SkillLift boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run analyses SkillLift as a model-dependent builder or improvement plane: software that generates and revises agent skill packages and their scoring rubrics, rather than the enclosing agent that executes the benchmark task. The intended use is to characterize the shipped improvement mechanism, retained state, admission checks and paths that consume that state. It is not to establish benchmark superiority or observed deployment reliability.

The target is in scope because its shipped runners create model clients for rubric planning, skill editing and verification, then connect them to a portfolio evolution coordinator. This is not merely a deterministic patch library. SRC-1, skilllift_eval/runners/skilllift_wildclaw.py and skilllift_eval/runners/skilllift_skillsbench_tasks.py supply implementation evidence for that classification. The coordinator declares its responsibilities as follows:

Extends the task-portfolio state machine with the SkillLift protocol: Mode A refines candidate branches with a rubric verifier (zero oracle calls), Mode B aligns the verifier ranking with the oracle ranking and revises the receipt. Benchmark differences stay in the injected adapter/planner/generator. --- skilllift/skilllift/coordinator/modes.py

The model-dependent edit path is explicit in SRC-1, skilllift_eval/runners/bounded_edits_skill_generator.py:

raw = self.client.call_text( system, user, temperature=0.0, response_format=_response_format_for( self.client, response_format(contract), ), max_completion_tokens=self.max_completion_tokens, ) --- skilllift_eval/runners/bounded_edits_skill_generator.py

The functional boundary includes the algorithm core; prompts and model-call plumbing; portfolio representation, patch validation and persistence; seed loaders; benchmark adapters; shipped CLI and batch drivers; recovery, audit, budget and token-accounting code; baseline implementations; and evaluation/test contracts. Callers and drivers in this repository are part of the target, not excluded applications. In particular, the domain-level tau2 training, fusion and held-out consumption path is included even though the root README foregrounds two other benchmarks. Its shipped driver imports training, rubric revision, fusion and held-out evaluation functions (SRC-1, scripts/skilllift_tau2_run_domain.py); the runner validates and reloads retained MASTER_SKILL content (SRC-1, skilllift_eval/runners/skilllift_tau2.py). Available but not initially inspected code remains eligible evidence at this same commit.

Initial tracing establishes these shipped paths:

Retained content is not only an output file: the store reconstructs an accepted parent from previously recorded candidates. SRC-1, skilllift/skilllift/portfolio/store.py:

for index, accepted in enumerate(state.get("accepted_candidates", [])): record = self.load_candidate(task_id, int(accepted["round_index"]), accepted["candidate_id"]) if record is None: raise PersistenceError(f"accepted candidate record is missing: {accepted}") current = self.materialize_candidate(current, record, workspace / f"accepted-{index:03d}") --- skilllift/skilllift/portfolio/store.py

The boundary kind is complete artifact, partial loop, not whole-system. The frozen evidence is the main repository artifact at the acquired commit; its two benchmark gitlinks are present, but their working directories have no source files. Their referenced commits are metadata, not additional registered sources. No submodule was fetched. The main repository's wrappers, local WildClaw execution engine and injection patch remain in scope; they must not be mistaken for evidence of the unavailable benchmark tasks, graders or host implementations. The target itself explains why these missing components matter (SRC-1, README.md):

The benchmark submodules are lightly patched forks of their upstreams (skill-injection hooks, seed-skill declarations, harness hardening) required to run skill evolution end-to-end; see each fork's git log for the exact deltas. --- README.md

External dependencies include model endpoints, Docker/container images, benchmark data and host runtimes such as OpenClaw and OpenHands. The tau2 driver also expects a sibling tau2-bench installation that is not part of the tracked artifact. These gaps prevent end-to-end conclusions about actual seed/task contents, host skill loading, grader correctness, environment isolation and execution outcomes. Host scheduling, tool execution and host memory behavior are not attributed to SkillLift. No target program, tests or model services were executed. Absent execution traces prevent claims of observed operation or causal validation. README result claims and checked-in figure data can support attributed reported results, not independently reproduced performance.

Repository coverage and inspection status

This table accounts for every top-level tracked entry. “Included; not yet inspected” is neither an exclusion nor an access gap. Rows describe the functional coverage intended for later jobs, not an assertion that every file was read for this boundary.

The frozen revision and applicability cutoff are respectively 599358b4d4c4a27c0e004df8228ab93026600653 and 2026-10-14. The cutoff is the supplied run date, not a claim that later changes were reviewed. The overall tier is code-grounded because executable entry-point, model-call, retention and consumption paths were inspected; individual documentation and result claims retain their weaker evidence layer.

Source register