thincoder 0.12.26 → 0.12.28

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thincoder",
-  "version": "0.12.26",
+  "version": "0.12.28",
   "description": "Thin coding agent - zero dependencies, no build step, Node.js native. Sharp code, zero bloat.",
   "keywords": [
     "ai",

package/src/agent/setup.mjs

@@ -1,6 +1,11 @@
 /**
  * agent/setup.mjs — runAgent pre-flight setup: context injection, system prompt construction, tool injection
  */
+/** Local time, second precision, for the per-run transient reminder. */
+function timeNowLocal() {
+  return new Date().toLocaleString("sv-SE")
+}
+
 import { search as memorySearch, docSearch } from "../memory.mjs"
 import { pushReal } from "../context.mjs"
 import { toOpenAISchema } from "../tools/index.mjs"
@@ -120,11 +125,6 @@ export async function prepareRun(agent, input, callbacks, {
       }
     }
     if (depth === 0) {
-      agent.history.push({
-        role: "user",
-        content: `[System reminder: current time is ${new Date().toISOString()}.]`,
-        transient: true,
-      })
       // Checklist injection: inject pending + in_progress items from .thincoder/checklist.md
       try {
         const { pendingItems } = await import("../tools/checklist.mjs")
@@ -139,6 +139,15 @@ export async function prepareRun(agent, input, callbacks, {
       } catch { /* checklist not available — suppress error */ }
     }
     pushReal(agent, { role: "user", content: input })
+    // Time grounding for EVERY agent depth, pushed LAST (after the user input): transient,
+    // dropped on persist, fresh at every run start. Tail position keeps the second-precision
+    // content out of any prefix — caches stay hit (plugin hit a position-drift regression
+    // when the reminder was interleaved before the input; CLI is aligned preemptively).
+    agent.history.push({
+      role: "user",
+      content: `[System reminder: current time is ${timeNowLocal()} (local; timezone ${Intl.DateTimeFormat().resolvedOptions().timeZone || "local"}).]`,
+      transient: true,
+    })
   }
 
   if (agent._pendingReminders.length > 0) {
@@ -221,14 +230,10 @@ export async function prepareRun(agent, input, callbacks, {
       ? `${base}\n\n${mainOverlay}`
       : base
 
-  // Local time + timezone on every agent (main, subagent, consult) — without it "today"/
-  // "just now"/"recent" in user messages and search freshness are ungrounded.
-  // MINUTE precision, deliberately: system prompts must stay byte-identical across runs
-  // within the same minute or provider prefix caches (DeepSeek cache_hit) never hit.
-  const timeZone = Intl.DateTimeFormat().resolvedOptions().timeZone || "local"
-  const now = new Date()
-  const mins = String(now.getMinutes()).padStart(2, "0")
-  systemPrompt += `\n\nCurrent time: ${now.getFullYear()}-${String(now.getMonth() + 1).padStart(2, "0")}-${String(now.getDate()).padStart(2, "0")} ${String(now.getHours()).padStart(2, "0")}:${mins} (${timeZone}).`
+  // Time injection deliberately does NOT live here: system prompts must be byte-identical
+  // across runs (provider prefix caches). The time rides a transient user reminder per turn
+  // (see the current-time injection below) — variable content belongs in the history, not
+  // the cached prefix.
 
   const projectRules = await loadProjectInstructions(agent.cwd)
   if (projectRules) {

package/src/provider/core.mjs

@@ -240,7 +240,9 @@ export function normalizeToolPairing(messages) {
       rest.push(m)
     }
   }
-  if (toolById.size === 0) return messages // no tool messages — nothing to enforce
+  if (toolById.size === 0 && !messages.some((m) => m.role === "assistant" && m.tool_calls?.length)) {
+    return messages // no tool messages AND no tool_calls declared — nothing to enforce
+  }
   const out = []
   for (const m of rest) {
     out.push(m)

package/src/session.mjs

@@ -325,8 +325,14 @@ export function saveSession(agent) {
   // _fullHistory is written at the source via pushReal — no flush needed here.
   // history        = FULL, never-compacted (human-readable; VS Code panel & CLI resume read this)
   // contextHistory = machine context (possibly compacted) so CLI resume keeps the token savings
+  // Human line: transient machine injections never enter the readable record.
   const history = (agent._fullHistory ?? agent.history).filter((m) => !m.transient && !isLegacyTransient(m))
-  const contextHistory = agent.history.filter((m) => !m.transient && !isLegacyTransient(m))
+  // Machine line (contextHistory): KEEP transient messages — resume must rebuild the
+  // machine line byte-identical to what the provider cache saw. Dropping them made every
+  // process restart diverge at the first injection position (git/OS/time reminders are
+  // re-injected with FRESH content) → whole-prefix cache miss on the CLI's very first
+  // request of each session (2026-08-16 cache-hit report; Kimi review).
+  const contextHistory = agent.history.filter((m) => !isLegacyTransient(m))
   const data = {
     version: 2,
     cwd: agent.cwd,