@jaguilar87/gaia 5.0.0-rc.2

package/CHANGELOG.md

@@ -0,0 +1,1298 @@
+# Changelog: CLAUDE.md
+
+All notable changes to the gaia-ops orchestration system are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [5.0.0-rc.2] - 2026-04-26
+
+### Release Candidate 2: Converger Identity, Session Liveness, Install-Gate Hardening
+
+Second release candidate for v5.0.0. Adds the orchestrator's Converger
+("Cerrar") conversational closure identity, real-PID session liveness in the
+registry, the `agent-creation` and `session-reflection` skills, and an
+end-to-end consumer-install validation harness that now actually exercises the
+gate. Three install-time bugs surfaced and were fixed alongside the harness
+that found them.
+
+#### Added
+- **Converger identity for orchestrator** — "Cerrar" conversational closure
+  framing. Brief-spec reframed as closure ritual (Size gate removed),
+  `planning_specs` surface routing narrowed to explicit artifact keywords,
+  architecture docs aligned with closure framing.
+- **session-reflection skill** — conversational session-close ritual. Surfaced
+  by orchestrator at session end; complements `gaia-compact`.
+- **agent-creation skill** — coach skill for designing new agents end-to-end:
+  identity, tool surface, contract, and verification.
+- **SessionEnd hook + PID liveness** — `session_end_hook.py` for clean
+  unregister; session_registry now uses real PID + `/proc` starttime to detect
+  liveness across sessions. `Stop` hook no longer mutates the registry (was
+  causing premature unregister mid-conversation).
+- **validate-sandbox.sh** — end-to-end consumer-install verification harness.
+  Two targets: `--target sandbox` (ephemeral fixture project) and
+  `--target local` (real workspace install with `--workspace` override). Eight
+  pass/fail checks: version, doctor, status, context show, memory stats,
+  memory search, scan, settings preservation. Wired into `publish.yml` so
+  every release smoke-tests the published tarball before notifying success.
+- **`gaia:verify-install:{local,rc,latest}` and `gaia:install-local`** scripts
+  in package.json for manual local validation against tarballs or registry.
+
+#### Changed
+- **REVIEW → APPROVAL_REQUEST** rename across active doctrine (state machine,
+  skills, hooks). Comments and references in `hooks/**` updated. The previous
+  `REVIEW` state caused confusion with the human review activity; the new name
+  reflects what the state actually represents (an agent requesting human
+  approval for a specific T3 operation).
+- **Stop hook decoupled from registry** — Stop event no longer mutates
+  session_registry. SessionEnd handles unregister cleanly; this avoids the
+  Stop-then-resume race where the registry would drop a still-active session.
+- **`publish.yml`** — sandbox harness step added after npm publish; waits for
+  registry propagation, then runs validate-sandbox.sh against the freshly
+  published tarball as a smoke test.
+
+#### Fixed
+- **Sandbox harness on noexec /tmp** — validate-sandbox.sh now detects
+  `noexec` mounts via `findmnt` (with `/proc/mounts` fallback) and falls back
+  to `$TMPDIR` → `/tmp` → `$HOME/.cache/gaia-sandbox`. Previously the harness
+  was unrunnable on WSL/Linux setups with `noexec /tmp` (rc=126 Permission
+  denied on the installed bin shims); the gate appeared to validate but never
+  actually ran.
+- **`gaia scan` harness check** — was invoking bare `gaia-scan --dry-run`,
+  which routes to `gaia-scan.py` whose argparse rejects `--dry-run`. Now uses
+  `gaia context scan --dry-run` (the higher-level CLI subcommand that does
+  accept `--dry-run`); drops the dead fallback.
+- **doctor `<lambda>` check** — `cmd_doctor` wrapped each check in a bare
+  `lambda`, so any exception surfaced as `'<lambda>'` in the JSON output
+  hiding which check actually failed. Replaced with `functools.partial` so
+  `__name__` resolves to the wrapped function (e.g. `check_project_dirs`).
+- **doctor `check_project_dirs` PosixPath/list TypeError** — code did
+  `project_root / dir_path` while iterating `paths.items()`; when a value was
+  a list (e.g. `"scan_targets": ["."]`), `Path / list` raised TypeError.
+  Values are now normalized to a flat sequence of `(label, str)` pairs before
+  joining; list values expand into `label[0]`, `label[1]`, ...
+- **postinstall FTS5 backfill on fresh install** — `maybeBackfillFts5()`
+  returned early when `search.db` was missing with comment "doctor --fix will
+  create it on first use", but nothing in the install flow runs `doctor --fix`
+  automatically. A consumer reinstalling after `gaia uninstall` (which scrubs
+  search.db) would have an empty FTS5 index until manual intervention. The
+  early return is gone; missing search.db now falls through to `doctor --fix`
+  which creates and populates the index.
+- **postinstall dynamic package resolution** — `gaia-update.js` now resolves
+  the gaia package name from `node_modules/@jaguilar87/` instead of
+  hardcoding, supporting both the v5+ `gaia` name and legacy `gaia-ops`. Also
+  detects and repairs symlinks pointing at the legacy path.
+- **memory sentinel return** — sentinel value returned with a surfaced warning
+  instead of a silent failure when memory paths fail to resolve.
+
+#### Internal
+- **Regenerated plugin artifacts** — `dist/gaia-ops/` and `dist/gaia-security/`
+  rebuilt for rc2.
+- **Cross-session liveness test** — real PID isolation in
+  `session_registry` test fixtures.
+
+## [5.0.0-rc1] - 2026-04-21
+
+### Release Candidate: Context Evals, Planner M1-M6, Memory CLI, Security Hardening
+
+First release candidate for v5.0.0. Consolidates the agentic-loop evaluation
+framework, the closed gaia-planner milestones, the unified `gaia memory` CLI,
+and a round of security hardening covering approval lifecycle, Gmail policy,
+and session compaction.
+
+#### Added
+- **Context-evals framework** — full pytest-driven evaluation suite for agent
+  context consumption. 5 graders (code, contract, trace, routing,
+  skill-injection), 3 backends (static, headless, live), 10 scenarios in
+  catalog, baseline snapshot with drift detection, and reporter for CI-friendly
+  output. Tests under `tests/evals/` with `baseline.json` tracked and
+  `{timestamp}-smoke.json` gitignored.
+- **gaia-planner M1-M6 closed** — brief-spec + gaia-planner agent pipeline
+  end-to-end. Includes plan state machine, REVIEW -> APPROVAL_REQUEST split,
+  session_registry liveness filter, and approvals-drift-fix closed 2026-04-20.
+- **gaia memory CLI** — `python3 bin/gaia memory` subcommand with search
+  (`gaia memory search`), episode inspection (`gaia memory show <id>`), FTS5
+  full-text index, scoring overhaul, and session context orientation.
+- **gaia-compact skill** — structured session compaction preserving decisions,
+  components, gaps, file map, and next steps. Invoked via `/compact` or
+  orchestrator-level "compacta" triggers.
+- **tools/__init__.py** — namespace marker for pytest rootdir parity. Resolves
+  8 collection errors when running full suite (tests goes to 3702 passed,
+  36 skipped, 0 errors).
+
+#### Changed
+- **Gmail policy** — macro-prefix fix: `+` in label prefixes now correctly
+  strips before state-machine classification. Reply classified as mutative
+  (was previously read-only, causing false negatives in T3 flow).
+- **Approval workflow docs** — documented that `permissionMode` does not
+  survive SendMessage resume. Subagents emitting APPROVAL_REQUEST mid-task
+  require orchestrator to re-dispatch fresh (mode does not inherit on resume).
+- **Package version** — `package.json` aligned with `pyproject.toml` at
+  `5.0.0-rc1` (previously drifted at `5.0.0-beta.9`).
+
+#### Fixed
+- **pytest collection** — `tools/__init__.py` prevents rootdir walk-up mismatch
+  between `tests/` and `tools/scan/tests/`. Full suite now collects cleanly.
+- **Evals smoke JSONs** — transient artifacts no longer tracked in git;
+  `tests/evals/results/*-smoke.json` gitignored, `baseline.json` preserved.
+
+### Unified Python CLI + JS CLI Deprecation (inherited from beta cycle)
+
+The JS CLIs (`gaia-status`, `gaia-doctor`, `gaia-cleanup`, `gaia-update`, `gaia-history`, `gaia-metrics`) are now deprecated in favor of the unified `bin/gaia` Python CLI. The JS CLIs remain functional but print deprecation warnings to stderr on every invocation.
+
+#### Migration: Old Command → New Command
+
+| Old JS command | New unified command |
+|---|---|
+| `npx gaia-status` | `python3 bin/gaia status` |
+| `npx gaia-doctor` | `python3 bin/gaia doctor` |
+| `npx gaia-cleanup` | `python3 bin/gaia cleanup` |
+| `npx gaia-update` | `python3 bin/gaia update` |
+| `npx gaia-history` | `python3 bin/gaia history` |
+| `npx gaia-metrics` | `python3 bin/gaia metrics` |
+
+#### New commands with no JS equivalent
+
+The unified CLI also provides subcommands that did not exist as standalone JS CLIs:
+
+| New command | Description |
+|---|---|
+| `python3 bin/gaia approvals list` | List pending T3 approval requests |
+| `python3 bin/gaia approvals show APPROVAL_ID` | Show approval detail |
+| `python3 bin/gaia approvals reject NONCE` | Reject a pending approval |
+| `python3 bin/gaia approvals clean` | Remove expired grants |
+| `python3 bin/gaia approvals stats` | Show approval statistics |
+| `python3 bin/gaia plans list` | List all feature briefs |
+| `python3 bin/gaia plans show BRIEF_NAME` | Show a brief and plan |
+| `python3 bin/gaia context show` | Display project-context.json summary |
+| `python3 bin/gaia context scan` | Invoke gaia-scan to refresh context |
+
+#### Deprecation timeline
+
+- **Now (M6):** JS CLIs print `[DEPRECATED]` warnings to stderr. All functionality remains intact.
+- **Future version (TBD):** JS CLIs will be removed from `package.json` bin field.
+
+#### Why a unified CLI?
+
+- Zero external dependencies (stdlib only, Python 3.9+)
+- Single entry point: `bin/gaia --help` for all subcommands
+- Machine-readable `--json` output on all subcommands
+- Consistent exit codes: 0=ok, 1=warnings, 2=errors
+- Extensible: add subcommands by dropping a `bin/cli/<name>.py` file
+
+---
+
+## [4.5.0] - 2026-03-24
+
+### Settings Architecture Redesign + Multi-Cloud Security
+
+Unified approach for permissions across NPM and plugin installation modes. Permissions now live in `settings.local.json` (union merge, preserves user config). `settings.json` contains only hooks.
+
+#### Added
+- **Azure deny rules** — 39 rules covering resource groups, networking, AKS, Key Vault, CosmosDB, Service Bus, and more
+- **Generic wildcard deny rules** — 20 rules that catch all present and future cloud services (`aws * delete-*`, `az * delete`, `gcloud * delete`, etc.)
+- **Indirect execution detection** — Catches `bash -c`, `eval`, `python3 -c`, `node -e`, `ruby -e`, `perl -e` wrappers that bypass regex patterns
+- **Managed settings template** — `templates/managed-settings.template.json` for enterprise deployment via Claude.ai Admin Console
+- **`updateLocalPermissions()`** in `gaia-update.js` — NPM postinstall now merges permissions into `settings.local.json` (same approach as plugin SessionStart)
+- **Plugin mode detection via `plugin.json`** — `plugin_setup.py` and `plugin_mode.py` now read `.claude-plugin/plugin.json` for reliable name/version/mode detection with `--plugin-dir`
+- **First-run welcome message** — `user_prompt_submit.py` detects first run and injects a welcome explaining that restart is needed to activate permissions
+
+#### Changed
+- **`settings.template.json`** — Removed permissions block; template now contains only hooks + environment
+- **`_DENY_RULES` centralized in Python** — Single source of truth in `plugin_setup.py`, shared by both OPS and SECURITY modes
+- **T3 approval flow** — All T3 mutative operations now use native `ask` dialog (both ops and security mode). Nonce workflow removed from direct conversation; kept for subagent use via skills.
+- **`approval_messages.py`** — Simplified T3 block message to minimal data (tier + nonce). Workflow instructions live in skills, not hook messages.
+- **`pre_tool_use.py`** — Simplified: passes through `block_response` from `bash_validator` directly, no more mode-specific branching
+- **`bash_validator.py`** — T3 mutative returns `ask` response directly (no nonce generation, no pending files)
+- **`session_start.py`** — Uses `mark_done=False` so `user_prompt_submit.py` can detect first-run and show welcome before marking initialized
+- **`gaia-update.js` registry path** — Fixed to write `plugin-registry.json` in `.claude/` (same path Python hooks expect)
+- **`gaia-doctor.js`** — Now checks permissions in `settings.local.json` (not just `settings.json`). Updated agent and config file lists.
+- **`gaia-update.js` health check** — Updated config files (`surface-routing.json`) and agent list (`gaia-system.md`, `speckit-planner.md`)
+
+#### Fixed
+- **Registry path mismatch** — `gaia-update.js` wrote to `.claude/project-context/`, Python read from `.claude/`. Now both use `.claude/`.
+- **Orphaned nonce files** — `bash_validator` no longer writes pending approval files for `ask` responses
+- **Plugin mode detection** — `--plugin-dir` now correctly detects `gaia-ops` vs `gaia-security` via `plugin.json` instead of path parsing
+- **First-run welcome race condition** — `SessionStart` no longer marks initialized; `UserPromptSubmit` marks after showing welcome
+- **`_build_welcome()` framing** — Rewritten to explain WHY the user needs to restart (permissions not active yet), making Claude naturally relay the message
+
+## [4.4.0-rc.5] - 2026-03-19
+
+### Identity Redesign
+
+Orchestrator identity is now minimal (~900 chars) and delegates to on-demand skills. CLAUDE.template.md deleted -- the UserPromptSubmit hook is the single source of truth for orchestrator identity.
+
+#### Added
+- **`skills/project-dispatch/SKILL.md`** (Reference type) -- agent routing table and dispatch rules, loaded on-demand via Skill tool
+- **`skills/agent-response/SKILL.md`** (Protocol type) -- contract status handling, loaded on-demand via Skill tool
+- Plugin distribution: `.claude-plugin/plugin.json` manifest with engines + categories for Claude Code native plugin system
+- Self-hosted marketplace: `.claude-plugin/marketplace.json` with 2 sub-plugin tiers (gaia-security, gaia-ops)
+- Adapter layer: `hooks/adapters/` with normalized types, abstract base, and Claude Code adapter
+- `hooks/hooks.json` for plugin-channel hook configuration
+- Distribution channel detection (`hooks/adapters/channel.py`)
+- Integration tests for adapter -> business logic -> response flow
+- Plugin manifest validation tests
+
+#### Changed
+- **`hooks/modules/identity/ops_identity.py`** -- reduced to ~900 chars; tells orchestrator to load skills on-demand instead of embedding all instructions inline
+- **SendMessage validation** -- moved from invalid hook event to PreToolUse matcher (agent ID format + nonce approval check)
+- **`hooks/modules/scanning/scan_trigger.py`** -- imports `tools.scan` directly (no `bin/` dependency), works in both npm and plugin mode
+- **Agent namespace support** -- accepts both `cloud-troubleshooter` and `gaia-ops:cloud-troubleshooter` forms
+- **`hooks/user_prompt_submit.py`** -- calls `ensure_plugin_registry()` as fallback if SessionStart didn't fire
+- **`hooks/modules/context/context_injector.py`** -- path fixes for plugin mode
+- **`hooks/modules/session/session_event_injector.py`** -- path fixes for plugin mode
+- Hook entry points (pre_tool_use.py, post_tool_use.py, subagent_stop.py) now use adapter layer for stdin/stdout
+- hook_response.py delegates to ClaudeCodeAdapter internally
+- npm dist-tag now derived from version suffix (rc -> next, beta -> beta, etc.)
+
+#### Removed
+- **`templates/CLAUDE.template.md`** -- identity now injected dynamically; no generated CLAUDE.md
+- **`copy_claude_md()`** in `tools/scan/setup.py` -- deprecated to no-op (callers still reference it for backward compat)
+
+## [4.0.0] - 2026-03-03
+
+### Breaking: Contracts as Single Source of Truth
+
+Contracts now fully control what context each agent receives. Removed the progressive disclosure layer that was silently overriding contract definitions, and cleaned up ~400 lines of dead code from context_provider.py.
+
+#### Changed
+- **context_provider.py**: Contracts are the single source of truth -- removed progressive disclosure filtering that overrode contract-defined sections
+- **context_provider.py**: Simplified output payload -- removed `enrichment` and `progressive_disclosure` keys from response
+- **contracts/terraform-architect.json**: Now reads `cluster_details` and `application_services` sections
+- **contracts/gitops-operator.json**: Now reads `gcp_services` section (GCP overlay)
+- **pre_tool_use.py**: Updated log message to show sections count and rules count
+- **templates/CLAUDE.template.md**: Synced agent routing descriptions with CLAUDE.md
+
+#### Fixed
+- **context_provider.py `get_contracts_dir()`**: Path traversal went up 2 levels instead of 3, producing wrong directory -- masked by legacy fallback that silently compensated
+
+#### Removed
+- **context_provider.py**: ~400 lines of dead code:
+  - Progressive disclosure engine (section filtering, phase-based visibility)
+  - `LEGACY_AGENT_CONTRACTS` dictionary (hardcoded fallback contracts)
+  - Semantic enrichment pipeline
+  - `validate_project_paths()` function
+  - Path resolution utility functions
+
+#### Tests
+- **tests/tools/test_context_provider.py**: Complete rewrite -- 8 tests covering all 6 agents, payload structure, and invalid agent handling
+
+## [3.15.1] - 2026-02-24
+
+### Fix: Cross-Layer Consistency & Dead Code Cleanup
+
+Comprehensive audit of skills, hooks, and security modules. Fixed inconsistencies between layers that caused silent failures (tests pass but system broken).
+
+#### Fixed
+- **bash_validator**: Check blocked commands BEFORE safe commands (defense-in-depth order was inverted)
+- **tiers.py**: Split `VALIDATION_PATTERNS` into `T1_PATTERNS` (validate, lint, fmt, check) and `T2_PATTERNS` (plan, template, diff) — aligns with security-tiers skill
+- **tiers.py**: Removed `terraform plan` from `ULTRA_COMMON_T0_COMMANDS` fast-path (was T0, should be T2)
+- **safe_commands.py**: Removed `terraform plan`/`terragrunt plan` from `ALWAYS_SAFE_MULTIWORD` (simulation, not read-only)
+- **safe_commands.py**: Removed `python3`, `python` from `always_safe` (can execute arbitrary code)
+- **safe_commands.py**: Removed `tar`, `gzip`, `gunzip`, `zip`, `unzip` from `always_safe` (modify filesystem)
+- **task_validator.py**: Removed legacy `APPROVAL_INDICATORS` (`'validation["approved"] == True'`, `"Phase 5: Realization"`)
+- **task_validator.py**: Added `speckit-planner` to `META_AGENTS`
+- **pre_tool_use.py**: Resume regex `{6,7}` → `{5,}` to accept real Claude Code agent IDs
+- **pre_tool_use.py**: Session events now inject BEFORE `# User Task` marker (was after)
+- **post_tool_use.py**: Added `fcntl.flock` to prevent race conditions on `context.json`
+- **post_tool_use.py**: Guard empty timestamps in retention filter
+- **subagent_stop.py**: Fixed indentation bug in consecutive failure detection
+- **subagent_stop.py**: Use `deque(f, maxlen=7)` instead of `f.readlines()` for metrics.jsonl
+- **settings.json**: Moved 7 T3 commands from `allow` → `ask`: kubectl exec/label/annotate/uncordon, helm rollback, flux suspend/resume
+- **settings.json**: Added `flux create` to `ask` list (was unprotected)
+- **agent-protocol skill**: Removed `CURRENT_PHASE` from AGENT_STATUS (redundant with `PLAN_STATUS`)
+- **agent-protocol skill**: `PLANNING` state now explicitly emitted in Phase 2
+- **execution skill**: Scope clarified as T3-only (was accidentally broadened to T2)
+- All 3 hooks: Removed `logging.StreamHandler()` (was sending noise to stderr)
+
+#### Removed
+- **`config_loader.py`** — Dead code, never imported by any module
+- **`discovery_classifier.py`** — Deprecated, replaced by context_writer.py (609 lines)
+- **`exhaustion_detector.py`** — Never worked (wrong glob pattern, wrong file format parsing, 200K thresholds obsolete with 1M context)
+- **`detect_speckit_milestone()`** in event_detector.py — Dead code (post_hook only runs for Bash, not Skill)
+- **`SPECKIT_MILESTONE`** enum value from EventType
+- **`test_config_loader.py`** — Tests for deleted module
+- **`test_discovery_classifier.py`** — Tests for deleted module
+- Slow execution detection in subagent_stop.py (duration_ms always None)
+
+#### Added
+- **`test_cross_layer_consistency.py`** — 24 tests validating consistency between settings.json ↔ safe_commands ↔ blocked_commands ↔ tiers ↔ skills ↔ task_validator
+
+#### Metrics
+- Dead code removed: ~1,500 lines (config_loader + discovery_classifier + exhaustion_detector + dead test files)
+- All 890 tests pass, 0 failures
+
+## [3.12.0] - 2026-02-17
+
+### Refactor: Principle-First Skills & Agent Deduplication
+
+Major redesign of skills and agents. Skills now teach principles instead of enumerating commands. Agents delegate process knowledge to skills, keeping only domain identity.
+
+#### Removed
+- **`skills/anti-patterns/`** - Merged into `command-execution` skill as defensive execution principles
+
+#### Changed
+- **`skills/command-execution/SKILL.md`** - Complete rewrite with defensive execution framework
+  - Timeout hierarchy (tool-native → shell wrapper → abort)
+  - Pre-flight checklist ("Can this hang?" / "Do I know the timeout?")
+  - 7 numbered rules: no pipes, one command per step, Claude Code tools over bash, validate before mutate, absolute paths, files over inline data, quote variables
+- **`skills/security-tiers/SKILL.md`** - Changed from command enumeration to decision framework
+  - Classification by question: "Does it modify live state?" → T3
+- **`skills/terraform-patterns/SKILL.md`** - Split into slim SKILL.md (86 lines) + reference.md
+- **`skills/gitops-patterns/SKILL.md`** - Split into slim SKILL.md (94 lines) + reference.md
+- **`skills/fast-queries/SKILL.md`** - Cut from 256 to 41 lines (essentials only)
+- **`skills/investigation/SKILL.md`** - Fixed to use Glob/Grep/Read tools, removed duplicated content
+- **`skills/output-format/SKILL.md`** - Removed dead escalation protocol
+- **`skills/execution/SKILL.md`** - Consolidated commit format to git-conventions reference
+- **`skills/approval/SKILL.md`** - Removed duplicated commit standards and AskUserQuestion section
+- **All 6 agents** - Removed duplicated Before Acting, Investigation Protocol, Pre-loaded Standards, and command enumeration tier tables
+
+#### Added
+- **`skills/reference.md`** - Agent template and npm release checklist (moved from gaia agent)
+- **`skills/terraform-patterns/reference.md`** - Full HCL examples
+- **`skills/gitops-patterns/reference.md`** - Full YAML examples
+- **`investigation` skill** assigned to cloud-troubleshooter, terraform-architect, gitops-operator, devops-developer, gaia
+- **`git-conventions` skill** assigned to terraform-architect, gitops-operator, devops-developer
+- **`agent-protocol` + `security-tiers` skills** assigned to speckit-planner
+
+#### Metrics
+- Skills: 1,865 → 725 lines (-61%)
+- Agents: 1,914 → 1,007 lines (-47%)
+- Total injected tokens significantly reduced
+- All 882 tests pass
+
+## [3.11.0] - 2026-02-16
+
+### feat: 3-Layer E2E Testing System
+
+Added Layer 1 prompt regression tests (86 tests) validating agent frontmatter, prompt content, skill cross-references, context contracts, security tier consistency, routing table, and skill content rules.
+
+## [3.7.0] - 2026-01-20
+
+### Refactor: Commit Validator Architecture
+
+Moved commit validation to hooks system for better encapsulation and clearer separation of concerns.
+
+#### Changed
+- **commit_validator.py location**: Moved from `tools/validation/` to `hooks/modules/validation/`
+- **bash_validator.py imports**: Updated to use relative import from sibling module
+- **Module structure**: commit_validator.py now exclusively used by bash_validator.py (no direct imports)
+- **Documentation**: Updated tools/validation/README.md to reflect new architecture
+
+#### Technical Details
+- bash_validator.py now uses relative import: `from ..validation.commit_validator import validate_commit_message`
+- commit_validator.py path resolution updated for new location (4 dirname calls instead of 3)
+- pre-publish-validate.js updated to validate new path
+- tools/validation/__init__.py no longer exports commit_validator (internal use only)
+
+#### Benefits
+- Better encapsulation: commit validation only accessible through bash_validator
+- Clearer architecture: validation logic properly contained within hooks system
+- No breaking changes: commit validation continues to work identically
+
+## [3.6.1] - 2026-01-20
+
+### Fix: Include skills/ directory in npm package
+
+#### Fixed
+- **package.json files array**: Added `"skills/"` to ensure skills directory is published to npm
+- This was preventing skills/standards/ from being available in v3.6.0
+
+## [3.6.0] - 2026-01-20
+
+### Standards Migration to Skills System
+
+Major architectural change: migrated from dual context system (standards + skills) to unified skills-based architecture.
+
+#### Added
+- **New skills directory**: `skills/standards/` with 4 standards skills:
+  - `security-tiers/` - T0-T3 operation classification (auto_load)
+  - `output-format/` - Global output contract for all agents (auto_load)
+  - `command-execution/` - Shell security rules and timeout guidelines (triggered)
+  - `anti-patterns/` - Common mistakes by tool: kubectl, terraform, gcloud, helm, flux, npm, docker (triggered)
+- **Standards loader in skill_loader.py**: New `_load_standards_skills()` method
+- **Standards config in skill-triggers.json**: New `standards` section with auto_load and triggers
+
+#### Changed
+- **Unified loading system**: All context now loaded via `skill_loader.py` (skills only)
+- **skill-triggers.json**: Added `standards` section with 4 skills configuration
+
+#### Removed
+- **build_standards_context()**: Removed 91 lines from `context_provider.py`
+- **Standards system**: Deleted `get_standards_dir()`, `read_standard_file()`, `should_preload_standard()`, `build_standards_context()`
+- **--no-standards flag**: Removed from context_provider.py (no longer needed)
+- **docs/ directory**: Eliminated symlink `.claude/docs` (standards now in skills/)
+- **Obsolete tests**: Removed 66 lines of standards-specific tests from `test_context_provider.py`
+- **Duplicate content**: Removed docs/standards reference from universal-protocol skill
+
+#### Migration Notes
+- **Breaking change**: Systems relying on `.claude/docs/standards/` must update to use skills system
+- **Skills auto-load**: `security-tiers` and `output-format` now load for ALL agents (not just PROJECT_AGENTS)
+- **No functional impact**: Same content, different delivery mechanism
+- **Benefits**: Single loading system, better versioning, no duplication
+
+## [3.3.2] - 2025-12-11
+
+### Read-Only Auto-Approval & Code Optimization
+
+Major improvements to the permission system with compound command support and code quality optimizations.
+
+#### Added
+- **Compound command auto-approval**: Safe compound commands (`cat file | grep foo`, `ls && pwd`, `tail file || echo error`) now execute WITHOUT ASK prompts
+- **Extended safe command list**: Added `base64`, `md5sum`, `sha256sum`, `tar`, `gzip`, `time`, `timeout`, `sleep` to always-safe commands
+- **Multi-word command support**: Added `kubectl get/describe/logs`, `helm list/status`, `flux check/get`, `docker ps/images`, `gcloud/aws describe/list` as always-safe
+
+#### Changed
+- **R1: Unified safe command configuration** (`SAFE_COMMANDS_CONFIG`) - Single source of truth for all safe commands, eliminating ~150 lines of duplicate patterns
+- **R2: Unified validation flow** - `classify_command_tier()` now uses `is_read_only_command()` for T0 classification
+- **R4: Singleton ShellCommandParser** - Single instance reused across all validations
+
+#### Removed
+- **R3: Dead code removal** - Removed unused `_contains_command_chaining()` method (~30 lines)
+- **Removed tenacity dependency** - Simplified capabilities loading (retry logic was over-engineering)
+- **Removed duplicate `allowed_read_operations`** - Now derived from `SAFE_COMMANDS_CONFIG`
+
+#### Fixed
+- Compound commands with safe components no longer trigger ASK prompts
+- More consistent tier classification between auto-approval and security validation
+
+#### Technical Details
+- **Lines reduced**: ~200 lines removed through deduplication
+- **Maintainability**: Single source of truth for safe commands
+- **Performance**: Singleton parser avoids repeated instantiation
+
+#### Test Results
+All previous tests continue to pass:
+- Simple read-only commands: NO ASK (auto-approved)
+- Safe compound commands: NO ASK (NEW - auto-approved)
+- Dangerous commands: BLOCKED correctly
+- Compound with dangerous components: BLOCKED correctly
+
+---
+
+## [3.3.1] - 2025-12-11
+
+### Granular AWS Permissions & Command Chaining Block
+
+Refined AWS permission patterns to read-only operations and blocked command chaining to ensure predictable permission evaluation.
+
+#### Changed
+- **AWS permissions**: Replaced broad service wildcards with granular read-only patterns
+  - `Bash(aws ec2:*)` → 40 specific `describe-*` and `get-*` commands
+  - `Bash(aws s3:*)` → `s3 ls`, `s3api get-*`, `s3api list-*`, `s3api head-*`
+  - `Bash(aws rds:*)` → `describe-*`, `list-tags-for-resource`
+  - `Bash(aws iam:*)` → `get-*`, `list-*`, `generate-*`, `simulate-*`
+  - Similar granular patterns for Lambda, Logs, CloudWatch, CloudFormation, ELB, Route53, SecretsManager, SSM, SNS, SQS, DynamoDB, ECR, EKS, ElastiCache
+
+#### Added
+- **Command chaining block** in `pre_tool_use.py`:
+  - Blocks `&&`, `;`, `||` operators to prevent bypassing permission checks
+  - Allows pipes `|` (don't affect permissions)
+  - Smart detection avoids false positives in quoted strings
+  - Clear error message: "Execute each command separately"
+
+#### Fixed
+- Moved `agents/README.md` files to `docs/` to resolve Claude Code parse errors
+
+#### Security Impact
+- Modification commands (create, start, stop) now properly require ASK confirmation
+- Chained commands can no longer bypass individual permission evaluation
+- Read-only operations execute without confirmation
+
+---
+
+## [3.2.3] - 2025-12-09
+
+### Service-Level Permission Wildcards
+
+Simplified permission patterns using service-level wildcards for better Claude Code compatibility.
+
+#### Changed
+- **AWS patterns**: Simplified from `Bash(aws rds describe-:*)` to `Bash(aws rds :*)`
+  - Service-level wildcards: `aws ec2`, `aws rds`, `aws s3`, `aws iam`, etc.
+  - Works around Claude Code pattern matching issues with hyphens
+- **GCP patterns**: Simplified to `Bash(gcloud compute :*)`, `Bash(gcloud container :*)`, etc.
+- **Format standardization**: Removed spaces before `:*` for commands without arguments
+
+#### Fixed
+- Agent README files renamed back to `README.md` (underscore prefix removed)
+- Pattern matching now works for `aws rds describe-db-instances` and similar commands
+
+#### Impact
+- **Read-only commands**: Execute automatically ✓
+- **Modification commands** (start/stop, upload, resize): Now execute automatically (Option A1)
+- **Destructive commands** (delete, terminate): Still blocked ✓
+
+#### Philosophy (Option A1 - Permissive with guardrails)
+- Wide `allow[]` for entire services (e.g., `aws ec2 :*`)
+- Strict `deny[]` for destructive operations
+- Trade-off: Modification commands no longer require confirmation
+
+---
+
+## [3.2.2] - 2025-12-09
+
+### Enhanced Permissions System
+
+Complete overhaul of the permissions configuration to implement "permissive-with-guardrails" strategy.
+
+#### Changed
+- **Comprehensive allow[] rules**: 331 specific read-only patterns for shell, git, kubernetes, helm, flux, terraform, aws, gcp, docker commands
+- **Granular ask[] rules**: 162 modification operations that require user confirmation
+- **Strict deny[] rules**: 73 destructive operations that are completely blocked
+
+#### Fixed
+- Removed duplicate patterns (`uname:*`, `xargs:*`)
+- Fixed `gsutil rm -r:*::*` → `gsutil rm -r:*` (incorrect double colon)
+- Added missing `git branch:*` to allow[] for `git branch -a`
+
+#### Added
+- **New test suite**: `tests/permissions-validation/test_permissions_validation.py`
+  - Emulates Claude Code's actual permission matching behavior
+  - 114 test cases across 13 categories
+  - Tests prefix matching with `:*` wildcard
+  - Validates precedence: Deny → Allow → Ask
+
+#### Philosophy
+- **Allow**: Read-only commands execute automatically (no confirmation)
+- **Ask**: Modification commands require user approval (can be approved)
+- **Deny**: Destructive commands are blocked (cannot be approved)
+
+---
+
+## [3.2.1] - 2025-12-06
+
+### Security Fix - Permission Bypass Bug
+
+**Critical security fix** for permission enforcement in `settings.template.json`.
+
+#### Fixed
+- **Removed generic `"Bash"` from `allow[]`**: The generic `"Bash"` permission was bypassing all specific `ask[]` rules like `"Bash(git push:*)"`, allowing T3 operations (git push, git commit) to execute without user confirmation.
+- **Changed hook matcher from `"BashTool"` to `"Bash"`**: The PreToolUse and PostToolUse hooks were configured with matcher `"BashTool"` but Claude Code invokes the tool as `"Bash"`, causing hooks to never execute.
+
+#### Root Cause Analysis
+- See post-mortem: Generic permission `allow: ["Bash"]` has higher precedence than specific `ask: ["Bash(git push:*)"]` in Claude Code's permission evaluation.
+- Hook matchers must match the exact tool name used by Claude Code.
+
+#### Impact
+- All git operations (push, commit, add) now correctly trigger "ask" confirmation
+- PreToolUse hooks now execute for bash commands
+- Security tier enforcement restored
+
+---
+
+## [3.2.0] - 2025-12-06
+
+### Added - Episodic Memory P0+P1 Enhancements
+
+Inspired by [memory-graph](https://github.com/gregorydickson/memory-graph) analysis, selective feature adoption.
+
+- **P0: Outcome Tracking** (`tools/4-memory/episodic.py`)
+  - New fields: `outcome`, `success`, `duration_seconds`, `commands_executed`
+  - Valid outcomes: "success", "partial", "failed", "abandoned"
+  - New method: `update_outcome()` - Update episode results after execution
+  - Search boost: 10% relevance increase for successful episodes
+
+- **P1: Simple Relationships** (`tools/4-memory/episodic.py`)
+  - New field: `related_episodes` - List of related episode IDs with types
+  - Relationship types: SOLVES, CAUSES, DEPENDS_ON, VALIDATES, SUPERSEDES, RELATED_TO
+  - New method: `add_relationship()` - Link episodes together
+  - New method: `get_related_episodes()` - Query related episodes (outgoing/incoming/both)
+  - Search enhancement: `include_relationships=True` parameter
+
+- **Statistics Enhancements**
+  - Outcome counts by type
+  - Total relationships count
+  - Relationship types breakdown
+
+- **CLI Commands**
+  - `store --outcome --duration` - Store with outcome tracking
+  - `update-outcome <id> <outcome>` - Update episode outcome
+  - `add-relationship <source> <target> <type>` - Create relationship
+  - `get-related <id>` - Query related episodes
+  - `search --include-relationships` - Search with relationship context
+
+### Design Decisions
+
+- Backward compatible: All new fields optional with None defaults
+- Audit trail: Relationship and outcome events logged to JSONL
+- Performance limits: 1000 episodes, 5000 relationships in index
+- No external dependencies: Pure Python implementation
+
+## [3.1.1] - 2025-12-06
+
+### Fixed
+
+- **package.json** - Added `docs/` to files array (was missing in 3.1.0)
+  - `docs/standards/` now included in npm package
+  - Required for hybrid pre-loading in `context_provider.py`
+
+## [3.1.0] - 2025-12-06
+
+### Added - Token Optimization & Consolidation
+
+- **NEW:** `docs/standards/` - Shared execution standards
+  - `security-tiers.md` - T0-T3 definitions
+  - `output-format.md` - Report structure
+  - `command-execution.md` - Execution pillars
+  - `anti-patterns.md` - Common mistakes by tool
+
+- **NEW:** Hybrid pre-loading in `context_provider.py`
+  - Always loads: security-tiers, output-format
+  - On-demand: command-execution
+  - **78% token reduction** per agent invocation
+
+- **NEW:** QuickTriage scripts
+  - `tools/fast-queries/cloud/aws/quicktriage_aws_troubleshooter.sh`
+  - `tools/fast-queries/appservices/quicktriage_devops_developer.sh`
+
+### Changed - Agent Optimization
+
+- **agents/*.md** - All 6 agents reduced by 78%
+  - terraform-architect: 916 → 183 lines
+  - gitops-operator: 1,238 → 217 lines
+  - gcp-troubleshooter: 600 → 156 lines
+  - aws-troubleshooter: 565 → 142 lines
+  - devops-developer: 641 → 173 lines
+
+### Removed - Session System Consolidation
+
+- **REMOVED:** Session management system (consolidated into Episodic Memory)
+  - `commands/save-session.md`
+  - `commands/restore-session.md`
+  - `commands/session-status.md`
+  - `hooks/session_start.py`
+  - `tools/5-task-management/session-manager.py`
+  - `tools/5-task-management/create_current_session_bundle.py`
+  - `tools/5-task-management/restore_session.py`
+
+### Changed - Episodic Memory Enhanced
+
+- **tools/4-memory/episodic.py** - Added `capture_git_state()` migrated from session system
+
+### Fixed - Test Suite
+
+- **359 tests passing (100%)**
+- Fixed import in `test_commit_validator.py`
+- Fixed import in `test_episodic_memory.py`
+- Updated `test_agent_definitions.py` for meta-agents
+- Changed `test_hook_blocks_docker_ps` to `test_hook_default_permit_for_docker_ps`
+- Fixed 11 warnings (return → assert)
+
+### Changed - Documentation
+
+- **README.md & README.en.md** - Updated to v3.1.0, reduced 41%
+- **All subdirectory READMEs** - Reduced 63% total (~2,025 lines removed)
+- Eliminated all references to session system
+
+---
+
+## [3.0.0] - 2025-12-05
+
+### Added - Agent Intelligence System (MAJOR)
+
+- **NEW:** `tools/10-agent-intelligence/` module for intelligent agent optimization
+  - `agent_writing_assistant.py` (24KB) - Assists in writing and improving agent definitions
+  - `workflow_optimizer.py` (29KB) - Applies the 7 LLM Engineering Principles to optimize workflows
+    - Binary Decision Trees
+    - Guards Over Advice
+    - Tool Contracts
+    - Failure Paths
+    - TL;DR First
+    - References Over Duplication
+    - Metrics Over Subjective Goals
+
+- **NEW:** `tools/4-memory/` Episodic Memory System
+  - `episodic.py` (23KB) - Persistent storage and retrieval of historical context
+  - `demo.py` - Demonstration script for episodic memory
+  - Features:
+    - Automatic episode storage with keywords and classifications
+    - Smart search with time decay and relevance scoring
+    - Auto-classification of episode types (deployment, troubleshooting, etc.)
+    - Index management with automatic trimming (1000 episode limit)
+    - Audit trail with append-only JSONL file
+
+- **NEW:** `tools/conversation/` Enhanced Conversation Management
+  - `enhanced_conversation_manager.py` (21KB) - Advanced conversation state management
+  - `agent_contract_builder.py` (19KB) - Dynamic agent contract generation
+  - `progressive_disclosure.py` (17KB) - Progressive context disclosure for token optimization
+
+- **NEW:** `tests/workflow/` directory for workflow-specific tests
+- **NEW:** `tests/test_agent_contract_integration.py` - Agent contract validation tests
+- **NEW:** `tools/agent_capabilities.json` - Centralized agent capabilities definition
+
+### Changed - Agent Enhancements
+
+- **agents/gaia.md** - Major refactoring (1707 lines changed)
+  - Streamlined agent definition
+  - Improved protocol definitions
+  - Better integration with new intelligence modules
+
+- **agents/gitops-operator.md** - Enhanced with 234 new lines
+  - Improved Kubernetes operation patterns
+  - Better Flux CD integration guidance
+  - Enhanced troubleshooting protocols
+
+- **agents/terraform-architect.md** - Enhanced with 47 new lines
+  - Improved Terragrunt support
+  - Better module design guidance
+  - Enhanced security scanning protocols
+
+- **agents/gcp-troubleshooter.md** - Enhanced with 52 new lines
+  - Improved GKE diagnostics
+  - Better IAM analysis patterns
+  - Enhanced networking troubleshooting
+
+### Changed - Tools & Infrastructure
+
+- **hooks/pre_tool_use.py** - Major enhancement (286+ lines)
+  - Improved security validations
+  - Better command blocking logic
+  - Enhanced credential detection
+
+- **hooks/subagent_stop.py** - Enhanced with 193 new lines
+  - Better result packaging
+  - Improved bundle generation
+  - Enhanced session integration
+
+- **tools/2-context/context_provider.py** - Enhanced (120+ lines changed)
+  - Better provider detection
+  - Improved contract validation
+  - Enhanced error handling
+
+- **tools/3-clarification/workflow.py** - Major enhancement (162+ lines)
+  - Episodic memory integration
+  - Improved ambiguity detection
+  - Better context enrichment
+
+- **tools/9-agent-framework/agent_orchestrator.py** - Enhanced (38+ lines)
+  - Better phase management
+  - Improved error recovery
+  - Enhanced logging
+
+### Changed - Fast Queries (Simplified)
+
+- **tools/fast-queries/README.md** - Simplified documentation (185 lines changed)
+- **tools/fast-queries/run_triage.sh** - Streamlined (152 lines changed)
+- **tools/fast-queries/terraform/quicktriage_terraform_architect.sh** - Enhanced (90+ lines)
+- **tools/fast-queries/gitops/quicktriage_gitops_operator.sh** - Enhanced (69+ lines)
+- **tools/fast-queries/cloud/gcp/quicktriage_gcp_troubleshooter.sh** - Enhanced (99+ lines)
+
+### Removed (BREAKING)
+
+- **REMOVED:** `tools/fast-queries/USAGE_GUIDE.md` (369 lines) - Consolidated into README
+- **REMOVED:** `tools/fast-queries/appservices/quicktriage_devops_developer.sh` (38 lines)
+- **REMOVED:** `tools/fast-queries/cloud/aws/quicktriage_aws_troubleshooter.sh` (45 lines)
+
+### Improved
+
+- **Token Efficiency:** New progressive disclosure system reduces context by up to 70%
+- **Agent Intelligence:** Workflows now validated against 7 engineering principles
+- **Memory System:** Historical context improves routing accuracy over time
+- **Conversation Management:** Multi-turn conversations with intelligent context carry-over
+- **Test Coverage:** New workflow and integration tests
+
+### Migration Guide for v3.0.0
+
+**Breaking Changes:**
+1. Removed `quicktriage_devops_developer.sh` - Use agent directly
+2. Removed `quicktriage_aws_troubleshooter.sh` - Use agent directly
+3. Removed `USAGE_GUIDE.md` - See README.md instead
+
+**New Features to Adopt:**
+```python
+# Episodic Memory
+from tools.4_memory.episodic import EpisodicMemory
+memory = EpisodicMemory()
+memory.store_episode(prompt="...", context={...})
+
+# Workflow Optimizer
+from tools.10_agent_intelligence.workflow_optimizer import WorkflowOptimizer
+optimizer = WorkflowOptimizer()
+result = optimizer.analyze(workflow_content)
+
+# Enhanced Conversation
+from tools.conversation.enhanced_conversation_manager import EnhancedConversationManager
+manager = EnhancedConversationManager()
+```
+
+**Recommended Actions:**
+- Review new agent definitions for improved patterns
+- Enable episodic memory for better context over time
+- Use workflow optimizer to validate custom workflows
+
+---
+
+## [2.6.2] - 2025-11-14
+
+### Added - Absolute Paths Support
+
+- **NEW:** `normalizePath()` function - Handles both absolute and relative paths transparently
+- **NEW:** CLI option `--project-context-repo` - Specify git repository for project context in non-interactive mode
+- **NEW:** Environment variable `CLAUDE_PROJECT_CONTEXT_REPO` - Alternative way to specify context repo
+
+### Changed
+
+- **`getConfiguration()`** - Now normalizes paths using `normalizePath()`
+- **`validateAndSetupProjectPaths()`** - Enhanced to handle absolute paths correctly
+- **CLI help and documentation** - Updated examples with absolute paths
+
+### Improved
+
+- Path handling is now more robust and user-friendly
+- Better error messages for path-related issues
+- Clearer documentation and examples
+
+### Examples
+
+```bash
+# Absolute paths without context repo
+npx gaia-init --non-interactive \
+  --gitops /home/user/project/gitops \
+  --terraform /home/user/project/terraform \
+  --app-services /home/user/project/services
+
+# Absolute paths with context repo
+npx gaia-init --non-interactive \
+  --gitops /path/to/gitops \
+  --terraform /path/to/terraform \
+  --project-context-repo git@bitbucket.org:org/repo.git
+```
+
+---
+
+## [2.3.0] - 2025-11-11
+
+### Added - Phase 0 Clarification Module
+
+- **NEW:** `tools/clarification/` module for intelligent ambiguity detection before routing
+  - `clarification/engine.py`: Core clarification engine (refactored from clarify_engine.py)
+  - `clarification/patterns.py`: Ambiguity detection patterns (ServiceAmbiguityPattern, NamespaceAmbiguityPattern, etc.)
+  - `clarification/workflow.py`: High-level helper functions for orchestrators (`execute_workflow()`)
+  - `clarification/__init__.py`: Clean public API
+- **Protocol G** in `agents/gaia.md`: Clarification system analysis and troubleshooting guide
+- **Rule 5.0.1** in `templates/CLAUDE.template.md`: Phase 0 implementation guide with code examples
+- **Phase 0 integration** in `/speckit.specify` command
+- **Regression tests** in `tests/integration/test_phase_0_regression.py`
+- **Clarification metrics** to Key System Metrics (target: 20-30% clarification rate)
+
+### Changed - Module Restructuring (BREAKING)
+
+- **BREAKING:** `clarify_engine.py` and `clarify_patterns.py` moved to `clarification/` module
+  - **Old imports:** `from clarify_engine import request_clarification`
+  - **New imports:** `from clarification import execute_workflow, request_clarification`
+- Updated `application_services` structure in project-context.json:
+  - Added `tech_stack` field (replaces `technology`)
+  - Added `namespace` field for service location
+  - **Removed** `status` field (dynamic state must be verified in real-time, not stored in SSOT)
+- Service metadata now shows only static information: `tech_stack | namespace | port`
+
+### Fixed
+
+- Import paths in `tests/tools/test_clarify_engine.py` updated to new module structure
+- Service metadata test updated to reflect removal of dynamic status field
+- All 20 unit tests passing with new module structure
+
+### Migration Guide for v2.3.0
+
+```python
+# Before (v2.2.x)
+from clarify_engine import request_clarification, process_clarification
+
+# After (v2.3.0)
+from clarification import execute_workflow
+
+# Simple usage
+result = execute_workflow(user_prompt)
+enriched_prompt = result["enriched_prompt"]
+```
+
+---
+
+## [2.2.3] - 2025-11-11
+
+### Fixed - Deterministic Project Context Location
+
+- **context_provider.py**
+  - Always reads `.claude/project-context/project-context.json` (no fallback to legacy paths)
+  - Removed legacy auto-detection logic and unused imports
+  - Prevents "Context file not found" errors when projects only use the new structure
+- **templates/CLAUDE.template.md**
+  - Rule 1 clarifies when to delegate vs. self-execute
+  - Rule 2 explicitly documents the `context_provider.py --context-file .claude/project-context/project-context.json …` invocation
+  - Workflow summary now references orchestration docs after the table (cleaner render)
+
+### Changed - CLI Documentation & Version Alignment
+
+- **README.md / README.en.md**
+  - Documented the exact `npx` commands (`npx gaia-init` / `npx @jaguilar87/gaia-ops`) and clarified installation steps
+  - Updated "Current version" badges to **2.2.3**
+- **package.json**
+  - Bumped package version to `2.2.3`
+
+### Benefits
+
+- No manual tweaks needed to point `context_provider.py` at the correct project context
+- CLAUDE template now tells the orchestrator exactly how to invoke the context provider
+- README instructions reflect the real CLI entry points, reducing confusion for new installs
+
+---
+
+## [2.2.2] - 2025-11-11
+
+### Added - Pre-generated Semantic Embeddings
+
+- **NEW:** Included pre-generated intent embeddings in package (74KB total)
+  - `config/intent_embeddings.json` (55KB) - Semantic vectors for intent matching
+  - `config/intent_embeddings.npy` (19KB) - Binary embeddings for fast loading
+  - `config/embeddings_info.json` (371B) - Metadata about embeddings
+
+### Changed - Semantic Routing Now Works Out-of-the-Box
+
+- **Semantic matching enabled by default:** No manual setup required
+- **Routing accuracy improved:** Ambiguous queries now route correctly using semantic similarity
+- **Example improvement:**
+  ```
+  Query: "puede decirme el estado de los servicios de tcm?"
+  Before: devops-developer (keyword "ci" - incorrect)
+  After: gitops-operator (semantic matching - correct)
+  ```
+
+### Fixed - Directory Structure Consistency
+
+- **Consolidated `configs/` into `config/`:** All configuration and data files now in single directory
+- **Updated tool references:**
+  - `tools/semantic_matcher.py`: Updated embeddings path (configs/ → config/)
+  - `tools/generate_embeddings.py`: Updated output path (configs/ → config/)
+  - All documentation updated to reference correct paths
+
+### Fixed - Test Suite (254 tests, 100% passing)
+
+- **tests/system/test_configuration_files.py:**
+  - Updated to validate `templates/settings.template.json` (package contains template, not installed settings.json)
+  - Tests now reflect npm package structure instead of installed project structure
+
+- **tests/system/test_directory_structure.py:**
+  - Completely rewritten for npm package validation
+  - Tests now verify package directories (agents/, tools/, config/, templates/, bin/)
+  - Removed tests for installed-project structure (session/, .claude/ name)
+  - Added comprehensive tests for all package subdirectories (agents, tools, hooks, config, speckit)
+
+- **tests/tools/test_clarify_engine.py:**
+  - Fixed import paths (tests/tools → gaia-ops/tools)
+  - Made emoji checks flexible (accepts any emoji, not just 📦)
+  - All 32 clarify_engine tests now pass
+
+- **tests/tools/test_context_provider.py:**
+  - Updated troubleshooter contract test (application_services is optional, not required)
+  - Fixed invalid_agent test expectation (now correctly exits with code 1)
+
+- **tools/context_provider.py:**
+  - Changed behavior for invalid agents: now exits with code 1 (was: warning + empty contract)
+  - Better error messages: "ERROR: Invalid agent" instead of "Warning: No contract found"
+
+### Benefits
+
+- Zero configuration: Semantic routing works immediately after installation
+- Better routing: Handles ambiguous queries with 6x higher confidence
+- Consistent structure: All config files in one place (`config/`)
+- Smaller package: Embeddings optimized for size (74KB vs 5MB unoptimized)
+- Regeneration optional: Users can regenerate with `python3 .claude/tools/generate_embeddings.py` if needed
+- Test coverage: 254 tests passing (0 failures)
+
+---
+
+## [2.2.1] - 2025-11-10
+
+### Fixed - Documentation Consistency
+
+- **README.md & README.en.md:**
+  - Updated version numbers from 2.1.0 → 2.2.0
+  - Corrected package structure (hooks/, templates/, commands/)
+  - Fixed hooks/ listing: now shows actual Python files (pre_tool_use.py, post_tool_use.py, etc.) instead of non-existent pre-commit
+  - Fixed templates/ listing: removed non-existent code-examples/, listed actual files (CLAUDE.template.md, settings.template.json)
+  - Added context-contracts.gcp.json and context-contracts.aws.json to config/ section
+  - Removed CLAUDE.md and AGENTS.md from package root (only templates exist)
+  - Added speckit/ directory to structure
+
+- **config/AGENTS.md:**
+  - Updated all references: `.claude/docs/` → `.claude/config/`
+  - Fixed quick links and support documentation paths
+
+- **config/agent-catalog.md:**
+  - Updated all 5 context contract references: `.claude/docs/` → `.claude/config/`
+
+- **index.js:**
+  - Deprecated `getDocPath()` function with console warning
+  - Function now redirects to `config/` directory instead of non-existent `docs/`
+  - Added JSDoc @deprecated annotation
+
+- **README.en.md (Documentation section):**
+  - Removed broken reference to `./CLAUDE.md` (file not in package)
+  - Fixed all documentation links: `./docs/` → `./config/`
+  - Updated to match actual config/ directory structure
+
+- **speckit/README.en.md:**
+  - Removed 3 non-existent commands: speckit.clarify, speckit.analyze-plan, speckit.constitution
+  - Updated command count: 9 → 7 actual commands
+  - Removed references to non-existent tasks-richer.py tool
+  - Removed entire sections for non-existent templates (data-model-template.md, contracts-template.md)
+  - Updated tool files list with actual tools (task_manager.py, clarify_engine.py, context_provider.py)
+  - Fixed all code examples to use only existing commands
+
+- **tools/context_provider.py:**
+  - Added auto-detection for project-context.json location
+  - Honors GAIA_CONTEXT_PATH environment variable
+  - Falls back through common locations (.claude/project-context.json, .claude/project-context/project-context.json)
+  - Fixes agent routing failures when project-context.json is in non-legacy location
+
+- **package.json:**
+  - Fixed `npm test` script (was calling non-existent pytest tests)
+  - Now echoes informative message about fixture availability
+
+- **Agent Branding Unification:**
+  - Renamed `agents/claude-architect.md` → `agents/gaia.md` (aligns with gaia-ops package name)
+  - Renamed `commands/gaina.md` → `commands/gaia.md` (unified as `/gaia` command)
+  - Updated all references in README.md, README.en.md, and agents/gaia.md
+  - Complete branding consistency: package name, agent name, and command name all use "gaia"
+
+### Benefits
+
+- Accurate documentation: All paths and structures match actual package contents
+- No broken links: References point to existing files
+- Clear API: Deprecated functions clearly marked
+- User trust: Documentation matches reality
+- npm test passes: No false failures
+
+---
+
+## [2.2.0] - 2025-11-10
+
+### Added - Unified Settings Template & Auto-Installation
+
+- **NEW:** Created unified `templates/settings.template.json` (214 lines)
+  - Merged functionality from `settings.json` + `settings.local.json`
+  - Includes all hooks (PreToolUse, PostToolUse, SubagentStop)
+  - Complete permissions (75+ allow, 9 deny, 27 ask entries)
+  - Full security tier definitions (T0-T3)
+  - Environment configuration
+
+- **Auto-Installation:** `gaia-init.js` now automatically generates `.claude/settings.json`
+  - Added `generateSettingsJson()` function
+  - Integrated into installation workflow (Step 6.5)
+  - Projects get complete settings from day 1
+
+### Removed - Dead Code Elimination
+
+- **CLAUDE.md** from package root (only template exists now)
+- **templates/code-examples/** (321 lines - never imported or executed)
+  - `commit_validation.py`
+  - `clarification_workflow.py`
+  - `approval_gate_workflow.py`
+- **templates/project-context.template.json** (126 lines - unused, installer generates programmatically)
+- **templates/project-context.template.aws.json** (128 lines - never used)
+- **package.json:** Removed `CLAUDE.md` from files array
+
+### Changed - Package Consistency
+
+- **templates/CLAUDE.template.md:**
+  - Updated all references: `.claude/docs/` → `.claude/config/`
+  - Updated package name: `@aaxis/claude-agents` → `@jaguilar87/gaia-ops`
+  - Removed code-examples reference (no longer exists)
+
+- **README.en.md:**
+  - Updated API examples to use `@jaguilar87/gaia-ops`
+  - Changed `getDocPath()` → `getConfigPath()` (correct function)
+
+- **index.js:**
+  - Updated header and JSDoc comments with new package name
+  - Updated example usage
+
+- **agents/gaia.md:**
+  - Updated system paths to reflect gaia-ops package structure
+  - Clarified symlink architecture and layout
+
+### Improved - Package Quality
+
+- **Reduced template bloat by 57%:** 882 lines → 378 lines (504 lines removed)
+- **Single source of truth:** One settings template instead of scattered config
+- **Cleaner architecture:** Only actual templates remain in `templates/`
+- **Better defaults:** Projects start with complete, production-ready settings
+
+### Benefits
+
+- Unified configuration: Everything in one settings.json file
+- Automatic setup: No manual settings configuration needed
+- Smaller package: 57% reduction in template code
+- Flexibility maintained: Users can still create `settings.local.json` for overrides
+- Package consistency: All references use correct package name
+
+---
+
+## [2.1.0] - 2025-11-10
+
+### Added - Provider-Specific Context Contracts
+
+- **NEW:** Created separate contract files per cloud provider
+  - `config/context-contracts.gcp.json` - GCP-specific contracts
+  - `config/context-contracts.aws.json` - AWS-specific contracts
+  - Ready for `context-contracts.azure.json` (future)
+
+- **Auto-Detection:** `context_provider.py` now automatically:
+  1. Detects cloud provider from `metadata.cloud_provider`
+  2. Falls back to inferring from field presence (`project_id` → GCP, `account_id` → AWS)
+  3. Loads the correct contract file
+  4. Validates against provider-specific requirements
+
+- **Test Fixtures:** Added sample contexts for testing
+  - `tests/fixtures/project-context.gcp.json`
+  - `tests/fixtures/project-context.aws.json`
+
+### Changed
+
+- **Context Provider:** Updated `tools/context_provider.py`
+  - Added `detect_cloud_provider()` function
+  - Added `load_provider_contracts()` function
+  - Updated `get_contract_context()` to accept provider contracts
+  - Legacy contracts remain for backward compatibility
+
+- **Field Names:** Standardized provider-specific fields
+  - GCP: `project_details.project_id` (no change)
+  - AWS: `project_details.account_id` (was `aws_account`)
+  - Installer updated to generate correct field names
+
+- **Templates:** Created AWS-specific template
+  - `templates/project-context.template.aws.json`
+  - Matches AWS naming conventions (EKS, RDS, ECR, etc.)
+
+- **Documentation:** Updated `config/context-contracts.md`
+  - Added "Provider-Specific Contracts" section
+  - Documented how provider detection works
+  - Explained benefits of provider-specific approach
+  - Version bumped to 2.1.0
+
+### Benefits
+
+- Clarity: Field names match cloud provider terminology
+- Simplicity: No complex conditional validation logic in agents
+- Extensibility: Adding Azure = create one JSON file (15 minutes)
+- Agents Stay Agnostic: Agents use pattern discovery, don't need provider logic
+- Single Source of Truth: Orchestrator selects the right contract
+
+### Backward Compatibility
+
+- Legacy support maintained: If provider-specific contracts don't exist, falls back to hardcoded contracts
+- Existing projects: Continue to work without changes
+- Migration: Optional, but recommended for clarity
+
+---
+
+## [1.4.0] - 2025-11-10
+
+### Changed - BREAKING: Complete Installer Redesign
+
+- **NEW FLOW:** Directories first, context second (much more logical!)
+  1. Ask for directories (gitops, terraform, app-services) - ALWAYS
+  2. Ask for project context repo - OPTIONAL
+  3. If NO context: Ask basic questions to create project-context.json
+  4. If YES context: Use that configuration and done!
+
+### Improved
+
+- **Clearer Purpose:** Context repo is now clearly optional
+- **Better Fallback:** If no context exists, creates a basic one with minimal info
+- **All Fields Optional:** Can leave everything empty if you don't know yet
+- **Logical Order:** Ask for what you always need first (paths), then optional context
+
+---
+
+## [1.3.6] - 2025-11-10
+
+### Fixed
+
+- **Installer:** Skip questions when project context already has the answers
+- **Smart Detection:** Only ask what's missing or needs confirmation (paths)
+- **User Experience:** Show config summary when context is loaded
+- **Directory Creation:** Auto-create missing directories without prompting
+
+### Changed
+
+- When project context loads successfully, only asks to confirm/adjust paths
+- Cloud provider, credentials, region, and cluster name auto-applied from context
+- Clearer feedback showing what was loaded from project context
+- Missing directories (gitops, terraform, app-services) now created automatically
+
+---
+
+## [1.3.5] - 2025-11-10
+
+### Added
+
+- **Smart Installer Flow:** Project context repo now asked FIRST, with auto-population of all config
+- **Input Sanitization:** Handles "git clone <url>" pastes automatically (extracts just URL)
+- **Auto-Configuration:** Parses project-context.json and pre-fills all wizard questions
+- **Better Error Messages:** Clear troubleshooting tips for git clone failures (SSH keys, access, URL)
+
+### Changed
+
+- **Wizard Question Order:** Project context moved from last to first question
+- **User Experience:** Reduced manual input when project context exists
+- **Clone Strategy:** Validates project context early, then sets up in final location
+- **Error Handling:** Installation continues even if project context clone fails
+
+---
+
+## [1.3.4] - 2025-11-10
+
+### Fixed
+
+- **Installer:** Removed incorrect AGENTS.md symlink creation in project root during installation
+- **Documentation:** AGENTS.md now only accessible via `.claude/config/AGENTS.md` as intended
+- **Package Quality:** Excluded Python cache files (`__pycache__/`) from published package
+
+### Changed
+
+- **README.md:** Updated project structure documentation to reflect correct AGENTS.md location
+- **README.en.md:** Updated project structure and corrected package references
+- **Package Size:** Reduced from 911.7 kB (93 files) to 660.7 kB (77 files) - 27% reduction
+
+### Added
+
+- **Package Metadata:** Added `homepage` and `bugs` fields to package.json for better npm discovery
+- **Badges:** Added npm version, license, and Node.js version badges to README files
+- **CI/CD:** Created GitHub Actions workflow for automated npm publishing
+- **.npmignore:** Added file to exclude development artifacts from package
+- **Cleanup Script:** Added `npm run clean` to remove Python cache files automatically
+- **Pre-publish Hook:** Added `prepublishOnly` script for automatic cleanup before publishing
+
+---
+
+## Versioning Policy
+
+### Version Number Format: MAJOR.MINOR.PATCH
+
+- **MAJOR:** Breaking changes to orchestrator behavior (requires agent updates, system changes)
+- **MINOR:** New features, sections, or substantial improvements (backward compatible)
+- **PATCH:** Bug fixes, clarifications, typos (backward compatible)
+
+### Examples
+
+- Adding new agent: MINOR (e.g., 2.0.0 → 2.1.0)
+- Changing core principle: MAJOR (e.g., 2.1.0 → 3.0.0)
+- Fixing typo in docs: PATCH (e.g., 2.1.0 → 2.1.1)
+- Refactoring structure (like 2.0.0): MAJOR (changed from monolith to modular)
+
+---
+
+## Maintainers
+
+- **Primary:** Jorge Aguilar (jorge.aguilar87@gmail.com)
+- **Contributors:** Claude Code Agent Swarm
+
+---
+
+## License
+
+Internal documentation for Aaxis RnD team. Not for external distribution.