Add pre-planning design contract generation and post-execution visual audit
for frontend phases. Closes the gap where execute-phase runs without a visual
contract, producing inconsistent spacing, color, and typography across components.
New agents: gsd-ui-researcher (UI-SPEC.md), gsd-ui-checker (6-dimension
validation), gsd-ui-auditor (6-pillar scored audit + registry re-vetting).
Third-party shadcn registry blocks are machine-vetted at contract time,
verification time, and audit time — three enforcement points, not a checkbox.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* fix: plan-phase Nyquist validation when research is disabled (#980)
plan-phase step 5.5 required Nyquist artifacts even when research was
disabled, creating an impossible state: no RESEARCH.md to extract
Validation Architecture from. Step 7.5 then told Claude to "disable
Nyquist in config" without specifying the exact key, causing Claude to
guess wrong keys that config-set silently accepted.
Three fixes:
- plan-phase step 5.5: skip when research_enabled is false
- plan-phase step 7.5: specify exact config-set command for disabling
- config-set: reject unknown keys with whitelist validation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* test: update config-set tests for key whitelist validation
Tests that used arbitrary keys (some_number, some_string, a.b.c) now
use valid config keys to test the same coercion and nesting behavior.
Adds new test asserting unknown keys are rejected with error.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Quick task IDs were sequential integers (001, 002...) computed by reading
the local .planning/quick/ directory max value. Two users running /gsd:quick
simultaneously would get the same number, causing directory collisions on git.
Replace with a collision-resistant format: YYMMDD-xxx where xxx is the
number of 2-second blocks elapsed since midnight, encoded as 3 lowercase
Base36 characters (000–xbz). Practical collision window is ~2 seconds per
user — effectively zero for any realistic team workflow.
- init.cjs: remove nextNum scan logic, generate quickId from wall clock
- quick.md: rename all next_num refs to quick_id, update directory patterns
- init.test.cjs: rewrite cmdInitQuick tests for new ID format
Co-authored-by: yanbing <yanbing@corp.netease.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
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>
* fix: add mandatory canonical_refs section to CONTEXT.md
CONTEXT.md is the bridge between user decisions and downstream agents
(researcher, planner). When projects have external specs, ADRs, or
design docs, these references were being silently dropped — mentioned
inline as "see ADR-019" but never collected into a section that agents
could find and read. This caused agents to plan and implement without
reading the specs they were supposed to follow.
Changes:
- templates/context.md: Add <canonical_refs> section to file template,
all 3 examples, and guidelines (marked MANDATORY)
- workflows/discuss-phase.md: Add step 1b (extract canonical refs),
add section to write_context template, add to success criteria
- workflows/plan-phase.md: Add canonical ref extraction to PRD express
path and its CONTEXT.md template
- workflows/quick.md: Add lightweight canonical_refs to --discuss mode
The section is mandatory but gracefully handles projects without external
docs ("No external specs — requirements fully captured in decisions above").
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: make canonical refs an accumulator across entire discussion
Refs come from 4 sources, not just ROADMAP.md:
1. ROADMAP.md Canonical refs line (initial seed)
2. REQUIREMENTS.md/PROJECT.md referenced specs
3. Codebase scout (code comments citing ADRs)
4. User during discussion ("read adr-014", "check the MCP spec")
Source 4 is often the MOST important — these are docs the user
specifically wants downstream agents to follow. The previous
version only handled source 1 and silently dropped the rest.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 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>
Surfaces gray areas and captures user decisions in CONTEXT.md before
planning, reducing hallucination risk for ambiguous quick tasks.
Composable with --full for discussion + plan-checking + verification.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: change nyquist_validation default to true and harden absent-key skip conditions
new-project.md never wrote the key, so agents reading config directly
treated absent as falsy. Changed all agent skip conditions from "is false"
to "explicitly set to false; absent = enabled". Default changed from false
to true in core.cjs, config.cjs, and templates/config.json.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: enforce VALIDATION.md creation with verification gate and Check 8e
Step 5.5 was narrative markdown that Claude skipped under context pressure.
Now MANDATORY with Write tool requirement and file-existence verification.
Step 7.5 gates planner spawn on VALIDATION.md presence. Check 8e blocks
Dimension 8 if file missing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add W008/W009 health checks and addNyquistKey repair for Nyquist drift detection
W008 warns when workflow.nyquist_validation key is absent from config.json
(agents may skip validation). W009 warns when RESEARCH.md has Validation
Architecture section but no VALIDATION.md file exists. addNyquistKey repair
adds the missing key with default true value.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add /gsd:validate-phase command and gsd-nyquist-auditor agent
Retroactively applies Nyquist validation to already-executed phases.
Works mid-milestone and post-milestone. Detects existing test coverage,
maps gaps to phase requirements, writes missing tests, debugs failing
ones, and produces {phase}-VALIDATION.md from existing artifacts.
Handles three states: VALIDATION.md exists (audit + update), no
VALIDATION.md (reconstruct from PLAN.md + SUMMARY.md), phase not yet
executed (exit cleanly with guidance).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: audit-milestone reports Nyquist compliance gaps across phases
Adds Nyquist coverage table to audit-milestone output when
workflow.nyquist_validation is true. Identifies phases missing
VALIDATION.md or with nyquist_compliant: false/partial.
Routes to /gsd:validate-phase for resolution. Updates USER-GUIDE
with retroactive validation documentation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor: compress Nyquist prompts to match GSD meta-prompt density conventions
Auditor agent: deleted philosophy section (35 lines), compressed execution
flow 60%, removed redundant constraints. Workflow: cut purpose bloat,
collapsed state narrative, compressed auditor spawn template. Command:
removed redundant process section. Plan-phase Steps 5.5/7.5: replaced
hedging language with directives. Audit-milestone Step 5.5: collapsed
sub-steps into inline instructions. Net: -376 lines.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The "depth" setting with values quick/standard/comprehensive implied
investigation thoroughness, but it only controls phase count. Renamed to
"granularity" with values coarse/standard/fine to accurately reflect what
it controls: how finely scope is sliced into phases.
Includes backward-compatible migration in loadConfig and config-ensure
that auto-renames depth→granularity with value mapping in both
.planning/config.json and ~/.gsd/defaults.json.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When a phase modifies server, database, seed, or startup files, verify-work
now prepends a "Cold Start Smoke Test" that asks the user to kill, wipe state,
and restart from scratch — catching warm-state blind spots.
Closes#904
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduce ephemeral `workflow._auto_chain_active` flag to separate chain
propagation from the user's persistent `workflow.auto_advance` preference.
Previously, `workflow.auto_advance` was set to true by --auto chains and
only cleared at milestone completion. If a chain was interrupted (context
limit, crash, user abort), the flag persisted in .planning/config.json
and caused all subsequent manual invocations to auto-advance unexpectedly.
The fix adds a "sync chain flag with intent" step to discuss-phase,
plan-phase, and execute-phase workflows: when --auto is NOT in arguments,
the ephemeral _auto_chain_active flag is cleared. The persistent
auto_advance setting (from /gsd:settings) is never touched, preserving
the user's deliberate preference.
Closes#857
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When running /gsd:update from $HOME, the local path ./.claude/ resolves
to the same directory as $HOME/.claude/. The local branch always wins,
triggering --local reinstall that corrupts all paths in a global install.
This is self-reinforcing — once corrupted, every subsequent update
perpetuates the corruption.
Compare canonical paths (via cd + pwd) and only report LOCAL when the
resolved directories differ. When they're the same, fall through to
GLOBAL detection.
Closes#721
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace general-purpose workaround pattern with named subagent types:
- plan-phase: researcher and planner now spawn as gsd-phase-researcher/gsd-planner
- new-project: 4 researcher spawns now use gsd-project-researcher
- research-phase: researcher spawns now use gsd-phase-researcher
- quick: planner revision now uses gsd-planner
- diagnose-issues: debug agents now use gsd-debugger (matches template spec)
Removes 'First, read agent .md file' prompt prefix — named agent types
auto-load their .md file as system prompt, making the workaround redundant.
Preserves intentional general-purpose orchestrator spawns in discuss-phase
and plan-phase (auto-advance) where the agent runs an entire workflow.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When users select "Other" to provide freeform input, Claude re-prompts
with another AskUserQuestion instead of dropping to plain text. This
loops 2-3 times before finally accepting freeform input.
Add explicit freeform escape rule to questioning.md reference, and
update new-milestone.md and discuss-phase.md to switch to plain text
when users signal freeform intent.
Closes#778
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Prevents re-asking questions already decided in earlier phases by reading
PROJECT.md, REQUIREMENTS.md, STATE.md, and all prior CONTEXT.md files
before generating gray areas. Prior decisions annotate options and skip
already-decided areas.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
zsh's built-in echo interprets \n escape sequences in JSON strings,
converting properly-escaped \\n back into literal newlines. This breaks
jq parsing with "control characters U+0000 through U+001F must be escaped".
printf '%s\n' preserves the JSON verbatim.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The auto-advance chain (discuss → plan → execute) was spawning each
subsequent phase as a Task(subagent_type="general-purpose"), creating
3-4 levels of nested agent sessions. At that nesting depth, the
Claude Code runtime hits resource limits or stdio contention, causing
the execute-phase to freeze or attempt to shell out to `claude` as a
subprocess (which is explicitly blocked).
The fix replaces Task spawns with Skill invocations for phase
transitions:
- discuss-phase auto-advance: Skill("gsd:plan-phase") instead of
Task(general-purpose)
- plan-phase auto-advance: Skill("gsd:execute-phase") instead of
Task(general-purpose)
The Skill tool runs in the same process context as the caller,
keeping the entire auto-advance chain flat at a single nesting level.
Each phase still spawns its own worker agents (gsd-executor,
gsd-planner, etc.) as Tasks, but the orchestration chain itself
no longer creates unnecessary depth.
Closes#686
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three files had hardcoded `.claude` paths that break OpenCode and
Gemini installations where the config directory is `.config/opencode`,
`.opencode`, or `.gemini` respectively.
Changes:
- hooks/gsd-check-update.js: add detectConfigDir() helper that checks
all runtime directories for get-shit-done/VERSION, falling back to
.claude. Used for cache dir, project VERSION, and global VERSION.
- commands/gsd/reapply-patches.md: detect runtime directory for both
global and local patch directories instead of hardcoding ~/.claude/
and ./.claude/
- workflows/update.md: detect runtime directory for local and global
VERSION/marker files, and clear cache across all runtime directories
Closes#682
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The SessionStart hook always writes to ~/.claude/cache/ via os.homedir()
regardless of install type. The update workflow previously only cleared
the install-type-specific path, leaving stale cache at the global path
for local installs.
Clear both ./.claude/cache/ and ~/.claude/cache/ unconditionally.
Co-authored-by: Claude Sonnet 4.5 <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>
Add lightweight codebase scanning before gray area identification:
- New scout_codebase step checks for existing maps or does targeted grep
- Gray areas annotated with code context (existing components, patterns)
- Discussion options informed by what already exists in the codebase
- Context7 integration for library-specific questions
- CONTEXT.md template includes code_context section
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* fix: use `.claude/skills/` instead of `.agents/skills/` in agent and workflow skill references
Claude Code resolves project skills from `.claude/skills/` (project-level)
and `~/.claude/skills/` (user-level). The `.agents/skills/` path is the
universal/IDE-agnostic convention that Claude Code does not resolve, causing
project skills to be silently ignored by all affected agents and workflows.
Fixes#758
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: support both `.claude/skills/` and `.agents/skills/` for cross-IDE compatibility
Instead of replacing `.agents/skills/` with `.claude/skills/`, reference both
paths so GSD works with Claude Code (`.claude/skills/`) and other IDE agents
like OpenCode (`.agents/skills/`).
Addresses review feedback from begna112 on #758.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Stephen Miller <Stephen@betterbox.pw>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
DEBUGGER_MODEL and CHECKER_MODEL used uppercase bash convention but the
Task() calls referenced {debugger_model} and {integration_checker_model}
(lowercase). The mismatch caused Claude to skip substitution and fall back
to the parent session model, ignoring the configured GSD profile.
Co-authored-by: Ethan Hurst <ethan.hurst@outlook.com.au>
* Fix /gsd:update to pin installer to latest
* Harden /gsd:update install detection with global fallback
---------
Co-authored-by: Colin <colin@solvely.net>
Records test generation in project state via state-snapshot call,
keeping add-tests consistent with other GSD workflows that track
their operations.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New command that generates unit and E2E tests for completed phases:
- Analyzes implementation files and classifies into TDD/E2E/Skip
- Enforces RED-GREEN gates for both unit and E2E tests
- Uses phase SUMMARY.md, CONTEXT.md, and VERIFICATION.md as test specs
- Presents test plan for user approval before generating
Test classification rules:
- TDD: pure functions where expect(fn(input)).toBe(output) is writable
- E2E: UI behavior verifiable by Playwright (keyboard, navigation, forms)
- Skip: styling, config, glue code, migrations, simple CRUD
Ref: Issue #302
The `--auto` flag on discuss-phase and plan-phase chains stages automatically:
discuss → plan → execute, each in a fresh context window. Currently the chain
is broken because auto-advance spawns `Task(prompt="Run /gsd:plan-phase --auto")`
which requires the Skill tool — but Skills don't resolve inside Task subagents.
Result: plan-phase never runs from discuss's auto-advance, execute-phase never
runs from plan's auto-advance, and gsd-executor subagents are never spawned.
Fix: Replace `Task(prompt="Run /gsd:XXX")` with Task calls that tell the
subagent to read the workflow .md file directly via @file references — the same
pattern that already works for gsd-executor spawning in execute-phase.
Changes:
- execute-phase.md: Add --no-transition flag handling so execute-phase can
return status to parent instead of running transition.md when spawned by
plan-phase's auto-advance
- plan-phase.md: Replace Skill-based Task call with direct @file reference
to execute-phase.md, passing --no-transition to prevent transition chaining
- discuss-phase.md: Replace Skill-based Task call with direct @file reference
to plan-phase.md, with richer return status handling (PHASE COMPLETE,
PLANNING COMPLETE, PLANNING INCONCLUSIVE, GAPS FOUND)
Nesting depth: discuss → Task(plan) → Task(execute) → Task(executor) = 3 levels
max. Each level gets clean 200k context.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Adds automated feedback architecture research to gsd-phase-researcher,
enforces it as Dimension 8 in gsd-plan-checker, and introduces
{phase}-VALIDATION.md as the per-phase validation contract.
Ensures every phase plan includes automated verify commands before
execution begins. Opt-out via workflow.nyquist_validation: false.
Closes#122 (partial), related #117
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Subagents now read project-level CLAUDE.md if it exists:
- Workflows: execute-phase, plan-phase, quick
- Agents: gsd-executor, gsd-planner, gsd-phase-researcher, gsd-plan-checker
Agents read ./CLAUDE.md in their fresh context, following project-specific
guidelines, security requirements, and coding conventions.
Fixes: #671
Co-Authored-By: Claude <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>
Closes five gaps where requirements could slip through unchecked at milestone
level: audit now cross-references VERIFICATION.md + SUMMARY frontmatter +
REQUIREMENTS.md traceability, integration checker receives req IDs, gap objects
carry plan-level detail, plan-milestone-gaps updates traceability, and
complete-milestone gates on requirements status.
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>
