Unstuck

/ouroboros:unstuck

Break through stagnation with lateral thinking personas. Two modes:

• Solo — one persona reframes the problem (fast, cheap).
• Debate — multiple personas run in parallel as sub-agents and the user picks the verdict (visual, thorough).

Usage

ooo lateral                       # debate (default) — all 5 lateral personas
ooo lateral <persona>             # solo — single persona
ooo lateral debate <p1> <p2> ...  # debate with explicit members
ooo lateral @<preset>             # debate with preset (Phase 1: only @all = 5 personas)

Trigger keywords: "I'm stuck", "think sideways", "ooo lateral", "/ouroboros:unstuck".

Personas (Lateral Pool)

The lateral pool is stateless mindset personas only — five reframing lenses. Stateful roles (evaluator, qa-judge, ontologist, socratic-interviewer, etc.) are NOT mixed into this pool; they have their own SKILLs.

When to Call

Direct user invocation — ooo lateral … from the prompt.

Autonomous chain from another SKILL — when you (the main session) are operating in another SKILL's persona (e.g., socratic-interviewer during ooo interview, or any agent role) and judge that the current question requires multi-perspective deliberation, you MAY invoke this SKILL on your own. No forced trigger; this is your self-assessment. After the debate, summarize the options for the user, return to the original SKILL's flow, and let the user decide. The user will see the sub-agent fan-out as it happens — that visibility is the point.

Instructions

Step 1 — Parse args → mode

Parse the user's argument string (or your autonomous-chain intent) into a mode (solo / debate):

Validate every persona name against the lateral pool above. If invalid, emit a brief error message naming the valid personas — do NOT silently coerce. Multiple persona tokens without the explicit debate keyword are rejected to keep the syntax unambiguous.

Step 2 — Gather required context before the MCP call

ouroboros_lateral_think hard-fails if either problem_context or current_approach is empty (evaluation_handlers.py:1324, 1333). A bare ooo lateral from a fresh session has neither — calling MCP directly would crash before any persona work. Resolve both fields before Step 3, in this order:

1. Reuse session state. If a parent SKILL is invoking this one (autonomous chain) or the current Claude Code / Codex session has clearly recent stuck-point context, extract it. Build:
   • problem_context — what the user is stuck on (1–3 sentences, current state of the world).
   • current_approach — what has been tried so far (1–3 sentences). For a brand-new attempt, this can be "none yet — first attempt".
   • failed_attempts (optional) — short list of prior failures, when the user has volunteered them.
2. If either field is unrecoverable, ask the user one short combined question before going further (this applies to both solo and debate — the handler requires both fields either way):

   "Two things I need before lateral thinking can run: (1) what are you stuck on right now, (2) what have you already tried? A sentence each is enough." Wait for the answer. Do not invent or paraphrase past turns into these fields if you are not certain — current_approach="not specified" is acceptable, fabricated content is not.

3. Only when both fields are populated, proceed to Step 3.

This pre-call branch applies to solo and debate. Skipping it makes ooo lateral (no args, fresh session) reliably error.

Step 3 — Always call ouroboros_lateral_think (routing contract)

Per the ooo routing contract in src/ouroboros/codex/ouroboros.md, every ooo lateral invocation MUST route through the MCP tool — solo and debate, in every runtime. This SKILL never substitutes a direct sub-agent fan-out for the MCP call.

1. Use the active runtime's tool-discovery capability with query "+ouroboros lateral" to load ouroboros_lateral_think (often prefixed, e.g., mcp__plugin_ouroboros_ouroboros__ouroboros_lateral_think). Deferred tools won't appear until runtime tool discovery runs.
2. Invoke the tool with the parsed mode and the context from Step 2:
   • Solo: persona=<one>, problem_context, current_approach, failed_attempts.
   • Debate: personas=[...], problem_context, current_approach, failed_attempts.
3. If the tool is not callable even after discovery — neither already exposed nor loadable (an empty discovery result for an already-exposed tool is expected, not a failure) — stop and report that the MCP dispatch surface is broken — same rule the contract applies to ooo auto. Do not improvise a sub-agent fan-out as a workaround; that bypasses the contract the bot review explicitly flagged.

The MCP call is cheap. The handler's inline path is a deterministic prompt builder — it constructs per-persona reframing prompts via LateralThinker.generate_alternative (src/ouroboros/mcp/tools/evaluation_handlers.py:1444+); it does not run an LLM rollout.

Step 4 — Branch on the handler's response shape

The handler picks one of two response shapes based on should_dispatch_via_plugin(...) (src/ouroboros/mcp/tools/subagent.py:186-218). You do not choose; you observe and act. The envelope key further depends on the mode you called with — solo and debate are not symmetric:

Inline dispatch block (debate, inline response only). The handler appends a versioned, sentinel-bracketed dispatch block to the end of content as a compatibility fallback for clients and runtimes that consume only textual content. MCP SDK v2 preserves structured metadata, while the inline block lets text-only consumers recover the same canonical payloads (src/ouroboros/mcp/server/adapter.py:923, src/ouroboros/mcp/tools/subagent.py:141-144). Format:

<!-- ouroboros-lateral-inline-dispatch-v1 base64
<base64-encoded-JSON>
-->

Two pieces matter:

• Hidden HTML comment — markdown viewers render nothing for <!-- ... -->, so the block doesn't pollute the human-visible output.
• Base64 body — base64's alphabet is [A-Za-z0-9+/=], which can never produce the sequence -->. So even if a user-supplied problem_context or current_approach contains --> (HTML/JS debugging is the obvious case), the encoded body cannot prematurely close the wrapper and leak the dispatch into the visible markdown.

Decoded, the body is JSON. An MCP host that did not declare subagent capability receives the neutral shape:

{"dispatch_mode": "host_decides", "host_action": "dispatch_subagents_if_supported", "execution_preference": "parallel", "fallback_strategy": "sequential", "persona_count": N, "payloads": [...]}

An explicitly sequential execution authority receives dispatch_mode="sequential" with legacy_dispatch_mode="inline_fallback" for compatibility. To recover: locate the substring between <!-- ouroboros-lateral-inline-dispatch-v1 base64\n and \n--> at the end of content, base64-decode it, then JSON.parse.

Shape A — dispatch_mode = "plugin" (OpenCode plugin mode only)

The plugin runtime spawns Task panes automatically from whichever envelope the handler emitted. You only need to read the right key and await the result(s):

Solo (plugin)

The response carries a single _subagent object (singular) — {tool_name, title, prompt, agent, model, context} — produced by build_subagent_result (evaluation_handlers.py:1554+). The plugin spawns one Task pane. Await its single result, then present the persona's reframing.

Debate (plugin)

The response carries a _subagents array (plural) — [{tool_name, title, prompt, agent, model, context}, ...] — produced by build_multi_subagent_result (evaluation_handlers.py:1419+). The plugin spawns N Task panes in parallel. Await all N results, then synthesize per the Synthesize block below.

If you expected plugin mode but the response is inline text (neither _subagent nor _subagents), you are not actually in plugin mode — fall through to Shape B; do not wait for an envelope that will not arrive.

Shape B — inline response (Claude Code, Codex CLI, OpenCode subprocess, every other runtime)

The handler ran the prompt builder internally and returned ready-to-use markdown:

• Solo response: a single # Lateral Thinking: <approach> block followed by the reframing prompt.
• Debate response: N such blocks concatenated with \n\n---\n\n separators. The dispatch block distinguishes declared host_driven, capability-neutral host_decides, and explicit sequential execution.

Solo (any runtime)

Present the persona's approach summary, reframing prompt, questions to consider, and a ◆ Current state → next: suggestion routing back to the workflow.

Debate, runtime supports sub-agent dispatch (Claude Code Task tool, Codex CLI sub-agent, etc.)

This is the default debate UX for Claude Code and Codex. The MCP call has already happened (Step 3), so the routing contract is satisfied; this step only changes how the already-built canonical prompts are rendered to the user.

Recover the structured dispatch from the inline dispatch block appended to content (see the "Inline dispatch block" note above the table). Each entry under payloads[] is a {tool_name, title, prompt, agent, model, context} dict; prompt is self-contained — it carries the same canonical reframing plus the "Task for you (subagent)" wrapper that plugin mode dispatches via _subagents (asking for a concrete plan, the biggest assumption challenged, and a one-line verdict). Same builder, byte-identical prompts across runtimes.

Driving fan-out from this dispatch block — instead of from the joined human-display text — avoids two pitfalls the bot review surfaced:

• Separator collision — \n\n---\n\n can legitimately appear inside a user-supplied problem_context or current_approach, so splitting the joined text would over-fragment and corrupt prompts. The dispatch block uses a unique versioned sentinel that user content cannot collide with.
• Behavioral drift across runtimes — the dispatch block carries the same canonical payloads _subagents carries, so debate results don't diverge by environment.

1. Locate the dispatch block at the end of the MCP response's content text. Base64-decode the captured body, then JSON.parse to get {dispatch_mode, host_action, persona_count, payloads}. If the block is missing, 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. If host_action="dispatch_subagents_if_supported", use the host's native parallel mechanism when available and otherwise process the same payloads sequentially. If host_action="spawn_subagents", spawn N persona subagents in parallel. Each child receives the payload's prompt verbatim plus its context, with strict isolation per child.
   • Claude Code → emit N Task calls (general-purpose subagent) in a single message so they run concurrently when the Task surface is available.
   • Codex → explicitly delegate one Codex subagent per payload in one turn; do not call a nonexistent multi_agent_v1.spawn_agent tool.
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.

Debate, constrained runtime without sub-agent dispatch

If the response is stamped with dispatch_mode="sequential", or with dispatch_mode="host_decides" and the host has no native parallel primitive, process each payload in order. Correlate every result by result_correlation_key (normally context.persona), then synthesize after the last payload. Treat legacy_dispatch_mode="inline_fallback" as compatibility metadata only.

Debate, runtime cannot dispatch sub-agents (constrained subprocess, no Task surface)

Present the concatenated markdown the handler returned, as-is — no parsing required, the user reads the whole text. There is no per-persona visualization and no Round 2 cross-attack — both require a sub-agent surface. Synthesize directly from the inline text.

Synthesize (debate, all shapes)

Do not auto-emit a verdict. Present:

## Debate result — N personas

### Options
- **Option A** (from <persona>): <one-liner>
- **Option B** (from <persona>): <one-liner>
- …

### Disagreements
- <persona X> vs <persona Y>: <what they disagree about>

### Recommended (with reasoning)
<your single best read of the debate, clearly labeled as your
recommendation, not a verdict>

◆ Current state → next: pick an option, or `ooo lateral debate <subset>` to drill into a disagreement

The verdict is the user's. Never auto-progress to the next workflow step on the user's behalf — wait for their choice.

Persona-selection heuristics (when args are empty and you must pick one for solo)

You only need this if a parent SKILL or the user explicitly requests one persona but doesn't name one. In debate mode, no selection needed (use all 5).

• Repeated similar failures → contrarian (challenge assumptions)
• Too many options → simplifier (reduce scope)
• Missing information → researcher (seek data)
• Analysis paralysis → hacker (just make it work)
• Structural issues → architect (redesign)

When MCP is unavailable

The contract is "fail loud, don't substitute": if ouroboros_lateral_think is not callable at all — neither already exposed nor loadable via discovery — stop and report that the MCP dispatch surface is broken. Do not improvise either solo or debate by reading persona files directly when MCP-driven invocation was requested — that re-introduces the contract bypass the bot review flagged.

The one exception, retained for documented offline use: a parent SKILL operating in degraded-offline mode that has already announced it cannot reach MCP MAY read src/ouroboros/agents/<persona>.md and adopt that persona inline for solo reframing. This is not a fallback for ooo lateral; it's a degraded helper for a parent SKILL that has already given up on MCP. Debate has no offline equivalent — report the broken surface and stop.

Examples

Solo

User: I'm stuck on the database schema design.
> ooo lateral simplifier

# Lateral Thinking: Reduce to Minimum Viable Schema
Start with exactly 2 tables. If you can't build the core feature
with 2 tables, you haven't found the core feature yet.

◆ Current state → next: try this, then `ooo run` — or `ooo interview` to re-examine.

Debate (default — Claude Code / Codex CLI)

> ooo lateral

[Step 2: Confirm problem_context + current_approach are present;
 ask the user one short combined question if not]
[runtime tool discovery loads ouroboros_lateral_think]
[Call ouroboros_lateral_think(personas=[hacker,researcher,simplifier,architect,contrarian], ...)]
[Handler returns: content = N persona blocks joined by ---,
                  followed by hidden <!-- ouroboros-lateral-inline-dispatch-v1 base64
                  <base64 body> --> sentinel block]
[Extract via the sentinel; base64-decode then JSON.parse → payloads[]]
[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]

## Debate result — 5 personas
### Options
- A (hacker): ship the 50-line version, defer correctness
- B (architect): the data model is wrong, redesign before coding
- C (simplifier): cut feature X, the rest fits in one file
- D (researcher): we don't have user data; instrument first
- E (contrarian): are users actually asking for this?

### Disagreements
- hacker vs architect: ship-first vs redesign-first
- contrarian vs all: whether the problem is real

### Recommended
Lean toward C+E: validate the need (E) before scoping (C); A and B
both assume the feature is wanted.

◆ Current state → next: pick an option, or `ooo lateral debate hacker contrarian` to drill in.

Autonomous chain from interview

[ooo interview mid-flow; main session is wearing the socratic-interviewer persona]
[Main session judges that the next question is too tangled — multiple
 reframings could be valid and asking the user to disambiguate would itself
 be confusing. It chains to this SKILL on its own.]

(autonomous) ooo lateral
[5-agent debate runs in parallel sub-agents; the user sees "Running 5 agents…"]
[Main session presents options + disagreements + a recommendation]
[User picks an option — or asks for a follow-up debate]
[Main session resumes the interview with the chosen framing]

RFC #1392 State Breadcrumb Footer

Your final response MUST end with exactly one breadcrumb footer line:

◆ <current state> → next: <recommended action>

Derive <current state> from live session state via ouroboros_session_status when that MCP projection is available; otherwise derive it from this skill's actual outcome. Never use a linear Step N of M footer because Ouroboros is an evolutionary loop. When the next action is genuinely a choice, list 2-3 honest options in the next: clause. The breadcrumb line must be the last line of the response.