llm.ts: V2 runner dropped V1's 'last step' signal — AI never warned to finalize work

[LifetimeVip]

Description

File: packages/core/src/session/runner/llm.ts:256

V1 (packages/opencode/src/session/prompt.ts:1349-1350,1461) sends max-steps.txt (which says "CRITICAL - MAXIMUM STEPS REACHED. Do NOT make any tool calls. Respond with text only.") as an assistant message on the last step:

const maxSteps = agent.steps ?? Infinity
const isLastStep = step >= maxSteps
// L1461:
messages: [...modelMsgs, ...(isLastStep ? [{ role: "assistant", content: MAX_STEPS }] : [])],

V2 has no equivalent. The AI is never told "this is your last turn." On step 24 (the actual last step), the AI creates tool calls that will never be settled — the session terminates right after. This compounds with bug #3 (hard MAX_STEPS cutoff) to waste the final turn entirely.

Fix: Before the last iteration, inject the max-steps.txt content as a system message.

OpenCode version

v1.16.0 (commit 3cf1cef)

[github-actions]

This issue might be a duplicate of or closely related to existing issues. Please check:

• llm.ts: MAX_STEPS=25 hard cutoff with no partial-result preservation #30865: Filed at the same time by the same author — covers the adjacent bug of MAX_STEPS hard cutoff with no partial-result preservation (the two issues are tightly coupled)
• [FEATURE]: customize the "max steps reached" prompt #28074: Feature request to customize the "max steps reached" prompt, which references the same V1 max-steps.txt behavior described here

If these are intentionally separate reports, that's fine — just worth cross-linking them.

[LifetimeVip]

// Reproduction for #30866: V2 runner dropped V1's "last step" signal
// Run: bun run repro_30866.ts
//
// Demonstrates: V1 injected max-steps.txt BEFORE the last turn, warning the AI.
// V2 only returns StepLimitExceededError AFTER the limit is hit — no warning.

import { readFileSync } from "fs"

let passed = 0
let failed = 0

function assert(condition: boolean, msg: string) {
if (condition) { passed++; console.log( ✅ ${msg}) }
else { failed++; console.log( ❌ ${msg}) }
}

// ──────────────────────────────────────────────────
// Scenario 1: V1 had a last-step warning, V2 doesn't
// ──────────────────────────────────────────────────
console.log("\n1. V1 max-steps.txt warns the AI BEFORE the last turn")
console.log( File: packages/opencode/src/session/prompt/max-steps.txt\n)

const v1MaxSteps = readFileSync(
"C:\Users\ABC\Documents\xiazaiqi\opencode\packages\opencode\src\session\prompt\max-steps.txt",
"utf-8"
)
console.log( Content of max-steps.txt:)
console.log( ---)
console.log(v1MaxSteps.split("\n").slice(0, 5).map(l => ${l}).join("\n"))
console.log( ---\n)

assert(
v1MaxSteps.includes("MAXIMUM STEPS REACHED"),
max-steps.txt tells the AI "MAXIMUM STEPS REACHED" (V1 signal)
)
assert(
v1MaxSteps.includes("Do NOT make any tool calls"),
max-steps.txt tells the AI to stop making tool calls
)

// ──────────────────────────────────────────────────
// Scenario 2: V2 runner doesn't warn the AI before the last step
// ──────────────────────────────────────────────────
console.log("\n2. V2 runner (llm.ts) behavior:")
console.log( File: packages/core/src/session/runner/llm.ts\n)

// The V2 runner code:
// Line 87: const MAX_STEPS = 25
// Line 383: for (let step = 0; step < MAX_STEPS; step++) {
// Line 384: needsContinuation = yield* runTurn(input.sessionID, promotion)
// Line 390: return yield* new StepLimitExceededError({ sessionID: input.sessionID, limit: MAX_STEPS })

const v2Code = const MAX_STEPS = 25 for (let step = 0; step < MAX_STEPS; step++) { needsContinuation = yield* runTurn(input.sessionID, promotion) if (!needsContinuation) break } if (needsContinuation) return yield* new StepLimitExceededError(...)

console.log( V2 loop logic:)
console.log(v2Code.split("\n").map(l => ${l}).join("\n"))
console.log()

assert(
!v2Code.includes("last") && !v2Code.includes("final") && !v2Code.includes("warning"),
V2 loop has NO "last step" or "final step" warning signal — AI never knows step 24 is the last chance
)

// ──────────────────────────────────────────────────
// Scenario 3: Impact analysis
// ──────────────────────────────────────────────────
console.log("\n3. Impact: AI never gets to finalize work")
console.log(" Without a 'step 24 of 25 — this is your last turn' signal:\n")

const impacts = [
"AI may be in the middle of reading a file when cutoff hits",
"AI cannot write a summary of completed work",
"AI cannot note what tasks remain unfinished",
"Partial edits are left in an inconsistent state",
"Next user message starts fresh with no context of what was cut short",
]
impacts.forEach((imp, i) => console.log( ${i+1}. ${imp}))
console.log()
assert(true, "V2 runner needs a pre-cutoff warning injected as a system message before step 24/25")

// ──────────────────────────────────────────────────
// Summary
// ──────────────────────────────────────────────────
console.log(\n${"=".repeat(50)})
console.log(Results: ${passed} passed, ${failed} failed, ${passed + failed} total)
console.log("\nConclusion:")
console.log("1. V1 (packages/opencode/src/session/prompt.ts:1349-1350,1461) injects max-steps.txt at step 24")
console.log("2. max-steps.txt tells the AI to finalize work, summarize, and stop using tools")
console.log("3. V2 runner (packages/core/src/session/runner/llm.ts:383-390) has NO equivalent signal")
console.log("4. The AI hits MAX_STEPS=25 without warning, and StepLimitExceededError is returned after the fact")
console.log("5. Fix: inject max-steps.txt (or equivalent) as a system message when step === MAX_STEPS - 1")

[lexlian]

I fixed the issue with PR #31328.

V2 was missing two things V1 had: (1) the agent's configured step limit was ignored — the runner used a hardcoded MAX_STEPS = 25 instead of agent.steps ?? 25 (see #30865); and (2) the synthetic assistant message from max-steps.txt was never appended to the final LLM request, so the model would emit tool calls on the cutoff step that never get settled.

The fix in packages/core/src/session/runner/llm.ts mirrors V1's pattern: it now resolves the agent's step limit at the top of run() for the for-loop bound, threads a step counter (Ref<number>) so runTurnAttempt can read the current iteration, and appends the max-steps prompt as an assistant message on the final iteration. Both regressions are covered by four new tests in packages/core/test/session-runner.test.ts.