v8.0.2 — Milestone 8 Archive & Core Fixes

Changed

• Commit Lifecycle Rule (ZAC) in system-prompt <constraints> — Added a <commit_lifecycle_rule> bullet to the top-level <constraints> section documenting the two commit-producing MCP tools (stage_and_inject_diff for development-time, commit_and_clean_task for closure-time), their distinct lifecycle semantics, the two-commit flow (feature + closure), and ZAC enforcement. Previously the ZAC intent was only visible in <bash_phase> implementation templates, allowing LLM agents to invoke commit_and_clean_task during implementation (Zen Router incident, Task 13). System prompt version bumped to 8.0.1 (PATCH).
• Milestone 8 archive — Compacted Tasks 69, 77–79 into docs/history/milestone-8-summary.md and moved completed task files to tasks/archive/. System prompt version bumped to 8.0.2.

Fixed

• Orphaned commit hash bug in commit_and_clean_task — The tool captured the commit hash before git commit --amend, so the hash stored in the task file pointed to a commit that became unreachable after the amend replaced it. Reworked the tool to a two-commit flow: the feature commit hash is captured and stored in the task file, then the cleaned task file is committed as a separate chore: close task N closure commit. The stored hash is now permanently reachable from HEAD, git show <hash> returns the real code diff, and no amend/orphaned commits are produced. The idempotency guard matches the exact cleaned-block structure (regex), so a raw injected diff that merely mentions "Stored in Commit Hash" (e.g. this very changelog entry or the guard's own source line) cannot false-positive and block a legitimate closure. Regression tests: test_commit_and_clean_task_stores_reachable_hash, test_commit_and_clean_task_guard_no_false_positive_on_diff_mention.
• stage_and_inject_diff crash when an ignored context-reports/ directory exists — the tool staged with git add . :!... negative pathspecs, which makes git exit 1 with "paths are ignored" whenever an excluded path actually exists on disk (the accumulated context-reports/ reports), blocking every closure until deleted. Replaced with plain git add -A . (gitignore-respected) plus defense-in-depth git reset -q -- <pattern> for the sensitive/ignored paths. Regression test: test_stage_and_inject_diff_with_ignored_context_reports.

v8.0.0 -- Cognitive Lead AI V8 Ecosystem Overhaul

[8.0.0] - 2026-08-04

Added

• P1 Quality Improvements (V8.0.0 Phase 5) — Added tests/test_mcp_servers.py for basic MCP server import and logic validation. Enhanced mcp-memory-server with YAML frontmatter support (pyyaml) for metadata tracking and improved search_memory with tag filtering and ranking. Created docs/system-prompt-modularization.md design document for V9.0.0 planning. Documented tree-sitter regex fallback for Swift, Ruby, PHP, and C# in code-search skill. System prompt version bumped to 7.5.0.
• New Lint MCP Server & Skill (V8.0.0 Phase 3) — Created mcp-lint-server/server.py providing lint_markdown, lint_task_file, and lint_all_tasks tools for structural validation. Registered server in opencode.json and LLM.txt global configs. Created task-lint skill template. Added task-lint to <agent_skills_registry> in system-prompt.md. System prompt version bumped to 7.4.0.

Changed

• Milestone 7 archive — Compacted tasks 70–76 into docs/history/milestone-7-summary.md and moved completed task files to tasks/archive/. System prompt version bumped to 8.0.0.
• Task Template Ecosystem Enforcement (V8.0.0 Phase 4 Revised) — Enforced new task template sections across the entire ecosystem. Updated mcp-lint-server to mandate ## Acceptance Criteria, ## Verification Evidence, and ## Risk & Rollback. Instructed telegram-issue-sync to populate acceptance criteria from message intent. Enhanced archive-tasks milestone summaries to extract and report criteria met. Added CRITICAL RULE 6 (Evidence Capture) to <bash_phase> in system-prompt.md to force verification evidence logging before summary. System prompt version bumped to 7.4.2.
• Task Template Enhancement & Lint Integration (V8.0.0 Phase 4) — Enhanced task-generator skill template with mandatory ## Acceptance Criteria, ## Verification Evidence, and ## Risk & Rollback sections for both unified and multi-phase task files. Integrated lint_task_file MCP tool into the <summary_phase> of both <opencode_implementation_task_template> and <opencode_combined_task_template> to enforce structural validation before diff injection. System prompt version bumped to 7.4.1.
• Input Processing Pipeline Enhancement (V8.0.0 Phase 1) — Enhanced <user_input_processing> in system-prompt.md with mandatory Input Validation Gate (Step 0.5), enriched Intent Expansion, and Prompt Refactor Gate (Step 5.5). Enhanced prompt-refactor skill with Step 0 validation and typo correction. Updated AGENTS.md guardrail to enforce Input Validation Pipeline. Created user-prompts/input-validation-test.md. System prompt version bumped to 7.3.0.

Fixed

• Fixed search_memory tag-only query edge case — When querying with only tag:xxx (no additional search terms), the function now correctly returns all files matching the tag filter instead of returning zero results. Previously, the fallback content matching would search for the literal string "tag:xxx" in file content, which would never match.
• P0 Consistency & Safety Fixes (V8.0.0 Phase 2) — Resolved AGENTS.md documentation-only contradiction by adding explicit exceptions for MCP servers and tooling. Hardened stage_and_inject_diff to exclude sensitive files (.env, .pem, etc.) from blind git add .. Hardened commit_and_clean_task with empty-staged checks and push-history amend warnings. Fixed version sync rules in versioning-and-release skill. Resolved DESIGN.md path conflict (root vs .stitch/). Converted archive-tasks to use git mv for history preservation. Secured memory deletion by changing delete_memory permission to ask in opencode.json and LLM.txt, and adding a Safety Gate to the project-memory skill.
• MCP servers crash on startup with MCP SDK 2.0 — Pinned mcp[cli]>=1.0,<2.0 in the # /// script dependency headers of mcp-context-server/server.py and mcp-memory-server/server.py. PyPI's latest mcp (2.0.0) removed mcp.server.fastmcp, causing ModuleNotFoundError on boot and disabling both custom_context and project_memory tools.

v7.2.1 — Milestone 6 Archive & Unified Task Template

What's New in v7.2.1

Release covering the 7.1.0 → 7.2.1 cycle: workflow friction fixes, V7 documentation consistency, the unified task template, and the Milestone 6 archive.

Unified Task Template (7.2.0)

• Unified canonical task-file template in task-generator with a polymorphic ## Source Context section supporting three provenances: orchestrator, telegram, manager.
• Source: metadata field in the task header for provenance tracking.
• Mandatory ## Goal and ## Local TODOs for all task files; # Task [NN]: [Title] title convention.
• telegram-issue-sync now references the unified template instead of defining an inline copy.
• archive-tasks milestone summaries now include a ## Source Distribution table.

Workflow Friction Fixes (7.1.0)

• Combined Discovery+Plan workflow — new <opencode_combined_task> template reduces Manager round-trips from 6 to 3, with explicit halt-on-architecture-mismatch conditions.
• Task numbering collision protection — Orchestrator-side pre-assignment validation plus a Collision Check in task-generator.
• Multi-phase task rule — single task file with inline ## Phase N: sections instead of separate subtask files.
• CHANGELOG Parse-Then-Append Protocol — prevents duplicate version/section headers.
• Memory hygiene — Memory Validation step in archive-tasks and Supersession Detection heuristic in project-memory.
• Pre-Commit Verification Gate (CRITICAL RULE 5) for DevOps/infrastructure tasks.
• Topic Shift Detection in <user_input_processing> for proactive context-switch notices.

Documentation Consistency (7.1.1)

• README now reflects V7: tree label, ## Key V7 Changes section, complete 28-skill skill-templates/ tree with standalone group labels.
• <execution_workflow> renumbered 1–9 (Phase 0 identity preserved), tags on own lines.

Milestone 6 Archive (7.2.1)

• Tasks 66–68 compacted into docs/history/milestone-6-summary.md and moved to tasks/archive/.
• System prompt version bumped to 7.2.1.

Full changelog: see CHANGELOG.md.

v7.0.1 — Milestone 5 Archive

Changed\n\n- Milestone 5 archive — Compacted task 65 into docs/history/milestone-5-summary.md and moved completed task file to tasks/archive/.\n\n### Full Changelog\n\nv7.0.0: Platform-Agnostic Rebrand — Removed all hardcoded references to Google AI Studio and Gemini. The workflow is now vendor-neutral.

v6.12.0 — audit-agents conventions governance

[6.12.0] — 2026-07-23

Added

• audit-agents conventions.md governance — The audit-agents skill now auto-generates and audits docs/conventions.md with Universal DateTime Standard and SOLID Programming Guidelines. Mode 1 (Phase 0) generates conventions.md alongside AGENTS.md. Mode 2 audits and patches conventions.md if missing or incomplete. Agent Audit Summary expanded with conventions.md compliance status.

Changed

• Milestone 4 archive — Compacted tasks 56–64 into docs/history/milestone-4-summary.md and moved completed task files to tasks/archive/.
• System prompt upgraded to V6.12.0 — <system_version> bumped from 6.11.0 to 6.12.0.

v6.5.1 — Skill Name Standardization & Milestone 3 Archive

What's Changed

Fixed

• Skill name standardization — Aligned YAML name: fields across all 27 skill templates to match folder names exactly. Removed architecture-prefixed names (backend-architecture-*, frontend-architecture-*, mobile-architecture-*, stitch::extract-design-md) in favor of clean short names.
• System prompt consistency — Added missing archive-tasks and migrate-kanban to the skills registry. Aligned all 27 skill descriptions verbatim with their YAML counterparts.

Changed

• Milestone 3 archive — Compacted tasks 49–55 into docs/history/milestone-3-summary.md and moved completed task files to tasks/archive/.

Full Changelog: v6.5.0...v6.5.1

v6.4.0 — Project Memory System

[6.4.0] — 2026-07-16

Added

• Project Memory MCP Server (mcp-memory-server/server.py) — new FastMCP server providing persistent, project-scoped memory via atomic-write markdown files under .opencode/memory/. Four tools: store_memory (with atomic writes using tempfile + os.replace), read_memory, search_memory (full-text across namespaces), and list_namespaces.
• project-memory Agent Skill (skill-templates/project-memory/SKILL.md) — companion skill teaching OpenCode when to proactively store Manager constraints and when to retrieve them during the Context Phase.
• opencode.json registration — project_memory MCP server registered alongside custom_context, with explicit permissions for store_memory, read_memory, search_memory, and list_namespaces.
• System prompt integration — project-memory added to <agent_skills_registry> under Global Workflow Skills. Software Architect and Senior Programmer personas updated to proactively save Manager constraints via the project-memory skill.

Changed

• System prompt upgraded to V6.4.0 — <system_version> bumped.
• audit-agents skill — Target Audit Criteria (Mode 1 and Mode 2) and AGENTS.md Template updated with Context Bootstrapping rule: agents must call search_memory/list_namespaces at task start.
• README.md — Roadmap item #7 (Memory Management) struck through and marked implemented in V6.4.0.

Security

• Path traversal prevention — mcp-memory-server/server.py: added _validate_and_resolve with regex ^[a-zA-Z0-9_-]+$ sanitization and is_relative_to boundary checks on all user-supplied namespace/key inputs.
• Temp file leak fix — store_memory atomic write wrapped in try/except with explicit os.unlink(temp_path) cleanup on failure.

v5.19.0

What's Changed

Added

• OpenCode docs mirror — all 36 documentation pages from opencode.ai/docs fetched into docs/opencode/
• LLM.txt — AI agent auto-configuration manual with platform-specific OpenCode setup
• docs/conventions.md — centralized syntax rules and automation conventions

Changed

• Rebranded repository to cognitive-lead-hq with updated metadata and polished README
• Enforced --body-file pattern for all gh issue create commands
• README Quick Start now references LLM.txt as canonical auto-setup source