The mod answers an agent's steps from the sidecar and never calls a model

actor

The Claude Code engine, running a subagent of type cc-agents:<harness>.

problem

The engine runs a real model loop for every subagent, and left alone it would call the type's fallback Claude model for the task.

outcome

Each step of such an agent is answered from the harness's turn, and an agent the mod cannot serve fails closed instead of reaching the fallback model.

value

The subagent's own model is never called, and the user's rows, transcripts and Workflow roles behave as for a Claude agent.

Open questions

• non-blocking · question.untestedPaths line 25 · specs/runner/mod-layer.sdp.md:25 · Built but not exercised by any test: a spawn that asks to be a team teammate is denied; the agent.spawn catch handler denies an agent the hook could not prepare; the internal progress tool answers "Still working." and is denied to agents that are not ours; a follow-up message to an ended agent starts a new run on the same thread; a step aborted mid-stream interrupts the turn; the turn-complete hook of an agent stopped before its first step asks the sidecar to interrupt it.Decider: delegatedProvisional: The code is as described; tests would pin each path.
• non-blocking · question.stepLabel line 28 · specs/runner/mod-layer.sdp.md:28 · The steps hand over through an internal progress tool call, so a step shows its progress only when it ends, every 20 seconds by default, never token by token.Decider: ownerProvisional: Kept as built; Claude Code buffers output until a step ends.

rule.ourTypesOnly line 34 · specs/runner/mod-layer.sdp.md:34

Only agent types named cc-agents:<harness> are handled; a spawn, step or message of any other agent passes through untouched.

no example exercises it

rule.headerThenCheck line 35 · specs/runner/mod-layer.sdp.md:35

A spawn reads an optional header of model, effort and mode lines ended by a line ---, refuses an unknown key before anything starts, asks the sidecar to check the settings, and starts the subagent with the task alone.

no example exercises it

rule.refusalNamesWhy line 36 · specs/runner/mod-layer.sdp.md:36

A check the sidecar refuses denies the spawn with its reason, and an OpenCode agent with no model and no default is refused with a pointer to the user's model guide.

no example exercises it

rule.defaultsAreOptions line 37 · specs/runner/mod-layer.sdp.md:37

A harness's default model comes from the option <harness>Model; a model in the header wins; with neither, the harness's own check decides.

no example exercises it

rule.slotBeforeStart line 38 · specs/runner/mod-layer.sdp.md:38

A spawn takes a slot from the sidecar before the subagent starts, and over the cap or on a critical machine it waits with a status line and a toast that name the reason once.

no example exercises it

rule.stepsFromEvents line 39 · specs/runner/mod-layer.sdp.md:39

A step shows the harness's events as the agent's text, and while the turn runs it ends by calling the internal progress tool; a step lasts the option stepSeconds, 20 by default and at least 5.

no example exercises it

rule.failClosed line 40 · specs/runner/mod-layer.sdp.md:40

An agent of this plugin that the mod has no record of after a reload, or whose spawn never recorded it, is answered with a message saying so and never handed to the type's Claude model, and Workflow agents in no list are recognised by a marker the mod wrote at spawn.

no example exercises it

rule.reloadReattaches line 41 · specs/runner/mod-layer.sdp.md:41

After the mod reloads, a step of an agent the sidecar knows attaches to its running turn from the event cursor, or hands back the answer that already came.

no example exercises it

rule.firstStepWaitsForSpawn line 42 · specs/runner/mod-layer.sdp.md:42

A first step waits for its own spawn to finish recording the agent, hands over to the next step when its share of five seconds is used, and gives up after two minutes.

no example exercises it

rule.lostTurnEnds line 43 · specs/runner/mod-layer.sdp.md:43

A turn whose sidecar restarted ends with a stated reason instead of polling forever.

no example exercises it

0 unkeyed rules.

sidecarEndpoints line 47 · specs/runner/mod-layer.sdp.md:47

/check, /reserve, /release, /register, /start, /events, /steer, /interrupt, /status, /agent, /resume-lookup, each described by the sidecar Spec.

no citation recorded

agentType line 48 · specs/runner/mod-layer.sdp.md:48

cc-agents:<harness>, where <harness> is also the adapter module's name and the agent file plugin/agents/<harness>.md.

no citation recorded

header line 49 · specs/runner/mod-layer.sdp.md:49

Lines key: value for model, effort and mode, then a line ---, then the task; a prompt without that shape is the task.

progressTool line 50 · specs/runner/mod-layer.sdp.md:50

An internal deferred tool named progress that the mod registers at session start; its calls mark a turn still running.

ownershipMarker line 51 · specs/runner/mod-layer.sdp.md:51

A file owned-<agent id> in the session's run directory, naming the harness, written at spawn.

Verifier bindings

• test:cc-agents.mod-suite · the mod against a scripted sidecar · anchored · test anchor binding present · plugin/tests/mod.test.ts:4

Requirement drawing

• Declared verifying examples: no declared verifying example.

Decisions

No decision record named.

• impl:cc-agents.mod · answers each step of an agent from the sidecar's events, hands back the exact answer, steers and stops, and returns the job state block · plugin/hooks/register.js:9

  Uses

  None.

Runtime observation: not tracked.

Runtime observation: not tracked.

Build attestations: not tracked.

Deployments: not tracked.

Bills of materials: not tracked.