feat(workflows): flow directives + sub-workflow invocation (lobster.run)

[rendrag-git]

What

Two features that transform Lobster's workflow file runtime from a linear step executor into a control-flow engine:

1. Flow Directives — branching, loops, skip-ahead

steps:
  - id: check
    command: exec --json --shell "cat /tmp/status.json"
    flow:
      - when: $check.json.has_questions == true
        goto: route-questions
      - when: $check.json.design_complete == true
        goto: approval
      - default: brainstorm

• goto forward (skip-ahead) or backward (loops)
• default unconditional fallback
• max_iterations per step (default 10) — hard error on breach
• Extended condition evaluator: deep property access ($step.json.field.sub), comparison operators (==, !=, >, <, >=, <=), truthiness checks
• Works with existing approval/halt/resume mechanism
• Visit counters persist across halt/resume

2. Sub-workflow Invocation (lobster.run)

steps:
  - id: brainstorm
    command: lobster.run --name brainstorm-loop --args-json '{"task": "${task}"}'

• In-process recursive runWorkflowFile() — not a subprocess
• File resolution: parent dir → $LOBSTER_WORKFLOW_PATH → cwd
• Args passing via --args-json (template-resolved)
• Child halt bubbles to parent via linked state keys
• Max nesting depth: 10 ($LOBSTER_MAX_WORKFLOW_DEPTH)
• Supports --name (search) and --file (explicit path)

Stats

• ~360 lines new code in src/workflows/file.ts
• ~920 lines of tests across 5 new test files
• 100% backward compatible — no flow field = linear execution as before
• All existing tests pass

Test files

• test/condition-evaluator.test.ts — extended condition evaluator
• test/workflow-flow-validation.test.ts — load-time flow validation
• test/workflow-flow.test.ts — flow execution (loops, branches, max_iterations)
• test/workflow-subworkflow-parsing.test.ts — lobster.run command parsing + file resolution
• test/workflow-subworkflow.test.ts — sub-workflow execution + halt propagation
• test/workflow-integration.test.ts — combined feature tests

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e1242bce44

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Include protocol/version fields in generated status resume tokens

getRunDetail() builds resumeToken without protocolVersion, v, or kind, but resume decoding requires those fields (src/resume.ts checks protocolVersion and v before accepting any token). In practice, the token returned by lobster status cannot be passed to lobster resume because it is rejected as an unsupported token format, so this new status/detail path exposes a non-functional resume token.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Skip visit-count increment when resuming halted sub-workflow

The step visit counter is incremented before resuming a halted child workflow, so any resume with childStateKey consumes an extra iteration even though the parent step is just continuing prior work. This causes false max_iterations failures for lobster.run steps that halt for approval (for example, max_iterations: 1 will fail on the first resume), which breaks resumable sub-workflows under tighter iteration limits.

Useful? React with 👍 / 👎.