cg
Skill44 starsChanged 8 months agoMITView on GitHub
Trace code and assess change impact via the pre-built symbol, call-graph and co-edit index.
- Sends data out
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 2 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 19
All responses are coordinates / file paths — never source bodies.
- Distinctive: line 54
Every endpoint speaks ONE node shape (call it NODE):
- Very specific: line 122
When the answer decides whether a change is safe, cross-check with grep, or use /affected — its file-level import closure does not depend …
- Very specific: line 124
riskdoes NOT reportambiguousIn: When several files define the same qname, callers / callees / impact / related list the others in … - Distinctive: line 152
When the six base endpoints' pure structural data isn't enough — especially when exploring code or evaluating change impact — use these to …
- Very specific: line 212
If related returns ambiguousIn, the same qname exists in multiple files — retry with &filePath=<rel>
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Enter project graph exploration mode (CodeGraph).
CodeGraph = pre-built symbol + call-graph index + git co-edit view. Six endpoints, each answers one class of question:
| Question | Endpoint |
| Where is X defined / which files share the name? | search?q=X |
| Who calls X? | callers?qname=X |
| What does X call? | callees?qname=X |
| Changing X affects which symbols? | impact?qname=X&depth=2 |
| What symbols does file F contain? | file?path=F |
| Which files are commonly edited alongside F? (conventional coupling / parallel registries) | coedit?filePath=F |
All responses are coordinates / file paths — never source bodies. (One exception: search&includeLiterals=true echoes each matched literal's own text in value.) More precise than grep's textual match, cheaper in tokens than Reading whole files.
The 6 graph endpoints ({{BASE_URL}})
# search: find symbols by name → hits carry the same node shape as every other
# endpoint, so a hit's startLine/endLine feed Read directly.
# q is normalized for naming style: user_profile / userProfile / user-profile / USER_PROFILE are equivalent
# Pass includeLiterals=true to also search identifier-shaped string literals (tool names, event names,
# config keys, route paths — the "looks like a name but isn't an identifier" strings). The response
# then carries an extra literals[] array with value / filePath / line / enclosingSymbol per hit.
curl -fsS "{{BASE_URL}}/api/projectGraph/search?cwd=$PWD&q=<NAME>"
curl -fsS "{{BASE_URL}}/api/projectGraph/search?cwd=$PWD&q=<NAME>&includeLiterals=true"
# callers / callees: 1-hop call relations
curl -fsS "{{BASE_URL}}/api/projectGraph/callers?cwd=$PWD&qname=<QNAME>"
curl -fsS "{{BASE_URL}}/api/projectGraph/callees?cwd=$PWD&qname=<QNAME>"
# impact: transitive callers BFS (depth 1-5, default 2; out-of-range is clamped
# silently, not rejected). `truncated: true` = node ceiling hit, list is partial.
curl -fsS "{{BASE_URL}}/api/projectGraph/impact?cwd=$PWD&qname=<QNAME>&depth=2"
# file: file symbol tree (no source). Hierarchical: nested symbols live in
# children[] — always an array, empty for leaves (never null / never absent).
curl -fsS "{{BASE_URL}}/api/projectGraph/file?cwd=$PWD&path=<REL_PATH>"
# coedit: files commonly edited alongside the target = git log history + current working-tree co-edits
# catches "conventional coupling" the call-graph can't see (parallel registries / double-writes / sibling .md configs)
# history[] entries may name PRE-RENAME paths (git --follow), so a path here need not exist today.
# `cooccurrence` is a RAW COUNT; the ratio ships alongside as `probability`.
curl -fsS "{{BASE_URL}}/api/projectGraph/coedit?cwd=$PWD&filePath=<REL_PATH>"Response shapes
Every endpoint speaks ONE node shape (call it NODE):
{ filePath, qualifiedName, name, kind, startLine, endLine, params?[] }
params is OPTIONAL: absent for non-callables (class / interface / type / enum /
const) and for languages without a tree-sitter grammar. params: [] is different —
it means "0 parameters", not "unknown".
search { files[], symbols[], literals?[] } # literals only with includeLiterals=true
files[] = { type:'file', label, hint?, target: { filePath } }
symbols[] = { type:'symbol', label, hint?, target: NODE }
literals[] = { type:'literal', value, filePath, line, enclosingSymbol? }
# narrow on the outer `type`
file { filePath, language, symbols[] }
symbols[] = NODE plus { contentHash, children[] }
# children[] is always an array, empty for leaves (never null / absent)
callers { qname, target: NODE|null, callers[]: { caller: NODE, callLines[] }, ambiguousIn? }
callees { qname, target: NODE|null, callees[]: { callee: NODE, callLines[] }, ambiguousIn? }
impact { qname, target: NODE|null, nodes[]: { symbol: NODE, depth }, truncated, ambiguousIn? }
coedit { target: string, totalCommits, uncommitted[],
history[]: { filePath, cooccurrence, probability, lastCoEdit } }
# probability = cooccurrence / totalCommits, precomputed. `cooccurrence`
# alone is a RAW COUNT — never compare it against a ratio threshold.
context { results[]: NODE + { score, signals[] }, seeds[]: { node, weight, from }, degraded }
related { target, results[]: NODE + { score, relations[] }, coedit[], degraded, ambiguousIn? }
risk { target, totalImpactedNodes, highRisk[]: NODE + { depth, risk{}, tags[] },
suggestedTests[], coedit[], degraded, degradedReason? } # NOTE: no ambiguousIn
affected { testFiles[], byInput[]: { filePath, reachable, reachableTests[] },
unresolved[], stats: { visited, bfsMs, truncated }, degraded }The coedit[] embedded in related / risk is a strict SUPERSET of a /coedit
history row (same fields plus a per-row totalCommits) — you never need a second
request to get recency.
Errors are { error, tag }. Tags seen from these routes: ValidationError (400),
NotFoundError (404), AppError (500 — every endpoint wraps its lookup in one, so
an index/git failure surfaces here), InternalError (500, uncaught defect).
Silent-failure traps
- An unknown qname returns HTTP 200, not 404.
callers/callees/impact/related/riskanswer a typo'd or non-indexed name withtarget: null(every endpoint spells this slottarget, callees included) and empty arrays — byte-identical to a real symbol that genuinely has no callers. Always checktarget !== nullbefore concluding "nothing calls this" / "this change is safe". Same forcoediton an unknown path (200, empty). Onlyfile404s. - A non-null
targetwithcallers: []does NOT mean "nothing calls this".obj.method()resolves only when the receiver's imported name directly names the container — a class (Klass.method()), a namespace import (import * as m), or a re-export chain. Two very common forms resolve to nothing and are dropped from the graph (kept asmethodCalls, which are visibility-only):- a call on an exported singleton instance —
export const repo = new Repo()thenrepo.method(). The receiver is the INSTANCE name, so the lookup misses the class'sRepo>method. In a singleton-style codebase this hides the whole repository / service layer from inbound call queries. - a re-export alias (
export const estimateTokens = countTokens) — calls are attributed to neither name: the alias indexes as aconstwith 0 callers, and the real function never sees them.callers/impact/riskfail outright here;callees(outbound) is fine.relatedonly DEGRADES — it loses thecallerrelation but still returns ppr-neighbor / sibling-in-community / callee neighbours (measured: 10 results for a singleton method whosecallerswas 0). It stays usable for "what else should I read", just not for "who calls this".riskis the dangerous one — it does not come back empty, it comes back confident:totalImpactedNodes: 0,highRisk: [],suggestedTests: [],degraded: false, i.e. "this change is safe and needs no tests".impactlikewise returns a plausible smallnodeslist (just the target itself) withtruncated: false. Neither carries any signal that resolution failed. When the answer decides whether a change is safe, cross-check with grep, or use/affected— its file-level import closure does not depend on call resolution.
- a call on an exported singleton instance —
riskdoes NOT reportambiguousIn. When several files define the same qname,callers/callees/impact/relatedlist the others inambiguousIn— butrisksilently scores an arbitrary one of them. Before trusting a risk report on a common name (render,handler,init), resolve the name withsearchorrelatedfirst, then pass&filePath=<rel>.frequent-coeditrelations quietly vanish in a restructured repo.coedithistory comes fromgit log --follow, so it can cite PRE-RENAME paths.relatedattaches afrequent-coeditrelation only when that path still resolves in the index, so after a large move/rename the relation is silently dropped — whilerelated.coedit[](the raw echo) still lists the stale path anddegradedstaysfalse. Measured on two repos: every coedit partner path was stale, so nofrequent-coeditrelation was reachable at all, despite co-edit strengths of 0.30-0.47 (well over the 0.2 emission threshold). The threshold is onprobability(= cooccurrence / totalCommits), which the API now returns directly.degraded: trueresults are still usable, just lower precision — but an emptycoedit/suggestedTestsundercoedit-unavailablemeans "signal missing", not "no coupling". Do not read it as evidence of safety.
Technical contract
- Endpoints return coordinates only. Fetch source with Read:
Read offset=startLine limit=endLine-startLine+1— works off any hit,searchincluded. - qname uses
Parent>Childform (not.); copyqualifiedNamefrom search's response directly - Cross-file name collisions are listed in
ambiguousIn— pass&filePath=<rel>to disambiguate. Present oncallers/callees/impact/relatedonly.
The 4 advanced endpoints (smart ranking / relatedness / risk)
When the six base endpoints' pure structural data isn't enough — especially when exploring code or evaluating change impact — use these to get scored, signal-annotated results.
| Question | Endpoint |
| Where is the code related to this question / cursor? | context?query=&cursor= |
| What else should I read while looking at X? | related?qname=X |
| Changing X — which few nodes truly matter? Which tests to run? | risk?qname=X |
| Changed these files — which tests should CI run? (conservative closure) | affected?files=… |
# context: multi-source LEXICAL + GRAPH retrieval — NOT semantic. `query` is matched
# as tf-idf over identifier TOKENS, then expanded through the graph (PPR / pagerank).
# There is no embedding step, so a natural-language question whose words are not
# identifiers in this repo returns confident noise: "how do users authenticate with
# oauth" ranks `withFileLock` first, matching the token "with". Feed it identifier-
# shaped terms, and treat a top hit carrying only ppr/pagerank (no query-match) as
# "the query matched nothing".
# (query / cursor / openFiles — at least one, else 400 at-least-one-required.)
# Returns Top-K coordinates + signals, each carrying
# its own payload field: query-match{tfidf} / ppr{pprScore} / pagerank{pagerank} / open{filePath}.
# seeds[] shows what drove retrieval: { node, weight, from: query|cursor|open }.
curl -fsS "{{BASE_URL}}/api/projectGraph/context?cwd=$PWD&query=<TEXT>&cursor=<FILE>::<QNAME>&topK=15"
# related: broader than callers/callees — includes coedit / PPR neighbours / Louvain community
# Each result carries relations[], each with its own payload: caller|callee{callLines[]} /
# ppr-neighbor{pprScore} / sibling-in-community{communityId} / frequent-coedit{cooccurrence,totalCommits}
# Cross-file name collisions are listed in ambiguousIn — pass &filePath=<rel> to disambiguate (same as callers/callees)
curl -fsS "{{BASE_URL}}/api/projectGraph/related?cwd=$PWD&qname=<QNAME>&topK=10"
# risk: risk-scored impact
# Returns highRisk (sorted by risk.score desc) + suggestedTests
# risk{} = { score, callFreq, coeditProb, hasTest, pagerank }
# tags[] = high-risk | untested | frequent-coedit | core | leaf
# suggestedTests[] = { filePath, reason, coveredNodes[] },
# reason = direct-test | coedit-history | sibling-test
# risk.score = callFreq + coeditProb + (hasTest ? 0 : penalty) + pagerank, decayed by depth
curl -fsS "{{BASE_URL}}/api/projectGraph/risk?cwd=$PWD&qname=<QNAME>&depth=2&topK=20"
# affected: file-level reverse-import closure → test files transitively affected
# Sister to /risk: risk is symbol-centric + precision-oriented (for analysis),
# affected is file-centric + recall-oriented (for CI / selective-test pipelines).
# depth is clamped to 1-20 (default 10). Inputs not in the index come back in unresolved[].
curl -fsS "{{BASE_URL}}/api/projectGraph/affected?cwd=$PWD&files=<a.ts,b.ts>&depth=10"
# affected via POST — for large file lists. cwd goes IN THE BODY; passing it as a
# query param returns 400. This is the one endpoint that does not take ?cwd=.
curl -fsS -X POST "{{BASE_URL}}/api/projectGraph/affected" -H 'content-type: application/json' \
-d '{"cwd":"'"$PWD"'","files":["a.ts","b.ts"],"depth":10}'
# format=plain → newline-separated test paths for xargs pipelines; diagnostics move to
# headers: X-Unresolved-Count / X-Visited / X-Truncated / X-Degraded / X-Degraded-Reason
curl -fsS "{{BASE_URL}}/api/projectGraph/affected?cwd=$PWD&files=<a.ts>&format=plain"Advanced endpoint contract
score/risk.scoreare for ranking only; absolute values have no meaningsignals/relations/tagsexplain WHY each result is relevant — feel free to cite them to the userdegraded: truemeans results are still usable but lower precision;degradedReasongives the cause (analytics-warming= backing index warming up, retry later;coedit-unavailable= git history signal unavailable, fall back to manually picking tests). The reason can flip between calls while the index warms.- risk / related responses already include a
coeditfield (target file's coedit history) — DO NOT issue a separate /coedit request for the same file. - If related returns
ambiguousIn, the same qname exists in multiple files — retry with&filePath=<rel> - These four endpoints also return coordinates only; fetch source with Read
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

