GitNexus/gitnexus/src/cli/tool.ts
Gergő Magyar a12f022b79 feat(mcp): add limit/offset/summaryOnly pagination to impact tool (#1818)
* feat(mcp): add limit/offset/summaryOnly pagination to impact tool (#414)

The impact tool returns unbounded byDepth arrays for hub symbols (base
error classes, shared utilities), producing 140KB+ responses that get
truncated by MCP clients. maxDepth alone does not help when most
dependents are at depth 1.

Add three new parameters:
- summaryOnly: returns counts/risk/processes/modules without byDepth
- limit: caps symbols per depth level (default 100)
- offset: skips symbols for pagination

Also adds byDepthCounts to all responses so agents can see total counts
even when the symbol list is paginated or omitted.

Closes #414

* fix(mcp): prevent pagination from silently truncating cross-repo impact

Address review findings on #1818:

- F1 (blocker): _runImpactBFS no longer defaults to limit 100 when
  limit is not set — only _impactImpl (MCP entry) applies the default.
  Internal callers (impactByUid, group impact) get complete results.
  GroupToolPort.impact interface gains optional limit param, and
  cross-impact.ts passes limit: 10000 for local UID collection.

- F2 (blocker): tool description updated — byDepth is now documented
  as paginated, not 'all affected symbols'.

- F3: impactByUid calls _runImpactBFS without limit, so Phase-2
  neighbor results are no longer capped at 100.

- F4: pagination metadata now appears when offset > 0 (head truncation),
  not just tail truncation. Pagination.limit is null when uncapped.

- F5: limit/offset schema types changed from number to integer;
  Math.trunc applied in implementation as defense-in-depth.

- F6: 7 new tests — multi-depth pagination, offset-only truncation,
  offset past end, float inputs, _runImpactBFS internal uncapped path,
  collectImpactSymbolUids with paginated vs complete data.

* fix(mcp): NaN guard on pagination params, complete GroupToolPort interface

- Add Number.isFinite guard to limit/offset in _runImpactBFS so NaN
  inputs fall through to uncapped/zero defaults instead of producing
  silent empty byDepth with no truncation signal.

- Add offset and summaryOnly to GroupToolPort.impact interface to
  match the implementation and prevent silent param loss at the
  port boundary.

- Replace bounds-only toBeLessThan assertion with exact byDepthCounts
  and pagination assertions per DoD §2.7.

* fix(mcp): address remaining review findings for impact pagination

- #3: Forward limit/offset/summaryOnly through callToolAtGroupRepo
  so group-mode MCP callers can use the new pagination params.

- #4: Extract GROUP_LOCAL_PHASE_LIMIT constant from magic 10000 in
  cross-impact.ts with a comment explaining the intent.

- #7: eval-server formatImpactResult uses byDepthCounts[depth] for
  the 'and N more' suffix instead of paginated slice length.

- #8: Extract ImpactParams interface from duplicate inline type
  definitions in impact() and _impactImpl().

- #9: Add --limit, --offset, --summary-only CLI flags to the impact
  command with i18n help strings (en + zh-CN).

- #10: Clarify in tool description that limit/offset apply per depth
  level, not per total result set.

* chore(autofix): apply prettier + eslint fixes via /autofix command

* @
fix(mcp): address Copilot review feedback on impact pagination

- Sanitize limit/offset with Number.isFinite in _impactImpl to prevent
  NaN passthrough from bypassing the default limit of 100
- Omit pagination.limit field instead of emitting null when paginationLimit
  is Infinity, keeping the response schema consistent
- Move GROUP_LOCAL_PHASE_LIMIT after all imports in cross-impact.ts
- Stop forwarding limit/offset/summaryOnly to group-mode impact since
  runGroupImpact overrides limit with GROUP_LOCAL_PHASE_LIMIT for UID
  collection and does not re-paginate
- Validate CLI parseInt results with Number.isFinite before passing to
  the backend, falling back to undefined so defaults apply
- Use byDepthCounts to decide whether to render depth sections in
  formatImpactResult, handling empty pages from offset past end
@

* @
fix(mcp): address code review findings on impact pagination

- Fix formatImpactResult "N more" count: use Math.min(items.length, 12)
  instead of hardcoded 12, so paginated pages with <12 items show the
  correct remaining count
- Detect summaryOnly responses (byDepth absent, byDepthCounts present)
  and show a summary-mode message instead of misleading "(0 items on
  this page — adjust offset)" per depth level
- Document that limit/offset/summaryOnly are single-repo only and
  ignored in group mode (@groupName) in MCP tool schema descriptions
- List byDepthCounts in summaryOnly description and note byDepth
  absence when summaryOnly is true
- Remove unused limit/offset/summaryOnly from GroupToolPort.impact
  interface since they are never forwarded to group impact
- Deduplicate parseInt calls in CLI tool.ts: extract to local variables
  with consistent optional-chain usage
@

* chore(autofix): apply prettier + eslint fixes via /autofix command

* @
fix(group): restore limit in GroupToolPort.impact interface

cross-impact.ts passes limit: GROUP_LOCAL_PHASE_LIMIT through the
GroupToolPort.impact interface for UID collection. Only offset and
summaryOnly were truly unused — limit must stay.
@

* @
docs: add limit/offset/summaryOnly to impact tool options in README
@

---------

Co-authored-by: Test <test@example.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
2026-05-26 08:48:24 +01:00

192 lines
5.4 KiB
TypeScript

/**
* Direct CLI Tool Commands
*
* Exposes GitNexus tools (query, context, impact, cypher) as direct CLI commands.
* Bypasses MCP entirely — invokes LocalBackend directly for minimal overhead.
*
* Usage:
* gitnexus query "authentication flow"
* gitnexus context --name "validateUser"
* gitnexus impact --target "AuthService" --direction upstream
* gitnexus cypher "MATCH (n:Function) RETURN n.name LIMIT 10"
*
* Note: Output goes to stdout via fs.writeSync(fd 1), bypassing LadybugDB's
* native module which captures the Node.js process.stdout stream during init.
* See the output() function for details (#324).
*/
import { writeSync } from 'node:fs';
import { LocalBackend } from '../mcp/local/local-backend.js';
import { cliErrorKey } from './cli-message.js';
import { formatDetectChangesResult } from './detect-changes-format.js';
let _backend: LocalBackend | null = null;
async function getBackend(): Promise<LocalBackend> {
if (_backend) return _backend;
_backend = new LocalBackend();
const ok = await _backend.init();
if (!ok) {
cliErrorKey('tool.noIndexed');
process.exit(1);
}
return _backend;
}
/**
* Write tool output to stdout using low-level fd write.
*
* LadybugDB's native module captures Node.js process.stdout during init,
* but the underlying OS file descriptor 1 (stdout) remains intact.
* By using fs.writeSync(1, ...) we bypass the Node.js stream layer
* and write directly to the real stdout fd (#324).
*
* Falls back to stderr if the fd write fails (e.g., broken pipe).
*/
function output(data: any): void {
const text = typeof data === 'string' ? data : JSON.stringify(data, null, 2);
try {
writeSync(1, text + '\n');
} catch (err: any) {
if (err?.code === 'EPIPE') {
// Consumer closed the pipe (e.g., `gitnexus cypher ... | head -1`)
// Exit cleanly per Unix convention
process.exit(0);
}
// Fallback: stderr (previous behavior, works on all platforms)
process.stderr.write(text + '\n');
}
}
export async function queryCommand(
queryText: string,
options?: {
repo?: string;
context?: string;
goal?: string;
limit?: string;
content?: boolean;
},
): Promise<void> {
if (!queryText?.trim()) {
cliErrorKey('tool.usage.query');
process.exit(1);
}
const backend = await getBackend();
const result = await backend.callTool('query', {
query: queryText,
task_context: options?.context,
goal: options?.goal,
limit: options?.limit ? parseInt(options.limit) : undefined,
include_content: options?.content ?? false,
repo: options?.repo,
});
output(result);
}
export async function contextCommand(
name: string,
options?: {
repo?: string;
file?: string;
uid?: string;
content?: boolean;
},
): Promise<void> {
if (!name?.trim() && !options?.uid) {
cliErrorKey('tool.usage.context');
process.exit(1);
}
const backend = await getBackend();
const result = await backend.callTool('context', {
name: name || undefined,
uid: options?.uid,
file_path: options?.file,
include_content: options?.content ?? false,
repo: options?.repo,
});
output(result);
}
export async function impactCommand(
target: string,
options?: {
direction?: string;
repo?: string;
depth?: string;
includeTests?: boolean;
limit?: string;
offset?: string;
summaryOnly?: boolean;
},
): Promise<void> {
if (!target?.trim()) {
cliErrorKey('tool.usage.impact');
process.exit(1);
}
try {
const backend = await getBackend();
const rawLimit = parseInt(options?.limit ?? '', 10);
const rawOffset = parseInt(options?.offset ?? '', 10);
const parsedLimit = Number.isFinite(rawLimit) ? rawLimit : undefined;
const parsedOffset = Number.isFinite(rawOffset) ? rawOffset : undefined;
const result = await backend.callTool('impact', {
target,
direction: options?.direction || 'upstream',
maxDepth: options?.depth ? parseInt(options.depth, 10) : undefined,
includeTests: options?.includeTests ?? false,
repo: options?.repo,
limit: parsedLimit,
offset: parsedOffset,
summaryOnly: options?.summaryOnly ?? undefined,
});
output(result);
} catch (err: unknown) {
// Belt-and-suspenders: catch infrastructure failures (getBackend, callTool transport)
// The backend's impact() already returns structured errors for graph query failures
output({
error:
(err instanceof Error ? err.message : String(err)) || 'Impact analysis failed unexpectedly',
target: { name: target },
direction: options?.direction || 'upstream',
suggestion: 'Try reducing --depth or using gitnexus context <symbol> as a fallback',
});
process.exit(1);
}
}
export async function cypherCommand(
query: string,
options?: {
repo?: string;
},
): Promise<void> {
if (!query?.trim()) {
cliErrorKey('tool.usage.cypher');
process.exit(1);
}
const backend = await getBackend();
const result = await backend.callTool('cypher', {
query,
repo: options?.repo,
});
output(result);
}
export async function detectChangesCommand(options?: {
scope?: string;
baseRef?: string;
repo?: string;
}): Promise<void> {
const backend = await getBackend();
const result = await backend.callTool('detect_changes', {
scope: options?.scope || 'unstaged',
base_ref: options?.baseRef,
repo: options?.repo,
});
output(formatDetectChangesResult(result));
}