adapter

type Adapter ¶

type Adapter interface {
	Name() string
	Resume(ctx context.Context, sess core.Session, prompt string) (Result, error)
	Validate() error
}

Adapter resumes a session with an injected prompt and reports the outcome. Resume returns an error only when the run could not even be attempted (binary missing, spawn failure); a nonzero agent exit is a Result, not an error.

type Claude ¶

type Claude struct{}

Claude resumes Claude Code sessions headlessly:

claude --resume <session-id> -p "<prompt>" --output-format json

run in the session's repoPath (restores the original project config — see docs/reference/claude-code.md). v1 passes NO permission flags (owner decision): the agent runs with its project-configured permissions, and unapproved tool calls are denied in headless mode.

func (Claude) Name() string

func (c Claude) Resume(ctx context.Context, sess core.Session, prompt string) (Result, error)

func (Claude) Validate() error

type Codex ¶

type Codex struct{}

Codex resumes Codex CLI sessions non-interactively:

codex exec resume <session-id> --skip-git-repo-check -- "<prompt>"

run in the session's repoPath. Default exec mode streams progress to stderr and prints only the final agent message to stdout (see docs/reference/codex-cli.md) — the shared runner keeps them apart. v1 passes NO sandbox/approval flags (owner decision); codex exec defaults to a read-only sandbox, a documented limitation.

func (Codex) Name() string

func (c Codex) Resume(ctx context.Context, sess core.Session, prompt string) (Result, error)

func (Codex) Validate() error

type Cursor ¶

type Cursor struct{}

Cursor resumes Cursor CLI sessions headlessly:

agent -p --resume <chat-id> --output-format json -- "<prompt>"

run in the session's repoPath (the CLI treats cwd as the repository root — see docs/reference/cursor-cli.md). The -p+--resume combination is not shown combined in official docs; Validate() existing is why. v1 passes NO permission flags (no --force/--trust — owner decision): in print mode the agent proposes edits without applying them unless the workspace allows it.

func (Cursor) Name() string

func (c Cursor) Resume(ctx context.Context, sess core.Session, prompt string) (Result, error)

func (Cursor) Validate() error

type Fake ¶

type Fake struct {
	// contains filtered or unexported fields
}

Fake is a controllable in-memory adapter for dispatcher tests.

func NewFake() *Fake

func (f *Fake) Calls() []FakeCall

func (f *Fake) Name() string

func (f *Fake) Resume(ctx context.Context, sess core.Session, prompt string) (Result, error)

func (f *Fake) SetDelay(d time.Duration)

func (f *Fake) SetOutcome(exitCode int, err error)

func (f *Fake) Validate() error

type FakeCall ¶

type FakeCall struct {
	SessionID string
	RepoPath  string
	Prompt    string
}

type OpenCode ¶

type OpenCode struct{}

OpenCode resumes OpenCode sessions non-interactively:

opencode run --session <ses_id> -- "<prompt>"

run in the session's repoPath (opencode stores sessions per-project keyed by directory — see docs/reference/opencode.md). Known caveat, documented there: opencode's CLI historically exits 0 on some errors; v1 trusts the exit code and records the output summary so failures are at least visible. The serve/HTTP path is a future improvement. v1 passes NO permission flags (owner decision).

func (OpenCode) Name() string

func (o OpenCode) Resume(ctx context.Context, sess core.Session, prompt string) (Result, error)

func (OpenCode) Validate() error

type Registry ¶

type Registry map[string]Adapter

Registry maps agent names (core.Session.Agent) to adapters.

type Result ¶

type Result struct {
	Command       string // the command line that ran (for the continuation record)
	ExitCode      int
	OutputSummary string
	DurationMs    int64
}

Result is what supervision records about one resume run.