≈609 tokens on demand. Your agent reads this file only when SKILL.md points to it.
Changelog
0.3.0 (2026-02-28)
Added
remember as a first-class pipeline skill (closes #328)
- Executes post-synthesis in Phase 4, not as a parallel subagent
- Uses LLM distillation to extract the key insight before writing to memory
- Falls back gracefully if
remembering skill is unavailable
persist=True parameter on orchestrate() — auto-injects a remember subtask
without requiring it in the planner's JSONmemory_ids field in the return dict — list of stored memory IDsPIPELINE_SKILLS dict in skill_library.py for skills that run outside Phase 3
Changed
collect_results() now accepts skills kwarg; pipeline-only skills are excluded
from synthesis input (storage artifacts, not analytical content)build_all_prompts() skips pipeline-only skills (they have no subagent prompt)skill_catalog() includes pipeline skills so the planner can emit themorchestrate() skills kwarg now merges with built-in skills instead of
replacing them entirely
0.2.0 (2026-02-27)
Breaking Changes
- Removed dependency on
orchestrating-agents skill and Anthropic SDK
- Removed
self_answer_ceiling parameter from orchestrate() and skill library
- Default
max_tokens reduced from 4096 to 2048 (subagents) and 8192 to 4096 (synthesis)
Added
client.py: Minimal HTTP client using httpx (~110 lines vs 800+ in claude_client.py)- Self-answering is now a pure LLM judgment call based on task complexity, not sentence counts
Improved
- Total pipeline time ~47% faster (152s → 80s on test document)
- Phase 1: 12s → 8s (cleaner planner prompt)
- Phase 3: 51s → 30s (tighter max_tokens, fewer subtasks)
- Phase 4: 88s → 43s (conciseness instructions, less input to synthesize)
- Planner produces fewer, better subtasks (4 → 3 on same test, dropping redundant synthesis subtask)
- Skill system prompts include conciseness instructions
- Synthesis prompt demands integration over expansion
0.1.0 (2026-02-27)
- Initial implementation: four-phase pipeline
- 8 built-in skills
- Section header context pointers
- Parallel execution via orchestrating-agents
[0.3.0] - 2026-02-28
Added
- add remember as first-class pipeline skill (closes #328)
Fixed
- add analysis memory type instead of overloading world (#328)
[0.2.0] - 2026-02-28
Other
- orchestrating-skills v0.2.0: drop SDK, fix self-answering, 47% faster
