocations
across 50 files). Added tests/bug-2767-gsd-sdk-commit-files-flag.test.cjs
as a permanent lint that scans every shipped .md file and asserts each
`gsd-sdk query commit[-to-subrepo]` invocation either uses `--files` or
carries no path arguments.
Closes#2767
* test(#2767): replace source-grep with behavioral SDK test
The original test walked every shipped .md file and regex-tokenized
`gsd-sdk query commit` invocations to assert `--files` was present.
CONTRIBUTING.md prohibits this source-grep pattern.
Rewrite as behavioral SDK tests against `sdk/dist/cli.js` over a real
tmp git project (createTempGitProject helper). Cover both the
well-formed (`--files <paths>`) form — clean subject, exactly-staged
files, .planning/ left untouched — and the buggy positional form,
asserting the documented misbehavior (paths leak into subject + the
`.planning/` wholesale-stage fallback at commit.ts:136). Also asserts
`commit-to-subrepo` rejects when `--files` is omitted (commit.ts:258).
The doc-lint is retained as a supplementary defense-in-depth guard
since agent-prompt markdown invocations cannot be exercised end-to-end
— but it is no longer the primary contract.
* docs(#2767): correct contradictory --files guidance in zh-CN/en docs + fix test docstring
All workflow, command, reference, template, and tool-output files that
surfaced /gsd:<cmd> as a user-typed slash command have been updated to
use /gsd-<cmd>, matching the Claude Code skill directory name.
Closes#2697
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(#2661): unconditional plan-checkbox sync in execute-plan
Checkpoint A in execute-plan.md was wrapped in a "Skip in parallel mode"
guard that also short-circuited the parallelization-without-worktrees
case. With `parallelization: true, use_worktrees: false`, only
Checkpoint C (phase.complete) then remained, and any interruption
between the final SUMMARY write and phase complete left ROADMAP.md
plan checkboxes stale.
Remove the guard: `roadmap update-plan-progress` is idempotent and
atomically serialized via readModifyWriteRoadmapMd's lockfile, so
concurrent invocations from parallel plans converge safely.
Checkpoint B (worktree-merge post-step) and Checkpoint C
(phase.complete) become redundant after A is unconditional; their
removal is deferred to a follow-up per the RCA.
Closes#2661
* fix(#2661): gate ROADMAP sync on use_worktrees=false to preserve single-writer contract
Adversarial review of PR #2682 found that unconditionally removing the
IS_WORKTREE guard violates the single-writer contract for shared
ROADMAP.md established by commit dcb50396 (PR #1486). The lockfile only
serializes within a single working tree; separate worktrees have
separate ROADMAP.md files that diverge.
Restore the worktree guard but document its intent explicitly: the
in-handler sync runs only when use_worktrees=false (the actual #2661
reproducer). Worktree mode relies on the orchestrator's post-merge
update at execute-phase.md lines 815-834, which is the documented
single-writer for shared tracking files.
Update tests to assert both branches of the gate:
- use_worktrees: false mode runs the sync (the #2661 case)
- use_worktrees: true mode does NOT run the in-handler sync
- handler-level idempotence and lockfile contention tests retained,
scope clarified to within-tree concurrency only
Commands are now installed as commands/gsd/<name>.md and invoked as
/gsd:<name> in Claude Code. The old hyphen form /gsd-<name> was still
hardcoded in hundreds of places across workflows, references, templates,
lib modules, and command files — causing "Unknown command" errors
whenever GSD suggested a command to the user.
Replace all /gsd-<cmd> occurrences where <cmd> is a known command name
(derived at runtime from commands/gsd/*.md) using a targeted Node.js
script. Agent names, tool names (gsd-sdk, gsd-tools), directory names,
and path fragments are not touched.
Adds regression test tests/bug-2543-gsd-slash-namespace.test.cjs that
enforces zero legacy occurrences going forward. Removes inverted
tests/stale-colon-refs.test.cjs (bug #1748) which enforced the now-obsolete
hyphen form; the new bug-2543 test supersedes it. Updates 5 assertion
tests that hardcoded the old hyphen form to accept the new colon form.
Closes#2543
Co-authored-by: Claude Opus 4.7 (1M context) <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>
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>
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>
* fix(worktree): use hard reset to correct file tree when branch base is wrong (#1981)
The worktree_branch_check mitigation detects when EnterWorktree creates
branches from main instead of the current feature branch, but used
git reset --soft to correct it. This only fixed the commit pointer —
the working tree still contained main's files, causing silent data loss
on merge-back when the agent's commits overwrote feature branch code.
Changed to git reset --hard which safely corrects both pointer and file
tree (the check runs before any agent work, so no changes to lose).
Also removed the broken rebase --onto attempt in execute-phase.md that
could replay main's commits onto the feature branch, and added post-reset
verification that aborts if the correction fails.
Updated documentation from "Windows" to "all platforms" since the
upstream EnterWorktree bug affects macOS, Linux, and Windows alike.
Closes#1981
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(worktree): update settings.md worktree description to say cross-platform
Aligns with the workflow file updates — the EnterWorktree base-branch
bug affects all platforms, not just Windows.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The existing MANDATORY acceptance_criteria instruction is purely advisory —
executor agents read it and silently skip criteria when they run low on
context or hit complexity. This causes planned work to be dropped without
any signal to the orchestrator or verifier.
Changes:
- Replace advisory text with a structured 5-step verification loop
- Each criterion must be proven via grep/file-check/CLI command
- Agent is BLOCKED from next task until all criteria pass
- Failed criteria after 2 fix attempts logged as deviation (not silent skip)
- Self-check step now re-runs ALL acceptance criteria before SUMMARY
- Self-check also re-runs plan-level verification commands
Closes#1958
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: harness engineering improvements — post-merge test gate, shared file isolation, behavioral verification
Three improvements inspired by Anthropic's harness engineering research
(March 2026) and real-world pain points from parallel worktree execution:
1. Post-merge test gate (execute-phase.md)
- Run project test suite after merging each wave's worktrees
- Catches cross-plan integration failures that individual Self-Checks miss
- Addresses the Generator self-evaluation blind spot (agents praise own work)
2. Shared file isolation (execute-phase.md)
- Executors no longer modify STATE.md or ROADMAP.md in parallel mode
- Orchestrator updates tracking files centrally after merge
- Eliminates the #1 source of merge conflicts in parallel execution
3. Behavioral verification (verify-phase.md)
- Verifier runs project test suite and CLI commands, not just grep
- Follows Anthropic's Generator/Evaluator separation principle
- Tests actual behavior against success criteria, not just file existence
Real-world evidence: In a session executing 37 plans across 8 phases with
parallel worktrees, we observed:
- 4 test failures after merge that all Self-Checks missed (models.py type loss)
- STATE.md/ROADMAP.md conflicts on every single parallel merge
- Verifier reporting PASSED while merged code had broken imports
References:
- Anthropic Engineering Blog: Harness Design for Long-Running Apps (2026-03-24)
- Issue #1451: Massive git worktree problem
- Issue #1413: Autonomous execution without manual context clearing
* fix: address review feedback — test runner detection, parallel isolation, edge cases
- Replace hardcoded jest/vitest with `npm test` (reads project's scripts.test)
- Add Go detection to post-merge test gate (was only in verify-phase)
- Add 5-minute timeout to post-merge test gate to prevent pipeline stalls
- Track cumulative wave failures via WAVE_FAILURE_COUNT for cross-wave awareness
- Guard orchestrator tracking commit against unchanged files (prevent empty commits)
- Align execute-plan.md with parallel isolation model (skip STATE.md/ROADMAP.md
updates when running in parallel mode, orchestrator handles centrally)
- Scope behavioral verification CLI checks: skip when no fixtures/test data exist,
mark as NEEDS HUMAN instead of inventing inputs
* fix: pass PARALLEL_MODE to executor agents to activate shared file isolation
The executor spawn prompt in execute-phase.md instructed agents not to
modify STATE.md/ROADMAP.md, but execute-plan.md gates this behavior on
PARALLEL_MODE which was never defined in the executor context. This adds
the variable to the spawn prompt and wraps all three shared-file steps
(update_current_position, update_roadmap, git_commit_metadata) with
explicit conditional guards.
* fix: replace unreliable PARALLEL_MODE env var with git worktree auto-detection
Address PR #1486 review feedback (trek-e):
1. PARALLEL_MODE was never reliably set — the <env> block instructed the LLM
to export a bash variable, but each Bash tool call runs in a fresh shell
so the variable never persisted. Replace with self-contained worktree
detection: `[ -f .git ]` returns true in worktrees (.git is a file) and
false in main repos (.git is a directory). Each bash block detects
independently with no external state dependency.
2. TEST_EXIT only checked for timeout (124) — test failures (non-zero,
non-124) were silently ignored, making the "If tests fail" prose
unreachable. Add full if/elif/else handling: 0=pass, 124=timeout,
else=fail with WAVE_FAILURE_COUNT increment.
3. Add Go detection to regression_gate (was missing go.mod check).
Replace hardcoded npx jest/vitest with npm test for consistency.
4. Renumber steps from 4/4b/4c/5/5/5b to 4a/4b/4c/4d/5/6/7/8/9.
* fix: address remaining review blockers — timeout, tracking guard, shell safety
- verify-phase.md: wrap behavioral_verification test suite in timeout 300
- execute-phase.md: gate tracking update on TEST_EXIT=0, skip on failure/timeout
- Quote all TEST_EXIT variables, add default initialization
- Add else branch for unrecognized project types
- Renumber steps to align with upstream (5.x series)
* fix: rephrase worktree success_criteria to satisfy substring test guard
The worktree mode success_criteria line literally contained "STATE.md"
and "ROADMAP.md" inside a prohibition ("No modifications to..."), but
the test guard in execute-phase-worktree-artifacts.test.cjs uses a
substring check and cannot distinguish prohibition from requirement.
Rephrase to "shared orchestrator artifacts" so the substring check
passes while preserving the same intent.
* fix: replace /gsd: command format with /gsd- skill format in all suggestions
All next-step suggestions shown to users were still using the old colon
format (/gsd:xxx) which cannot be copy-pasted as skills. Migrated all
occurrences across agents/, commands/, get-shit-done/, docs/, README files,
bin/install.js (hardcoded defaults for claude runtime), and
get-shit-done/bin/lib/*.cjs (generate-claude-md templates and error messages).
Updated tests to assert new hyphen format instead of old colon format.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: migrate remaining /gsd: format to /gsd- in hooks, workflows, and sdk
Addresses remaining user-facing occurrences missed in the initial migration:
- hooks/: fix 4 user-facing messages (pause-work, update, fast, quick)
and 2 comments in gsd-workflow-guard.js
- get-shit-done/workflows/: fix 21 Skill() literal calls that Claude
executes directly (installer does not transform workflow content)
- sdk/prompt-sanitizer.ts: update regex to strip /gsd- format in addition
to legacy /gsd: format; update JSDoc comment
- tests/: update autonomous-ui-steps, prompt-sanitizer to assert new format
Note: commands/gsd/*.md frontmatter (name: gsd:xxx) intentionally unchanged
— installer derives skillName from directory path, not the name field.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(plan-phase): preserve --chain flag in auto-advance sync and handle ui-phase gate in chain mode
Bug 1: step 15 sync-flag check only guarded against --auto, causing
_auto_chain_active to be cleared when plan-phase is invoked without
--auto in ARGUMENTS even though a --chain pipeline was active. Added
--chain to the guard condition, matching discuss-phase behaviour.
Bug 2: UI Design Contract gate (step 5.6) always exited the workflow
when UI-SPEC was missing, breaking the discuss --chain pipeline
silently. When _auto_chain_active is true, the gate now auto-invokes
gsd-ui-phase --auto via Skill() and continues to step 6 without
prompting. Manual invocations retain the existing AskUserQuestion flow.
* fix: remove <sub>/clear</sub> pattern and duplicate old-format command in discuss-phase.md
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
When EnterWorktree creates a branch from main instead of the current HEAD
(a known issue on Windows), executor agents now detect the mismatch and
reset their branch base to the correct commit before starting work.
- execute-phase: capture EXPECTED_BASE before spawning, inject
<worktree_branch_check> block into executor prompts
- execute-plan: document Pattern A worktree_branch_check requirement
- quick.md: inject worktree_branch_check into executor prompt
- diagnose-issues: inject worktree_branch_check into debugger prompts
- settings: add workflow.use_worktrees option so Windows users can
disable worktree isolation via /gsd:settings without editing files
Closes#1510
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds `workflow.use_worktrees` config option (default: `true`) that
allows users to disable git worktree isolation for executor agents.
When set to `false`:
- Executor agents run without `isolation="worktree"`
- Plans execute sequentially on the main working tree
- No worktree merge ordering issues or orphaned worktrees
- Normal git hooks run (no --no-verify needed)
This provides an escape hatch for solo developers and users who
experience worktree merge conflicts, as worktree ordering issues
are inherently difficult when parallel agents modify overlapping
files.
Usage:
/gsd:settings → set workflow.use_worktrees to false
Or directly:
gsd-tools config-set workflow.use_worktrees false
Changes:
- config.cjs: add workflow.use_worktrees to valid keys
- planning-config.md: document the option
- execute-phase.md: read config, conditional worktree + sequential mode
- execute-plan.md: conditional worktree in Pattern A
- quick.md: conditional worktree for quick executor
- diagnose-issues.md: conditional worktree for debug agents
- 2 new tests (config set + workflow structural check)
Closes#1451
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
PR #1139 added <available_agent_types> sections to execute-phase.md and
plan-phase.md to prevent /clear from causing silent fallback to
general-purpose. However, 14 other workflows and 2 commands that also
spawn named GSD agents were missed, leaving them vulnerable to the same
regression after /clear.
Added <available_agent_types> listing to: research-phase, quick,
audit-milestone, diagnose-issues, discuss-phase-assumptions,
execute-plan, map-codebase, new-milestone, new-project, ui-phase,
ui-review, validate-phase, verify-work (workflows) and debug,
research-phase (commands).
Added regression test that enforces every workflow/command spawning
named subagent_type must have a matching <available_agent_types>
section listing all spawned types.
Fixes#1357
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Address 4 root causes of Windows + Claude Code reliability issues:
1. Workflow shell robustness: add || true guards to informational commands
(ls, grep, find, cat) that return non-zero on "no results", preventing
workflow step failures under strict execution models. Guard glob loops
with [ -e "$var" ] || continue to handle empty glob expansion.
2. Hook stdin handling: replace readFileSync('/dev/stdin') with async
process.stdin + timeout in agent templates (gsd-verifier.md). Existing
JS hooks already have timeout guards.
3. project_root detection: fix isInsideGitRepo() to check .git at the
candidate parent level (not just below it), enabling correct detection
when .git and .planning/ are siblings at the same directory level —
the common single-repo case from a subdirectory.
4. @file: handoff: add missing @file: handlers to autonomous.md and
manager.md workflows that call gsd-tools init but lacked the handler
for large output payloads.
Fixes#1343
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
All code-writing agents (gsd-executor, gsd-debugger) were dispatched
without isolation: "worktree", causing branch pollution when agents
switched branches in the shared working tree during concurrent work.
Added isolation="worktree" to all Task() dispatch sites:
- execute-phase.md: executor agent dispatch
- execute-plan.md: Pattern A executor reference
- quick.md: quick task executor dispatch
- diagnose-issues.md: debugger agent dispatch
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add support for workspaces with multiple independent git repositories.
When configured, GSD routes commits to the correct sub-repo and ensures
.planning/ stays at the project root.
Core features:
- detectSubRepos(): scans child directories for .git to discover repos
- findProjectRoot(): walks up from CWD to find the project root that
owns .planning/, preventing orphaned .planning/ in sub-repos
- loadConfig auto-syncs sub_repos when repos are added or removed
- Migrates legacy "multiRepo: true" to sub_repos array automatically
- All init commands include project_root in output
- cmdCommitToSubrepo: groups files by sub-repo prefix, commits independently
Zero impact on single-repo workflows — sub_repos defaults to empty array.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The agent was telling users to run '/gsd:transition' after phase completion,
but this command does not exist. transition.md is an internal workflow invoked
by execute-phase during auto-advance.
Changes:
- Add <internal_workflow> banner to transition.md declaring it is NOT a user command
- Add explicit warning in execute-phase completion section that /gsd:transition
does not exist
- Add 'only suggest commands listed above' guard to prevent hallucination
- Update resume-project.md to avoid ambiguous 'Transition' label
- Replace 'ready for transition' with 'ready for next step' in execute-plan.md
Fixes#1081
GSD executor agents ignore MCP tools (e.g. jCodeMunch) even when CLAUDE.md
explicitly instructs their use. Agents default to Grep/Glob because those
are explicitly referenced in workflow patterns.
Added MCP tool instructions to:
- execute-phase.md: <mcp_tools> section in executor agent prompt telling
agents to prefer MCP tools over Grep/Glob when available
- execute-plan.md: Step 2 in execute section with MCP tool fallback guidance
Agents now:
1. Check if CLAUDE.md references MCP tools
2. Prefer MCP tools for code navigation when accessible
3. Fall back to Grep/Glob if MCP tools are not available
Fixes#973
After committing task changes, the executor now checks for untracked
files (git status --short | grep '^??') and handles them: commit if
intentional, add to .gitignore if generated/runtime output.
This prevents generated artifacts (build outputs, .env files, cache
files) from being silently left untracked in the working tree.
Changes:
- execute-plan.md: Add step 6 to task commit protocol
- gsd-executor.md: Add step 6 to task commit protocol
Fixes#957
Executor agents often produce shallow work because plans say "align X with Y"
without specifying what the aligned result looks like. The executor changes one
value and calls it done.
This adds three mandatory fields to every task:
- `<read_first>`: Files the executor MUST read before editing. Ensures ground
truth is loaded, not assumed.
- `<acceptance_criteria>`: Grep-verifiable conditions checked after each task.
No subjective language ("looks consistent"), only concrete checks
("file contains 'exact string'").
- `<action>` guidance: Must include concrete values (identifiers, signatures,
config keys), never vague references like "update to match production".
Adds `<deep_work_rules>` to planner instructions with mandatory quality gate
checks. Executor workflow enforces both gates: read before edit, verify after
edit. Adds generic pre-commit hook failure handling guidance.
Co-authored-by: Dammerzone <dammerzone@users.noreply.github.com>
When a task fails verification, the executor now attempts structured
repair before interrupting the user:
- RETRY: adjusts approach and re-attempts
- DECOMPOSE: splits the task into smaller verifiable sub-tasks
- PRUNE: skips with justification when infeasible
Only escalates to the user when the repair budget is exhausted or an
architectural decision is required (Rule 4). Configurable via
workflow.node_repair (bool) and workflow.node_repair_budget (int).
Defaults to enabled with budget=2.
Inspired by the NODE_REPAIR operator in STRUCTUREDAGENT (arXiv:2603.05294).
Co-authored-by: buftar <buftar@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
When gsd-tools init output exceeds 50KB, core.cjs writes to a temp file
and outputs @file:<path>. No workflow handled this prefix, causing agents
to hallucinate /tmp paths that fail on Windows (C:\tmp doesn't exist).
Add @file: resolution line after every INIT=$(node ...) call across all
32 workflow, agent, and reference files.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Claude Code subagents sometimes rewrite ~/. paths to relative paths,
causing MODULE_NOT_FOUND when CWD is the project directory. $HOME is a
shell variable resolved at runtime, immune to model path rewriting.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
gsd-executor had no instruction to update ROADMAP or REQUIREMENTS after
completing a plan — both stayed unchecked throughout milestone execution.
- Add `roadmap update-plan-progress` call to executor state_updates
- Add `requirements mark-complete` CLI command to gsd-tools
- Wire requirement marking into executor and execute-plan workflows
- Include ROADMAP.md and REQUIREMENTS.md in executor final commit
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Tighten all requirements references to use MUST/REQUIRED/CRITICAL language
instead of passive suggestions. Close the verification loop by extracting
phase requirement IDs from ROADMAP and passing them through the full chain:
researcher receives IDs → planner writes to PLAN frontmatter → executor
copies to SUMMARY → verifier cross-references against REQUIREMENTS.md with
orphan detection.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add optional phase archival to milestone completion and a standalone
/gsd:cleanup command for retroactive use. Phase dirs move to
.planning/milestones/v{X.Y}-phases/, reducing phases/ clutter after
multiple milestones.
Core changes:
- getArchivedPhaseDirs() and searchPhaseInDir() helpers in gsd-tools
- findPhaseInternal() searches archives when phase not found in current
- cmdPhasesList() accepts --include-archived flag
- cmdHistoryDigest() scans both current and archived phases
- cmdMilestoneComplete() accepts --archive-phases flag
- Workflow globs replaced with find-phase/phases-list CLI calls
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When init commands include many file contents (state, roadmap, requirements,
etc.), the JSON output can exceed Claude Code's Bash tool buffer (~50KB),
causing parse errors. The output() function now auto-detects large payloads
and writes to a tmpfile, returning @file:/path instead. All workflows that
consume init output handle both formats.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Projects with "type": "module" in package.json cause Node to treat
gsd-tools.js as ESM, crashing on require(). The .cjs extension forces
CommonJS regardless of the host project's module configuration.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace NL instructions in execute-plan and execute-phase workflows with
deterministic `roadmap update-plan-progress` command that counts PLAN vs
SUMMARY files on disk. Prevents LLM from miscounting plan progress.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(gsd-tools): add frontmatter CRUD, verification suite, template fill, and state progression
Four new command groups that delegate deterministic operations from AI agents to code:
- frontmatter get/set/merge/validate: Safe YAML frontmatter manipulation with schema validation
- verify plan-structure/phase-completeness/references/commits/artifacts/key-links: Structural checks agents previously burned context on
- template fill summary/plan/verification: Pre-filled document skeletons so agents only fill creative content
- state advance-plan/record-metric/update-progress/add-decision/add-blocker/resolve-blocker/record-session: Automate arithmetic and formatting in STATE.md
Adds reconstructFrontmatter() + spliceFrontmatter() helpers for safe frontmatter roundtripping,
and parseMustHavesBlock() for 3-level YAML parsing of must_haves structures.
20 new functions, ~1037 new lines.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: wire gsd-tools commands into agents and workflows
- gsd-verifier: use `verify artifacts` and `verify key-links` instead of
manual grep patterns for stub detection and wiring verification
- gsd-executor: use `state advance-plan`, `state update-progress`,
`state record-metric`, `state add-decision`, `state record-session`
instead of manual STATE.md manipulation
- gsd-plan-checker: use `verify plan-structure` and `frontmatter get`
for structural validation and must_haves extraction
- gsd-planner: add validation step using `frontmatter validate` and
`verify plan-structure` after writing PLAN.md
- execute-plan.md: use gsd-tools state commands for position/progress updates
- verify-phase.md: use gsd-tools for must_haves extraction and artifact/link verification
This makes the gsd-tools commands from PR #485 actually used by the system.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Claude Code v2.1.27+ has a bug where all Task tool agents report
"failed" due to `classifyHandoffIfNeeded is not defined` — a function
called but never defined in the cli.js bundle. The error fires AFTER
all agent work completes, so actual work is always done.
This adds spot-check fallback logic to execute-phase, execute-plan,
and quick workflows: when an agent reports this specific failure,
verify artifacts on disk (SUMMARY.md exists, git commits present).
If spot-checks pass, treat as successful.
Tracked upstream: anthropics/claude-code#24181
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* fix(execute-phase): pass file paths to subagents instead of content
The --include flag added in fa81821 caused orchestrator context bloat
by reading STATE, config, and plan files into the orchestrator's context,
then embedding all content in Task() prompts.
With multiple plans, this consumed 50-60%+ of context before execution.
Fix: Pass file paths only. Subagents read files themselves in their
fresh 200k context. Orchestrator stays lean (~10-15% as intended).
Fixes#479
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* fix: respect commit_docs=false in execute-plan and debugger workflows
Two code paths bypassed the commit_docs configuration check, causing
.planning files to be intermittently committed when commit_docs=false:
1. execute-plan.md: update_codebase_map step ran `git add .planning/codebase/*.md`
unconditionally — now gated behind commit_docs check
2. gsd-debugger.md: used `git add -A` which stages .planning/ files — replaced
with explicit individual file staging and proper commit_docs conditional
Fixes#478https://claude.ai/code/session_013yS1F2VR3Jn2pdwqr5NuDo
* fix: route all .planning commits through gsd-tools.js CLI
Instead of wrapping direct git commands in markdown conditionals,
both bypass paths now use gsd-tools.js commit which has the
commit_docs check built in:
1. execute-plan.md: uses `gsd-tools.js commit --amend` for codebase
map updates (new --amend flag added to CLI)
2. gsd-debugger.md: code commit uses direct git (no .planning files),
planning docs commit uses gsd-tools.js commit
Also added --amend support to gsd-tools.js commit command so the
execute-plan codebase map step can amend the previous metadata commit.
Fixes#478https://claude.ai/code/session_013yS1F2VR3Jn2pdwqr5NuDo
* docs: update reference docs to use gsd-tools.js CLI for all .planning commits
Reference documentation showed direct git add/commit patterns for
.planning files, which agents copy-paste and bypass the commit_docs
check. Updated all three reference files to show gsd-tools.js commit
as the canonical pattern:
- git-planning-commit.md: replaced manual bash conditionals with CLI
- git-integration.md: replaced direct git add/commit in initialization,
plan-completion, and handoff examples
- planning-config.md: replaced conditional git example with CLI call
https://claude.ai/code/session_013yS1F2VR3Jn2pdwqr5NuDo
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Workflows were calling init commands to get parsed JSON metadata, then
immediately reading the same files again with cat to pass raw content
to agents. This wastes context tokens.
Changes:
- Add --include flag to init execute-phase, plan-phase, and progress
- Support includes: state, config, roadmap, requirements, context,
research, verification, uat, project
- Update plan-phase.md to use --include (removes 6 cat calls)
- Update execute-phase.md to use --include (removes 2 cat calls)
- Update execute-plan.md to use --include (removes 2 cat calls)
- Update progress.md to use --include (removes 4 cat calls)
- Add 7 tests for --include functionality
Token savings: ~5,000-10,000 tokens per plan-phase execution,
~1,500-3,000 per other workflow executions.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Executor self-check after SUMMARY.md creation verifies key-files exist
on disk and commit hashes exist in git log. Orchestrator spot-checks
SUMMARY claims before trusting and proceeding to next wave. Segmented
execution (execute-plan.md) gets the same self-check inline.
Trim redundant documentation while preserving actionable instructions.
-173 lines of prose that restated what templates already show.
Kept: size constraints, mandatory fields, security guidance.
Removed: "when to create/read/update" lifecycle prose, section
descriptions that duplicate template structure.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Rolled back the intel system due to overengineering concerns:
- 1200+ line hook with SQLite graph database
- 21MB sql.js dependency
- Entity generation spawning additional Claude calls
- Complex system with unclear value
Removed:
- /gsd:analyze-codebase command
- /gsd:query-intel command
- gsd-intel-index.js, gsd-intel-session.js, gsd-intel-prune.js hooks
- gsd-entity-generator, gsd-indexer agents
- entity.md template
- sql.js dependency
Preserved:
- Model profiles feature
- Statusline hook
- All other v1.9.x improvements
-3,065 lines removed
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
