Workflows

syndicate run anatomy

syndicate run <workflow-name> [flags]

Source: cli_command_reference.md §6.

Autonomy modes

Mode is selected per run with --autonomy. Policy may restrict or mandate stricter behavior than the requested mode (cli_command_reference.md §6.2).

Workflow lifecycle (normative 10-stage model)

This page describes the full implementation lifecycle (governed_execution_spec.md §5). The six-stage overview model remains a conceptual summary.

1. Proposal intake — control plane receives structured proposal metadata and objective context.
2. Canonicalization — proposal is converted into deterministic canonical form before any policy or approval logic.
3. Policy evaluation — control plane determines allow/deny, envelope policy, side-effect classes, and checkpoint triggers.
4. Context and evidence preparation — control plane finalizes identifiers and evidence inputs for reconstruction.
5. Approval determination — control plane decides whether execution proceeds, needs operator approval, or is denied.
6. Approval recording — if approval is required, a bound approval record is durably persisted.
7. Permit issuance — control plane mints permit bound to canonical target and scope constraints.
8. Tool-layer validation — tool layer validates permit authenticity, expiry, scope, side-effect class, lineage, trust, and checkpoint status.
9. Execution attempt — side effects run only after successful validation.
10. Result recording — execution result evidence is appended with status, side effects, timing, permit ID, lineage, and trust impact.

Enforcement invariants:

• Stage 8 validation failure denies execution immediately (governed_execution_spec.md GP-003).
• If required evidence cannot be durably written, execution hard-blocks. No degraded mode permitted (governed_execution_spec.md GP-004).

Streaming output

Streaming is incremental output-fragment rendering, not buffered full-response rendering.

• TUI mode: fragments render in main pane from turnOutput subscription.
• --no-tui mode: fragments write to stdout as received.

syndicate run <workflow-name> --no-tui --objective "<text>" 2>&1 | tee workflow.log

Source: operator_interface_spec.md §18.7 and cli_command_reference.md §6.

Self-correction loop

After file-modifying execution, LSP diagnostic validation runs for modified files (lsp_integration_spec.md §8.1).

When result class is new_errors, diagnostics are fed into the next model context window for self-correction (lsp_integration_spec.md §8.4).

• Default maximum iterations: 2
• Policy-controlled via lsp.self_correction_max_iterations
• If iterations are exhausted, unresolved diagnostics are surfaced to operator

Workflow interruption

Interruption paths:

• Ctrl+C in TUI
• process termination in headless mode

Interruption behavior preserves evidence (session_and_turn_model.md ST-007):

• partial turn records through the last completed step
• all emitted audit events before interruption
• trust state snapshot at interruption point

Session termination invalidates in-flight permits and active execution envelopes.

Exit code 3 indicates interruption (cli_command_reference.md §6.3).

v1 deployment note This page describes v1 local-embedded behavior. Team and enterprise deployments use syndicate-server; control-plane topology differs while permit validation, checkpoint rules, and evidence requirements stay authoritative.

Non-guarantee callout Governed workflows guarantee authorization and evidence controls, not correctness of model output and not absence of hallucinations.