configure-subagents

name	configure-subagents
description	User-invoked setup action that configures the model and reasoning effort of the installed Oh No Harness Claude Code subagents. Run it explicitly with the slash command; it is never auto-invoked by the model.
argument-hint	[check]
disable-model-invocation	true

Configure Subagents for Claude Code

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

Generated Runtime Composition

Source order:

• ../../docs/skill-core/configure-subagents.md
• ../../docs/platforms/claude-code-runtime.md
• ../../docs/platforms/claude-code-configure-subagents.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/configure-subagents.md

Configure Subagents

Configure Subagents is a user-invoked setup action. It walks you through choosing a model and reasoning effort for each of the 9 Oh No Harness Claude Code subagents, then applies those choices to the installed runtime agent Markdown so the next Claude Code session dispatches each role with your chosen model and effort.

This skill is human-invoke-only. It is never auto-invoked by the model: its frontmatter sets disable-model-invocation: true, and it is intentionally left out of the SessionStart routing map. Only an explicit /oh-no-harness:configure-subagents from the user runs it.

It is Claude Code-only: it configures Claude Code subagent frontmatter and ships no Codex wrapper. It does not change the generator-owned canonical agents/*.md in the source checkout and never touches Codex custom-agent TOMLs.

Software Development Stage

Configure Subagents is a one-time environment-setup action, not a software development stage. It should not gather requirements, plan, edit project code, debug failures, clean code, or verify completion. It does not invoke workflow skills.

When To Use

Use only when the user explicitly asks to configure, set, or change which model or reasoning effort the Oh No Harness subagents run with. There is no model-driven trigger for this skill; do not use it as part of any automated workflow.

What Gets Configured

The bundled configurator edits the installed runtime agents under the active plugin root's agents/ directory (the platform runtime overlay names the concrete ${CLAUDE_PLUGIN_ROOT} path). For each agent it replaces the model: value and inserts or replaces a single effort: line immediately after model:, preserving every other frontmatter and body byte.

Your choices are also saved as durable, schema-versioned preferences outside the plugin cache (in the Oh No Harness data directory), so they can be reapplied after a plugin update restores the canonical runtime agents. The same preferences store top_tier_models, the machine's top-tier model set, and the optional native-only secondary_top_model used as the diversity leg for paired reviews and fusion. No proxy URL or token value is ever stored or printed.

Status-Only Check

When the user invokes the command with the check argument (the [check] argument-hint), run only the configurator's read-only check, report the resulting STATUS: line, and stop — do not begin the interview and write nothing. A recovery-required status means a prior apply was interrupted and is recovered on the next apply or SessionStart reapply. Run the full interview below only when the command is invoked with no argument.

Interview Flow

Follow this order to the end and write no files before the final confirmation.

1. Run the bundled configurator's read-only check to confirm the active plugin root and locate the 9 installed agent files.
2. Ask first whether the user has CLIProxyAPI installed, as an explicit yes/no question. Do not let any auto-detection stand in for the user's answer.
3. When the answer is yes, diagnose only the presence of the ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN environment variables (never print their values); warn if the proxy wiring looks incomplete.
4. Configure these 9 agents one at a time, in this exact order: explore, analyst, planner, plan-reviewer, executor, debugger, verifier, code-reviewer, fusion-rescue-analyst.
5. For each agent, decide both a model and an effort before moving to the next agent.
   • Model choices when proxy is no: fable, opus, sonnet.
   • Model choices when proxy is yes: the three native models plus GPT via CLIProxyAPI; if the user picks GPT, ask a follow-up to choose an available GPT model alias. Splitting the GPT choice into a follow-up keeps every question within the host's 4-option limit.
   • Effort choices: max, xhigh, high, medium.
6. Ask for top_tier_models, the space-separated models that count as top tier on this machine. Propose the native default fable opus; when proxy is yes, the user may also include GPT aliases offered by the per-role primary-model flow.
7. Ask whether to set secondary_top_model, the diversity model for paired reviews and fusion. If yes, offer only the native aliases fable, opus, sonnet, and haiku — never GPT — and require the selection to be present in top_tier_models.
8. Summarize all 9 model+effort selections and both diversity settings, then ask for a single final apply or cancel confirmation. No file is written before that confirmation.
9. On apply, invoke the configurator exactly once so all 9 agents and the diversity settings change in one all-or-nothing transaction.

Apply And Activation

• The configurator applies all 9 agents as one transaction: it stages and re-validates every result, backs up the originals, writes a recovery journal, and swaps each file in with an atomic rename. Any mid-transaction failure rolls every file back, and a stale journal from an interrupted run is recovered on the next invocation. Each individual rename is atomic and the whole operation is recoverable through the backup and journal.
• After a manual apply, a new Claude Code session (or /clear) is the supported activation boundary; report that to the user.
• After a plugin update, the SessionStart hook reapplies stored preferences best-effort. Because agent metadata may already be snapshotted for the current session, a /clear or one more new session may still be needed for a SessionStart repair to take effect.

Safety Rules

• Collect every choice first; write nothing until the user confirms the final summary.
• The configurator refuses to apply inside a Git source checkout, so the generator-owned canonical agents/*.md are never changed with user preferences.
• GPT primary models are offered only after an explicit CLIProxyAPI yes; when the answer is no, only native primary models are valid and a GPT choice is rejected before any write. secondary_top_model is always native-only, regardless of the proxy answer.
• Never print or store the proxy base URL or auth token.
• Do not invoke workflow skills from this setup skill.

Source: docs/platforms/claude-code-runtime.md

Claude Code Runtime Rules

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

Skill Loading

Claude Code-facing public skills live under skills-claude/. Generated skills-claude/<skill>/SKILL.md files compose the matching skill core, this compact runtime section, and any Claude Code skill-specific overlay such as docs/platforms/claude-code-<skill>.md. Slash commands must delegate to the matching generated skill document.

User Approval, Tasks, And Prompting

Use the host's structured question tool when available for approval, preference, scope, or next-step selection; otherwise ask one focused plain-text question and wait. Present options as actions the host agent will take.

When a core skill has a multi-phase approval handoff and the host exposes task tracking, create one task per phase and complete them sequentially.

Keep Claude prompts explicit and sectioned: state scope, non-goals, constraints, approval gates, expected evidence, and output format. Preserve long-running context in artifacts before compaction, task handoff, or subagent dispatch.

Role Dispatch

Dispatch only after the active skill's trigger fires, then read docs/platforms/claude-code.md ## Role Dispatch for the full host contract. Prefer oh-no-harness:<role>, request the whole independent batch before waiting, capture every final result, and clean up only after integration. An approved-plan handoff is dispatch authorization for eligible isolated roles; plugin-agent unavailability uses the documented embedded-role fallback.

Model Diversity Pair

This mechanism is trigger-loaded, not embedded in every workflow decision. For any dispatched plan-reviewer or code-reviewer pair (every dispatched review), or a named THOROUGH debugger pair, both legs MUST be requested in a single batch: issue both subagent tool calls in the same assistant turn (or with Background: yes for both) BEFORE waiting on either result; a serial dispatch-wait-dispatch sequence is not a valid pair. The two legs' packet bodies MUST be identical except the single Assigned perspective: line (Lens A on the primary leg, Lens B on the diversity leg); leg identity (primary vs diversity) is carried ONLY by the host dispatch metadata (the description field and the model override), never inside the packet text. Read the role's concrete stored primary and validated secondary top-tier model from the session <OH_NO_MODEL_DIVERSITY> block. The primary leg is unoverridden and uses the declared-frontmatter primary; the secondary leg carries an explicit NATIVE model override. Claim model-diversity-pair only when the primary is not host-default and the secondary differs from it. Otherwise default to two independent same-model instances as same-model-parallel-fallback with the reason recorded; an explicit require-model-diversity demand transitions to PAUSED when the diversity leg is unavailable. Fusion Rescue uses its Claude Code overlay's three-panel assignment instead of this two-leg shape.

Source: docs/platforms/claude-code-configure-subagents.md

Claude Code Configure Subagents Rules

This platform overlay is source content for the generated Claude Code-facing configure-subagents runtime document, after the shared core and docs/platforms/claude-code-runtime.md.

The configurator edits the installed runtime agents under the active plugin root's agents/ directory. When CLAUDE_PLUGIN_ROOT is set, run the bundled configurator directly:

"${CLAUDE_PLUGIN_ROOT}/scripts/configure-subagents" check

If the plugin root is not exposed, locate the installed script first; the script derives its own physical plugin root from that location, so no root path is passed to it:

tab="$(printf '\t')"
plugins="${CLAUDE_CONFIG_DIR:-$HOME/.claude}/plugins"
reg="$plugins/installed_plugins.json"
# 1) Host-exposed plugin root wins (stays first-preferred).
script="${CLAUDE_PLUGIN_ROOT:+$CLAUDE_PLUGIN_ROOT/scripts/configure-subagents}"
# 2) Else the authoritative installPath for the EXACT oh-no-harness@oh-no-harness
#    registry key (no jq; block-scoped so a config-home path or another plugin that
#    merely contains the substring never matches; newest version among that key's
#    scope entries). Trust only one validly closed array; truncated, compact, or
#    ambiguous input yields no candidate. installPath is a hint only, and a missing
#    or unreadable registry must not abort under strict mode: the [ -r ] guard +
#    `|| true` keep tier 3 reachable, and the [ -x ] guard lets a Windows drive-letter
#    path fall through.
if [ ! -x "$script" ] && [ -r "$reg" ]; then
  root="$(awk '
    /"oh-no-harness@oh-no-harness":[[:space:]]*\[/ {
      keys++
      if (keys != 1 || inblk || closed) { bad=1; next }
      inblk=1; next
    }
    inblk && /^[[:space:]]*\]/ { inblk=0; closed=1; next }
    inblk && /"installPath":/ {
      p=$0
      if (p !~ /"installPath":[[:space:]]*"[^"]+"/) { bad=1; next }
      sub(/.*"installPath":[[:space:]]*"/,"",p); sub(/".*/,"",p)
      v=p; sub(/.*\//,"",v); paths[++n]=p; versions[n]=v
    }
    END {
      if (keys == 1 && closed && !inblk && !bad)
        for (i=1; i<=n; i++) print versions[i] "\t" paths[i]
    }' "$reg" 2>/dev/null | LC_ALL=C sort -t"$tab" -k1,1V -k2,2 | tail -n1 | cut -f2- || true)"
  script="${root:+$root/scripts/configure-subagents}"
fi
# 3) Else the newest INSTALLED semver in the cache — never the marketplace checkout.
#    Sort on the VERSION path component (full path only as tie-break) so a second
#    marketplace identity cannot let an older version win.
if [ ! -x "$script" ]; then
  script="$(find "$plugins/cache" -path '*/oh-no-harness/*/scripts/configure-subagents' 2>/dev/null \
    | awk -F/ '{for(i=NF;i>0;i--) if($i=="scripts"){print $(i-1)"\t"$0; break}}' \
    | LC_ALL=C sort -t"$tab" -k1,1V -k2,2 | tail -n1 | cut -f2-)"
fi
"$script" check

Never apply against a Git source checkout; the configurator refuses that so the generator-owned canonical agents/*.md are never rewritten with preferences.

The check (status-only) branch

When the command is invoked as /oh-no-harness:configure-subagents check (the [check] argument-hint), run only the read-only status check and report its STATUS: line to the user, then stop. Do not start the interview and write nothing. Its statuses are: unconfigured, matching, drifted, invalid-preferences, invalid-agents, source-checkout, and recovery-required. A recovery-required status means a prior run was interrupted; re-running an apply (or the SessionStart reapply) recovers it.

Only when the command is invoked with no argument do you run the full interview below. Every run — including the interview's first step — begins with this same read-only check to confirm the plugin root and the 9 installed agent files.

Asking The Questions

Use the host's structured question tool (AskUserQuestion) for every choice, one focused question at a time.

1. First ask whether the user has CLIProxyAPI installed as a yes/no question. This answer alone decides whether GPT model aliases are offered; do not substitute environment auto-detection for the user's answer.

2. When the answer is yes, check only whether ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN are set (presence only, never echo their values) and warn if either is missing. Use a presence-only test that never prints a value:

   # Reports only set/missing; never expands the variable into output.
   [ -n "${ANTHROPIC_BASE_URL:-}" ] && printf 'ANTHROPIC_BASE_URL: set\n' || printf 'ANTHROPIC_BASE_URL: missing\n'
   [ -n "${ANTHROPIC_AUTH_TOKEN:-}" ] && printf 'ANTHROPIC_AUTH_TOKEN: set\n' || printf 'ANTHROPIC_AUTH_TOKEN: missing\n'
   

3. For each of the 9 agents in order, ask the model question:

   • Proxy no: offer exactly fable, opus, sonnet.
   • Proxy yes: offer fable, opus, sonnet, and GPT via CLIProxyAPI. When the user picks the GPT option, ask a second question offering the available GPT aliases. Keeping GPT behind a follow-up keeps each AskUserQuestion within its four-option limit.

4. Then ask the effort question offering max, xhigh, high, medium.

5. Ask for the space-separated top_tier_models value, proposing the native default fable opus. GPT aliases may be included only when the proxy answer was yes.

6. Ask whether to configure secondary_top_model. If yes, offer exactly fable, opus, sonnet, and haiku; do not offer GPT for this key, and require the selection to be present in top_tier_models.

Applying

Collect all 9 model+effort selections plus the diversity settings, show a final summary table, and ask a single apply-or-cancel confirmation with AskUserQuestion. Only after an explicit apply, run the configurator once with the collected settings so all 9 agents change in one transaction. Write nothing before that confirmation.

Build the apply invocation as a Bash argument array so each option and role=model,effort token is passed as one literal argument (never interpolated into a single string and never through eval). The 9 role tokens must be in the canonical role order:

args=(apply --proxy no
  --secondary-top-model fable
  --top-tier-models "fable opus"
  explore=sonnet,high
  analyst=opus,xhigh
  planner=opus,max
  plan-reviewer=opus,xhigh
  executor=opus,high
  debugger=opus,xhigh
  verifier=sonnet,high
  code-reviewer=opus,xhigh
  fusion-rescue-analyst=opus,xhigh)
"$script" "${args[@]}"    # or "${CLAUDE_PLUGIN_ROOT}/scripts/configure-subagents" "${args[@]}"

Omit --secondary-top-model when the user chooses no secondary. Use --proxy yes when CLIProxyAPI was confirmed; only then may a per-role primary or a top_tier_models entry use one of the GPT aliases offered by the wizard. The configurator exits non-zero and writes nothing if the role tokens are incomplete or reordered, if a model is not allowed by the proxy answer, or if the secondary is not native and present in the top-tier list.

After a successful apply, tell the user the change takes effect in the next Claude Code session; /clear or a new session guarantees it. Stored preferences are reapplied best-effort by the SessionStart hook after a plugin update, and a /clear may still be needed for that repair to take effect in the current session.