agentsope/SkillAlchemy / skills/agentsop-trajectory-to-semantic-ops
agentsop-trajectory-to-semantic-ops
VerifiedSOP for distilling reusable semantic operators (data-science operation skills) from agent execution trajectories: read action/observation traces, classify per-task failures, bind data-semantic constraints (input semantics / operation decision / output constraints / validation), and feed bounded edits back through an optimizer such as SkillOpt. Use when mining agent runs for cross-task operation knowledge, when a data-science skill keeps failing on the same operation family (join cardinality, group granularity, unit, time alignment), or when wiring trajectory history into a skill optimizer without polluting the skill body.
| File | Kind | Size |
|---|---|---|
| README.md | markdown | 3.7 KB |
| SKILL.md | skill md | 11.6 KB |
| intermediate/operation_candidates.json | config | 8.6 KB |
| references/R1-architecture.md | markdown | 5.8 KB |
| references/R2-sop-workflow.md | markdown | 4.5 KB |
| references/R3-dilemma-cases.md | markdown | 3.8 KB |
| references/R4-anti-patterns.md | markdown | 3.1 KB |
| references/R5-ecosystem-context.md | markdown | 4.1 KB |
README.md
View on GitHub# trajectory-to-semantic-ops — Distilling Semantic Operators from Agent Execution Trajectories
A distilled SOP skill for turning an agent's execution trajectory (action/observation/exit-code stream + scores) into reusable **semantic operators**: data-science operation skills that carry explicit data-semantic constraints (input semantics, operation decision, output constraints, validation). Grounded in the local **AgenticDataBench** repo, whose `testbed/` contains a working trajectory → skill-optimization loop (SkillOpt + history-material injection).
## What this skill teaches
- The three-way distinction between LOTUS-style semantic operators, task-capability labels, and runtime SKILL.md skills — and why only the last is the carrier.
- The real **execution-trajectory schema** (`{kind: message|command, command, exit_code, output}`) and the three rules for reading it without being fooled by the agent's self-narration.
- The **five-field semantic-operator template** (applicability / input semantics / operation decision / output constraints / validation) and the eight high-error operation families.
- The **seven-phase extraction pipeline**: collect → compress → classify → synthesize → inject → optimize → strip.
- The grounded decision rules: prioritize `exit≠0` steps, five-way failure classification, bundle-over-flat history injection, gate-noise limits, and the "invocation count ≠ adoption" trap.
## File map
```
agentsop-trajectory-to-semantic-ops/
├── SKILL.md # main skill (7 sections, Chinese-language SOP)
├── README.md # this file
├── references/
│ ├── R1-architecture.md # trajectory schema + semantic-op schema + pipeline topology
│ ├── R2-sop-workflow.md # 7-phase workflow + function cheat-sheet
│ ├── R3-dilemma-cases.md # grounded decision cases
│ ├── R4-anti-patterns.md # when NOT to extract; common mistakes; hard boundaries
│ └── R5-ecosystem-context.md # vs SkillOpt / LOTUS / EvoDS / Skill Blocks / …
└── intermediate/
└── operation_candidates.json # extracted operation/decision atoms for synthesis
```
## How to consume
- **Coder-agent runtime**: read `SKILL.md`. The 7 sections are self-contained.
- **Horizontal synthesis**: load `intermediate/operation_candidates.json` — each entry has `operation`, `trigger`, `decision_rule`, `evidence_cite`.
- **Deep dive on one dimension**: open `references/R{1-5}-*.md`.
## Provenance
- Trajectory schema, pipeline functions, and experiment numbers are read from the local `AgenticDataBench` repo (`testbed/opt_common.py`, `testbed/SKILLOPT_HISTORY_PATCH.md`, `testbed/SkillOpt_两种优化方式分析.md`, `docs/research/IDEA_ASSESSMENT_2026-09-19.md`) and sample `trajectory.json` / `agent_trace.txt` / `results.jsonl` artifacts.
- Related-work citations come from the project's own survey (`IDEA_ASSESSMENT_2026-09-19.md`), which cites arxiv IDs verbatim.
- No fabricated commands, APIs, or benchmark numbers. Where the project itself marks an estimate as an estimate (e.g. token counts from `tiktoken` vs gateway-reported), the skill preserves that caveat.
## Caveats
- The experiment numbers (baseline 0.315, v2 0.512, gate ±0.22, etc.) are single-run / small-sample observations on 10–13 tasks; treat them as directional evidence, not published results.
- The five-field operator template and eight operation families are the project's *prototype scope*, not a claim of optimal granularity.
- The optimizer-specific flags (SkillOpt `exec_timeout`, `num_epochs`, `use_slow_update`) drift with SkillOpt versions; cross-check live configs for production use.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |