CHANGELOG.md

Changelog

All notable changes to the claude-plugins project will be documented in this file.

The format is based on Keep a Changelog.

[Unreleased]

code v1.2.2

Changed

• Updated plan-with-codex command argument-hint to use positional syntax instead of optional bracket notation

platform v1.0.2

Added

• New "Refactoring Existing Prompts" section in context-engineering skill covering pitfalls for stale cross-references, over-abstraction, lost preconditions, and silent behavior changes

code v1.2.1

Changed

• plan-agent now verifies Codex findings against the codebase before acting -- rejects findings that don't hold up with evidence, writes a revision summary for cross-round context
• codex-review skill accepts --revisions-file parameter, injecting Claude's revision summary into Codex's prompt on rounds > 1 so rejected findings are not re-raised

Fixed

• Fixed plan-with-codex resume path triggering a redundant user review checkpoint when the user had already confirmed by choosing "resume with existing plan"

code v1.2.0

Added

• New plan-agent agent for creating and revising implementation plans via codebase exploration
• New plan-with-codex command for iterative plan refinement through Claude + Codex debate loops
• New codex-review skill to run Codex plan reviews and return structured verdict feedback
• New debate-loop.sh script providing standalone CLI for Claude + Codex debate orchestration
• New plan-review.sh hook that triggers Codex review when Claude exits plan mode

code-review v1.2.0

Added

• New resolve-scope subcommand in code_review_helpers.py -- deterministic scope resolution replacing inline shell logic for PR branch lookup, git fetch, base-ref overrides, and path filter preservation
• New fetch-intent subcommand -- fetches PR description or commit messages as intent context for the Premise Reviewer
• New classify-intent subcommand -- classifies diff intent (feature, fix, refactor, mixed) from PR metadata and file statuses for model routing
• New collect-findings subcommand -- merges agent_*.json files and hygiene findings into a single findings.json, replacing inline Python-in-Bash merge logic
• New verdict subcommand -- computes deterministic PR verdict (approve, needs_attention, decline) from validated findings, replacing inline orchestrator logic
• New prep-assets subcommand -- copies shared_prompt.txt and bha_suffix.txt from plugin to CR_DIR in a single step, consolidating scattered cp commands
• New extract-patches subcommand -- extracts per-partition and full-diff patches to disk with batched extraction for large diffs (>200 files)
• New bha_suffix.txt prompt file -- Bug Hunter A persona and focus areas extracted from inline heredoc in start.md
• Intent-aware model routing: Premise Reviewer uses Opus for fix/refactor/mixed intents, Sonnet for feature intents; BHA uses Opus for implementation partitions, Sonnet for test-only partitions
• Mixed-partition splitting in partition subcommand -- separates test files from implementation files when impl LOC exceeds threshold
• Agent cap enforcement via --max-bha-agents parameter in partition, computed from route output
• Trivial partition merging -- partitions below 20 LOC are absorbed into same-type normal partitions
• Cache status message (status_kind, status_message) appended to cache_result.json by cache-check, replacing orchestrator-side message formatting
• --exclude-test-partitions flag on cache-update to skip caching files from Sonnet-reviewed test-only partitions
• Self-discard validation rule (check 7) in shared_prompt.txt -- agents must discard findings they conclude are not actually problems

Changed

• Refactored start.md orchestrator to delegate workflow steps to Python subcommands instead of inline shell logic
• setup subcommand now accepts --cr-dir-prefix and creates CR_DIR with random suffix, removing the need for the orchestrator to generate random directory names
• route subcommand now accepts --intent parameter and outputs max_bha_agents for downstream partition cap enforcement
• Reduced default partition LOC budget from 800 to 500

judges v1.3.1

Changed

• run-judges skill now accepts a --workdir <path> parameter for standalone use outside run-loop.sh; resolved in order: --workdir arg → $CLOSEDLOOP_WORKDIR env var → .closedloop-ai/judges default (directory created automatically if absent)

code v1.1.4

Changed

• run-loop.sh judge invocations (plan_judges, code_judges) now pass --workdir $workdir explicitly in the claude -p prompt, aligning with the updated run-judges skill parameter contract

judges v1.3.0

Added

• New prd artifact type support in run-judges skill — 4 dedicated PRD judges executed in 2-phase execution, output to prd-judges.json, validated with --category prd
• New prd-auditor agent — structural completeness auditor for draft PRDs; checks US/AC coverage, success metrics table completeness, critical open questions, scope section structure, kill criteria presence, and template section inventory
• New prd-dependency-judge agent — evaluates PRD dependency completeness and risk assessment; flags missing dependencies, underdefined integration points, and unacknowledged cross-team risks
• New prd-testability-judge agent — evaluates whether PRD acceptance criteria are testable and measurable; flags vague or unverifiable criteria and missing success metrics
• New prd-scope-judge agent — evaluates PRD scope discipline and hypothesis traceability; flags stories with no traceable origin, out-of-scope overlaps, story count exceeding 8, and unacknowledged dependencies; emits review-delta JSON
• New prd_preamble.md in skills/artifact-type-tailored-context/preambles/ — artifact-type-tailored context preamble injected before PRD judge prompts
• validate_judge_report.py: Added prd category to JUDGE_REGISTRY with 4 expected judges (prd-auditor, prd-dependency-judge, prd-testability-judge, prd-scope-judge)
• validate_judge_report.py: Replaced valid_suffixes list with VALID_SUFFIXES dict mapping each category to its accepted report_id suffixes (prd maps to ["-prd-judges"])
• validate_judge_report.py: Reconciled JUDGE_REGISTRY plan set — removed phantom entries efficiency-judge and informativeness-relevance-judge; added brownfield-accuracy-judge, codebase-grounding-judge, and convention-adherence-judge
• judge-input.schema.json: Added "prd" to the evaluation_type enum

code v1.1.3

Added

• stream_formatter.py now accumulates per-model token usage from assistant events and prints a summary in the format the harness expects, fixing zero token counts for PLAN/EXECUTE loops

Fixed

• stream_formatter.py returns early on BrokenPipeError before printing usage summary, preventing tracebacks when used in pipelines with early-exit consumers

judges v1.2.0

Added

• New brownfield-accuracy-judge agent — evaluates how accurately a plan accounts for existing code (reuse vs reimplementation, integration-point accuracy, scope accuracy against investigation findings)
• New codebase-grounding-judge agent — detects hallucinated file paths, nonexistent modules, and fabricated APIs by comparing plan claims against the investigation log
• New convention-adherence-judge agent — evaluates whether a plan follows the conventions, patterns, and style found in the actual codebase as documented in the investigation log

Changed

• Updated run-judges skill to support 16 plan judges (up from 13), adding the three new grounding/brownfield/convention judges in Batch 4
• brownfield-accuracy-judge and convention-adherence-judge now invoke @code:pre-explorer to generate investigation-log.md when absent, instead of immediately scoring 0.5; fall back to 0.5 only if pre-explorer fails or the file remains absent
• codebase-grounding-judge: add validation step to ensure net-new code does not duplicate existing functionality (e.g., utilities/helpers already in codebase)

code v1.1.2

Fixed

• Restored boolean semantics for has_code_changes in run-loop.sh and updated judge gating to skip code judges when no implementation changes are detected, without relying on numeric stdout parsing

judges v1.1.0

Added

• New context-manager-for-judges agent (moved from code plugin) to orchestrate context compression for judge evaluation
• New judge-input.schema.json — formal JSON schema defining the standard judge input contract with source_of_truth field
• Investigation log (investigation-log.md) reuse in plan judge context with pre-explorer fallback when no CLOSEDLOOP_WORKDIR is set

Changed

• Generalized judge input contract to use orchestrator-provided judge-input.json (task + context envelope) instead of hardcoded artifact assumptions
• Standardized all judge agents to read judge-input.json from $CLOSEDLOOP_WORKDIR and load mapped artifacts via source-of-truth ordering
• Centralized judge input-read requirements into shared preamble common_input_preamble.md; judge-specific files no longer duplicate input-contract boilerplate
• Enforced strict SSOT by removing residual per-agent Input Contract stubs; common_input_preamble.md is now the single runtime source for input-loading guidance

Fixed

• Added source_of_truth to required array in judge-input.schema.json — schema now matches SKILL.md and judge agent expectations for evidence prioritization

code v1.1.0

Changed

• Migrated session/hook data directory from .claude/.closedloop/ to .closedloop-ai/ across all hooks (session-start, session-end, subagent-start, subagent-stop, pretooluse, loop-stop) and setup-closedloop.sh, with legacy fallback for mid-upgrade sessions
• Added legacy directory cleanup in session-end-hook.sh — removes stale PID mappings, expired session files, and deletes empty legacy directory on session end

self-learning v1.0.3

Fixed

• Fixed pattern cap trimming to sort by staleness flags only instead of confidence — low-confidence patterns were always dropped before being observed, preventing them from ever earning higher confidence
• Fixed extraneous f-string prefix lint warning in write_merged_patterns.py default header

Changed

• Updated process-learnings cap strategy to trim [PRUNE] then [STALE] then [REVIEW], with seen_count as tiebreaker

code v1.1.1

Added

• Integrated investigation-log.md into judge context assembly, sourced from $CLOSEDLOOP_WORKDIR

Fixed

• Fixed judges agents path resolution in run-loop.sh to support monorepo, cache, and marketplace installation layouts via a four-level fallback strategy (CLOSEDLOOP_JUDGES_AGENTS_DIR env override → repo-relative path → non-versioned sibling → latest semver-versioned sibling)
• Fixed agent snapshot to read judge agents from the judges plugin rather than the code plugin, and corrected plugin field in manifest to "judges"

code-review v1.1.0

Breaking

• Removed github-review slash command — /code-review:github-review is no longer a valid entry point. Use /code-review:start --github instead.
• Renamed review.md → start.md — slash command is now /code-review:start
• Moved github-review.md from commands/ to prompts/ — callers using ${CLAUDE_PLUGIN_ROOT}/commands/github-review.md must update to ${CLAUDE_PLUGIN_ROOT}/prompts/github-review.md

Changed

• Unified session directory path for all modes — removed $RUNNER_TEMP override in GitHub CI, now uses .closedloop-ai/code-review/cr-<RANDOM> everywhere
• Replaced Bash heredoc/cat usage with Write and Read tools for PR metadata file operations in github-review.md
• Updated temp file path references from $RUNNER_TEMP/cr-review/ to <CR_DIR>/* in GitHub mode constraints
• Fixed usage examples to use /start to match the command filename
• Fixed internal references from code-review-github.md to github-review.md

Added

• Compound Bash command prohibition in GitHub mode — no &&, ||, ;, or | pipes allowed

code v1.0.5

Changed

• Updated review-delta.schema.json description to reference "code hybrid workflow" instead of "impl-plan hybrid workflow"
• Updated compliance-checkpoint.md to reference /code instead of /impl-plan
• Removed Bash from visual-qa-subagent tool list to prevent shell access during visual QA

Security

• Added credential theft blocklist to pretooluse-hook.sh: denies Bash commands and file access targeting macOS Keychain, browser cookie databases, SSH private keys, and cloud credentials
• Blocklist applies to all Claude sessions, not just ClosedLoop-managed sessions

bootstrap v1.1.0

Added

• Schema-aligned constraints in AGENT_FORMAT.md: tools, skills, permissionMode fields, name kebab-case/64-char limit, description 1024-char limit, expanded 8-color enum with cyan/pink
• Context-engineering activation in agent-prompt-generator via platform:context-engineering skill
• Tools/skills inline format validation in agent-prompt-validator (BLOCKING on block array syntax)
• additionalProperties violation detection and skills→Skill tool cross-check
• Critic Review Schema Alignment (Check 8) and critic-gates.json Structure Validation (Check 9) in generation-validator
• critic-gates.json schema validation in bootstrap-validator
• Context-engineering compliance warnings in anti-pattern detection

Changed

• description max raised from 120 → 1024 chars (warn >200)
• model enum now accepts inherit
• color field changed from required to optional; enum expanded to 8 values
• Removed legacy prd2plan/ directory namespace — agent output now writes to .claude/agents/ (flat)
• Moved .bootstrap-metadata.json from .claude/agents/prd2plan/ to .closedloop-ai/bootstrap-metadata.json
• Replaced all /impl-plan command references with /code
• Removed DAG validation infrastructure (deleted impl-plan-dag.schema.json, removed Check 2 from bootstrap-validator)
• Updated default --target-command from impl-plan to code
• Updated default --output-dir from .claude/agents/prd2plan/ to .claude/agents/

code v1.0.4

Changed

• Generalized prd-creator skill description and replaced analytics discovery step with risks assessment
• Updated PRD template to add compliance checkpoint and remove event instrumentation section
• Revised story patterns and examples references to align with compliance-focused workflow

Removed

• Deleted event-instrumentation.md reference

code v1.0.3

Changed

• Migrated learnings path from ~/.claude/.learnings/ to ~/.closedloop-ai/learnings/ in pretooluse-hook.sh and subagent-start-hook.sh with legacy fallback

self-learning v1.0.2

Changed

• Migrated learnings path from ~/.claude/.learnings/ to ~/.closedloop-ai/learnings/ across commands, tools, and skills with legacy fallback

bootstrap v1.0.0

Added

• Initial release
• Bootstrap plugin for ClosedLoop agent creation and validation

code v1.0.2

Added

• Step 8.5 in run-loop.sh for deterministic TOON writing via write_merged_patterns.py

code v1.0.1

Added

• New prd-creator skill for drafting lightweight PRDs through conversational workflow

code v1.0.0

Added

• Initial release

code-review v1.0.0

Added

• Initial release

judges v1.0.0

Added

• Initial release

platform v1.0.1

Added

• New claude-creator skill for scaffolding and creating new skills from scratch

platform v1.0.0

Added

• Initial release

self-learning v1.0.1

Added

• New write_merged_patterns.py tool for deterministic JSON-to-TOON conversion

Changed

• Refactored process-learnings command to output merge-result.json instead of writing TOON directly
• Updated process-chat-learnings.sh to run deterministic TOON write step after classification

self-learning v1.0.0

Added

• Initial release