Build initial graph structure. Args: skills_dir: path to skills directory to discover nodes from build_edges: if True

Return True if ``skilldag graph edit-edge ...`` is being invoked.

Load embeddings cache file, failing loudly when a path is specified.

Compose the e_self text for a node (matches initialize.py).

For each row, pick neighbors: adaptive threshold (mean+1σ clipped) with top-K cap. Floor: each skill gets at least TOP_K_MIN neighbors even if th

Asymmetric cosine: cos(anchor.e_needs, other.e_self). For each anchor i with non-None e_needs, score every other skill j by cos(e_needs[i], e

Map the upstream alfworld_run.py argparse Namespace to the fields the SkillDAG runner expects. Upstream-defined skilldag flags (prefixed with

Cosine similarity vs every node's e_self embedding.

Return ``[(neighbor_id, edge_type)]`` reachable from ``skill_id`` via walkable edges, both directions, excluding ``conflicts_with``.

Inline copy of terminus_2's anthropic_caching.add_anthropic_caching. Marks the last 2 messages' content as ``ephemeral`` cache breakpoints for

Scan a completed run directory and produce a query manifest. The manifest maps task_id -> list of (query_string, gold_skill_ids).

Canonical Dockerfile snippet injected between GRAPH_MARKER_START/END. The generated docker-compose file bind-mounts: * the SkillDAG package

Prefer the stable canonical task under the bundled data tree.

Return curated SkillsBench skill ids before generated variants strip them.

Rerun queries on each scale graph and compute Ret@K / MRR.

Compare cold vs edited graph on the same query set.

Ensure skillgraph.json exists at graph_path; bootstrap if missing.

Extract search query strings from a SkillsBench trial directory. Parses trajectory.json or agent/prompt.txt files to find the queries the age

Concatenate every shared section in canonical order. Used directly by SkillsBench's ``inject_skilldag_instruction_protocol`` to inline the same

Append a universal online SkillDAG protocol to generated task instructions. This is the only agent-agnostic prompt channel we can rely on across

litellm completion with anthropic ephemeral caching. Wrapped with retry Backoff schedule: 5s → 10s → 20s → 40s → 80s plus jitter, 5 tries tota

Load ground-truth skill IDs for a trial from the task's gold_skills.json. Args: trial_dir: directory of a single trial run. gold_

Normalize OpenRouter usage to local token buckets. OpenRouter reports prompt caching under ``prompt_tokens_details``. We store uncached promp

Map local OpenRouter aliases to the model id sent to OpenRouter.

Parse one structured agent turn. Expected format: {"thought": "...", "action": "..."} or {"thought": "...", "command": "..."}

Inject the skilldag block between GRAPH_MARKER_START/END. Order matters: 1. Strip any upstream `COPY skills …` lines from the *original* Do

Set `docker_image = "<image_tag>"` in task.toml's [environment] block. Idempotent: if the field already exists it is overwritten; if missing it

Copy canonical task, overlay _skilldag_template files, drop workspace.

Rewrite relative mount paths in _skilldag_template/docker-compose.yaml to be resolvable from the rendered env dir.

Revert committed edit(s). Without ``task_id`` reverts the most recent ``steps`` reversible entries (LIFO); with ``task_id`` reverts ev

Execute one shell command. Return (stdout, return_code). stdout/stderr is truncated to ``COMMAND_OUTPUT_MAX_CHARS`` so a noisy command can't

Dispatched from alfworld_run.py when ``--mode skilldag``.

Trim common reasoning leakage before structured parsing.

Run search() over N queries with ONE embedding API call. Returns a list aligned with input ``queries``; each element is the same ``{m