fusion-rescue

Fusion Rescue for Codex

This generated file is the Codex-facing runtime skill document. Codex should read this file directly; maintainers edit the source documents listed below instead.

Generated Runtime Composition

Source order:

• ../../docs/skill-core/fusion-rescue.md
• ../../docs/platforms/codex-runtime.md
• ../../docs/platforms/codex-fusion-rescue.md

The sections below are already composed for this platform. Do not ask the runtime model to load another platform's runtime document or invocation syntax.

Source: docs/skill-core/fusion-rescue.md

Fusion Rescue

Fusion Rescue is a bounded deliberation skill. It borrows the compound-model shape of inference-time ensemble systems: multiple panel analyses first, then a single judge/synthesis pass by the current host main agent.

It is not weight fusion, model merging, OpenRouter API integration, a bridge daemon, or a hidden runtime. It is a Markdown workflow for difficult cases where one normal analysis path has stopped producing useful progress.

Software Development Stage

Fusion Rescue is an escalation and synthesis stage.

Use it standalone when the user explicitly asks for fusion-rescue or asks for multi-agent/multi-host rescue analysis. Inside another workflow, use it only after the selected workflow's ordinary analysis, debugging, or review loop is blocked, inconclusive, or repeatedly failing.

When To Use

Use for:

• a hard implementation, debugging, architecture, or verification problem that has stalled
• a plan or fix where independent critique could change the next action
• contradictory evidence that needs synthesis rather than another linear pass
• ralph or systematic-debugging escalation after their own gates show the current path is not resolving the problem
• explicit user requests for fusion-style rescue analysis

Do not use for:

• ordinary implementation work that ralph can execute directly
• normal first-pass debugging before reproduction and hypotheses exist
• vague requirements before interview
• planning gaps that ralplan should handle
• quick one-file edits or tiny checks
• default multi-host deliberation for every task

Agent Roles

Agent	Dispatch (when)
fusion-rescue-analyst	Use for current-host panel slots. Give each analyst one lens, the same evidence packet, non-goals, and the output fields below.

If the active platform cannot dispatch the role, run separate inline panel blocks and record the fallback reason. The current host main agent still owns judge/synthesis.

Panel Contract

Run exactly three default panel slots:

1. primary: the strongest constructive diagnosis or solution path.
2. adversarial: the strongest critique, failure mode search, and assumption attack.
3. pragmatic: the simplest viable next action, verification path, and rollback boundary.

When the opposite host is reachable, at least one of these three slots must use an actual opposite-host response. Do not satisfy Fusion Rescue with three current-host-only panels when the opposite host can provide bounded panel evidence. The active platform-specific Fusion Rescue rules may pin a lens to the current host or the opposite host. If no platform-specific Fusion Rescue rule pins a lens, the current host may choose which panel slot uses the opposite host.

When a platform-specific Fusion Rescue rule assigns one panel to collect the opposite-host response, that panel receives exactly one permitted cross-host consult. Other panels must remain current-host-only and must not call another host.

Each panel receives:

• the caller's Direction Contract, including required AC IDs, non-goals, protected assumptions, and direction-change approval rule
• the exact blocked decision that ordinary Ralph or Systematic Debugging could not resolve
• the remaining process budget for diagnostics, review, tests, and additional panel work
• problem statement and current workflow context
• relevant evidence, commands, logs, diffs, or plan excerpt after redaction and minimization
• non-goals and forbidden behavior
• any known budget, auth, safety, or environment constraints
• explicit instruction not to invoke nested rescue, fusion-rescue, another workflow skill, or any cross-host call except the single assigned cross-host consult when this panel owns the opposite-host response slot; same-host read-only subagents and read-only tools remain allowed
• explicit read-only instructions: do not edit files, run mutating commands, write state, or install plugins from a panel; same-host read-only analysis tools and subagents are allowed, but make no cross-host call beyond the single assigned cross-host consult
• explicit instruction not to create a new proof architecture, scheduler, state machine, protocol, oracle, fixture system, or review layer outside the Direction Contract's goal and non-goals

Each panel returns:

• lens name
• strongest finding
• evidence used
• assumption under test
• likely failure mode
• recommended next action
• confidence and why
• what would change the conclusion

Use these exact field labels in panel output. Do not omit a field even when the answer is short, synthetic, or read-only.

Cross-Host Consult

Cross-host consultation is attempted for at least one panel in default mode and is required when the caller explicitly asks for require-cross-host behavior. The success condition is that a panel result includes a real assigned-lens analysis from the opposite host, and the synthesis names which panel used that response in panel availability/fallback notes.

Use the active platform-specific Fusion Rescue rules for the consult mechanism, command or plugin capability, permission preflight, foreground or response proof, and any lens pinning. A launch notice, queued-job message, background acknowledgement, deferred status pointer, or proof that only says a job started is not a valid opposite-host response. The consult call itself must return the assigned panel analysis unless the platform-specific Fusion Rescue rules define a stricter foreground response path.

The outbound prompt must request only the assigned lens fields. It must not ask the opposite host to invoke public Fusion Rescue, another workflow skill, a slash command, Task, Agent, Workflow, subagent, or a further host consult unless the active platform-specific Fusion Rescue rules explicitly identify that named capability as the single allowed opposite-host consult path. Even then, the result must be the assigned panel output, not another nested rescue transcript.

If platform consult controls are unavailable, if they reject the read-only boundary, if foreground response proof fails, or if the consult cannot return a panel response, treat the cross-host consult as unavailable. The consult prompt must include one assigned lens, a redacted and minimized problem packet, the recursion guard, and the instruction to avoid nested rescue or host-to-host ping-pong.

Before sending any cross-host consult packet:

• remove credentials, tokens, API keys, cookies, private keys, payment data, personal contact details, and unrelated user data
• replace secret-like values with labels such as [REDACTED_TOKEN]
• include only the minimal code/log/diff excerpts needed for the assigned lens
• omit raw auth/config file contents and environment dumps
• state read-only consult: no edits, no writes, no installs; read-only analysis tools are allowed only when the active opposite host permits them
• include fusion depth: 1

Do not hard-code absolute host binary paths. When command availability, auth, plan, plugin install state, budget, or policy blocks cross-host consultation, record only the failure class, command or plugin name, path/auth status, and next local fallback. Do not record credential values, config contents, or full environment output.

Fallback Behavior

Default mode degrades instead of blocking:

• If the opposite host, auth, command/plugin, or response collection path is unavailable or cannot be proven, run all three panel slots on the current host and include a panel availability/fallback note that says no opposite-host response evidence was available.
• If a platform-specific permission, auth, budget, command, plugin, or foreground-response preflight fails, record the failure class and continue on the current host.
• If the opposite-host call returns only a launch notice, queued-job message, background acknowledgement, or status pointer instead of assigned panel analysis, record the missing response proof, treat the slot as having no opposite-host response, and run it on the current host in default mode.
• If platform-specific Fusion Rescue rules pin a lens to a host that is unavailable, run the pinned lens on the current host in default mode and state that it is not opposite-host evidence.

Require-cross-host mode blocks when the requested host cannot be reached. The blocking output must include which host was required, what command or plugin was attempted, failure class, and the next local fallback the user can approve.

Recursion Guard

Every ordinary panel and every outbound cross-host consult packet must state:

fusion depth: 1
Do not invoke rescue, fusion-rescue, cross-host consult, or another host from inside this panel.
Same-host read-only subagents and read-only tools are allowed; the prohibition above is the cross-host hop, not same-host fan-out.
Return your assigned lens analysis to the caller (a same-host read-only subagent or tool you used to produce it is fine).

fusion depth: 1 is a cross-host-hop count: this panel sits one cross-host hop from the caller and must not add another cross-host hop. The "from inside this panel" prohibition above scopes to cross-host calls (rescue, fusion-rescue, cross-host consult, or another host); a panel MAY use same-host read-only subagents or read-only tools to form its assigned-lens analysis.

When a panel is assigned to collect the opposite-host response, its panel prompt must state that the assigned consult is the only permitted cross-host call, and the outbound consult prompt itself must contain the strict guard above.

This is a one-hop guard. The current host must not call the opposite host and allow that host to call back into the current host or another host. The cross-host block applies transitively: a same-host read-only subagent spawned by a panel inherits the same no-further-cross-host-hop rule.

Judge And Synthesis

The current host main agent is the judge. Do not spawn a judge role.

The judge compares the three panel outputs and produces a synthesis with these fields:

• consensus
• contradictions
• unique insights
• blind spots
• recommended next action
• confidence and why
• panel availability/fallback notes
• fusion depth: 1

panel availability/fallback notes must state which panel used the opposite-host response. If none did, it must state why the opposite host was unavailable or unproven and whether default fallback or require-cross-host blocking applied.

The synthesis must compare, decompose, and recombine the panels. It must not only concatenate the answers. When panel findings conflict, name the conflict, state which evidence decides it, and identify the smallest check that would change the recommendation.

The synthesis may recommend a smaller diagnostic, an explicit approved direction-change request, or a block. It must not spend beyond the recorded remaining process budget or silently replace the Direction Contract with new proof architecture.

Semantic Scenario Checks

Before treating a Fusion Rescue update as verified, inspect the contract against these scenarios:

• Opposite host available: default mode must include at least one panel with an actual opposite-host response. Three current-host-only panels are insufficient unless availability or proof failed and fallback is disclosed.
• Intentional contradiction: primary recommends a path, adversarial shows why that path may violate a constraint, and pragmatic suggests a smaller reversible action. The synthesis must name the contradiction, decide what evidence matters, and recommend the smallest next check instead of merging incompatible advice.
• Missing opposite host: cross-host consult is unavailable in default mode. The workflow must still produce three current-host panel slots and include panel availability/fallback notes.
• Platform preflight denied: if the active platform-specific Fusion Rescue rules require permission, auth, budget, command, plugin, foreground, or response-proof preflight, a denied or missing preflight must prevent the consult. Default mode must use three current-host panel slots and disclose the failure class in panel availability/fallback notes.
• require-cross-host unavailable: the required host, command, plugin, auth, or budget is unavailable, or the opposite-host response cannot be collected or proven. The workflow must block with failure class, attempted command or plugin, path/auth status, proof status, and next local fallback, without exposing secret values.
• Recursive consult: a panel attempts to call rescue, fusion-rescue, or another host. The workflow must reject the nested call using fusion depth: 1 and the one-hop guard. This rejection targets the nested CROSS-HOST call; a panel using same-host read-only subagents or read-only tools to form its assigned-lens analysis is not a recursive consult and is allowed.

Caller Return

Standalone mode returns analysis and recommendations only. Do not edit files directly from standalone Fusion Rescue.

When called from ralph, return control to ralph with the synthesis, recommended next story or verification step, and any blocked/residual risk. Ralph remains responsible for edits, TDD, review, cleanup, and final verification.

When called from systematic-debugging, return control to systematic-debugging with the synthesized root-cause direction, competing hypotheses, and smallest diagnostic or fix step. Systematic Debugging remains responsible for reproduction, causal-chain closure, fix evidence, and verification-before-completion.

When called from ultrawork, return the synthesis to the active phase instead of changing the workflow stage directly.

Output

Return:

• Mode: standalone or caller workflow.
• Panel availability and fallback notes.
• Opposite-host response path used, or the unavailable/unproven fallback reason.
• Panel summaries by primary, adversarial, and pragmatic.
• Judge synthesis with all required fields.
• Recommended next action.
• Confidence and why.
• Caller return target, when any.
• Residual risks or blocked cross-host requirements.

Anti-Patterns

• Calling Fusion Rescue before the ordinary selected workflow has enough facts.
• Treating cross-host output as automatically better than local evidence.
• Hiding auth, path, plan, or plugin failures.
• Using a bridge hook, daemon, ledger, or background state to trigger rescue.
• Creating a separate consult role or judge role.
• Expanding from three default slots without a new approved plan.
• Claiming OpenRouter Fusion equivalence or weight fusion.

Source: docs/platforms/codex-runtime.md

Codex Runtime Rules

This compact platform section is embedded in generated Codex-facing skill documents.

Skill Loading

Codex-facing public skills live under skills/. Generated skills/<skill>/SKILL.md files compose the matching skill core, this compact runtime section, and any Codex skill-specific overlay such as docs/platforms/codex-<skill>.md.

User Approval And Prompting

Ask approval, preference, scope, or next-step questions directly in the Codex conversation. Keep prompts outcome-first: state the desired outcome, acceptance criteria, non-goals or side effects, expected evidence, and output shape before detailed steps.

Use compact final answers unless the active skill requires a plan, review, or verification report. Preserve durable state in written artifacts before long work, compaction, or handoff.

Role Dispatch

Dispatch only after the active skill's trigger fires, then read docs/platforms/codex.md ## Role Dispatch for the full host contract. Use spawn_agent(agent_type="oh-no-<role>", ...) first, do not combine it with fork_context=true, and use generic prompt embedding only after the custom agent is actually rejected. The task packet carries scope, ownership, expected output, and lifecycle.

Every dispatched result is a dependency: wait_agent must reach final status, the caller captures and uses the output, and only then performs lifecycle cleanup. Timeout, empty output, or "No agents completed yet" is not final; do not close, redo inline, or use missing output as evidence.

Generic Role Prompt Fallback

After confirmed custom-agent unavailability, embed docs/agent-core/<role>.md; see the full platform doc for the fallback shape.

Cross-Host Consult Channel

This channel is trigger-loaded, not embedded in every workflow decision. When a named THOROUGH paired-review or Fusion Rescue trigger fires, read and apply docs/platforms/codex.md ## Cross-Host Consult Channel before dispatch. Until then, do not preload opposite-host invocation details.

Source: docs/platforms/codex-fusion-rescue.md

Codex Fusion Rescue Rules

This platform overlay is source content for the generated Codex-facing fusion-rescue runtime document, after the shared core and docs/platforms/codex-runtime.md.

Lens Ownership

Codex remains responsible for the adversarial lens when Codex is available. From Codex, when the Claude consult preflight succeeds, assign exactly one non-adversarial panel slot to collect the Claude response. That slot may be owned by a Codex fusion-rescue-analyst panel subagent, and it may call Claude Code exactly once to collect the Claude response for that panel. Claude Code does not need to spawn another nested subagent for the panel to count. If the preflight fails, use the documented default fallback or require-cross-host block instead of pretending an opposite-host response was collected.

Claude Consult Path

From Codex, ask Claude Code through ${CLAUDE_BIN:-claude} -p when available. Before assigning a Claude consult panel, the Codex main agent must inspect the active Codex permission/sandbox context. Claude consult is allowed only when the current Codex permission state is exactly danger-full-access. If the state is missing, unknown, read-only, workspace-write, or anything other than danger-full-access, do not call Claude. State that Claude is unavailable because the Codex permission state is not danger-full-access, then use three current-host Codex panel agents in default mode. In require-cross-host mode, block instead of pretending an opposite-host response was collected, and name the current-host three-panel fallback as the next local option the user can approve.

When the Codex permission preflight confirms danger-full-access, build the Claude command as an argument vector, not through shell string interpolation. The argument vector is ${CLAUDE_BIN:-claude}, --print, --model, opus, --permission-mode, dontAsk, --no-session-persistence, then the prompt packet, unless the user explicitly supplied a different Claude model for this rescue. Do not specify a Claude tools override by default: Claude Code may need its own permitted tools to perform the read-only analysis for the assigned lens. The read-only boundary is enforced by the redacted prompt packet and the active host permissions, not by stripping tools from the consult. The Claude prompt and active host permissions must still forbid file edits, writes, installs, mutating commands, Codex calls, nested rescue, and any host-to-host ping-pong.

From Codex, this is direct Opus panel review, not a request for Claude Code to run its public Fusion Rescue workflow. Claude Opus must answer the assigned panel directly. The Claude prompt must not ask Claude Code to invoke /oh-no-harness:fusion-rescue, oh-no-harness:fusion-rescue, /codex:rescue, codex:codex-rescue, Task, Agent, Workflow, subagents, or any Claude-side skill or slash command. It must request only the assigned lens analysis fields from Claude Opus.

For Codex-hosted Fusion Rescue, the cross-host slot may be a Codex panel subagent whose only special responsibility is to run that Claude command and return Claude's response as its panel output. If the active Claude binary rejects those controls, cannot enforce them, or cannot return a panel response, treat the cross-host consult as unavailable. The Claude prompt must include one assigned lens, a redacted and minimized problem packet, the shared recursion guard, and the instruction to avoid nested rescue, Codex calls, or host-to-host ping-pong.

Fallback Notes

Default mode degrades instead of blocking:

• If Claude is unavailable from Codex, record the command/path/auth failure class or missing response proof and continue.
• If the Codex permission state is not exactly danger-full-access, record Claude unavailable: Codex permission state is not danger-full-access and run all three panel slots on the current Codex host in default mode.

Require-cross-host mode blocks when the Claude consult path cannot return the assigned panel output. The blocking output must include the attempted command, permission state class, response-proof status, and current-host three-panel fallback.