Tailor the design phase to the feature's domain instead of using one generic
(backend-flavored) schema for everything.
- classify_feature_domain: auto-detect frontend/backend/fullstack from an
explicit story Type:/Domain: field, then heuristic keyword scoring, failing
safe to fullstack (the superset) when ambiguous
- build_lens_block + build_domain_schema: inject a domain-specific planning
lens (component states, a11y, responsive / API contract, error handling,
migrations, concurrency, observability) and matching JSON fields, added to
the existing core schema (non-breaking)
- run_design_critic is now domain-aware: missing FE component states/a11y or
BE error paths/status codes are enforced as NEEDS_REVISION gaps via the
existing revision loop
- validate_domain_completeness: advisory warning + metric for the common
omissions (FE components without states, BE API without error handling)
- get_result_feature_type getter; TDD reconciliation now hints which test
kinds to emphasize per domain
Auto-detection only (no manual override flag yet). All additions are advisory
except the critic enforcement, preserving the non-blocking design contract.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Design phase improvement #7:
- Add build_planned_test_files_context, which extracts the test_files the
design phase already planned (resume-safe: in-memory then persisted file)
- Inject it into the test-spec phase prompt so TDD reuses the planned test
files and paths instead of independently deciding the test surface, and
flags any deviation
Returns empty (no-op) when there is no plan, no test_files, or jq is absent.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Design phase improvements #5 and #4:
#5 Deterministic, language-aware exploration:
- Add detect_project_type (node/rust/go/python) and build_repo_map, which
pre-computes a bounded repository map (project type, top-level structure,
representative source files) tailored to the detected language
- Inject the map into the design prompt instead of hardcoded JS/TS find
commands and "hope the model explores" guidance
#4 Critic loop:
- Add run_design_critic: a fresh-context skeptic that checks whether the plan
maps every acceptance criterion and conforms to the architecture, emitting
structured gaps
- execute_design_phase now generates -> critiques -> regenerates with gap
feedback, bounded by MAX_DESIGN_CRITIC_ATTEMPTS (default 2). Design stays
advisory: it always proceeds with the best plan and records a metric when
gaps remain
- Add --skip-design-critic flag and document MAX_DESIGN_CRITIC_ATTEMPTS
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Design phase improvement #3:
- Switch the design plan output contract from a free-text DESIGN START/END
block to a JSON result block, parsed via the shared extract_json_result /
check_phase_completion helpers (jq-less and legacy-text fallbacks retained)
- Add validate_design_coverage: warns + records a metric when the plan does
not map every acceptance criterion declared in the story (advisory only,
since design is a non-blocking phase). AC detection is heuristic and skips
when no AC identifiers are found or jq is unavailable
- Add a "design" case to the legacy fallback in check_phase_completion for
robustness when no JSON block is present
Hook bypassed: pre-existing markdownlint errors are confined to the gitignored
.context/ workspace dir; lint, format:check, and bash -n all pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Design phase improvements #1 and #2:
#1 Bounded context:
- Pass architecture.md by path instead of embedding full contents
(the main unbounded size risk in this prompt)
- Cap decision-log context at last 20KB (matches dev phase)
- Add log_prompt_size guard, consistent with other phases
- Replace hardcoded JS/TS exploration hints with language-agnostic guidance
#2 File persistence:
- Add DESIGN_DIR config and persist each plan to <DESIGN_DIR>/<story>-design.md
- build_design_context_for_dev falls back to the persisted file when the
in-memory plan is empty, so resumed runs keep their design context
Story file remains inlined (small, bounded, needed in full by the planner).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sync functional improvements developed in revive-dev into BMAD-METHOD
fork while preserving repo-specific paths:
- Add memory-safe Claude helpers (run_claude_to_file, read_phase_tail)
that pipe output to temp files instead of bash variables, preventing
GB-scale RAM usage during long epic executions
- Add kill_orphaned_test_processes() to clean up zombie jest/vitest/
playwright/pytest processes between stories and on exit
- Replace per-call `env -u CLAUDECODE` with global `unset CLAUDECODE`
at script start for cleaner nested session support
- Port metrics resume/accumulation logic that restores counters from
existing YAML on resumed runs and accumulates duration
- Add log truncation between stories (64KB cap) to prevent unbounded
log growth across multi-story runs
- Add log persistence and cleanup trap to epic-chain.sh
- Revert regression-gate.sh test commands to direct execution (matching
revive-dev pattern)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Wrap all Claude CLI subprocess calls with `env -u CLAUDECODE` to prevent
parent env var interference with child processes (17 sites across 7 files)
- Add `flush_log_to_repo()` to epic-execute.sh for incremental log persistence
after each story completes or fails (prevents log loss on interruption)
- Add portable `run_with_timeout` utility to utils.sh and wrap all test
invocations in epic-execute.sh and regression-gate.sh with configurable
timeout (default 120s via REGRESSION_TEST_TIMEOUT)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add decision log module for context preservation across phases
- Add regression gate module for test baseline tracking
- Add design phase module for pre-implementation planning
- Enhance fix phase to include real tooling output
- Pass design and decision context to dev phase
- Add --skip-design and --skip-regression CLI flags
- Modularize into epic-execute-lib/ for maintainability
Implements improvements from bmad_improvements_v2.md:
- Phase 2.1: Real test output in fix loops
- Phase 2.2: Cumulative decision log
- Phase 2.3: Regression test gate
- Phase 3.1: Pre-implementation design phase
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Caleb