P0 - Must Have

Backlog

Tracked features, ideas, and deferred work for grooming and future sessions.

P0 - Must Have

(Empty)

P1 - Should Have

gitlab-skill: Remove hardcoded corporate URL

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P1 Description: validate_glfm.py lines 152-153 hardcode https://sourcery.assaabloy.net as the default GitLab instance URL. gitlab-ci-local-guide.md line 51 also references this URL. This leaks a corporate internal URL into a public repository. Replace with a generic placeholder (e.g., https://gitlab.example.com) or make the URL a required argument with no default. Files:

• plugins/gitlab-skill/skills/gitlab-skill/scripts/validate_glfm.py (lines 152-153)
• plugins/gitlab-skill/skills/gitlab-skill/references/gitlab-ci-local-guide.md (line 51)

bash-development: Fix bash-53-features inaccuracies and task_output bug

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P1 Status: FACT-CHECKED 2026-02-21 — All 11 Bash 5.3 features verified against CHANGES file, NEWS file, and bash.1 manpage from official tarball. Features exist but SKILL.md has detail inaccuracies. Original review agent "fabrication" claim was wrong. Description: Two issues:

1. bash-53-features/SKILL.md inaccuracies (features exist but details wrong):
   • GLOBSORT: :asc/:desc suffixes fabricated — actual syntax is +/- prefix; date not a valid specifier (should be mtime); missing specifiers: blocks, atime, ctime, numeric, nosort
   • ${ command; } examples: missing required space after { (bash.1 requires space/tab/newline/| after {)
   • ${| command; } REPLY examples: misleading — REPLY is local within substitution, restored after completion
   • C23 claim overstated: build minimum is C90, not C23; C23 changes are about conformance
   • kv/strptime usage examples: unverifiable from official sources (existence confirmed, interface speculative)
2. log_functions.sh bug (line 1401): task_output() function references ${task_output} variable on lines 1410/1412, but only raw_task_output is assigned (line 1406). Output will be empty.

Citations:

• Bash CHANGES: https://tiswww.case.edu/php/chet/bash/CHANGES lines 850-860 (accessed 2026-02-21)
• Bash NEWS: https://ftp.gnu.org/gnu/bash/bash-5.3.tar.gz extracted NEWS lines 48-57 (accessed 2026-02-21)
• Bash manpage: bash-5.3/doc/bash.1 lines 2525-2566 (GLOBSORT), 4146+ (command substitution) (accessed 2026-02-21) Files:
• plugins/bash-development/skills/bash-53-features/SKILL.md (GLOBSORT, examples, C23)
• plugins/bash-development/skills/bash-logging/scripts/log_functions.sh (lines 1401, 1410, 1412)

commitlint: Verify --last flag and exit codes against primary sources RESOLVED

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Resolved: 2026-02-21 Priority: P1 Status: FACT-CHECKED 2026-02-21 — --last flag VERIFIED across 4 independent sources. Flag exists in source code (cli.ts with alias -l), official CLI reference docs, commitlint help output, and raw documentation. Exit codes verified against ExitCode enum in cli-error.ts. Original review agent claim that --last was fabricated was wrong. Citations:

• commitlint source: @commitlint/cli/src/cli.ts (accessed 2026-02-21)
• commitlint docs: https://commitlint.js.org/reference/cli.html (accessed 2026-02-21)
• commitlint source: @commitlint/cli/src/cli-error.ts ExitCode enum (accessed 2026-02-21)

clang-format: Fix broken YAML frontmatter

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P1 Description: SKILL.md line 3 has description:"Configure clang-format..." — missing the required space after the description: key. This causes YAML parsing failures. The frontmatter should be description: "Configure clang-format...". Files:

• plugins/clang-format/skills/clang-format/SKILL.md (line 3)

agent-orchestration: Remove phantom /is-it-done command references

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P1 Description: The /is-it-done command is referenced 12 times across 4 files in the agent-orchestration plugin, but the command does not exist in the repository. No is-it-done command file exists in any commands/ directory. All references should be removed or replaced with the actual verification mechanism. Also contains orphaned "Clavix" references in clear-framework.md (lines 3, 15-20) — Clavix is an external tool not present in this repo. Files:

• plugins/agent-orchestration/skills/how-to-delegate/SKILL.md
• plugins/agent-orchestration/skills/agent-orchestration/SKILL.md
• plugins/agent-orchestration/skills/agent-orchestration/post-completion-validation-protocol.md
• plugins/agent-orchestration/skills/agent-orchestration/references/synthesis-improvements-from-research.md
• plugins/agent-orchestration/skills/agent-orchestration/clear-framework.md (Clavix references)

perl-development: Fix shell injection vulnerability in example template

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P1 Description: The Perl example template in perl_example_file.pl references undefined App::Logger module and contains patterns vulnerable to shell injection. Example templates used by AI to generate code should demonstrate secure coding practices. Missing shebangs on example scripts also violate repo conventions. Files:

• plugins/perl-development/skills/perl-development/references/perl_example_file.pl

hallucination-detector: Fix dead backtick evidence marker and multi-occurrence "because" bug

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P1 Description: The hallucination-audit-stop.js hook has two critical bugs: (1) Evidence markers using backticks are not properly detected, causing false negatives when evidence is inline-code formatted. (2) The "because" pattern matcher has a multi-occurrence bug where only the first match is processed, missing subsequent hallucination indicators in the same response. Files:

• plugins/hallucination-detector/scripts/hallucination-audit-stop.js

python3-development: Fix 3 malformed frontmatter files

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P1 Description: Three frontmatter files in the python3-development plugin have malformed YAML frontmatter (similar missing-space-after-colon pattern). Also contains a ghost agent reference (referencing an agent that doesn't exist in the plugin). Template comment loss indicates content was stripped during automated processing. Files: TBD — run uv run plugins/plugin-creator/scripts/plugin_validator.py plugins/python3-development/ to identify exact files

the-rewrite-room: Fix nonexistent script reference and 6 broken links

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P1 Description: validators.yaml line 12 references plugins/plugin-creator/scripts/validate_frontmatter.py as a script path — this file exists but the path may not resolve correctly from the rewrite-room context. research-utilities.md references a --json flag on file_metrics.py (lines 62, 71) that may not exist. 6 additional broken links need verification and fixing. Files:

• plugins/the-rewrite-room/skills/the-rewrite-room/registry/validators.yaml (line 12)
• plugins/the-rewrite-room/skills/the-rewrite-room/workflows/research-utilities.md (lines 62, 71)

fastmcp-creator: Add citations for 1200+ lines of FastMCP 3.x API documentation

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P1 Status: FACT-CHECKED 2026-02-21 — require_auth hallucination flag was WRONG. The string require_auth does not appear anywhere in the plugin. The actual auth APIs documented (require_scopes, restrict_tag, AuthContext, AuthMiddleware) are all VERIFIED against installed FastMCP 3.0.0rc2 source code (fastmcp/server/auth/authorization.py, fastmcp/server/middleware/authorization.py). Citation need still valid — documentation lacks source attribution. Description: The plugin contains over 1200 lines of FastMCP 3.x API documentation derived from a release candidate version, with no source citations. Per CLAUDE.md citation requirements, all factual claims must have cited sources. Auth API claims verified but citations still needed for all documentation. Citations:

• FastMCP auth: fastmcp/server/auth/__init__.py lines 8-13, authorization.py lines 48, 78, 106 (accessed 2026-02-21)
• FastMCP middleware: fastmcp/server/middleware/authorization.py line 51 (accessed 2026-02-21) Files:
• plugins/fastmcp-creator/skills/fastmcp-creator/references/ (multiple files)

Consolidate validate_frontmatter.py into plugin_validator.py

Source: Frontmatter validation bug-fix session 2026-02-20 Added: 2026-02-20 Completed: 2026-02-20 Status: DONE — commit bfb03f1 on branch claude/fix-frontmatter-validation-u2iV2 Priority: P1 Description: plugin_validator.py (4190 lines) copy-pastes frontmatter validation logic from validate_frontmatter.py (1341 lines) instead of importing it. Comments acknowledge this: "PYDANTIC FRONTMATTER MODELS (from validate_frontmatter.py)" and "Complexity preserved from validate_frontmatter.py for behavioral parity." This creates two maintenance surfaces: any change must be applied to both scripts (as seen when reversing the name-field bug workaround).

Required work:

1. Audit both scripts for all divergences (validation checks, Pydantic models, helper functions, CLI flags, scan patterns).
2. Extract shared code (Pydantic models, _fix_skill_name*, extract_frontmatter, detect_file_type, validate_and_normalize) into a shared module plugins/plugin-creator/scripts/frontmatter_core.py.
3. Refactor validate_frontmatter.py and plugin_validator.py to import from frontmatter_core.py.
4. Port any validation steps present in validate_frontmatter.py but missing from plugin_validator.py (e.g. skill-directory-name check, name-mismatch warning added in 2026-02-20).
5. Update frontmatter_utils.py if overlap exists.
6. Update all documentation: CLAUDE.md, reference files, script docstrings.
7. Update tests to import from the correct locations.
8. Run full test suite; verify pre-commit hooks still pass.

Suggested location: plugins/plugin-creator/scripts/

Validate and verify orchestrator-discipline plugin hooks and processes

Source: Plugin creation session 2026-02-19 Added: 2026-02-19 Completed: 2026-02-21 Status: DONE — commits 49e0ae5, ea3e737 on branch claude/bulk-backlog-grooming-LIQDi Plan: plan/tasks-4-validate-orchestrator-discipline.md Description: T1-T4 complete. Plugin passes claude plugin validate. Hook directory detection added. user-invocable: true added to SKILL.md. All 5 hook behavior tests pass. Suggested location: plugins/orchestrator-discipline/

SAM: Error Recovery / Rollback Procedures

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Define explicit procedure when a task fails irrecoverably. How to undo artifact changes? How to restore artifact plane to consistent state after failure? Research first: How do GSD, BMAD-METHOD, AutoGPT, and traditional CI/CD handle rollback? What patterns exist for transactional artifact updates? Suggested location: stateless-software-engineering-framework.md (new Appendix or Part 6 addition)

SAM: Human Escalation Criteria

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Define explicit triggers for escalating to human at each stage (not just Discovery). When should agents block and ask vs attempt repair vs fail? Research first: How do GSD deviation rules work? How does BMAD-METHOD handle human checkpoints? What escalation patterns exist in agent frameworks? Suggested location: stateless-software-engineering-framework.md (each Agent Specification section)

SAM: Timeout/Stall Detection

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Define mechanism to detect when an agent is stuck or has stalled. Include timeout thresholds per stage, health check patterns, and recovery actions. Research first: How do orchestration frameworks (Temporal, Prefect, Airflow) handle task timeouts? What heartbeat patterns exist? How does Gas Town handle session recycling? Suggested location: stateless-software-engineering-framework.md (Orchestrator section 3.8)

SAM: Artifact Schema Validation

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Define formal validation rules or JSON schemas for artifact formats. Currently only templates provided. Enable automated validation at stage boundaries. Research first: How do GSD artifacts (STATE.md, ROADMAP.md) enforce structure? What validation approaches exist in BMAD-METHOD? JSON Schema vs YAML validation vs custom parsers? Suggested location: sam-artifact-schemas/ (new directory with schema files)

SAM: Scope Creep Detection

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Define mechanism to detect when execution diverges from plan. How does Forensic Review detect that the execution agent solved a different problem than planned? Research first: How does GSD plan-checker detect deviation? What diff/comparison techniques exist? How do code review tools detect scope creep in PRs? Suggested location: stateless-software-engineering-framework.md (section 3.6 Forensic Review)

Meta-Process Capture — Expert Panel Dataset Builder

Source: ARL expert panel process (sessions 2026-02-12 to 2026-02-13) Added: 2026-02-13 Description: Document the multi-agent expert panel methodology as a reusable system for building datasets that inform skills and systems. The process — assign framework experts to repositories, ask structured questions, cross-examine, synthesize, map to requirements, validate — produced high-quality sourced findings. Capture this as a repeatable pattern. Key elements to document:

• Expert assignment protocol (one agent per source repo, source-code-only evidence standard)
• Question group design (themed questions → cross-examination → synthesis)
• Cross-examination as adversarial validation (experts challenge each other's claims)
• Phased output (discussion → requirement mapping → synthesis → validation)
• Traceability chain (claim → expert citation → file:line evidence)
• Session continuity handling (state file enables cross-session resumption) Input artifacts: plugins/plugin-creator/skills/assessor/references/ARL/ARL-agent-instructions.md, qa-expert-panel.md Suggested location: New skill or methodology document — captures the meta-process, not the ARL content

SAM Extension — Integrate ARL General Theory

Source: ARL expert panel Phase 3 output Added: 2026-02-13 Description: Integrate the 7 universal principles from synthesis-general-theory.md into SAM methodology documents. These principles (structure over instruction, front-loading reduces gates, AI cannot self-evaluate, compression is architectural, iteration-aware state required, parallelism enables independent verification, failure paths need more compression) extend SAM's scope to cover autonomous refinement loops. Input artifacts: plugins/plugin-creator/skills/assessor/references/ARL/synthesis-general-theory.md Target files: stateless-agent-methodology.md, stateless-software-engineering-framework.md Dependencies: None — general theory is framework-agnostic Related backlog items: SAM gap items (error recovery, human escalation, scope creep detection) — the general theory findings inform several of these

ARL Skill Development

Source: ARL expert panel Phase 3 output Added: 2026-02-13 Description: Build the Autonomous Refinement Loop as a skill using synthesis-arl-applicable.md as its reference foundation. The ARL is a logical process (Assess → Plan → Implement → Review → Repeat) for autonomous skill refinement with R1-R10 requirement gates. Input artifacts: plugins/plugin-creator/skills/assessor/references/ARL/synthesis-arl-applicable.md, synthesis-general-theory.md Scope: Logical process design — what gates fire when, what each gate checks, success/failure criteria. NOT implementation artifacts (schemas, thresholds, pseudocode). Dependencies: Benefits from SAM extension being done first (ARL would be a SAM-based skill) Suggested location: New skill under plugins/plugin-creator/skills/ or standalone plugin

Extract claude-plugin-lint to standalone PyPI package

Source: Gap analysis - no existing Claude Code plugin linters exist Added: 2026-02-01 Description: Extract and enhance validate_frontmatter.py into a standalone open-source project. First dedicated linter for Claude Code plugin frontmatter (SKILL.md, agents/.md, commands/.md). Official claude plugin validate only checks plugin.json structure. Features to include:

• YAML frontmatter schema validation with Pydantic models
• Auto-fix capabilities (arrays → comma-separated, multiline → single-line)
• Token-based complexity metrics (tiktoken) instead of line counts
• Cross-reference validation (agent references non-existent skill)
• Marketplace readiness scoring
• Pre-commit hook integration
• CLI with --fix and --report modes Current source: plugins/plugin-creator/scripts/validate_frontmatter.py Suggested repo name: claude-plugin-lint or cc-plugin-validator

P2 - Could Have

Add ty support alongside mypy in distributed plugins COMPLETED

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Completed: 2026-02-21 Description: Completed as full mypy-to-ty migration across all active documentation, skills, agents, and reference files. All stale mypy references updated to ty (Astral ecosystem). Third-party reference docs (mypy-docs/, rules/mypy/) retained as-is. Historical plan/ files left unchanged.

conventional-commits: Fix CHANGELOG references to nonexistent files

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Description: CHANGELOG references files that do not exist in the repository. Additionally, related skills referenced in the plugin do not exist. All dead references need to be either created or removed. Files: plugins/conventional-commits/ (CHANGELOG and skill cross-references)

dasel: Reconcile 265 -f flag occurrences with reference documentation

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Description: Reference documentation states dasel uses a specific flag pattern, but 265 occurrences of the -f flag across the skill contradict this documentation. The hook file exists on disk but is not registered in the plugin manifest (plugin.json). Run auto_sync_manifests.py --reconcile to fix manifest drift, and audit -f flag usage against official dasel documentation. Files:

• plugins/dasel/ (skill files with -f flag usage)
• plugins/dasel/.claude-plugin/plugin.json (missing hook registration)

litellm: Remove private API documentation and update verification date RESOLVED

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Resolved: 2026-02-21 Status: FACT-CHECKED 2026-02-21 — Review claim REFUTED. litellm._should_retry() is NOT a private/internal API. It is explicitly documented in official litellm Exception Mapping docs (https://docs.litellm.ai/docs/exception_mapping) with code examples showing litellm._should_retry(e.status_code). Despite the underscore prefix, this is an intentionally public utility function exported in __init__.py. The function exists at utils.py:6506. Description: Original review flagged _should_retry() as private API. Fact-checking verified it is documented public API in official docs. Stale verification date may still need updating. Citations:

• litellm docs: https://docs.litellm.ai/docs/exception_mapping (accessed 2026-02-21)
• litellm source: litellm/utils.py line 6506 (accessed 2026-02-21)
• litellm source: litellm/__init__.py type stub exports _should_retry (accessed 2026-02-21) Files:
• plugins/litellm/skills/litellm/SKILL.md (line 257)

verification-gate: Remove unsubstantiated 95% confidence claim

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Description: The skill contains an unsubstantiated claim about "95% confidence" with no supporting data, citation, or methodology. Per CLAUDE.md verification protocol, claims must be cited. Either add supporting evidence or remove the specific percentage. Also fix missing code fence language specifiers and stale dates. Files: plugins/verification-gate/ (SKILL.md and reference files)

development-harness: Remove hardcoded machine path and fix version drift

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Description: Contains a hardcoded machine-specific path (likely /home/user/... or similar) that won't work on other systems. Version references have drifted from actual tool versions. Role table has inconsistencies between documented and actual roles. Files: plugins/development-harness/ (SKILL.md and reference files)

llamafile: Fix HuggingFace model URLs (wrong org name + fabricated repos)

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Priority: P2 Status: FACT-CHECKED 2026-02-21 — SourceForge URL VERIFIED (legitimate auto-mirror with explicit disclaimer). HuggingFace model URLs REFUTED — wrong org name (Mozilla should be mozilla-ai) and model repo names (gemma-3-3b-it-gguf, Qwen3-0.6B-gguf, Mistral-7B-gguf, Llama-3.1-8B-gguf) all return 404. GitHub URLs using mozilla-ai/llamafile are correct (old Mozilla-Ocho redirects). Description: SourceForge mirror is confirmed legitimate (auto-mirror with disclaimer). HuggingFace URLs need two fixes: (1) org name Mozilla must be changed to mozilla-ai, (2) model repo names need verification against actual mozilla-ai org repos — current names appear fabricated. The llava-v1.5-7b-llamafile URL works only via redirect. Citations:

• SourceForge mirror: https://sourceforge.net/projects/llamafile.mirror/files/0.9.3/ — disclaimer states "exact mirror of the llamafile project" (accessed 2026-02-21)
• HuggingFace 404s: Mozilla/Qwen3-0.6B-gguf, Mozilla/Mistral-7B-gguf, Mozilla/gemma-3-3b-it-gguf, Mozilla/Llama-3.1-8B-gguf (accessed 2026-02-21)
• HuggingFace redirect: Mozilla/llava-v1.5-7b-llamafile redirects to mozilla-ai/llava-v1.5-7b-llamafile (accessed 2026-02-21)
• GitHub: mozilla-ai/llamafile resolves correctly; Mozilla-Ocho/llamafile redirects (accessed 2026-02-21) Files:
• plugins/llamafile/skills/llamafile/SKILL.md (lines 78, 87-92 model URLs; line 69 SourceForge OK)

prompt-optimization: Fix unreachable reference files and raw JSX/MDX markup

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Description: Two reference files are unreachable (not linked from SKILL.md or any other file). Contains raw Anthropic JSX/MDX markup that should be converted to standard markdown for compatibility with Claude Code's markdown rendering. Files: plugins/prompt-optimization-claude-45/ (reference files)

brainstorming-skill: Remove orphaned bibliography entry and cross-reference headings

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Description: Contains an orphaned bibliography entry (referenced nowhere) and orphaned cross-reference section headings that point to removed or renamed content. Files: plugins/brainstorming-skill/ (SKILL.md and reference files)

uv: Fix incorrect script paths in README

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Description: README contains incorrect script paths that don't match the actual file locations. The README is also disproportionately large for what is essentially a thin wrapper plugin. Files: plugins/uv/ (README.md)

plugin-creator: Remove dead code and triplicated regex

Source: Plugin code review session 2026-02-21 Added: 2026-02-21 Description: Contains triplicated regex patterns (same regex defined 3 times), a dead skipped list that is populated but never read, an unused sum() call, and HK005 warning is incorrectly treated as an error in certain code paths. Also has a noqa BLE001 suppression that should be addressed per CLAUDE.md linting policy. Files: plugins/plugin-creator/ (scripts and skill files)

Add PR003/PR004 test coverage to plugin registration validator

Source: Code review session 2026-02-21 Added: 2026-02-21 Description: PluginRegistrationValidator defines PR003 (missing metadata fields: repository, homepage, author) and PR004 (repository URL mismatches git remote URL) at lines 276-277 of plugin_validator.py, and emits them at lines 2815 and 2834. Tests exist for PR001 (unregistered) and PR002 (missing file), but not PR003/PR004. Add tests to plugins/plugin-creator/tests/test_plugin_registration_validator.py covering: (1) PR003 emitted when metadata fields absent; (2) PR004 emitted when repo URL mismatches remote.

kaizen: MCP consolidation analysis

Source: Design session 2026-02-20 Added: 2026-02-20 Description: The plugin currently runs two MCP servers (kaizen-duckdb via mcp-server-motherduck, kaizen-analysis via server.py) plus a standalone CLI script (sentiment-score.py). Investigate: (1) What does each MCP server provide that the other cannot? Can they be merged into a single server? (2) Why is sentiment-score.py a standalone script rather than an MCP tool inside server.py? What would be gained or lost by moving scoring into the MCP server (always-on scoring, no manual invocation, lock ownership)? (3) Is there a clean boundary between "batch processing" (script) and "query/serve" (MCP) that should be preserved? Decision needed: Consolidate vs. keep separate, with rationale. Suggested location: plugins/agentskill-kaizen/

SAM: Parser regex false positive on "## Task Summary Statistics"

Source: Migration proof-of-concept (2026-02-13) Added: 2026-02-13 Description: The widened task header regex ^#{2,3}\s+Task:?\s+([A-Za-z0-9.]+)[:\s-]+(.+)$ in implementation_manager.py matches ## Task Summary Statistics as task ID "Summary" with title "Statistics". The regex needs a negative lookahead or post-parse filter to exclude non-task sections. Observed when parsing plan/tasks-1-plugin-linter.md. File: plugins/python3-development/skills/implementation-manager/scripts/implementation_manager.py line 645

SAM: Replace validate-task-file.sh with Python validator

Source: Task format standardization plan (2026-02-13) Added: 2026-02-13 Description: The bash validator at plugins/plugin-creator/scripts/validate-task-file.sh validates a different schema (tasks-refactor-*.md) and doesn't understand YAML frontmatter. Replace with Python validator that uses the shared task_format.py module. File: plugins/plugin-creator/scripts/validate-task-file.sh

SAM: Parallel Execution Details

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Detail safe parallelization within SAM pipeline. When can tasks run in parallel? How to handle merge conflicts? Reference GSD wave execution pattern. Research first: How does GSD wave execution work in detail? How do task orchestrators (Temporal, Prefect) handle parallel dependencies? What conflict resolution patterns exist? Suggested location: stateless-software-engineering-framework.md (new section 2.4 or Appendix)

SAM: Multi-Model Strategy

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Define guidance for using different models for different agent types. E.g., cheaper/faster models for simple verification, stronger models for planning. Research first: How do agent frameworks handle model selection? What cost/quality tradeoffs exist? How does Claude Code's haiku/sonnet/opus selection work? Suggested location: stateless-software-engineering-framework.md (Implementation Roadmap or new Appendix)

SAM: Audit Trail / Observability

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Beyond artifacts, define logging/metrics/tracing guidance. How to diagnose pipeline issues? What telemetry to capture? Research first: How do GSD and BMAD-METHOD handle logging? What observability patterns exist in agent frameworks? OpenTelemetry for LLM workflows? Suggested location: stateless-software-engineering-framework.md (new Appendix I)

SAM: Partial Success Handling

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Define how to represent and handle partial task success. Task completes some DoD items but not all. How is this state represented in artifacts? Research first: How do GSD checkpoints represent partial progress? How do CI/CD systems handle partial test passes? What state machine patterns exist? Suggested location: stateless-software-engineering-framework.md (section 3.5 Execution Agent output)

SAM: Context Size Management

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Define explicit guidance for measuring and managing context size per agent. What's the target token budget? How to detect context pressure? Research first: How do agent frameworks measure context usage? What token counting approaches exist? How does Claude Code handle context limits internally? Suggested location: stateless-software-engineering-framework.md (section 2.1 or Appendix C)

SAM: Conflicting Review Findings

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Define protocol when forensic review and self-verification disagree. Which takes precedence? How to adjudicate conflicts? Research first: How do code review systems handle conflicting reviewers? What adjudication patterns exist in multi-agent systems? How does GSD handle verification disagreements? Suggested location: stateless-software-engineering-framework.md (section 3.6 Forensic Review)

Multi-session build state lost during context compaction

Source: agentskill-kaizen plugin build (2026-02-18), 3 sessions with 1 compaction Added: 2026-02-18 Description: During the agentskill-kaizen build (8-phase /plugin-dev:create-plugin workflow), context compaction mid-build converted structured task state into a narrative summary. The resuming session had to reconstruct "what's done vs pending" from prose rather than a checklist. Background agent results that were already consumed and applied reappeared as late notifications after compaction, requiring manual deduplication ("did I already handle this?"). No persistent artifact tracked phase completion, commit SHAs per phase, deferred items, or agent result consumption status. Observed symptoms:

• Phase completion status existed only in ephemeral context — lost on compaction
• Background agent notifications arrived after their findings were already applied (3 duplicate notifications)
• Plan committed early (87a0b93) diverged from actual implementation but was never updated
• No mechanism to mark agent results as "consumed" — each notification required re-evaluation

/plugin-dev:create-plugin workflow lacks intra-phase parallelism tracking

Source: agentskill-kaizen plugin build (2026-02-18) Added: 2026-02-18 Description: The 8-phase create-plugin workflow treats each phase as a serial step, but Phase 5 (Implementation) actually consisted of 6 parallel sub-tasks and Phase 6 (Validation) spawned 3 parallel review agents. The workflow provides no structure for tracking parallel work within a phase — no task dependencies, no completion gates, no way to know which sub-tasks are done after compaction. Batching validation fixes by file rather than by finding would also have been more efficient (SKILL.md was edited 3 separate times when one pass would have sufficed).

Background agent result deduplication after compaction

Source: agentskill-kaizen plugin build sessions 2-3 (2026-02-18) Added: 2026-02-18 Description: Background agents launched in session N may complete after context compaction or session restart. The system delivers their results as <task-notification> messages, but there is no mechanism to mark results as already consumed. During the kaizen build, 3 review agents (plugin-validator, 2x skill-reviewer) completed during Phase 6 and their findings were applied in commit 0d61480. After compaction, all 3 re-delivered their notifications in session 3, requiring manual evaluation each time ("was this already handled?"). A persistent state file (e.g., .planning/kaizen/agent-results.json tracking task IDs → consumed/pending) would eliminate this waste.

/plugin-dev:create-plugin Phase 6 validation should batch fixes by file, not by finding

Source: agentskill-kaizen plugin build Phase 6 (2026-02-18) Added: 2026-02-18 Description: Phase 6 collected findings from 3 parallel review agents, then applied fixes one finding at a time. This resulted in SKILL.md being edited 3 separate times (description rewrite, SQL removal, MCP server name fix) when a single pass through the file would have applied all fixes together. The workflow should group all findings by file, then make one editing pass per file. Reduces Edit tool calls and context consumed by repeated reads.

Plan artifact diverges from implementation without update mechanism

Source: agentskill-kaizen plugin build (2026-02-18), plan committed as 87a0b93 Added: 2026-02-18 Description: The research/plan from Phase 2-3 was committed early as a markdown file. During implementation (Phase 5), decisions changed — the MCP server grew from planned scope, analysis dimensions were rebalanced, hook patterns shifted. The plan was never updated to reflect actual implementation. After compaction, the stale plan became a potential source of confusion. Two options to address: (1) update the plan artifact after each phase, or (2) treat the plan as disposable and track only the living state (what's done, what's pending, what deferred).

Evaluate scikit-learn dependency weight for agentskill-kaizen cluster_sessions tool

Source: agentskill-kaizen MCP server review (2026-02-18) Added: 2026-02-18 Description: The cluster_sessions tool in plugins/agentskill-kaizen/mcp/server.py uses scikit-learn (KMeans, CountVectorizer) for session clustering. scikit-learn pulls in ~40MB of transitive dependencies (numpy, scipy, joblib, threadpoolctl). The typical use case is clustering dozens of sessions, not thousands. The Phase 1-2 research produced no durable artifact evaluating library choices — scikit-learn was assumed from the backlog without validation. Investigate whether a lighter alternative (e.g., pyclustering, stdlib-based implementation, or just numpy directly) would suffice for this scale, and whether the KMeans-on-bag-of-words approach is even appropriate for tool-call sequence similarity. Research first: What clustering approaches work for short categorical sequences? Is cosine similarity on bag-of-words tool vectors meaningful for workflow comparison? What lightweight Python clustering libraries exist that don't pull in scipy?

Ideas

Configurable Token Thresholds for plugin_validator.py

Source: Session 2026-02-18, token threshold analysis Added: 2026-02-18 Description: plugin_validator.py token complexity thresholds (4400 warn, 8800 error) are hardcoded constants. Allow per-project or per-plugin configuration via a config file (e.g., .claude-plugin/validator.json or a [tool.plugin-validator] section in pyproject.toml) so different projects can set their own limits. Suggested location: plugins/plugin-creator/scripts/plugin_validator.py

SAM: Cost/Token Management

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Explore token budgets and cost controls per agent/stage. Track API costs. Set limits per task. Research first: How do LLM cost management tools work (LangSmith, Helicone)? What budget enforcement patterns exist?

SAM: Team Coordination Protocols

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Explore how multiple humans interact with the pipeline concurrently. Locking? Ownership? Notifications? Research first: How do collaborative editing systems handle concurrent users? What patterns exist in git-based workflows?

SAM: External System Integration Patterns

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Explore patterns for integrating with issue trackers (GitHub Issues, Jira), CI/CD pipelines, git hooks. Research first: How does GSD integrate with external tools? What MCP servers exist for issue trackers? How do agent frameworks bridge to CI/CD?

SAM: Migration Strategy Guide

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Explore how to migrate existing projects to SAM. Incremental adoption? Parallel running? Artifact bootstrap? Research first: How do organizations adopt new methodologies incrementally? What migration patterns exist for process changes?

SAM: Training/Onboarding Materials

Source: Gap analysis of SAM framework Added: 2026-02-01 Description: Explore creating materials for new team members learning the methodology. Tutorials, examples, quick-start guides. Research first: What training materials exist for GSD, BMAD-METHOD? What makes effective agent methodology onboarding?

SAM: Non-Code Workflow Guidance

Source: Gap analysis of SAM framework Added: 2026-02-01 **Desc

…(truncated)