config

Everything non-secret about how Shipyard behaves in this repo lives in a layered config file that the sy MCP server's resolver (sy_tools/config.py) is the only reader of, reached through that server's config tools and nothing else. This skill is the human front end to them. It never reads a credential: secrets stay in the environment, and the resolver refuses a credential-shaped key in either direction. Resolve each tool's exposed name from the tools actually available to you rather than typing a literal identifier, per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/config-values.md.

$ARGUMENTS

With no argument, do show.

show

show_config {}

Report it as a status update. Lead with the three layers and which exist, because "my change did nothing" is nearly always a layer question — a value set in the committed layer while the gitignored local layer overrides it, or a user-global default that no repo layer touches. Then give the resolved values with their provenance, and call out anything resolving from derived-default (computed from the repo, not written anywhere) so the user knows it is not in a file they can grep.

For one value, get_config. show_config is already machine-readable and reports the digest of the resolved values and the layer chain on disk alongside them, so there is no separate dump to ask for.

validate

validate_config {}

valid: true comes back with the fingerprint. valid: false comes back with an errors list naming every problem, each with its key and the layer it came from. Relay the errors verbatim — they are written to be actionable — and do not paraphrase a floor refusal into "the model is too weak": the message already names the agent, the floor, and why that floor exists.

Four failure classes are worth recognising on sight, because each has a different fix:

• A required key is unset. Set it in the committed layer, or run /sy:init-repo.
• A retired SY_* variable is still in the environment. This is an error, not an override, deliberately: two resolution paths for one key is what made the old secret/config boundary illegible. Move the value into config and unset the variable. Note that a variable which merely agrees with config is still an error.
• A floor refusal. A config layer tried to drop an agent below config/floors.json. Cost-scaling may raise a floor, never lower it; the fix is to accept the floor, not to edit the floors file, which is plugin-shipped precisely so it is not locally negotiable.
• CLAUDE_CODE_SUBAGENT_MODEL is set. It outranks the per-invocation model parameter on Claude Code before 2.1.251, or when CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1, so it silently reroutes every agent off whatever the resolver decided. Unset it.
• CLAUDE_CODE_EFFORT_LEVEL is set. It outranks the per-invocation effort parameter, so it silently runs every agent at one effort. Unset it.

reload

A config edit needs no restart, because nothing caches a resolved value on disk: every consumer reads through the resolver. What goes stale is the resolver's own hot copy inside the running server, the resolved block sitting in this session's context, and any preflight cache keyed on the old values.

reload_config {}
validate_config {}
show_config {}

reload_config re-resolves the layer chain from disk, which is what makes the edit visible to every later tool call in this session without a restart.

Then state the scope boundary plainly, because it is the part users get wrong: workers already dispatched keep what they were given. The boundary is the next dispatch. A reload mid-/sy:ship does not retroactively change the model a running BUILD is using, and it must not silently change the reviewer between fix rounds — see below.

Report the new fingerprint and name which keys changed provenance or value since the previous show in this session, rather than reprinting everything unchanged. The preflight cache invalidates itself: its fingerprint folds in the resolved config, so a changed setting produces a cache miss and the next tracker call re-verifies liveness for real.

A config change mid-ship is refused, not absorbed

/sy:ship stamps the resolved-config fingerprint into its recorded state, the same discipline as its pinned base/head SHAs. If the fingerprint at a fix round does not match the one recorded at START, the run stops and reports rather than continuing on quietly changed settings: a reviewer whose model changed between rounds invalidates the comparison the gate exists to make. Re-running /sy:ship after an intentional change is the supported path; silently absorbing it is not.

set

get_config {"key": "<key>"}     # confirm the current value and layer first

There is no set tool on the resolver, deliberately — writing config is a small, legible JSON edit, and a generated writer would only obscure which layer changed. Edit the file directly with Edit, choosing the layer by what the value is, not by convenience:

Include "$schema" in a newly created file so editors validate it. After any edit, run validate, then show the affected keys so the user sees the change actually landed in the layer they meant. If the user asks to set something credential-shaped, refuse and say where it belongs: the environment, never a config file — a config file is greppable, and any skill that cats it burns the value into permanent transcript history.

agent

agent_model {"name": "<name>"}

Answers "what will sy:hunt actually run at, and why". The binding reports the resolved model and effort to dispatch with, what config (or a passed model/effort override) requested, and whether either was clamped up to a floor. Use it to explain a surprise: a model_clamped: true means config asked for something below the floor and the floor won.

When effort comes up: effort in config is applied at dispatch, the same way model is — the resolved value is passed as the Agent invocation's effort parameter, so a config edit changes the next dispatch, never a running agent. Frontmatter effort: is only the fallback when a dispatch passes none. The full explanation is ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/model-dispatch.md.

Interaction

One mode per turn — status update, AskUserQuestion, or an isolated ## Action needed block — per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/user-interaction.md. A validation failure the user must fix outside this session (unset a variable, export a credential) is an ## Action needed block naming the exact command, not prose mixed into a status report. This skill never calls the tracker, so it runs no tracker preflight.