diff --git a/README.md b/README.md index d5b3ec87..5a6d2ae3 100644 --- a/README.md +++ b/README.md @@ -130,11 +130,13 @@ Each level overrides the previous, so project settings take priority over global // Soft upper threshold: above this, DCP keeps injecting strong // compression nudges (based on nudgeFrequency), so compression is // much more likely. Accepts: number or "X%" of model context window. - "maxContextLimit": 100000, + // Defaults to 70% of the model window so large-context models are + // not nudged as if they were close to a fixed 100k limit. + "maxContextLimit": "70%", // Soft lower threshold for reminder nudges: below this, turn/iteration // reminders are off (compression less likely). At/above this, reminders // are on. Accepts: number or "X%" of model context window. - "minContextLimit": 50000, + "minContextLimit": "35%", // Optional per-model override for maxContextLimit by providerID/modelID. // If present, this wins over the global maxContextLimit. // Accepts: number or "X%". diff --git a/dcp.schema.json b/dcp.schema.json index 474d5bab..40e0d3da 100644 --- a/dcp.schema.json +++ b/dcp.schema.json @@ -158,7 +158,7 @@ }, "maxContextLimit": { "description": "Soft upper threshold. Above this, DCP keeps sending strong compression nudges (based on nudgeFrequency), so the model is pushed to compress. Accepts number or \"X%\" of the model context window.", - "default": 100000, + "default": "70%", "oneOf": [ { "type": "number" @@ -171,7 +171,7 @@ }, "minContextLimit": { "description": "Soft lower threshold for reminder nudges. Below this, turn/iteration reminders are off (compression is less likely). At or above this, reminders are on. Accepts number or \"X%\" of the model context window.", - "default": 50000, + "default": "35%", "oneOf": [ { "type": "number" @@ -254,8 +254,8 @@ "permission": "allow", "showCompression": false, "summaryBuffer": true, - "maxContextLimit": 100000, - "minContextLimit": 50000, + "maxContextLimit": "70%", + "minContextLimit": "35%", "nudgeFrequency": 5, "iterationNudgeThreshold": 15, "nudgeForce": "soft", diff --git a/lib/config.ts b/lib/config.ts index 792b7ca3..46ea4380 100644 --- a/lib/config.ts +++ b/lib/config.ts @@ -691,8 +691,8 @@ const defaultConfig: PluginConfig = { permission: "allow", showCompression: false, summaryBuffer: true, - maxContextLimit: 100000, - minContextLimit: 50000, + maxContextLimit: "70%", + minContextLimit: "35%", nudgeFrequency: 5, iterationNudgeThreshold: 15, nudgeForce: "soft", diff --git a/tests/context-limit-defaults.test.ts b/tests/context-limit-defaults.test.ts new file mode 100644 index 00000000..01ecf93e --- /dev/null +++ b/tests/context-limit-defaults.test.ts @@ -0,0 +1,134 @@ +import assert from "node:assert/strict" +import { readFileSync } from "node:fs" +import { join } from "node:path" +import test from "node:test" +import type { PluginConfig } from "../lib/config" +import { isContextOverLimits } from "../lib/messages/inject/utils" +import { createSessionState, type WithParts } from "../lib/state" + +// `lib/config` cannot be imported from tests (its `jsonc-parser` ESM subpath has no +// named exports under tsx), so the shipped defaults are asserted against the schema +// and the config source instead of against the runtime object. +const repoRoot = join(import.meta.dirname, "..") + +function percentageConfig(maxContextLimit: string, minContextLimit: string): PluginConfig { + return { + compress: { + mode: "range", + permission: "allow", + showCompression: false, + summaryBuffer: true, + maxContextLimit: maxContextLimit as `${number}%`, + minContextLimit: minContextLimit as `${number}%`, + nudgeFrequency: 5, + iterationNudgeThreshold: 15, + nudgeForce: "soft", + protectedTools: [], + protectTags: false, + protectUserMessages: false, + }, + } as PluginConfig +} + +function buildMessages(inputTokens: number): WithParts[] { + const sessionID = "ses_context_limit_defaults" + + return [ + { + info: { + id: "msg-user", + role: "user", + sessionID, + agent: "assistant", + time: { created: 1 }, + model: { providerID: "opencode-go", modelID: "big-window" }, + } as WithParts["info"], + parts: [ + { + id: "msg-user-part", + messageID: "msg-user", + sessionID, + type: "text" as const, + text: "continue please", + }, + ], + }, + { + info: { + id: "msg-assistant", + role: "assistant", + sessionID, + agent: "assistant", + time: { created: 2 }, + tokens: { + input: inputTokens, + output: 10, + reasoning: 0, + cache: { + read: 0, + write: 0, + }, + }, + } as WithParts["info"], + parts: [ + { + id: "msg-assistant-part", + messageID: "msg-assistant", + sessionID, + type: "text" as const, + text: "done", + }, + ], + }, + ] +} + +function limitsFor(modelContextLimit: number | undefined, tokens: number) { + const state = createSessionState("ses_context_limit_defaults") + state.modelContextLimit = modelContextLimit + return isContextOverLimits( + percentageConfig("70%", "35%"), + state, + "opencode-go", + "big-window", + buildMessages(tokens), + ) +} + +test("shipped default limits are percentages, not absolute token budgets", () => { + const schema = JSON.parse(readFileSync(join(repoRoot, "dcp.schema.json"), "utf-8")) + const compress = schema.properties.compress.properties + + assert.equal(compress.maxContextLimit.default, "70%") + assert.equal(compress.minContextLimit.default, "35%") + assert.equal(schema.properties.compress.default.maxContextLimit, "70%") + assert.equal(schema.properties.compress.default.minContextLimit, "35%") + + const configSource = readFileSync(join(repoRoot, "lib", "config.ts"), "utf-8") + assert.match(configSource, /maxContextLimit: "70%",/) + assert.match(configSource, /minContextLimit: "35%",/) +}) + +test("percentage defaults stay quiet well past 100k on a 1M window", () => { + const at250k = limitsFor(1_000_000, 250_000) + assert.equal(at250k.overMaxLimit, false) + assert.equal(at250k.overMinLimit, false) + + const at750k = limitsFor(1_000_000, 750_000) + assert.equal(at750k.overMaxLimit, true) + assert.equal(at750k.overMinLimit, true) +}) + +test("percentage defaults keep the historical thresholds on a 128k window", () => { + const at80k = limitsFor(128_000, 80_000) + assert.equal(at80k.overMaxLimit, false) + assert.equal(at80k.overMinLimit, true) + + const at95k = limitsFor(128_000, 95_000) + assert.equal(at95k.overMaxLimit, true) +}) + +test("percentage defaults fail open when the model window is unknown", () => { + const unknown = limitsFor(undefined, 500_000) + assert.equal(unknown.overMaxLimit, false) +})