@reticlehq/server 2.3.0 → 2.5.0

package/dist/tools/profiles.js

@@ -1,139 +0,0 @@
-import { ReticleTool } from './tool-names.js';
-/**
- * Which MCP tool surface to expose. The advertised tool DEFINITIONS are re-sent to the model on
- * every turn, so a smaller surface is a per-turn token saving that compounds across a loop. Fewer
- * tools also makes the model wander less (fewer turns, higher accuracy). See bench/agent-loop-and-replay.md.
- *
- * MEASURED surface sizes (assert them with profile-reachability.test.ts rather than trusting prose —
- * every count previously written here was wrong, and the token figures below derive from those counts):
- *   core 15 · standard 32 · hybrid 15 · full 43 · dynamic 2
- * core and hybrid are both 14 because a trimmed profile now also advertises the two meta-tools, which
- * is what keeps an un-advertised tool reachable through reticle_run.
- *
- * core — the verify loop a coding agent actually needs: navigate→look→act→observe→assert,
- * WITH direct network + console + state observability (the highest-signal checks).
- * The recommended profile for agent-driven verification.
- * standard — core + common extras (inspect, sequences, animations, flows, session lifecycle,
- * scroll, baselines, …). For agents that need more than the bare loop.
- * hybrid — THE DEFAULT: core verify+oracle tools advertised directly + 2 meta-tools for on-demand
- * reach to everything else. Core accuracy/detection at ~64% less schema tax than full.
- * full — all tools advertised directly. Opt in via RETICLE_TOOL_PROFILE=full for hard-call scripts.
- */
-export const TOOL_PROFILE = {
-    /** dynamic — advertise only 2 meta-tools (reticle_tools + reticle_run); load real tools on demand.
-     * Fixed ~hundreds of tokens/turn regardless of how many tools exist. See dynamic-tools.ts. */
-    DYNAMIC: 'dynamic',
-    /** hybrid — the core verify tools advertised directly (so the agent acts reliably) PLUS the 2
-     * meta-tools for on-demand reach to every other tool. Core accuracy + full reach at ~core cost. */
-    HYBRID: 'hybrid',
-    CORE: 'core',
-    STANDARD: 'standard',
-    FULL: 'full',
-};
-export const TOOL_PROFILE_ENV = 'RETICLE_TOOL_PROFILE';
-// The set an agent needs to verify a change end-to-end. Tool DEFINITIONS are re-sent every turn, so a
-// smaller surface compounds.
-//
-// MEASURED per-turn `tools/list` cost — the bytes an MCP client re-sends EVERY turn — taken off the
-// real wire (spawn `mcp`, read tools/list, measure the serialized result), not estimated:
-//
-//   dynamic     2 tools    1,531 B   ~383 tok/turn
-//   core       15 tools   14,104 B  ~3,526 tok/turn
-//   hybrid     15 tools   14,104 B  ~3,526 tok/turn
-//   standard   32 tools   26,769 B  ~6,692 tok/turn
-//   full       43 tools  105,061 B ~26,265 tok/turn
-//
-// These REPLACE the previous figures (core 6,479 / standard 13,951 / full 20,441), which the comment
-// itself flagged as stale. Two of the three moved in opposite directions, which is why re-measuring
-// mattered: the lean profiles got much CHEAPER because they now drop the advertised outputSchema,
-// while `full` got much more EXPENSIVE because it still carries it — the gap between lean and full is
-// 7.4x, not the ~3x the old numbers implied.
-//
-// Where the remaining cost sits, on the hybrid default: inputSchema is 76% of the payload (parameter
-// descriptions are half of that), tool descriptions are 12%, and outputSchema is already ~0. So the
-// next real saving is in parameter prose, not in dropping more tools.
-//
-// There is a floor, though: an 8-tool cut (dropping act/navigate/wait_for/sessions) was MEASURED to
-// regress real-agent accuracy 5/5 -> 3/5, because the model loses scaffolding and wanders on harder
-// flows. Direct network/console stay (far more discoverable than observe-with-filters -> fewer turns,
-// better verdicts).
-//
-// Evidence status: the 5/5 figure came from a single gpt-4o run and is STALE as a justification.
-// Current-model evidence is indirect but real — the cost-delta run (bench/fix-loop/COST-DELTA.md)
-// drove this hybrid default on a current model and fixed 4/4 cells with ~25% FEWER tool calls than
-// the baseline. A formal core-vs-hybrid A/B on a current model is still UNRUN; do not quote the 5/5
-// number as if it were current. See bench/agent-loop-and-replay.md.
-export const CORE_TOOL_NAMES = new Set([
-    ReticleTool.SESSIONS,
-    ReticleTool.NAVIGATE,
-    ReticleTool.SNAPSHOT,
-    ReticleTool.QUERY,
-    ReticleTool.ACT,
-    ReticleTool.ACT_AND_WAIT,
-    ReticleTool.OBSERVE,
-    ReticleTool.NETWORK,
-    ReticleTool.CONSOLE,
-    ReticleTool.WAIT_FOR,
-    ReticleTool.ASSERT,
-    ReticleTool.STATE,
-    // INSPECT is what turns a finding into an EDIT: it maps a DOM node to `src/App.tsx:104`. Finding a
-    // bug is half the job; knowing which file to open is the half that makes the agent useful, and it
-    // is the one capability here with no substitute in any other verification tool. It sat in
-    // `standard`, so under the default profile an agent had to already know it existed and reach it
-    // through reticle_run — which, observed over a full 43-tool drive, means it never gets called.
-    // The measured floor that justifies a lean core was about CUTTING to 8 (accuracy 5/5 → 3/5), not
-    // about holding at 12; one tool of schema tax to close the find→fix loop is the right trade.
-    ReticleTool.INSPECT,
-    // FEEDBACK is in every profile for the same reason INSPECT is: a tool an agent has to already know
-    // about, and reach through reticle_run, is a tool that never gets called. That is fatal here in a way
-    // it is not elsewhere — an unadvertised feedback channel collects nothing, which is indistinguishable
-    // from not having built one. It is also the cheapest tool on the surface to carry (three params) and
-    // the only one whose whole purpose is telling us which of the other fifteen are failing.
-    ReticleTool.FEEDBACK,
-]);
-const STANDARD_TOOL_NAMES = new Set([
-    ...CORE_TOOL_NAMES,
-    ReticleTool.ACT_SEQUENCE,
-    ReticleTool.NETWORK,
-    ReticleTool.CONSOLE,
-    ReticleTool.ANIMATIONS,
-    ReticleTool.FLOW_SAVE,
-    ReticleTool.FLOW_REPLAY,
-    ReticleTool.FLOW_VERIFY,
-    ReticleTool.FLOW_HEAL,
-    ReticleTool.SESSION, // merged lifecycle/human-channel family
-    ReticleTool.FLOW, // merged flow list/load/delete
-    ReticleTool.RECORD, // merged record start/stop
-    ReticleTool.BASELINE, // merged baseline save/list/diff
-    ReticleTool.SCROLL_TO,
-    ReticleTool.CRAWL,
-    ReticleTool.REPLAY,
-    ReticleTool.EXPLORE,
-    ReticleTool.CONTRACT_SAVE,
-    ReticleTool.NETWORK_MOCK,
-    ReticleTool.VIEWPORT,
-]);
-export function resolveToolProfile(explicit) {
-    const raw = explicit ?? process.env[TOOL_PROFILE_ENV];
-    if (raw === TOOL_PROFILE.DYNAMIC)
-        return TOOL_PROFILE.DYNAMIC;
-    if (raw === TOOL_PROFILE.HYBRID)
-        return TOOL_PROFILE.HYBRID;
-    if (raw === TOOL_PROFILE.CORE)
-        return TOOL_PROFILE.CORE;
-    if (raw === TOOL_PROFILE.STANDARD)
-        return TOOL_PROFILE.STANDARD;
-    if (raw === TOOL_PROFILE.FULL)
-        return TOOL_PROFILE.FULL;
-    // Default: hybrid — the core verify+oracle tools advertised directly (no detection loss, verified
-    // 10/10 on the regression bench) PLUS the 2 meta-tools for on-demand reach to every other tool. ~64%
-    // less per-turn schema than `full` at the same accuracy. Explicit `full` still opts into all tools.
-    return TOOL_PROFILE.HYBRID;
-}
-export function filterTools(tools, profile) {
-    if (profile === TOOL_PROFILE.CORE)
-        return tools.filter((t) => CORE_TOOL_NAMES.has(t.name));
-    if (profile === TOOL_PROFILE.STANDARD)
-        return tools.filter((t) => STANDARD_TOOL_NAMES.has(t.name));
-    return tools;
-}