--- name: routing-tracedecay description: Use when you know the kind of code question but not which TraceDecay tool answers it; names the one call to make. Skip when the task administers TraceDecay itself (projects, configuration, storage). Load this skill before the first tool call. metadata: opencode/slash: "true" --- # Routing TraceDecay Use this when you know the kind of question and not the tool. Start with the opening call for that question, then follow only the evidence needed to answer it. Stop when the question is answered. Locate a file before opening it. | Moment | Skill | Opening call | | --- | --- | --- | | Landed in a repo and need the code that matters for a task | exploring-code | `tracedecay_context` with the task text | | Who calls this symbol | tracing-functions | `tracedecay_callers` | | What this symbol calls | tracing-functions | `tracedecay_callees` | | Is it safe to change this symbol | assessing-impact | `tracedecay_impact` | | Find an exact string or config key | exploring-code | `tracedecay_grep` | | Recover what a prior session said | managing-session-context | `tracedecay_message_search` | | Recall a durable project fact | project-memory | `tracedecay_fact_store_search` | | Facts about one entity | project-memory | `tracedecay_fact_store_probe` | | Facts naming several entities together | project-memory | `tracedecay_fact_store_reason` | | Facts about an entity's neighbors | project-memory | `tracedecay_fact_store_related` | | Find a symbol by its bare name | exploring-code | `tracedecay_find_exact_symbol` | | Find a symbol by its full qualified name | exploring-code | `tracedecay_by_qualified_name` | | Read a known symbol's signature without its body | exploring-code | `tracedecay_signature` | | Read a known symbol's body | exploring-code | `tracedecay_source_body` with its `node_id` | | Read a committed source span, including deleted code | exploring-code | `tracedecay_source_lines` with full `revision`, `file`, `start_line`, and `end_line` | | Summarize exact staged or net HEAD-to-worktree declarations | reviewing-changes | `tracedecay_commit_context` (`staged_only: true` for the index) | | Understand or review a PR or committed ref range | reviewing-changes | `tracedecay_pr_context` | | Inspect working-tree or staged file/hunk metadata | reviewing-changes | `tracedecay_git_diff` with `scope: working_tree` or `scope: staged` | | Read the raw working-tree or staged patch | reviewing-changes | Shell `git diff` or `git diff --cached` | | Inspect possible consumers and tests for specified files | assessing-impact | `tracedecay_diff_context` with `files` | | Select test files for a changed-file manifest | assessing-impact | `tracedecay_affected` with `files` | | Read a completed feedback cycle's test selection | assessing-impact | `tracedecay_affected_tests` with its `request_handle` | | A build or type error | fixing-build-and-type-errors | `tracedecay_diagnose` | A name you already have is a lookup, not a tour of the repository. A question one `grep` in a file you can already name would answer does not need the graph. For PR review, read the briefing's ranked entrypoints before widening the diff. Hosted evidence is retained by default; `refresh_github: true` requests the configured matching PR's provider refresh. Retained CI is not a live inventory of required checks. The brief's output budget is approximate and reports cuts. `diff_context` is prospective: it seeds analysis with every indexed symbol in the supplied files. `commit_context` selects changed declarations from Git: its default is the net HEAD-to-worktree result, including untracked files; `staged_only: true` selects immutable index blobs. Worktree captures are checked per file, not an atomic checkout snapshot. Read comparison and coverage fields. `git_diff` returns structured metadata without hunk bodies. Its `working_tree` scope means index-to-worktree; use shell `git diff` or `git diff --cached` for patch text. `source_lines` reads exact historical commit/path spans as well as indexed selectors. Continue historical reads at `next_line`, keeping the original revision, file, and end line. This uses normal admitted source-service availability and path policy. Raw staged blob IDs are not commit selectors. **Report:** the answer supported by the returned evidence, plus its freshness line when it is not `fresh` and any coverage needed to qualify the conclusion. Do not chain calls the evidence did not call for.