🇺🇸English
Synthesis Orchestrator
Drives the full synthesis flow: create_strategy -> run_slice (per slice) -> validate. Uses abd-story-synthesizer for instructions and validation. Supports checkpointing and stop-on-repeated-errors.
When to Apply
- User asks to run the synthesis orchestrator
- User wants to orchestrate story synthesis from source material
- User mentions "run orchestrator", "synthesis workflow", or "synthesis orchestrator"
CRITICAL: Orchestrator as Single Entry Point
When the orchestrator is active, all instructions for the synthesizer MUST go through the orchestrator to the agent.
- Do NOT create
strategy.md, interaction-tree.md, or state-model.md yourself.
- Do NOT run
build.py get_instructions directly and hand output to the agent.
- The orchestrator generates instructions and saves them to
runs/slice-N-run-M-instructions.md (or create_strategy instructions). The agent receives those instructions from the orchestrator's output.
- Flow: Orchestrator → instructions file → Agent. Never bypass the orchestrator.
Prerequisites
-
Synthesizer skill (abd-story-synthesizer) installed
-
Skill space folder set to a folder that exists
Run with --skill-space <path>: python scripts/run_orchestrator.py --skill-space C:/dev/agile_bots_demo/mm3e
Config Location (IMPORTANT)
abd-config.json MUST live in the synthesizer skill, NOT in the skill space.
| Location | Contains | Config? |
|---|
abd-story-synthesizer/conf/abd-config.json | Engine config (skills, context_paths, skill_space_path) | YES — config goes here |
Skill space (e.g. mm3e/) | Content to synthesize (goal.md, context/, docs/) | NO — never put abd-config in the skill space |
The skill space is the workspace/content root. The engine reads config from the synthesizer skill and uses skill_space_path in that config to find the workspace.
Run the Orchestrator
From abd-synthesis-orchestrator:
cd .../agilebydesign-skills/skills/abd-synthesis-orchestrator
python scripts/run_orchestrator.py --skill-space <path-to-skill-space>
Options
| Option | Default | Description |
|---|
--skill-space | required | Path to skill space folder; orchestrator passes to synthesizer (orchestrator does not set it) |
--slice N | all | Run only slice N (1-based) |
--checkpoint | stage | When to pause: never, stage, slice, run |
--max-consecutive-failures | 3 | Stop after N consecutive validation failures |
--max-inner-iterations |
Typical Workflow
- Run with default checkpoint (stage): orchestrator pauses after each slice for human verification
- At checkpoint: run agent with saved instructions in
skill-space/story-synthesizer/runs/slice-N-run-M-instructions.md
- Press Enter to continue after agent produces output
- Use
--checkpoint never for fully automated runs (e.g. CI)
Test Mode (sync run with agent tracking)
- Run
--test-mode: stops at every step (create_strategy, each run_slice, each validate)
- At each checkpoint: run agent, then press Enter to continue 1 run, or type a number (e.g. 5) to do N runs then exit
- State is saved to
skill-space/story-synthesizer/runs/orchestrator-state.json when you exit
- Re-run with
--test-mode to resume from where you left off
- Use
--max-runs N to do N runs then exit (e.g. agent says "proceed with 3 runs" → --max-runs 3)
Self-Correction Pipeline
The orchestrator integrates with the synthesizer's three-layer correction system. Use --correct-level to control depth:
| Level | What happens on validation failure |
|---|
none | No correction — just log and re-run |
run (default) | correct_run: capture DO/DO NOT in run log with wrong/correct examples |
session | run + correct_session: fold corrections into session strategy for future runs |
skill | run + + : promote corrections to skill rules (cross-project) |
Correction instructions are saved to runs/slice-N-run-M-corrections.md alongside the run instructions.
Session Archival
At the end of every session, the orchestrator copies the story-synthesizer/ output to an archive directory for history. Default: C:\dev\synchronizer_runs\<skill-space-name>\<timestamp>\.
- Use
--archive-dir <path> to change the archive location
- Use
--no-archive to skip archival
Skill Version Control
Use --branch <name> to enable git-based version control for the synthesizer skill:
- On validation pass with
--correct-level skill or all: skill changes are committed to the branch
- On validation fail with
--rollback-on-regression: skill changes are rolled back to the pre-correction state
- Use
--no-rollback to keep skill changes even when output degrades (for manual review)
Typical usage:
python scripts/run_orchestrator.py --skill-space C:/dev/mm3e --correct-level all --branch orchestrator
This creates a traceable history of skill evolution driven by session corrections.
Workflow (Essential Instructions)
Lifecycle stages: shaping → discovery → exploration → specification. Each slice goes through all four stages in order.
Strategy validation: (1) Does the strategy make sense from an identification perspective? (2) Does it slice to cover the full domain model? Slices must validate the entire domain (concepts, effects, attributes; commonality or diversity).
Validation flow: Run validate from the synthesizer skill (includes scanners). On failure, the correction pipeline runs automatically at the configured level (--correct-level). Correction instructions are saved for the agent.
Re-runs: When re-running the same stage, use corrections from the previous run as input.
Output storage: Per-run files in runs/; consolidated output in interaction-tree.md and state-model.md. Strategy and output are workspace-specific (skill-space); the synthesizer reads from and writes to the workspace.
Output
All output is in the skill space folder:
- Instructions:
skill-space/story-synthesizer/runs/slice-N-run-M-instructions.md
- Run logs:
skill-space/story-synthesizer/runs/run-N.md
- Artifacts:
skill-space/story-synthesizer/interaction-tree.md, state-model.md
Weekly Installs
1
Repository
agilebydesign/a…n-skills
First Seen
1 day ago
Security Audits
Gen Agent Trust HubPassSocketPassSnykPass
Installed on
amp1
cline1
opencode1
cursor1
kimi-cli1
codex1
