Skip to content

feat(pov): Parzival 2.1 — Skill Architecture + Dispatch System#65

Open
Hidden-History wants to merge 22 commits intomainfrom
feature/parzival-2.1
Open

feat(pov): Parzival 2.1 — Skill Architecture + Dispatch System#65
Hidden-History wants to merge 22 commits intomainfrom
feature/parzival-2.1

Conversation

@Hidden-History
Copy link
Owner

@Hidden-History Hidden-History commented Mar 16, 2026

Summary

Parzival 2.1 introduces the skill-based dispatch architecture, enabling Parzival to activate and manage parallel agent teams directly via Claude Code teams throughout the full project lifecycle.

Core Changes

  • Skill shim architecture: Skill content lives in _ai-memory/pov/skills/ (7 skills), with thin shim files in .claude/skills/ that point to them. This keeps the POV module self-contained while maintaining Claude Code skill discovery.

  • Dispatch skills: Team Builder (parallel team design), Agent Dispatch (instruction preparation), Agent Lifecycle (monitoring/review/shutdown), BMAD Dispatch (agent selection), Model Dispatch (optional multi-provider LLM routing)

  • Constraints GC-19 and GC-20: Agents must be spawned as teammates (not standalone subagents). BMAD activation and instruction must be separate messages.

  • Updated identity: Parzival now activates and manages all agents directly — the user interacts with Parzival only, Parzival manages everything else.

  • Installer updates: generate_parzival_skill_shims() creates shims dynamically from skill frontmatter. setup_model_dispatch() provides optional multi-provider setup. Stale 2.0 files cleaned up automatically.

Documentation

  • Updated all 9 existing Parzival docs for 2.1 accuracy
  • Created 4 new skill guides: Team Builder, Agent Dispatch, BMAD Dispatch, Model Dispatch
  • Updated main README Parzival section with full lifecycle description and dispatch skills table

Files Changed

  • 155+ files across _ai-memory/pov/, .claude/skills/, docs/parzival/, scripts/install.sh
  • Decisions: DEC-114 through DEC-120

Test Plan

  • Install from feature branch (Option 1) — verify skill shims generated correctly
  • Verify /pov:parzival activates with 15-item menu
  • Verify teams_enabled: true in installed config
  • Verify model-dispatch prompt appears during install
  • Verify stale 2.0 files cleaned up (teams/, team-prompt workflow, old templates)
  • Verify Parzival can dispatch a dev agent via [DA] menu item

WB Solutions and others added 7 commits March 15, 2026 17:57
@claude
feat(pov): Parzival 2.1 — shim architecture, dispatch skills, GC-19/G…
2b0f694 
…C-20

Deploy Parzival 2.1 integration from verified staging (3 rounds adversarial review,
zero issues). Key changes:

- Shim architecture: skill content in _ai-memory/pov/skills/, thin pointers in
  .claude/skills/ (7 skills: agent-dispatch, agent-lifecycle, bmad-dispatch,
  model-dispatch, parzival-bootstrap, parzival-constraints, parzival-team-builder)
- GC-19 (spawn agents as teammates) and GC-20 (no instruction in activation)
  added to global constraints with full 5-point enforcement integration
- Parzival identity rewritten for direct agent management model
- Agent-dispatch workflow inlines selection guide and combination sequences
- Installer: generate_parzival_skill_shims(), setup_model_dispatch(), stale cleanup
- Config parameterized with {USER_NAME} for installer substitution
- Superseded files removed: teams/, team-prompt workflow, moved templates/data

Decisions: DEC-114 through DEC-120
Staging: oversight/staging/parzival-2.1/ (356 files, PM #195)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
docs(pov): update all Parzival docs for 2.1 + add 4 new skill guides
4d07c0b 
Updated 9 existing docs:
- PARZIVAL-SESSION-GUIDE: slash commands, menu items, GC-04 compliance, shim refs
- README-POV: full rewrite — 17 constraints, 15-item menu, current file structure
- CHANGELOG-POV: v2.1.0 entry with upgrade guide
- CONSTRAINT-ENFORCEMENT-SYSTEM: full constraint table, Layer 1/3 self-check
- ESCALATION-ADOPTION-GUIDE: removed stale file refs
- INSTALL-GUIDE-POV: unified installer, shim generation, config sync
- BMAD-Multi-Agent-Architecture: 2.1 context, GC-19/20, current paths
- Multi-Agent-Research-Tracker: fixed duplicate BP IDs, teammate pattern
- SHARDING_STRATEGY: fixed stale bmad-parzival-module paths

Created 4 new skill guides:
- TEAM-BUILDER-GUIDE: tier selection, sizing, presets, conflict avoidance
- AGENT-DISPATCH-GUIDE: 9-step lifecycle, instruction template, corrections
- BMAD-DISPATCH-GUIDE: agent selection, activation sequence, DC-08 rule
- MODEL-DISPATCH-GUIDE: provider setup, model selection, dispatch modes

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(docs): adversarial review fixes + README Parzival section rewrite
becfca3 
Adversarial review by 5 parallel verifiers found 26 issues across 8 files:

README.md:
- Parzival section rewritten for 2.1 (full lifecycle, dispatch skills table)
- Added Agent Lifecycle to dispatch table (was missing)
- "optional but highly recommended" with cross-session memory rationale

CONSTRAINT-ENFORCEMENT-SYSTEM.md (6 fixes):
- GC-04 compliance: Parzival dispatches agents, not user
- Stale CONSTRAINTS.md/PROCEDURES.md paths replaced
- Self-check layer structure corrected

README-POV.md (7 fixes):
- Self-check "three layers" → "two layers"
- Config field oversight_folder → oversight_path
- Activation command /pov:agents:parzival → /pov:parzival
- Constraint range "GC-01 to GC-20" → "GC-01 through GC-15 + GC-19 + GC-20"

CHANGELOG-POV.md: Removed incorrect teams_enabled claim
INSTALL-GUIDE-POV.md (6 fixes): Phantom function name, stale paths, slash commands
BMAD-DISPATCH-GUIDE.md: Added missing Maintenance + QA phases
MODEL-DISPATCH-GUIDE.md: Corrected Vertex AI URL
Multi-Agent-Research-Tracker.md: Fixed 16 broken BP source links → Qdrant reference

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(docs): round 2 adversarial review — 8 fixes + restore teams_enabled
cc78a6c 
Round 2 review (5 verifiers) found 8 remaining issues, all fixed:

- config.yaml: Restore teams_enabled: true (required for Claude Code teammates)
- CONSTRAINT-ENFORCEMENT-SYSTEM: Fix layer count contradiction (15+2, not 13+4)
- MODEL-DISPATCH-GUIDE: Add missing "general" category, video-gen triggers, audio signal
- BMAD-DISPATCH-GUIDE: Fix PM workflow phase scope to "Discovery (or any phase)"
- INSTALL-GUIDE-POV: Label "dispatch skills" → "skills" (2 of 7 are core, not dispatch)
- Multi-Agent-Research-Tracker: Fix stale methodology to reference Qdrant storage

R1: 26 issues → R2: 8 issues → converging toward zero.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(docs): round 3 adversarial review — 5 fixes, convergence confirmed
5568dbd 
R3 review (4 verifiers) found 5 remaining issues:

- INSTALL-GUIDE-POV: 3 stale links to removed CONSTRAINTS.md/PROCEDURES.md
  replaced with current constraint-enforcement-system and workflows refs
- BMAD-Multi-Agent-Architecture: Footer date synced to header (2026-03-15)
- README-POV: GC-04 behavioral contradiction — Duties sections 2/4/5 still
  described v1.x prompt-provider model. Updated to v2.1 dispatch model
  (Parzival activates and manages agents, user manages Parzival only)

Convergence: R1=42 → R2=8 → R3=5 (3 stale links + 1 date + 1 behavioral)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(docs): round 4 review — 4 remaining v1.x prompt-passing fixes
afefd95 
R4 found 4 more instances of v1.x "provide prompt / user runs" language:
- README-POV: "Provides review agent prompts" → "Dispatches review agents"
- CONSTRAINT-ENFORCEMENT-SYSTEM: Layer 4 diagram + 2 test scenarios updated
  from prompt-provider model to v2.1 dispatch model

All same class as R3 finding. Convergence: R1=42 → R2=8 → R3=5 → R4=4

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(installer): verify all system dependencies at install time
8960d8e 
Add tiktoken, anthropic, langfuse, numpy to critical package verification
(these are eagerly imported — missing any crashes the memory module).

Add tree-sitter grammars (js, ts, go, rust) and spacy to optional package
verification (these have graceful fallbacks but should warn if missing).

Previously only 5 critical + 2 optional were checked. Now 9 critical + 7
optional — matches every package the memory module actually imports.

Prevents silent degradation: spaCy and tree-sitter gaps went unnoticed
for weeks because pip partial failures weren't caught by verification.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
WB Solutions and others added 15 commits March 16, 2026 03:07
@claude
fix: PLAN-018 Phase 1 — 6 bugs fixed, TD-296 thin shims, 15 review fi…
f62d498 
…ndings resolved

BUG-225: Classifier SKIP_RECLASSIFICATION_TYPES expanded to 9 (frozenset), error_pattern regex tightened
BUG-218: RRF score clamped with min(0.95, ...) in search.py
BUG-226: Edit hook reads full file via Path.read_text() with 200KB guard and OSError fallback
BUG-227: Installer Option 1 copies docker/.env.example with error handling
BUG-219: Scanner source_type="user_session" added to store_async.py
BUG-222: Verified step-03-create-handoff.md already exists (QA false positive)
TD-296: 15 full-content skill shims replaced with thin routing shims (82KB→5KB, 93.5% reduction)
  - 7 POV skill shims removed (dynamically generated by installer)
  - 2 orphaned canonical directories deleted
  - Installer deploy_ai_memory_skills() + trigger extraction added

15 review findings from dual-model adversarial review (Sonnet + Opus) also fixed.
2541 tests passing. Zero legitimate issues at convergence.

Ref: PLAN-018, PM #200

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix: PLAN-018 Phase 2 — observability, config, security, 25 review fi…
5bc06f7 
…ndings fixed

Phase 2 items:
- TD-275/289: Cherry-picked tags on all 108 emit_trace_event calls
- TD-290: Classifier LLM calls wrapped with @observe(as_type="generation")
- TD-291: freshness_scan → freshness_scan_complete event naming
- TD-292: Quality gate skip Prometheus metric in both hook scripts
- TD-262: BMAD_LOG_LEVEL → AI_MEMORY_LOG_LEVEL with deprecation alias
- TD-189: Langfuse moved to optional dependency ([observability] extra)
- TD-220: SQL injection fix in langfuse_setup.sh (psql -v parameterized)
- TD-221: Dead fallback fixed in langfuse_setup.sh:606
- TD-293/294/295: Doc/spec fixes, parzival.md standards + dispatch refs
- TD-219/222: Verified already fixed (screenshots dir, macOS sed)
- BUG-221: Verified not a bug (MagicMock bypasses frozen)

Review findings (25 total, all fixed):
- BUG-228: Copy-paste tags in injection.py (bootstrap/format, not greedy_fill)
- BUG-229: hook_type label mismatch in user_prompt_store_async.py
- BUG-230/231: langfuse_setup.sh dead fallback + hardcoded port
- BUG-232: parzival.md dispatch skill paths marked [PLANNED]
- BUG-233: E2E conftest.py videos dir + absolute paths
- BUG-234: 8 flaky caplog tests fixed (propagate=False, BUG-209 pattern)
- BUG-235: test_logging.py deprecation warning coverage
- TD-297: injection.py project_id threading to trace events
- TD-298: freshness.py per-invocation session_id
- TD-299/300: llm_classifier.py top-level imports + narrowed exceptions
- TD-301: langfuse_stop_hook.py targeted ImportError guard
- TD-302: langfuse_instrument.py removed dead _NoOpGeneration
- TD-303: logging_config.py formatter update on reconfig

14 agents across 4 rounds (3 dev → 6 review → 3 fix → 2 re-review).
Dual-model adversarial review (Sonnet + Opus) converged at zero issues.
2542 tests passing. Zero failures on both random and deterministic ordering.

Ref: PLAN-018, PM #201

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(tests): resolve BUG-234 final caplog instance in test_classificat…
922963d 
…ion_worker

setup_hook_logging() sets logger to INFO at import time, overriding
caplog.at_level(DEBUG) from the autouse fixture. Add explicit setLevel(DEBUG)
after each local import to ensure DEBUG-level log assertions work.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(lint): combine nested with statements in test_token_metrics (SIM117)
c114989 
Ruff SIM117 requires single with statement with multiple contexts
instead of nested with blocks. Three instances at lines 62, 94, 273.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
style: apply black 26.3.0 formatting to 8 PLAN-018 files
f9394ae 
Black version mismatch (local 26.1.0 vs CI 26.3.0) caused formatting
differences in files touched by PLAN-018 Phase 1+2 commits. Aligned
local to pinned version and reformatted.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix: PLAN-018 sprint fixes — BUG-219 source_type, M-12 gitignore, eva…
4488bb3 
…luator config, CHANGELOG v2.3.0

- BUG-219: add source_type="user_session" to store_async.py (all 5 hooks now consistent)
- M-12: add _ai-memory/ to .gitignore template in install.sh (existing-file and create-new paths)
- Evaluator: re-comment base_url in evaluator_config.yaml, remove /v1 from compose default,
  add /v1 auto-append in provider.py
- CHANGELOG: add v2.3.0 release entry

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix: single docker/.env source of truth — TD-308 .env architecture cl…
7d2eeb5 
…eanup

Restructure .env handling so docker/.env is the ONLY config file:

- Rewrite docker/.env.example with 5-section layout (API Keys, Auto-Generated,
  Feature Toggles, Configuration, Internal) — 248 lines, zero missing keys
- Gut import_user_env() in install.sh (deprecated, warns if legacy root .env exists)
- Fix upgrade.sh to read from $INSTALL_DIR/docker/.env (was $INSTALL_DIR/.env)
- Fix rollback.sh to restore to $INSTALL_DIR/docker/ with legacy fallback
- Fix classifier/config.py to search ~/.ai-memory/docker/.env (was ~/.ai-memory/.env)
- Fix config.py pydantic env_file to use absolute path via AI_MEMORY_INSTALL_DIR

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix: isolate tests from real docker/.env + update docs for TD-308
930c0ca 
- Add conftest.py AI_MEMORY_INSTALL_DIR isolation (set before class import)
- Fix test_config.py test_env_file_loading to use _env_file= parameter
- Fix test_agent_memory.py test_parzival_config_defaults with _env_file=None
- Fix test_evaluator_provider.py 3 ollama tests: clear OLLAMA_API_KEY env
- Fix test_sops_encryption.py test_secrets_backend_in_env_file with _env_file=
- Update CHANGELOG.md v2.3.0 with TD-308 entries
- Update INSTALL.md, CONFIGURATION.md, TROUBLESHOOTING.md: docker/.env paths

2543 passed, 0 env-related failures. 3 pre-existing flaky (test ordering).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(lint): remove unused original_cwd variable in test_env_file_loading
87a1831 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
style: apply black formatting to evaluator/provider.py
4460932 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix: remove quotes from JIRA_PROJECTS comment that broke Docker Compo…
9c0eab7 
…se parsing

Docker Compose .env parser rejects inline comments containing quote characters.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix: remove inline comments from .env.example — Docker Compose and py…
f82df0e 
…dantic parse failure

Move all comments to line above KEY=value (never inline). Docker Compose .env
parser and pydantic-settings DotEnvSettingsSource both treat inline # comments
as part of the value, causing parse errors.

Also set GITHUB_SYNC_ENABLED, JIRA_SYNC_ENABLED, LANGFUSE_ENABLED to false
in template — these require credentials that aren't present on fresh install.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(docker): add tiktoken to github-sync requirements — BUG-236
91fbfbf 
memory.__init__ → storage → chunking → truncation → tiktoken import
chain causes github-sync container to crash loop after rebuild.
Added tiktoken>=0.5.0,<1.0.0 matching main requirements.txt spec.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
fix(config): uncomment and add missing compose-referenced vars in .en…
3f0dce2 
…v.example

Uncommented 8 vars that were commented but referenced by docker-compose
with ${VAR:-default} syntax: 5 GitHub sync timeouts, 3 embedding retry
params. Added 3 new vars: QDRANT_TIMEOUT, QDRANT_USE_HTTPS,
AI_MEMORY_QUEUE_DIR. All compose-referenced vars now documented.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@claude
docs: update CHANGELOG and TROUBLESHOOTING for BUG-236 and container …
f5a2f65 
…procedures

Add BUG-236 fix entry, container rebuild instructions, and post-update
troubleshooting section. Document critical rules for docker compose operations.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@Hidden-History