Using pilotfish

September 19, 2026 · View on GitHub

This guide collects day-to-day model, delegation, compatibility, and opt-out questions. Installation and file mutation rules remain in the install runbook; exact orchestration behavior remains in the policy template.

繁體中文

Contents

Model routing

NeedSetting or actionEffect
Default main sessionmodel: "opus"Uses the provider-resolved Opus family alias; the installer preserves an existing choice unless you approve a change
Explicit Fable session/model fableOpts into Fable without changing role-agent bindings
Lower main-session quota use/model opusplanUses Opus for planning turns and Sonnet for the main session's execution turns
Explicit 1M context requestmodel: "opus[1m]"Requests the documented 1M Opus alias where the provider supports it
Primary model unavailablefallbackModel: ["sonnet"]Falls back on overload or unavailability; it does not catch authentication, billing, or rate-limit failures

Each role's model and effort live in its agent frontmatter. Do not set CLAUDE_CODE_SUBAGENT_MODEL unless you intentionally want to override every role, including the Opus review and security roles.

Delegation behavior

Higher-priority Claude Code instructions can suppress Agent dispatch. When the pilotfish lifecycle matters, make the request explicit:

Use pilotfish. Follow its dispatch brake: keep direct work in the main session
and call the named agents only when the policy selects delegation.

The optional activation guide offers a user-invocable /pilotfish skill and a one-line CLI wrapper, with an approval-gated AI install contract. Both remain explicit opt-ins, not cue-free dispatch; the wrapper can optionally mention an already installed Baton skill.

Work shapeExpected owner
Small, local, stable work or one tightly coupled unknown bugMain session
Stable multi-file mechanical repetition with a complete one-shot briefmech-executor
Approved implementation requiring local judgmentexecutor
Security-sensitive implementation after approvalsecurity-executor
Risk-triggered Plan or outcome challengeplan-verifier, security-reviewer, or verifier

The cue-free evidence records where automatic dispatch did and did not occur. Those observations are bounded examples, not a dispatch rate or proof of the active system-prompt bytes.

Configuration and compatibility

SituationWhat to check
Custom configuration rootEvery ~/.claude/ path moves under CLAUDE_CONFIG_DIR; the installer resolves it before writing
Project-level CLAUDE.mdClaude Code stacks project and user memory; pilotfish never writes into the project
Custom Explore roleLegacy global install only — the Plugin cannot override the built-in (why). It pins reconnaissance to Haiku, but unlike the built-in role it loads user memory; the policy self-disables inside subagent roles to limit that overhead
availableModels allowlistInclude opus, fable, sonnet, haiku, and the selected main-model value or role aliases may silently inherit the main model
Managed or enterprise settingsManaged models, allowlists, and same-name agents outrank the user-level install; pilotfish does not bypass them
claude-routerKeep forceRoute off because it overrides agent frontmatter; restoreDelegation strips the separately tracked delegation injection
Delegation-planning skillsTools such as Baton may shape work topology; pilotfish still owns named roles, model routing, approval, and verifier contracts

Tuning

GoalAdjustment
Reduce quota useUse /model opusplan; keep reconnaissance and mechanical roles at their shipped low effort
Increase main-session judgmentStart at high effort and lower it only when quota or latency matters more
Change one role's tierEdit only that agent file's model: frontmatter; the policy names roles, not models
Keep more work inlineAsk the main session to work inline; this disables optional execution delegation, not mandatory risk review
Understand spawn overheadEvery agent starts a fresh context and pays reconstruction plus integration cost; dispatch only when the combined benefit is positive

Model economics, official mechanisms, and measured limitations are documented in research, the design rationale, and the behavioral benchmarks.

Long runs and verification

StateMeaning
AUTOContinues reversible work inside approved scope; it grants no new commit, publish, install, destructive, external-action, or spending authority
ASKPauses for decisions through native input or PAUSED_NEEDS_USER
P0Freezes the affected slice and its dependants
P1 or introduced P2Must be fixed within approved scope or paused; P3/P4 are advisory
verifier resultEvidence for main-session judgment, never automatic scope or implementation authority

Normal verification is one complete pass plus one targeted recheck after a reproduced blocker is fixed. Long-running commands stay owned by the main session; leaf agents return the exact command and working context instead of detaching a process.

Disable, update, or uninstall

ActionMethod
UpdateRe-run the pinned install prompt and follow the runbook's Updating an existing install section
Disable optional execution delegationAsk the main session to work inline
Disable pilotfish for one repositoryLaunch that repository with a separate CLAUDE_CONFIG_DIR that has no pilotfish policy block
Disable the policy globallyRemove or comment the pilotfish:begin/end block, then start a new session
UninstallFollow the runbook's Uninstall section so agent files and settings backups are handled safely