feat: migrate hades, eight-gates, red-blue-review, council to Teams API (#146)

ANcpLua · claude · web-flow · commit 1b534f47942f · 2026-02-26T23:05:46.000+01:00
* feat(exodia/hades): migrate Teams API from vague references to explicit tool usage

Replace half-baked "Create an agent team" and "MESSAGE smart-audit-deadcode"
instructions with explicit TeamCreate, SendMessage, TaskCreate/TaskUpdate,
and TeamDelete calls with proper parameters. Add team context preamble and
shutdown_response protocol to all 4 teammate templates. Remove fallback
subagent path and duplicate STEP -1 block.

Co-Authored-By: Claude Opus 4.6 &lt;noreply@anthropic.com&gt;

* feat: migrate council, red-blue-review, eight-gates Gate 7 to Teams API

Council: researcher + synthesizer cross-pollinate via SendMessage, clarity
asks live follow-ups instead of one-shot read. 10-step orchestration flow.

Red-blue-review: Red attackers coordinate attacks, Blue defenders claim
findings from shared task list, full TeamCreate→shutdown→TeamDelete lifecycle.

Eight-gates Gate 7: removed dual Mode A/B, Teams-only execution. Lane
workers coordinate via SendMessage and claim work via TaskCreate/TaskUpdate.

Co-Authored-By: Claude Opus 4.6 &lt;noreply@anthropic.com&gt;

---------

Co-authored-by: Claude Opus 4.6 &lt;noreply@anthropic.com&gt;
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
@@ -52,8 +52,8 @@
     },
     {
       "name": "exodia",
-      "description": "Multi-agent workflow orchestration (9 commands + 2 skills: eight-gates, hades): fix, turbo-fix, fix-pipeline, tournament, mega-swarm, deep-think, batch-implement, red-blue-review, baryon-mode.",
-      "version": "2.0.0",
+      "description": "Multi-agent workflow orchestration (9 commands + 2 skills: eight-gates, hades): fix, turbo-fix, fix-pipeline, tournament, mega-swarm, deep-think, batch-implement, red-blue-review, baryon-mode. Hades, eight-gates, and red-blue-review use Teams API for reactive collaboration.",
+      "version": "2.1.0",
       "source": "./plugins/exodia"
     }
   ]
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -8,9 +8,12 @@ and the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 
 ### Changed
 
-- **`exodia/skills/hades`**: Migrated from vague Teams references to explicit Teams API usage. SKILL.md now uses `TeamCreate`, `TeamDelete`, `SendMessage` (with type shutdown_request/shutdown_response), `TaskCreate`, `TaskList`, `TaskUpdate` with explicit parameters. Removed fallback subagent path and duplicate STEP -1 block. All 4 teammate templates (auditors, eliminators, verifiers, goggles) updated: vague `MESSAGE` replaced with `SendMessage (recipient: "...")`, vague `Create tasks in shared list` replaced with `TaskCreate`/`TaskUpdate`, team context preamble and shutdown_response protocol added to each
-- **`exodia`**: Bumped 2.0.0 → 2.1.0
+- **`exodia/skills/hades`**: Migrated from vague Teams references to explicit Teams API. SKILL.md now uses `TeamCreate`, `TeamDelete`, `SendMessage` (shutdown_request/shutdown_response), `TaskCreate`/`TaskList`/`TaskUpdate` with explicit parameters. Removed fallback subagent path and duplicate STEP -1 block. All 4 teammate templates (auditors, eliminators, verifiers, goggles) updated: vague `MESSAGE` → `SendMessage (recipient: "...")`, vague task list → `TaskCreate`/`TaskUpdate`, team context preamble and shutdown protocol added
+- **`exodia/eight-gates` Gate 7 EXECUTE**: Removed dual Mode A (Task subagents) / Mode B (Agent Teams) pattern. Teams API is now the single execution mode. Lane workers coordinate via `SendMessage` and claim work via `TaskCreate`/`TaskUpdate`. Collision avoidance uses teammate messaging
 - **`exodia/skills/hades` allowed-tools**: Added `TeamCreate`, `TeamDelete`, `TaskCreate`, `TaskList`, `TaskUpdate`, `SendMessage` to frontmatter
+- **`exodia`**: Bumped 2.0.0 → 2.1.0
+- **`exodia/red-blue-review`**: Migrated from fire-and-forget subagents to Teams API. Red attackers coordinate attacks via `SendMessage`, Blue defenders claim findings from shared `TaskCreate`/`TaskUpdate`, re-attackers mark verdicts. Full TeamCreate→shutdown→TeamDelete lifecycle across 3 adversarial phases
+- **`council`**: Bumped 1.1.0 → 1.2.0. Migrated from subagents to Teams API. Researcher + synthesizer cross-pollinate via `SendMessage`. Clarity asks live follow-ups instead of one-shot read. 10-step orchestration flow. Cost profile ~2.5x → ~3x
 
 ### Removed
 
diff --git a/plugins/council/.claude-plugin/plugin.json b/plugins/council/.claude-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "name": "council",
-  "version": "1.1.0",
-  "description": "Five-agent council: Opus captain decomposes and synthesizes, researcher + synthesizer run in parallel, clarity checks their output, Haiku janitor flags bloat. Each agent identity is inlined in its agent file — passive context, zero activation cost.",
+  "version": "1.2.0",
+  "description": "Five-agent council using Teams API: Opus captain orchestrates, researcher + synthesizer cross-pollinate via SendMessage, clarity asks live follow-ups, Haiku janitor flags bloat. Reactive collaboration instead of fire-and-forget subagents.",
   "author": {
     "name": "ANcpLua",
     "email": ""
diff --git a/plugins/council/commands/council.md b/plugins/council/commands/council.md
@@ -1,8 +1,8 @@
 ---
 description: >-
   Invoke the five-agent council on any complex task. Opus captain decomposes and synthesizes.
-  Researcher and synthesizer run in parallel. Clarity reads their raw output.
-  Haiku janitor flags bloat. Captain removes cuts and delivers.
+  Researcher and synthesizer run in parallel and cross-pollinate via SendMessage.
+  Clarity reads their output and asks follow-ups. Haiku janitor flags bloat. Captain removes cuts and delivers.
 argument-hint: [task description]
 ---
 
@@ -20,21 +20,53 @@ Invoke the council on `[task]`.
 ## How it runs
 
 ```text
-opus-captain receives task
+captain (lead) creates team "council"
   │
-  ├── sonnet-researcher  (parallel) → FINDING/SOURCE/CONFIDENCE/GAPS
-  └── sonnet-synthesizer (parallel) → REASONING/CONCLUSION/CONFIDENCE/BREAKS
+  ├── researcher   (teammate, parallel) ─┐
+  └── synthesizer  (teammate, parallel) ─┤── cross-pollinate via SendMessage
+  │                                       │
+  │   captain waits for convergence       │
+  │                                       │
+  └── clarity (teammate) reads task list + messages researcher/synthesizer
+        → flags GAPS/ASSUMPTIONS/MISALIGNMENT via SendMessage
   │
-  └── sonnet-clarity reads researcher + synthesizer raw output
-        → GAPS/ASSUMPTIONS/MISALIGNMENT/RESEARCHER_SYNTHESIZER_CONFLICT
-  │
-  └── opus-captain reads all three → produces draft answer
+  └── captain reads all messages → produces draft
         │
-        └── haiku-janitor → BLOAT_FLAG + CUTS list
+        └── janitor (teammate) → BLOAT_FLAG + CUTS via SendMessage
+              │
+              └── captain removes cuts → final output
               │
-              └── opus-captain removes cuts → final output
+              └── TeamDelete
 ```
 
+## Orchestration
+
+1. **TeamCreate:** `team_name="council"`, description = "Council: [task summary]"
+2. **Spawn researcher + synthesizer** (both in ONE message, parallel):
+   - `Task: team_name="council", name="researcher", subagent_type="council:sonnet-researcher"`
+   - `Task: team_name="council", name="synthesizer", subagent_type="council:sonnet-synthesizer"`
+   - They cross-pollinate via `SendMessage`: "My sources say X" / "That contradicts my reasoning on Y"
+3. **Wait for convergence** — both go idle, no new messages for sustained period
+4. **Spawn clarity** (researcher + synthesizer stay alive):
+   - `Task: team_name="council", name="clarity", subagent_type="council:sonnet-clarity"`
+   - Clarity reads team message history AND messages researcher/synthesizer for follow-ups
+   - Researcher/synthesizer respond to clarifying questions in real time
+5. **When clarity converges:** shutdown researcher + synthesizer + clarity
+   (`SendMessage type="shutdown_request"` to each)
+6. **Captain synthesizes** from all team messages into draft answer
+7. **Spawn janitor:**
+   - `Task: team_name="council", name="janitor", subagent_type="council:haiku-janitor"`
+   - Janitor sends `BLOAT_FLAG` + `CUTS` via `SendMessage`
+8. **Shutdown janitor** (`SendMessage type="shutdown_request"`)
+9. **Captain applies cuts** → final output
+10. **TeamDelete** — clean up team
+
+**Why researcher + synthesizer stay alive through clarity:**
+Clarity's value comes from asking follow-up questions — "Your source X contradicts synthesizer's
+assumption Y, can you clarify?" — which requires live teammates. Shutting them down early
+saves tokens but eliminates the reactive collaboration that justifies using Teams over
+fire-and-forget subagents.
+
 ## Usage
 
 ```text
@@ -55,10 +87,11 @@ opus-captain receives task
 | Agent | Model | Relative cost |
 |-------|-------|---------------|
 | opus-captain | opus-4.6 | High (runs three times: dispatch + clarity read + synthesis) |
-| sonnet-researcher | sonnet-4.6 | Medium |
-| sonnet-synthesizer | sonnet-4.6 | Medium |
-| sonnet-clarity | sonnet-4.6 | Low — reads output, no tool calls |
+| sonnet-researcher | sonnet-4.6 | Medium-High (stays alive through clarity phase for follow-ups) |
+| sonnet-synthesizer | sonnet-4.6 | Medium-High (stays alive through clarity phase for follow-ups) |
+| sonnet-clarity | sonnet-4.6 | Medium — reads output, asks follow-ups, receives responses |
 | haiku-janitor | haiku-4.5 | Minimal |
 
-Total: ~2.5x a single Opus pass. Researcher and synthesizer run in parallel,
-clarity and haiku are lightweight sequential passes on their output.
+Total: ~3x a single Opus pass. Higher than fire-and-forget (~2.5x) because researcher
+and synthesizer stay alive through the clarity phase. The cost buys reactive
+cross-pollination and real follow-up conversations instead of one-shot reads.
diff --git a/plugins/exodia/commands/red-blue-review.md b/plugins/exodia/commands/red-blue-review.md
@@ -31,21 +31,24 @@ allowed-tools: Task, Bash, TodoWrite
 ```text
 REVIEW LEAD (You — Orchestrator)
 │
-├─ Phase 1: RED ATTACK (3 agents parallel)
-│  ├── red-crash-hunter
-│  ├── red-security-attacker
-│  └── red-api-breaker
-│  └── GATE → validate findings
+│  TeamCreate: "red-blue-review"
 │
-├─ Phase 2: BLUE DEFENSE (1 agent per MODULE — grouped findings)
-│  └── blue-defender-N (one per affected module, all its findings)
-│  └── GATE → fixes collected
+├─ Phase 1: RED ATTACK (3 teammates, coordinate via SendMessage)
+│  ├── red-crash-hunter ──────┐
+│  ├── red-security-attacker ─┼── SendMessage: share attack vectors
+│  └── red-api-breaker ───────┘   TaskCreate: each finding → shared list
+│  └── GATE → validate findings, shutdown_request → Red
 │
-├─ Phase 3: RED RE-ATTACK (1 agent per module's fixes)
-│  └── red-reattacker-N (one per BLUE module)
-│  └── VERDICT: DEFEATED / BYPASSED / INCOMPLETE
+├─ Phase 2: BLUE DEFENSE (1 teammate per MODULE)
+│  └── blue-defender-N ── TaskUpdate: claim findings, SendMessage: cross-module fixes
+│  └── GATE → fixes collected, shutdown_request → Blue
 │
-└─ RELEASE: SAFE / BLOCK
+├─ Phase 3: RED RE-ATTACK (1 teammate per module)
+│  └── red-reattacker-N ── TaskUpdate: DEFEATED/BYPASSED/INCOMPLETE
+│  └── shutdown_request → re-attackers
+│
+├─ RELEASE: SAFE / BLOCK
+└─ TeamDelete
 ```
 
 ---
@@ -58,13 +61,45 @@ Inject findings into Red Team prompts as attack surface hints. Do NOT give to Bl
 
 **THIS IS AN ADVERSARIAL EXERCISE.**
 
-1. Launch 3 Red Team agents in ONE message
-2. Validate findings (reject false positives), then GROUP by target module/file
-3. Launch 1 Blue defender per MODULE (with all that module's findings)
-4. Launch 1 Red re-attacker per Blue module
-5. Score and generate release recommendation
+**STEP 0 — Create Team:**
+TeamCreate: team_name = "red-blue-review", description = "Adversarial review: $0"
+
+**STEP 1 — Red Attack Phase:**
+Spawn 3 Red attackers as teammates (ALL in ONE message):
+Task tool: team_name="red-blue-review", name="red-crash-hunter", subagent_type="deep-debugger", model="opus"
+Task tool: team_name="red-blue-review", name="red-security-attacker", subagent_type="general-purpose", model="opus"
+Task tool: team_name="red-blue-review", name="red-api-breaker", subagent_type="general-purpose", model="opus"
+
+Red attackers use SendMessage to coordinate: "I found SQL injection in handler X, check for XSS too"
+Red attackers use TaskCreate for each finding (shared task list).
+
+**STEP 2 — Validate & Transition:**
+When Red converges (idle, no new messages): validate findings, reject false positives.
+SendMessage type="shutdown_request" to all Red attackers.
+Group validated findings by module.
+
+**STEP 3 — Blue Defense Phase:**
+Spawn 1 Blue defender per module as teammates:
+Task tool: team_name="red-blue-review", name="blue-defender-[module]", subagent_type="general-purpose", model="opus"
+
+Blue defenders claim findings from shared task list via TaskUpdate.
+Blue defenders use SendMessage to coordinate fixes across modules.
+
+**STEP 4 — Validate & Transition:**
+SendMessage type="shutdown_request" to all Blue defenders.
+
+**STEP 5 — Red Re-Attack Phase:**
+Spawn 1 Red re-attacker per module:
+Task tool: team_name="red-blue-review", name="red-reattacker-[module]", subagent_type="deep-debugger"
+
+Re-attackers use TaskUpdate to mark findings as DEFEATED/BYPASSED/INCOMPLETE.
+
+**STEP 6 — Score & Cleanup:**
+Score results, generate release recommendation.
+SendMessage type="shutdown_request" to all re-attackers.
+TeamDelete.
 
-**YOUR NEXT MESSAGE: 3 Red Team Task tool calls. NOTHING ELSE.**
+**YOUR NEXT MESSAGE: TeamCreate + 3 Red Team Task tool calls. NOTHING ELSE.**
 
 </CRITICAL_EXECUTION_REQUIREMENT>
 
@@ -76,24 +111,36 @@ Launch ALL 3 in ONE message.
 
 ### red-crash-hunter
 
-> subagent: deep-debugger | model: opus
+> teammate: red-crash-hunter | team: red-blue-review | subagent_type: deep-debugger | model: opus
 > RED TEAM — Crash Hunter. TARGET: $0 | SCOPE: $1
+> You are a teammate in the red-blue-review team.
+> Use SendMessage to coordinate with other Red team members (red-security-attacker, red-api-breaker).
+> Use TaskCreate for each finding you discover.
+> When you receive a shutdown_request, approve it.
 > Find ways to CRASH the code: Null refs, invalid input, resource exhaustion, race conditions, overflow.
 > Format: CRASH-001: [title] | Severity | Reproduction | Location
 > Real bugs only — false alarms cost -5 points.
 
 ### red-security-attacker
 
-> subagent: feature-dev:code-reviewer | model: opus
+> teammate: red-security-attacker | team: red-blue-review | subagent_type: general-purpose | model: opus
 > RED TEAM — Security Attacker. TARGET: $0 | SCOPE: $1
+> You are a teammate in the red-blue-review team.
+> Use SendMessage to coordinate with other Red team members (red-crash-hunter, red-api-breaker).
+> Use TaskCreate for each finding you discover.
+> When you receive a shutdown_request, approve it.
 > Find SECURITY vulnerabilities: Injection, path traversal, data exposure, unsafe deserialization, SSRF/CSRF.
 > Format: SEC-001: [title] | Severity | Attack Input | Exploitation | Impact
 > Proof of concept required. Theoretical issues = 0 points.
 
 ### red-api-breaker
 
-> subagent: feature-dev:code-explorer
+> teammate: red-api-breaker | team: red-blue-review | subagent_type: general-purpose | model: opus
 > RED TEAM — API Breaker. TARGET: $0 | SCOPE: $1
+> You are a teammate in the red-blue-review team.
+> Use SendMessage to coordinate with other Red team members (red-crash-hunter, red-security-attacker).
+> Use TaskCreate for each finding you discover.
+> When you receive a shutdown_request, approve it.
 > Find ways to BREAK the API contract: Behavior != docs, edge cases, missing validation, breaking changes.
 > Format: BREAK-001: [title] | Severity | Documented | Actual | Proof
 > Real contract violations only, not style preferences.
@@ -108,9 +155,13 @@ Launch ONE defender per MODULE (not per finding):
 
 ### blue-defender-N (one per module)
 
-> subagent: feature-dev:code-architect | model: opus
+> teammate: blue-defender-[module] | team: red-blue-review | subagent_type: general-purpose | model: opus
 > BLUE TEAM — Defend MODULE: [MODULE_PATH]
 > FINDINGS IN THIS MODULE: [PASTE ALL RED FINDINGS FOR THIS MODULE]
+> You are a teammate in the red-blue-review team.
+> Claim findings from the shared task list using TaskUpdate (set status to "in_progress").
+> Use SendMessage to coordinate with other Blue defenders when fixes span modules.
+> When you receive a shutdown_request, approve it.
 >
 > **FILE OWNERSHIP:** You own ONLY files in [MODULE_PATH]. Do not modify files outside your module.
 >
@@ -132,9 +183,13 @@ Launch ONE re-attacker per Blue module (mirrors Phase 2 grouping):
 
 ### red-reattacker-N (one per module)
 
-> subagent: deep-debugger
+> teammate: red-reattacker-[module] | team: red-blue-review | subagent_type: deep-debugger | model: opus
 > RED RE-ATTACK — Try to bypass ALL fixes in MODULE: [MODULE_PATH]
 > BLUE FIXES: [PASTE ALL BLUE FIXES FOR THIS MODULE]
+> You are a teammate in the red-blue-review team.
+> Use TaskUpdate to mark each finding as DEFEATED/BYPASSED/INCOMPLETE.
+> Use SendMessage to share bypass techniques with other re-attackers.
+> When you receive a shutdown_request, approve it.
 >
 > For EACH fix: VERDICT: **DEFEATED** (Blue +5) | **BYPASSED** (Red +3, Blue -3) | **INCOMPLETE** (list gaps)
 
diff --git a/plugins/exodia/skills/eight-gates/SKILL.md b/plugins/exodia/skills/eight-gates/SKILL.md
@@ -214,7 +214,7 @@ Token costs tracked via OTel, not here.
 **FALLBACK MODES:**
 
 - Smart scripts unavailable → inline checkpointing via TodoWrite
-- Agent Teams unavailable → Task tool with `subagent_type: general-purpose`
+- Teams API unavailable → fall back to `Task` tool without `team_name` (no shared task list or messaging)
 - Trivial scope (S estimate) → compress Gates 3-5 into minimal checkpoints
   (mark as "bypassed-trivial" with rationale), then proceed to Gate 6-7
 
diff --git a/plugins/exodia/skills/eight-gates/templates/gate-07-execute.md b/plugins/exodia/skills/eight-gates/templates/gate-07-execute.md

Original file line number	Diff line number	Diff line change
`@@ -52,8 +52,8 @@`
`52`	`52`	`},`
`53`	`53`	`{`
`54`	`54`	`"name": "exodia",`
`55`		`- "description": "Multi-agent workflow orchestration (9 commands + 2 skills: eight-gates, hades): fix, turbo-fix, fix-pipeline, tournament, mega-swarm, deep-think, batch-implement, red-blue-review, baryon-mode.",`
`56`		`- "version": "2.0.0",`
	`55`	`+ "description": "Multi-agent workflow orchestration (9 commands + 2 skills: eight-gates, hades): fix, turbo-fix, fix-pipeline, tournament, mega-swarm, deep-think, batch-implement, red-blue-review, baryon-mode. Hades, eight-gates, and red-blue-review use Teams API for reactive collaboration.",`
	`56`	`+ "version": "2.1.0",`
`57`	`57`	`"source": "./plugins/exodia"`
`58`	`58`	`}`
`59`	`59`	`]`