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]

bootstrap v1.2.0

Changed

• Migrated critic-gates configuration path from .claude/settings/critic-gates.json to .closedloop-ai/settings/critic-gates.json across agent-decomposer, agent-prompt-validator, generation-validator, and agent-bootstrap command
• Migrated schema validation path from .claude/schemas/ to .closedloop-ai/schemas/ in agent-prompt-validator
• Updated agent output path references from .claude/runs/ to .closedloop-ai/runs/ in agent-prompt-generator
• Updated bootstrap configuration documentation in agent-bootstrap.md to reference .closedloop-ai/ state directory

code v1.7.0

Added

• Multi-repo planning and exploration support via new --add-dir flag in run-loop.sh, exposing CLOSEDLOOP_ADD_DIRS and CLOSEDLOOP_REPO_MAP env vars to downstream agents
• pre-explorer agent produces per-repo code maps (code-map-{name}.json) when secondary repos are supplied
• plan-draft-writer agent emits multi-repo plans with a ## Repositories table and @{repo}:path task prefixes
• New multi-repo.overlay.md overlay assembled onto prompt.md at runtime for cross-repository planning workflows
• repositories map field added to the plan root schema in plan-schema.json for multi-repo plan traceability, keyed by repo short-name with path and isPrimary metadata
• Tier 0 explicit-directory discovery and dedup helpers in discover-repos.sh, with structured JSON output
• Enhancements to cross-repo-coordinator and cross-repo-prd-writer agents for multi-repo context
• Tests for discover-repos.sh and setup-closedloop.sh (test_discover_repos.py, test_setup_closedloop.py) plus new multi-repo cases in test_validate_plan.py

Fixed

• run-loop.sh now scans the full per-iteration stream for the <promise> completion marker instead of only inspecting the final type==result record, preventing missed completion signals when the orchestrator emits the promise in an intermediate message followed by additional tool_use or wrap-up output
• discover-repos.sh now filters add-dirs that are ancestors of the workdir and deduplicates repo entries to prevent duplicate discovery results
• Fixed multi-repo execution ordering and prompt definition string assembly in overlay resolution

Changed

• Consolidated Tier 0 discover-repos.sh tests into a single scenario-driven harness, replacing the prior fragmented per-case test files
• Migrated workdir internal state directory from .closedloop/ to .closedloop-ai/ across hooks, setup scripts, and loop state management
• Established CLOSEDLOOP_STATE_DIR constant as single source of truth for state directory name across shell scripts
• Added Skill to plan-evaluator agent's allowed tools to enable code:plan-validate skill execution

code v1.6.0

Changed

• Migrated all remaining .claude/ path references to .closedloop-ai/ across hooks, scripts, agents, skills, and orchestrator prompt -- completes the directory migration started in v1.1.0
• Replaced gawk FPAT-based TOON parser with portable csv_split() function in pretooluse-hook.sh and subagent-start-hook.sh, removing the hard dependency on GNU awk
• Refactored awk array usage from associative patterns[n]["key"] to parallel flat arrays for POSIX awk compatibility
• Updated install-dependencies.sh to verify any awk instead of requiring gawk with FPAT support
• Updated org learnings copy path in run-loop.sh to use .closedloop-ai/learnings/ with workdir-adjacent state directory resolution

Removed

• Removed all legacy .claude/.closedloop/ session/workdir/env fallback paths from loop-stop-hook.sh, pretooluse-hook.sh, session-end-hook.sh, subagent-start-hook.sh, subagent-stop-hook.sh, and setup-closedloop.sh
• Removed legacy ~/.claude/.learnings/org-patterns.toon fallback from pretooluse-hook.sh and subagent-start-hook.sh
• Removed legacy cleanup logic from session-end-hook.sh (PID cleanup, stale session removal, legacy directory deletion)

Added

• Tests for legacy path ignorance in pretooluse and subagent-start hooks, setup-closedloop, and self-learning flag tests
• Tests for portable awk injection (test_injects_when_only_plain_awk_is_available) in both hook test suites

code-review v1.4.0

Changed

• Migrated GitHub mode output file paths from .claude/ to .closedloop-ai/: code-review-findings.json, code-review-threads.json, and code-review-summary.md
• Updated route subcommand to read critic-gates from .closedloop-ai/settings/critic-gates.json
• Simplified fast-path routing to total_loc <= 200 threshold only (was <= 150 LOC AND <= 5 files AND no domain critics); domain critics are now folded into the fast-path agent as an additional pass

Added

• Structured reasoning protocol for Premise Reviewer: AUTHOR'S CLAIM / COUNTER-EVIDENCE / ALTERNATIVE CHECK / CONCLUSION validation gate before reporting premise findings
• Reasoning certificate for Bug Hunter A: PREMISE / TRACE / DIVERGENCE / GUARD CHECK / CONCLUSION trace-based bug confirmation gate with emission filtering
• Domain critic pass injection in fast-path reviewer via {DOMAIN_CRITIC_PASS} placeholder, enabling domain expert review within single-agent fast-path runs
• Replaced shared prompt reasoning checklist with structured PREMISE / EVIDENCE / GUARD CHECK / SEVERITY CHECK analysis framework

judges v1.5.1

Changed

• Migrated perf-substep state paths from .closedloop/ to .closedloop-ai/ in run-judges skill telemetry instrumentation

judges v1.5.0

Changed

• Migrated threshold override paths from .claude/settings/threshold-overrides.json to .closedloop-ai/settings/threshold-overrides.json in run-judges skill (both run-specific and repo-level locations)

platform v1.1.0

Changed

• Version bump to align with cross-plugin .closedloop-ai/ directory migration

self-learning v1.1.1

Changed

• Established CLOSEDLOOP_STATE_DIR constant as single source of truth for state directory name in bootstrap-learnings.sh, compute_success_rates.py, and write_merged_patterns.py

self-learning v1.1.0

Changed

• Migrated org learnings paths from .claude/learnings/ to .closedloop-ai/learnings/ across pull-learnings, push-learnings, and bootstrap-learnings.sh
• Migrated run path references from .claude/runs/ to .closedloop-ai/runs/ in process-learnings command
• Simplified preflight-check.sh to verify awk availability instead of requiring gawk with FPAT support

Removed

• Removed legacy ~/.claude/.learnings/org-patterns.toon fallback from compute_success_rates.py and write_merged_patterns.py
• Removed legacy session file lookup path from evaluate_goal.py

Added

• Test verifying CLI ignores legacy home TOON path in test_compute_success_rates.py

code v1.5.10

Changed

• Enhanced plan-agent with verification-before-proposing requirements: must Read every function, type, and validator before writing tasks that modify them; must check receiving validators/schemas when tasks construct events or payloads
• plan-agent now requires explicit task dependency declarations ("Depends on T-A.B"), null/empty/missing edge case specification for every new field, and accurate summary language (no overclaiming)
• Added multi-repository plan guidelines to plan-agent: absolute file paths for cross-repo references, per-repo file existence verification, repo labels on tasks, and cross-repo contract documentation
• Added self-check gates to plan-agent: modification targets verified, validators audited, edge cases specified, dependencies declared, summary accuracy confirmed -- with a concrete good-vs-bad task example

code v1.5.9

Fixed

• stream_formatter.py now uses Optional[str] instead of str | None union syntax, making it import-safe on Python 3.9 and preventing silent JSONL pipeline truncation on macOS systems using the default system Python

code v1.5.8

Removed

• Deleted feedback-explorer agent and removed its integration from plan-with-codex debate loop -- plan-agent now receives feedback directly without pre-fetched context briefs
• Removed {stem}.context sidecar file from plan-with-codex debate loop

Changed

• Updated default Codex model from gpt-5.4 to gpt-5.3-codex in plan-with-codex command and debate-loop.sh (completes model migration started in v1.5.5)
• Reduced Codex reasoning effort from xhigh to high in run_codex_review.sh

code v1.5.7

Added

• Ghost loop detection in run-loop.sh -- tracks consecutive empty iterations and aborts after 3 to prevent infinite loops with no output
• Session/context limit detection from is_error flag in Claude JSONL result records, with immediate abort and context_limit run log entry
• Session/context limit detection from stderr pattern matching (prompt is too long, context limit reached, etc.), with immediate abort on non-zero exit

code-review v1.3.0

Added

• PR auto-detection in local mode: when the current branch has an open PR, resolve-scope now auto-detects it via gh pr view and scopes the review to the PR diff instead of main...HEAD
• Small-diff fast path: diffs with <=150 LOC and <=5 files now route to a single fast-path reviewer agent instead of spawning the full 5-agent fleet, reducing review time and token usage
• Fast-path reviewer performs three scoped passes (Bug Hunter, Bug Hunter B / Unified Auditor, Premise) in a single agent run
• Partition cap enforcement with unconditional force-merge fallback when budget-respecting merges cannot reduce partition count below the cap

Changed

• Deferred cache-status printing from Task 6 to Task 8 (standard flow) or Task 7 (hygiene-only exit) to allow fast-path routing to suppress cache output
• extract-patches --partitions-file is now optional; omitting it produces only patches_all.txt
• Reviewer/model routing lines in local output and GitHub summary are now conditional on fast_path
• Footer omits --cache-result on fast-path runs (cache intentionally bypassed)
• Renamed Step 4 to Step 4A (standard flow) and added Step 4B (fast-path flow); Step 5.5 now gated on fast_path == false

code v1.5.6

Added

• Severity gate for Codex debate rounds 5+ in run_codex_review.sh -- only flags findings that would cause functionally wrong behavior (incorrect output, data loss, crashes, security holes); suppresses wording ambiguities, hypothetical misimplementations, and style suggestions

Changed

• Split Codex debate round handling into three tiers: round 1 (initial review), rounds 2-4 (standard re-review), rounds 5+ (severity-gated re-review with elevated approval bar)
• Codex responses with no verdict AND no findings now emit CODEX_EMPTY instead of defaulting to NEEDS_CHANGES, distinguishing truncated/empty responses from genuine review feedback

code v1.5.5

Changed

• Updated default Codex model from gpt-5.4 to gpt-5.3-codex in codex-review skill parameter docs and run_codex_review.sh default
• Migrated remaining .claude/work path references to .closedloop-ai/work in orchestrator prompt example and extract-plan-md skill usage examples

code v1.5.4

Removed

• Removed self-learning write references from agent prompts: implementation-subagent, plan-importer, plan-writer, plan-draft-writer, generic-discovery, cross-repo-coordinator, build-validator, verification-subagent, plan-validator, code-reviewer -- learning capture sections, Organization Learnings sections, and self-learning:learning-quality skill references
• Deleted learning prompt files: plan-writer-learning.md, implementation-learning.md, discovery-learning.md

code v1.5.3

Changed

• Migrated work directory paths from .claude/ to .closedloop-ai/ across run-loop.sh (state file, progress log, directory creation), amend-plan command (default workdir), and cancel-code command (loop state file path)
• Enhanced codex-review prompt with 6 new analysis criteria: canonical state preservation, task specificity, behavioral precision, order-of-operations, lifecycle symmetry, and test fidelity -- plus implementability-focused preamble instructions

code v1.5.2

Added

• Rule 8 in build-validator agent: never use pkill, killall, or broad kill patterns — use timeout to bound hung commands and report stuck processes as failures instead of killing them

Security

• Added pkill and killall to credential-theft blocklist in pretooluse-hook.sh — broad process killing is now globally denied to prevent worktree agents from killing processes outside their context

self-learning v1.0.4

Changed

• Migrated .claude/work path reference to .closedloop-ai/work in process-chat-learnings.sh usage documentation

[Releases]

code v1.5.1

Removed

• Removed judge integration from run-loop.sh — run_judges_if_needed, has_code_changes, resolve_judges_agents_dir, ensure_agents_snapshot, store_agents_snapshot, and check_completion functions removed along with Step 11 judge invocation in post_iteration_processing
• Deleted run_judges_test_helper.sh and test_run_loop_imported_plan.py (tests for removed judge functions)

Changed

• Refactored run-loop.sh workdir references to use a single effective_workdir local variable instead of repeated ${workdir:-$WORKDIR} expansions

judges v1.4.0

Added

• Agents snapshot pre-step in run-judges skill — creates $CLOSEDLOOP_WORKDIR/agents-snapshot/ with all judge agent .md files and a manifest.json before judge execution begins (skipped if snapshot already exists)
• New ensure_agents_snapshot.sh script in run-judges skill scripts

Changed

• Renamed plan evaluation output from judges.json to plan-judges.json for consistency with code-judges.json and prd-judges.json
• Updated validate_judge_report.py default filename for plan category to plan-judges.json

code v1.5.0

Added

• --self-learning opt-in flag for run-loop.sh -- self-learning is now disabled by default
• CLOSEDLOOP_SELF_LEARNING config propagation via config.env and state frontmatter
• Self-learning guard in subagent-start-hook.sh to skip learning injection when disabled
• Self-learning guard in subagent-stop-hook.sh to skip entire learning region when disabled
• Self-learning guard in pretooluse-hook.sh to skip tool-specific pattern injection when disabled

Changed

• post_iteration_processing() skips steps 2-10 when self-learning is off; step 1 (changed-files.json) and step 11 (judges) always run
• bootstrap_learnings() skips .learnings/ directory creation when self-learning is off
• run_background_pruning() skips pruning when self-learning is off
• Resume restores SELF_LEARNING from state frontmatter and re-exports to hooks

code v1.4.1

Added

• New feedback-explorer agent (haiku) for pre-fetching codebase context referenced in reviewer feedback, reducing redundant exploration during plan revisions with delta caching across debate rounds
• Deferral detection in plan-with-codex -- scans plans for "Deferred", "Out of Scope", "Future Work" items and requires explicit user approval before excluding work from scope
• Exclusions sidecar file ({stem}.exclusions) in plan-with-codex to persist user-confirmed deferral decisions across debate rounds

Changed

• plan-with-codex argument-hint updated to positional syntax instead of optional bracket notation
• plan-with-codex uses Write tool for state persistence instead of Bash printf
• plan-with-codex launches feedback-explorer before plan-agent revision rounds to pre-fetch context
• plan-agent enforces "no silent deferrals" rule -- must not create deferred/out-of-scope sections without explicit user approval
• plan-agent supports pre-fetched context briefs from feedback-explorer, reads brief before revision to skip redundant exploration
• Added Write tool to plan-agent tools list

Fixed

• Fixed plan-with-codex to use fully qualified agent name code:feedback-explorer

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