ouroboros-ai 0.43.2.dev1__tar.gz → 0.43.3.dev1__tar.gz

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
     {
       "name": "ouroboros",
       "description": "Transform vague ideas into validated seed specifications through Socratic questioning and ontological analysis",
-      "version": "0.43.2",
+      "version": "0.43.3",
       "author": {
         "name": "Q00",
         "email": "jqyu.lee@gmail.com"

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "ouroboros",
-  "version": "0.43.2",
+  "version": "0.43.3",
   "description": "Self-improving AI workflow system. Crystallize requirements before execution with Socratic interview, ambiguity scoring, and 3-stage evaluation.",
   "author": {
     "name": "Q00",

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/.claude-plugin/skills/interview/SKILL.md

@@ -171,10 +171,13 @@ MCP (question generator) ←→ You (answerer + router) ←→ User (human judgm
    it must not hide, replace, or delay the question itself.
 
    Run the advisory lanes in parallel using your host's native subagent
-   primitive. In Claude Code that primitive is the **Task/Agent tool** — spawn
-   one subagent per lane in a single batch so they run concurrently. Only fall
-   back to the request's `sequential_fallback` semantics when the host has no
-   subagent primitive at all. The standard lanes are:
+   mechanism. In Claude Code that mechanism is the **Task/Agent tool** — spawn
+   one subagent per lane in a single batch so they run concurrently. In Codex,
+   explicitly start a native subagent workflow in natural language: spawn one
+   Codex subagent per lane, pass that lane's payload prompt, wait for all
+   agents, then synthesize. Only fall back to the request's
+   `sequential_fallback` semantics when the host has no subagent mechanism at
+   all. The standard lanes are:
    - `code_context` — inspect repo-local facts and reuse
      `meta.code_investigation_request` when present.
    - `web_context` — browse/search only when current external facts genuinely
@@ -193,9 +196,11 @@ MCP (question generator) ←→ You (answerer + router) ←→ User (human judgm
    When `meta.question_advisory_subagents` is present you MUST fan out: treat
    each entry as a spawn-ready advisory payload with `title`, `agent`, `prompt`,
    and `context`, and dispatch every payload through your host's native subagent
-   primitive (Claude Code → one Task/Agent call per payload, all in one parallel
-   batch) instead of reconstructing prompts from prose. This is required
-   regardless of dispatch mode: the payloads themselves are the spawn signal.
+   mechanism (Claude Code → one Task/Agent call per payload in one parallel
+   batch; Codex → explicitly spawn one Codex subagent per payload, wait for all
+   results, then synthesize) instead of reconstructing prompts from prose. This
+   is required regardless of dispatch mode: the payloads themselves are the
+   spawn signal.
    Treat `meta.question_advisory_host_action=spawn_subagents`, when present, as
    a reinforcing cue for host-driven runtimes such as Codex or Claude Code, not
    as a prerequisite. The only time you skip spawning is when the host has no

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/.claude-plugin/skills/unstuck/SKILL.md

@@ -151,7 +151,9 @@ Driving fan-out from this dispatch block — instead of from the joined human-di
 
 1. Locate the dispatch block at the end of the MCP response's `content` text — the substring between `<!-- ouroboros-lateral-inline-dispatch-v1 base64\n` and `\n-->`. Base64-decode the captured body, then `JSON.parse` to get `{dispatch_mode, persona_count, payloads}`. If the block is missing (older handler that pre-dates the v1 sentinel), fall through to the constrained-runtime path below — *do not* split the joined text.
 2. Surface a short "what the lateral toolkit suggests" header to the user, with the markdown above the dispatch block, so the MCP call is visible as a real product surface, not silent.
-3. In a **single message**, emit N parallel `Task` calls (`general-purpose` subagent) — one per entry in `payloads`. Each Task receives the payload's `prompt` verbatim plus the payload's `context` so the persona is grounded. Strict isolation per Task. The user sees "Running N agents…".
+3. Spawn N persona subagents in parallel — one per entry in `payloads` — using your host's native subagent mechanism. Each child receives the payload's `prompt` verbatim plus the payload's `context` so the persona is grounded, with strict isolation per child. The user sees "Running N agents…".
+   - **Claude Code** → emit N `Task` calls (`general-purpose` subagent) in a **single message** so they run concurrently.
+   - **Codex** → Codex subagents are triggered by explicit natural-language delegation, not a callable tool name (do NOT call `multi_agent_v1.spawn_agent` — it does not exist). In one turn, explicitly spawn one Codex subagent per payload, hand each child its payload `prompt` + `context`, wait for all children, then continue.
 4. Wait for all N to return.
 5. (Optional) **Round 2 cross-attack** — only if Round 1 answers diverge meaningfully. Dispatch a second N-fan-out where each persona receives short summaries of the other answers and is asked: "Identify one weakness in each. ≤200 words." Skip if Round 1 already converges.
 6. Synthesize per the **Synthesize** block below.
@@ -226,8 +228,9 @@ with 2 tables, you haven't found the core feature yet.
                   followed by hidden <!-- ouroboros-lateral-inline-dispatch-v1 base64
                   <base64 body> --> sentinel block]
 [Extract via the sentinel; base64-decode then JSON.parse → payloads[]]
-[Single message: 5 parallel Task calls — one per payloads[] entry,
- each Task receives payload.prompt + payload.context verbatim;
+[Spawn 5 parallel persona subagents — one per payloads[] entry — via the
+ host's native mechanism (Claude: Task calls; Codex: explicit NL delegation);
+ each child receives payload.prompt + payload.context verbatim;
  the visible markdown is shown to the user as the lateral scaffold]
 [User sees "Running 5 agents…"]
 [Round 1 returns]

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ouroboros-ai
-Version: 0.43.2.dev1
+Version: 0.43.3.dev1
 Summary: Specification-first workflow engine for AI coding agents. Works with Claude Code and Codex CLI.
 Author-email: Q00 <jqyu.lee@gmail.com>
 License-File: LICENSE

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/scripts/install.sh

@@ -4,7 +4,7 @@
 #
 # Runtime selection (first match wins):
 #   1. OUROBOROS_INSTALL_RUNTIME env var
-#      (claude|codex|opencode|hermes|gemini|kiro|copilot|pi|all)
+#      (claude|codex|opencode|hermes|gemini|goose|kiro|copilot|pi|gjc|all)
 #   2. Existing ~/.ouroboros/config.yaml runtime — preserved on upgrade
 #      unless OUROBOROS_INSTALL_RECONFIGURE=1 (or --reconfigure flag) is set.
 #   3. Interactive prompt when stdin is a TTY.
@@ -214,7 +214,7 @@ if [ "$HAS_UV" = false ]; then
 fi
 
 # 2. Detect runtimes
-_step "2/4  Choosing an agent backend" "Codex, Claude, Hermes, OpenCode, Gemini, Kiro, Copilot, and Pi are supported."
+_step "2/4  Choosing an agent backend" "Codex, Claude, Hermes, OpenCode, Gemini, Goose, Kiro, Copilot, Pi, and GJC are supported."
 EXTRAS=""
 RUNTIME=""
 HAS_CODEX=false
@@ -222,9 +222,11 @@ HAS_CLAUDE=false
 HAS_HERMES=false
 HAS_OPENCODE=false
 HAS_GEMINI=false
+HAS_GOOSE=false
 HAS_KIRO=false
 HAS_COPILOT=false
 HAS_PI=false
+HAS_GJC=false
 if command -v codex &>/dev/null; then
   _ok "Codex found: $(which codex)"
   HAS_CODEX=true
@@ -245,6 +247,10 @@ if command -v gemini &>/dev/null; then
   _ok "Gemini found: $(which gemini)"
   HAS_GEMINI=true
 fi
+if command -v goose &>/dev/null; then
+  _ok "Goose found: $(which goose)"
+  HAS_GOOSE=true
+fi
 if command -v kiro-cli &>/dev/null; then
   _ok "Kiro found: $(which kiro-cli)"
   HAS_KIRO=true
@@ -257,6 +263,10 @@ if command -v pi &>/dev/null; then
   _ok "Pi: $(which pi)"
   HAS_PI=true
 fi
+if command -v gjc &>/dev/null; then
+  _ok "GJC found: $(which gjc)"
+  HAS_GJC=true
+fi
 
 RUNTIME_COUNT=0
 [ "$HAS_CLAUDE" = true ] && RUNTIME_COUNT=$((RUNTIME_COUNT + 1))
@@ -264,9 +274,11 @@ RUNTIME_COUNT=0
 [ "$HAS_HERMES" = true ] && RUNTIME_COUNT=$((RUNTIME_COUNT + 1))
 [ "$HAS_OPENCODE" = true ] && RUNTIME_COUNT=$((RUNTIME_COUNT + 1))
 [ "$HAS_GEMINI" = true ] && RUNTIME_COUNT=$((RUNTIME_COUNT + 1))
+[ "$HAS_GOOSE" = true ] && RUNTIME_COUNT=$((RUNTIME_COUNT + 1))
 [ "$HAS_KIRO" = true ] && RUNTIME_COUNT=$((RUNTIME_COUNT + 1))
 [ "$HAS_COPILOT" = true ] && RUNTIME_COUNT=$((RUNTIME_COUNT + 1))
 [ "$HAS_PI" = true ] && RUNTIME_COUNT=$((RUNTIME_COUNT + 1))
+[ "$HAS_GJC" = true ] && RUNTIME_COUNT=$((RUNTIME_COUNT + 1))
 
 # Map a runtime name to (EXTRAS, RUNTIME) pair.
 # Used after explicit/preserved runtime resolution to derive install extras.
@@ -282,14 +294,16 @@ _runtime_to_extras() {
     opencode) EXTRAS="[tui]"; RUNTIME="opencode" ;;
     hermes)  EXTRAS="[mcp,tui]"; RUNTIME="hermes" ;;
     gemini)  EXTRAS="[tui]"; RUNTIME="gemini" ;;
+    goose)   EXTRAS="[tui]"; RUNTIME="goose" ;;
     kiro)    EXTRAS="[tui]"; RUNTIME="kiro" ;;
     copilot) EXTRAS="[tui]"; RUNTIME="copilot" ;;
     pi)      EXTRAS="[tui]"; RUNTIME="pi" ;;
+    gjc)     EXTRAS="[tui]"; RUNTIME="gjc" ;;
     all)     EXTRAS="[all]"; RUNTIME="" ;;
     "")      EXTRAS="[tui]"; RUNTIME="" ;;
     *)
       _err "unsupported runtime '$1'"
-      _info "Expected one of: claude, codex, opencode, hermes, gemini, kiro, copilot, pi, all"
+      _info "Expected one of: claude, codex, opencode, hermes, gemini, goose, kiro, copilot, pi, gjc, all"
       exit 1
       ;;
   esac
@@ -306,7 +320,7 @@ FRESH_CONFIG=true
 if [ -z "$EXPLICIT_RUNTIME" ] && [ -z "$RECONFIGURE" ] && [ -f "$EXISTING_CONFIG" ] && command -v python3 &>/dev/null; then
   EXISTING_RUNTIME=$(EXISTING_CONFIG="$EXISTING_CONFIG" python3 -c "
 import os, re
-supported = {'claude', 'codex', 'opencode', 'hermes', 'gemini', 'kiro', 'copilot', 'pi'}
+supported = {'claude', 'codex', 'opencode', 'hermes', 'gemini', 'goose', 'kiro', 'copilot', 'pi', 'gjc'}
 try:
     lines = open(os.environ['EXISTING_CONFIG']).read().splitlines()
     in_orchestrator = False
@@ -344,10 +358,12 @@ elif [ "$RUNTIME_COUNT" -gt 1 ]; then
     _choice 3 "Hermes" "Hermes agent guides + MCP server (${PACKAGE_NAME}[mcp,tui])"
     _choice 4 "OpenCode" "OpenCode commands and agent files (${PACKAGE_NAME}[tui])"
     _choice 5 "Gemini" "Gemini CLI integration (${PACKAGE_NAME}[tui])"
-    _choice 6 "Kiro" "Kiro CLI integration (${PACKAGE_NAME}[tui])"
-    _choice 7 "Copilot" "GitHub Copilot integration (${PACKAGE_NAME}[tui])"
-    _choice 8 "Pi" "Pi CLI bridge and instruction artifacts (${PACKAGE_NAME}[tui])"
-    _choice 9 "All" "Install every optional integration (${PACKAGE_NAME}[all])"
+    _choice 6 "Goose" "Goose CLI integration (${PACKAGE_NAME}[tui])"
+    _choice 7 "Kiro" "Kiro CLI integration (${PACKAGE_NAME}[tui])"
+    _choice 8 "Copilot" "GitHub Copilot integration (${PACKAGE_NAME}[tui])"
+    _choice 9 "Pi" "Pi CLI bridge and instruction artifacts (${PACKAGE_NAME}[tui])"
+    _choice 10 "GJC" "GJC CLI bridge and instruction artifacts (${PACKAGE_NAME}[tui])"
+    _choice 11 "All" "Install every optional integration (${PACKAGE_NAME}[all])"
     _prompt "Select [1]: "
     read -r choice
     case "${choice:-1}" in
@@ -355,10 +371,12 @@ elif [ "$RUNTIME_COUNT" -gt 1 ]; then
       3) _runtime_to_extras "hermes" ;;
       4) _runtime_to_extras "opencode" ;;
       5) _runtime_to_extras "gemini" ;;
-      6) _runtime_to_extras "kiro" ;;
-      7) _runtime_to_extras "copilot" ;;
-      8) _runtime_to_extras "pi" ;;
-      9) _runtime_to_extras "all" ;;
+      6) _runtime_to_extras "goose" ;;
+      7) _runtime_to_extras "kiro" ;;
+      8) _runtime_to_extras "copilot" ;;
+      9) _runtime_to_extras "pi" ;;
+      10) _runtime_to_extras "gjc" ;;
+      11) _runtime_to_extras "all" ;;
       *) _runtime_to_extras "claude" ;;
     esac
   else
@@ -376,12 +394,16 @@ elif [ "$HAS_OPENCODE" = true ] && [ "$RUNTIME_COUNT" -eq 1 ]; then
   _runtime_to_extras "opencode"
 elif [ "$HAS_GEMINI" = true ] && [ "$RUNTIME_COUNT" -eq 1 ]; then
   _runtime_to_extras "gemini"
+elif [ "$HAS_GOOSE" = true ] && [ "$RUNTIME_COUNT" -eq 1 ]; then
+  _runtime_to_extras "goose"
 elif [ "$HAS_KIRO" = true ] && [ "$RUNTIME_COUNT" -eq 1 ]; then
   _runtime_to_extras "kiro"
 elif [ "$HAS_COPILOT" = true ] && [ "$RUNTIME_COUNT" -eq 1 ]; then
   _runtime_to_extras "copilot"
 elif [ "$HAS_PI" = true ] && [ "$RUNTIME_COUNT" -eq 1 ]; then
   _runtime_to_extras "pi"
+elif [ "$HAS_GJC" = true ] && [ "$RUNTIME_COUNT" -eq 1 ]; then
+  _runtime_to_extras "gjc"
 else
   # No runtime CLI on PATH yet — first install. Always prompt when interactive
   # so the user picks deliberately rather than silently defaulting to claude.
@@ -393,10 +415,12 @@ else
     _choice 3 "Hermes" "Hermes agent guides + MCP server (${PACKAGE_NAME}[mcp,tui])"
     _choice 4 "OpenCode" "OpenCode commands and agent files (${PACKAGE_NAME}[tui])"
     _choice 5 "Gemini" "Gemini CLI integration (${PACKAGE_NAME}[tui])"
-    _choice 6 "Kiro" "Kiro CLI integration (${PACKAGE_NAME}[tui])"
-    _choice 7 "Copilot" "GitHub Copilot integration (${PACKAGE_NAME}[tui])"
-    _choice 8 "Pi" "Pi CLI bridge and instruction artifacts (${PACKAGE_NAME}[tui])"
-    _choice 9 "All" "Install every optional integration (${PACKAGE_NAME}[all])"
+    _choice 6 "Goose" "Goose CLI integration (${PACKAGE_NAME}[tui])"
+    _choice 7 "Kiro" "Kiro CLI integration (${PACKAGE_NAME}[tui])"
+    _choice 8 "Copilot" "GitHub Copilot integration (${PACKAGE_NAME}[tui])"
+    _choice 9 "Pi" "Pi CLI bridge and instruction artifacts (${PACKAGE_NAME}[tui])"
+    _choice 10 "GJC" "GJC CLI bridge and instruction artifacts (${PACKAGE_NAME}[tui])"
+    _choice 11 "All" "Install every optional integration (${PACKAGE_NAME}[all])"
     _choice 0 "None" "Base CLI only; choose a backend later"
     _prompt "Select [1]: "
     read -r choice
@@ -406,17 +430,19 @@ else
       3) _runtime_to_extras "hermes" ;;
       4) _runtime_to_extras "opencode" ;;
       5) _runtime_to_extras "gemini" ;;
-      6) _runtime_to_extras "kiro" ;;
-      7) _runtime_to_extras "copilot" ;;
-      8) _runtime_to_extras "pi" ;;
-      9) _runtime_to_extras "all" ;;
+      6) _runtime_to_extras "goose" ;;
+      7) _runtime_to_extras "kiro" ;;
+      8) _runtime_to_extras "copilot" ;;
+      9) _runtime_to_extras "pi" ;;
+      10) _runtime_to_extras "gjc" ;;
+      11) _runtime_to_extras "all" ;;
       *) _runtime_to_extras "claude" ;;
     esac
   else
     # Pipe mode (curl | bash): install base package, skip runtime-specific setup.
     _blank
     _warn "No runtime detected in non-interactive mode; installing the base package."
-    _info "Pick a backend afterwards with: ouroboros setup --runtime <claude|codex|opencode|hermes|gemini|kiro|copilot|pi>"
+    _info "Pick a backend afterwards with: ouroboros setup --runtime <claude|codex|opencode|hermes|gemini|goose|kiro|copilot|pi|gjc>"
     _runtime_to_extras ""
   fi
 fi
@@ -530,10 +556,16 @@ if [ -z "$OUROBOROS_BIN" ] && ! command -v ouroboros &>/dev/null; then
   done
 fi
 
+OUROBOROS_SETUP_CMD=""
+if [ -n "$OUROBOROS_BIN" ]; then
+  OUROBOROS_SETUP_CMD="$OUROBOROS_BIN"
+elif command -v ouroboros &>/dev/null; then
+  OUROBOROS_SETUP_CMD="ouroboros"
+fi
+
 # 4. Setup (ouroboros CLI configures runtime-specific integration)
 _step "4/4  Wiring local integrations" "Creates config and runtime-specific files when a backend was selected."
-if [ -n "$RUNTIME" ] && { [ -n "$OUROBOROS_BIN" ] || command -v ouroboros &>/dev/null; }; then
-  OUROBOROS_SETUP_CMD="${OUROBOROS_BIN:-ouroboros}"
+if [ -n "$RUNTIME" ] && [ -n "$OUROBOROS_SETUP_CMD" ]; then
   _info "Running: $OUROBOROS_SETUP_CMD setup --runtime $RUNTIME --non-interactive"
   "$OUROBOROS_SETUP_CMD" setup --runtime "$RUNTIME" --non-interactive || true
 elif [ -n "$RUNTIME" ]; then
@@ -542,9 +574,24 @@ else
   _info "No backend selected; skipping runtime setup."
 fi
 
+# Refresh Codex rules/skills whenever this machine appears to use Codex, even
+# if the preserved primary runtime is another backend. Setup already does this
+# for `--runtime codex`, but upgrades and `all` installs should not leave stale
+# ~/.codex artifacts behind.
+if [ -n "$OUROBOROS_SETUP_CMD" ] && {
+  [ "$RUNTIME" = "codex" ] ||
+    [ "$EXTRAS" = "[all]" ] ||
+    [ "$HAS_CODEX" = true ] ||
+    [ -d "$HOME/.codex" ]
+}; then
+  _info "Refreshing Codex rules and skills"
+  "$OUROBOROS_SETUP_CMD" codex refresh || _warn "Codex artifact refresh skipped; run: ouroboros codex refresh"
+fi
+
 # 5. Claude Code integration (MCP + skills)
-# Only apply Claude-specific integration when Claude was the selected runtime,
-# or when the user explicitly asked for the all-runtimes install.
+# MCP registration changes Claude's tool wiring, so keep it tied to Claude/all.
+# Skill refresh is artifact-only and should happen whenever Claude is present;
+# otherwise a Codex-primary upgrade can leave Claude Code reading stale skills.
 if command -v claude &>/dev/null && { [ "$RUNTIME" = "claude" ] || [ "$EXTRAS" = "[all]" ]; }; then
   _blank
   _say "${BLUE}◆${RESET} ${BOLD}Claude Code extras${RESET}"
@@ -616,7 +663,13 @@ with open(mcp_file, 'w') as f:
   else
     _warn "MCP skipped: no python3 found. Add the entry manually to $MCP_FILE."
   fi
+fi
 
+if command -v claude &>/dev/null; then
+  if ! { [ "$RUNTIME" = "claude" ] || [ "$EXTRAS" = "[all]" ]; }; then
+    _blank
+    _say "${BLUE}◆${RESET} ${BOLD}Claude Code skills${RESET}"
+  fi
   # 5b. Install/update Ouroboros skills (claude plugin)
   _info "Installing Ouroboros skills via Claude plugin marketplace..."
   claude plugin marketplace add Q00/ouroboros 2>/dev/null || true
@@ -637,7 +690,7 @@ _info 'Or from the terminal: ouroboros init start "your idea here"'
 if [ -n "$RUNTIME" ]; then
   _info "Current backend: $RUNTIME"
 fi
-_info "Switch backend later: ouroboros setup --runtime <claude|codex|opencode|hermes|gemini|kiro|copilot|pi>"
+_info "Switch backend later: ouroboros setup --runtime <claude|codex|opencode|hermes|gemini|goose|kiro|copilot|pi|gjc>"
 _say "${BOLD}Settings GUI — pick per-stage agents & models${RESET}"
 _info 'Inside your AI agent: > ooo config   (opens in your browser)'
 _info 'From this terminal:  ouroboros config   (full-screen TUI)'

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/skills/interview/SKILL.md

@@ -171,10 +171,13 @@ MCP (question generator) ←→ You (answerer + router) ←→ User (human judgm
    it must not hide, replace, or delay the question itself.
 
    Run the advisory lanes in parallel using your host's native subagent
-   primitive. In Claude Code that primitive is the **Task/Agent tool** — spawn
-   one subagent per lane in a single batch so they run concurrently. Only fall
-   back to the request's `sequential_fallback` semantics when the host has no
-   subagent primitive at all. The standard lanes are:
+   mechanism. In Claude Code that mechanism is the **Task/Agent tool** — spawn
+   one subagent per lane in a single batch so they run concurrently. In Codex,
+   explicitly start a native subagent workflow in natural language: spawn one
+   Codex subagent per lane, pass that lane's payload prompt, wait for all
+   agents, then synthesize. Only fall back to the request's
+   `sequential_fallback` semantics when the host has no subagent mechanism at
+   all. The standard lanes are:
    - `code_context` — inspect repo-local facts and reuse
      `meta.code_investigation_request` when present.
    - `web_context` — browse/search only when current external facts genuinely
@@ -193,9 +196,11 @@ MCP (question generator) ←→ You (answerer + router) ←→ User (human judgm
    When `meta.question_advisory_subagents` is present you MUST fan out: treat
    each entry as a spawn-ready advisory payload with `title`, `agent`, `prompt`,
    and `context`, and dispatch every payload through your host's native subagent
-   primitive (Claude Code → one Task/Agent call per payload, all in one parallel
-   batch) instead of reconstructing prompts from prose. This is required
-   regardless of dispatch mode: the payloads themselves are the spawn signal.
+   mechanism (Claude Code → one Task/Agent call per payload in one parallel
+   batch; Codex → explicitly spawn one Codex subagent per payload, wait for all
+   results, then synthesize) instead of reconstructing prompts from prose. This
+   is required regardless of dispatch mode: the payloads themselves are the
+   spawn signal.
    Treat `meta.question_advisory_host_action=spawn_subagents`, when present, as
    a reinforcing cue for host-driven runtimes such as Codex or Claude Code, not
    as a prerequisite. The only time you skip spawning is when the host has no

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/skills/setup/SKILL.md

@@ -282,7 +282,7 @@ A backup will be created: CLAUDE.md.bak
 **If "Preview first", show:**
 ````markdown
 <!-- ooo:START -->
-<!-- ooo:VERSION:0.43.2 -->
+<!-- ooo:VERSION:0.43.3 -->
 # Ouroboros — Specification-First AI Development
 
 > Before telling AI what to build, define what should be built.

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/skills/unstuck/SKILL.md

@@ -151,7 +151,9 @@ Driving fan-out from this dispatch block — instead of from the joined human-di
 
 1. Locate the dispatch block at the end of the MCP response's `content` text — the substring between `<!-- ouroboros-lateral-inline-dispatch-v1 base64\n` and `\n-->`. Base64-decode the captured body, then `JSON.parse` to get `{dispatch_mode, persona_count, payloads}`. If the block is missing (older handler that pre-dates the v1 sentinel), fall through to the constrained-runtime path below — *do not* split the joined text.
 2. Surface a short "what the lateral toolkit suggests" header to the user, with the markdown above the dispatch block, so the MCP call is visible as a real product surface, not silent.
-3. In a **single message**, emit N parallel `Task` calls (`general-purpose` subagent) — one per entry in `payloads`. Each Task receives the payload's `prompt` verbatim plus the payload's `context` so the persona is grounded. Strict isolation per Task. The user sees "Running N agents…".
+3. Spawn N persona subagents in parallel — one per entry in `payloads` — using your host's native subagent mechanism. Each child receives the payload's `prompt` verbatim plus the payload's `context` so the persona is grounded, with strict isolation per child. The user sees "Running N agents…".
+   - **Claude Code** → emit N `Task` calls (`general-purpose` subagent) in a **single message** so they run concurrently.
+   - **Codex** → Codex subagents are triggered by explicit natural-language delegation, not a callable tool name (do NOT call `multi_agent_v1.spawn_agent` — it does not exist). In one turn, explicitly spawn one Codex subagent per payload, hand each child its payload `prompt` + `context`, wait for all children, then continue.
 4. Wait for all N to return.
 5. (Optional) **Round 2 cross-attack** — only if Round 1 answers diverge meaningfully. Dispatch a second N-fan-out where each persona receives short summaries of the other answers and is asked: "Identify one weakness in each. ≤200 words." Skip if Round 1 already converges.
 6. Synthesize per the **Synthesize** block below.
@@ -226,8 +228,9 @@ with 2 tables, you haven't found the core feature yet.
                   followed by hidden <!-- ouroboros-lateral-inline-dispatch-v1 base64
                   <base64 body> --> sentinel block]
 [Extract via the sentinel; base64-decode then JSON.parse → payloads[]]
-[Single message: 5 parallel Task calls — one per payloads[] entry,
- each Task receives payload.prompt + payload.context verbatim;
+[Spawn 5 parallel persona subagents — one per payloads[] entry — via the
+ host's native mechanism (Claude: Task calls; Codex: explicit NL delegation);
+ each child receives payload.prompt + payload.context verbatim;
  the visible markdown is shown to the user as the lateral scaffold]
 [User sees "Running 5 agents…"]
 [Round 1 returns]

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/src/ouroboros/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.43.2.dev1'
-__version_tuple__ = version_tuple = (0, 43, 2, 'dev1')
+__version__ = version = '0.43.3.dev1'
+__version_tuple__ = version_tuple = (0, 43, 3, 'dev1')
 
 __commit_id__ = commit_id = None

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/src/ouroboros/backends/__init__.py

@@ -5,6 +5,7 @@ from ouroboros.backends.capabilities import (
     RuntimeSubagentOrchestrationContract,
     SkillExecutionCapability,
     SubagentDispatchMode,
+    SubagentSpawnTriggerMechanism,
     backend_supports_tool_envelope,
     build_runtime_subagent_orchestration_contract,
     get_backend_capability,
@@ -25,6 +26,7 @@ __all__ = [
     "RuntimeSubagentOrchestrationContract",
     "SkillExecutionCapability",
     "SubagentDispatchMode",
+    "SubagentSpawnTriggerMechanism",
     "backend_supports_tool_envelope",
     "build_runtime_subagent_orchestration_contract",
     "get_backend_capability",

{ouroboros_ai-0.43.2.dev1 → ouroboros_ai-0.43.3.dev1}/src/ouroboros/backends/capabilities.py

@@ -43,6 +43,10 @@ class BackendCapability:
     # backend can have the latter without the former. Do NOT conflate them.
     supports_native_parallel_subagents: bool = False
     supports_host_driven_subagents: bool = False
+    host_driven_subagent_mechanism: SubagentSpawnTriggerMechanism | None = None
+    host_driven_subagent_requires_explicit_request: bool = False
+    host_driven_callable_spawn_tool_name: str | None = None
+    prohibited_subagent_spawn_tool_names: tuple[str, ...] = ()
     skill_execution_capabilities: tuple[SkillExecutionCapability, ...] = ()
 
     @property
@@ -73,6 +77,15 @@ class SubagentDispatchMode(StrEnum):
     SEQUENTIAL = "sequential"
 
 
+class SubagentSpawnTriggerMechanism(StrEnum):
+    """Concrete runtime mechanism used after dispatch mode is resolved."""
+
+    PASSIVE_BRIDGE_ENVELOPE = "passive_bridge_envelope"
+    CODEX_NATURAL_LANGUAGE_DELEGATION = "codex_natural_language_delegation"
+    CLAUDE_TASK_AGENT_TOOL = "claude_task_agent_tool"
+    SEQUENTIAL_FALLBACK = "sequential_fallback"
+
+
 @dataclass(frozen=True, slots=True)
 class RuntimeSubagentOrchestrationContract:
     """Runtime handling contract for MCP subagent directive metadata."""
@@ -80,6 +93,10 @@ class RuntimeSubagentOrchestrationContract:
     backend_name: str
     supports_native_parallel_subagents: bool
     dispatch_mode: str
+    spawn_trigger_mechanism: str
+    requires_explicit_spawn_request: bool
+    callable_spawn_tool_name: str | None
+    prohibited_spawn_tool_names: tuple[str, ...]
     mcp_directive_keys: tuple[str, ...]
     sequential_fallback: Mapping[str, Any]
     runtime_instruction_handling: str
@@ -91,6 +108,10 @@ class RuntimeSubagentOrchestrationContract:
             "backend_name": self.backend_name,
             "supports_native_parallel_subagents": self.supports_native_parallel_subagents,
             "dispatch_mode": self.dispatch_mode,
+            "spawn_trigger_mechanism": self.spawn_trigger_mechanism,
+            "requires_explicit_spawn_request": self.requires_explicit_spawn_request,
+            "callable_spawn_tool_name": self.callable_spawn_tool_name,
+            "prohibited_spawn_tool_names": list(self.prohibited_spawn_tool_names),
             "mcp_directive_keys": list(self.mcp_directive_keys),
             "sequential_fallback": dict(self.sequential_fallback),
             "runtime_instruction_handling": self.runtime_instruction_handling,
@@ -186,9 +207,12 @@ _CODEX_SKILL_EXECUTION_CAPABILITIES: tuple[SkillExecutionCapability, ...] = (
             "`question_advisory_dispatch_mode=host_driven` in the result `meta`, or "
             "the `ouroboros-lateral-inline-dispatch-v1` content block), do NOT just "
             "read the inline text: spawn one subagent per entry in the payload array "
-            "(`payloads` or `question_advisory_subagents`) using your native "
-            "multi-agent spawn primitive (`multi_agent_v1.spawn_agent`), passing "
-            "each payload's `prompt`. Correlate every child result by the "
+            "(`payloads` or `question_advisory_subagents`) using Codex's native "
+            "subagent workflow. Codex subagents are triggered by an explicit "
+            "natural-language delegation, not by a callable tool name: explicitly "
+            "spawn one Codex subagent per payload, give each child the payload's "
+            "`prompt`, wait for all children, then summarize the results. Correlate "
+            "every child result by the "
             "payload-specific key named in the result `meta` "
             "(`result_correlation_key`): lateral payloads use `context.persona`, "
             "interview advisory payloads use `context.lane_id` (their `persona` is "
@@ -293,6 +317,9 @@ _CAPABILITIES: tuple[BackendCapability, ...] = (
         cli_config_key="cli_path",
         skill_execution_capabilities=_CLAUDE_SKILL_EXECUTION_CAPABILITIES,
         supports_host_driven_subagents=True,
+        host_driven_subagent_mechanism=SubagentSpawnTriggerMechanism.CLAUDE_TASK_AGENT_TOOL,
+        host_driven_subagent_requires_explicit_request=True,
+        host_driven_callable_spawn_tool_name="Task/Agent",
     ),
     BackendCapability(
         name="codex",
@@ -305,6 +332,12 @@ _CAPABILITIES: tuple[BackendCapability, ...] = (
         cli_config_key="codex_cli_path",
         skill_execution_capabilities=_CODEX_SKILL_EXECUTION_CAPABILITIES,
         supports_host_driven_subagents=True,
+        host_driven_subagent_mechanism=(
+            SubagentSpawnTriggerMechanism.CODEX_NATURAL_LANGUAGE_DELEGATION
+        ),
+        host_driven_subagent_requires_explicit_request=True,
+        host_driven_callable_spawn_tool_name=None,
+        prohibited_subagent_spawn_tool_names=("multi_agent_v1.spawn_agent",),
     ),
     BackendCapability(
         name="copilot",
@@ -478,6 +511,10 @@ def build_runtime_subagent_orchestration_contract(
     dispatch_mode = mode.value
 
     if mode is SubagentDispatchMode.PLUGIN_PASSIVE:
+        spawn_trigger_mechanism = SubagentSpawnTriggerMechanism.PASSIVE_BRIDGE_ENVELOPE
+        requires_explicit_spawn_request = False
+        callable_spawn_tool_name = None
+        prohibited_spawn_tool_names: tuple[str, ...] = ()
         runtime_instruction_handling = (
             "Consume MCP `_subagent` or `_subagents` directive payloads with the "
             "runtime's native parallel subagent primitive. For OpenCode this "
@@ -486,6 +523,13 @@ def build_runtime_subagent_orchestration_contract(
             "surfaces."
         )
     elif mode is SubagentDispatchMode.HOST_DRIVEN:
+        spawn_trigger_mechanism = (
+            capability.host_driven_subagent_mechanism
+            or SubagentSpawnTriggerMechanism.SEQUENTIAL_FALLBACK
+        )
+        requires_explicit_spawn_request = capability.host_driven_subagent_requires_explicit_request
+        callable_spawn_tool_name = capability.host_driven_callable_spawn_tool_name
+        prohibited_spawn_tool_names = capability.prohibited_subagent_spawn_tool_names
         runtime_instruction_handling = (
             "This runtime has a native subagent primitive but no passive "
             "Ouroboros bridge. Consume the inline `host_driven` dispatch payloads "
@@ -496,6 +540,10 @@ def build_runtime_subagent_orchestration_contract(
             "parallel primitive is available."
         )
     else:
+        spawn_trigger_mechanism = SubagentSpawnTriggerMechanism.SEQUENTIAL_FALLBACK
+        requires_explicit_spawn_request = False
+        callable_spawn_tool_name = None
+        prohibited_spawn_tool_names = ()
         runtime_instruction_handling = (
             "This runtime has no native parallel subagent primitive. Follow the "
             "MCP `sequential_fallback` contract and process each structured "
@@ -509,6 +557,10 @@ def build_runtime_subagent_orchestration_contract(
         # report False here and carry their capability via ``dispatch_mode``.
         supports_native_parallel_subagents=mode is SubagentDispatchMode.PLUGIN_PASSIVE,
         dispatch_mode=dispatch_mode,
+        spawn_trigger_mechanism=spawn_trigger_mechanism.value,
+        requires_explicit_spawn_request=requires_explicit_spawn_request,
+        callable_spawn_tool_name=callable_spawn_tool_name,
+        prohibited_spawn_tool_names=prohibited_spawn_tool_names,
         mcp_directive_keys=("_subagent", "_subagents"),
         sequential_fallback=dict(sequential_fallback),
         runtime_instruction_handling=runtime_instruction_handling,
@@ -642,6 +694,7 @@ __all__ = [
     "RuntimeSubagentOrchestrationContract",
     "SkillExecutionCapability",
     "SubagentDispatchMode",
+    "SubagentSpawnTriggerMechanism",
     "backend_supports_tool_envelope",
     "build_runtime_subagent_orchestration_contract",
     "get_backend_capability",