#5781 Dream runs for 1–2 h looping on the same read_file calls; dream.maxIterations is deprecated/ignored so the global 200-iteration cap applies

Bug Description

Scheduled Dream consolidation runs turn into very long agent loops (25–111 minutes each, up to ~200 tool calls), during which the model alternates between re-reading the same two files dozens of times. dream.maxIterations in config has no effect: the source marks it deprecated (config/schema.py: max_iterations: int = Field(default=15, ge=1) # Deprecated: no longer used) and cmd_dream in command/builtin.py calls loop.process_direct(...) without a max_iterations, so Dream inherits the global agents.defaults.maxToolIterations (200).

Over the last ~26 hours of gateway logs:

dream 09-14 10:16  ran ~25 min   202 tool calls
dream 09-14 12:42  ran ~26 min   201 tool calls
dream 09-14 15:10  ran ~36 min   201 tool calls
dream 09-14 17:47  ran ~95 min    31 tool calls
dream 09-15 05:50  ran ~70 min    12 tool calls
dream 09-15 07:50  ran ~111 min   31 tool calls
dream 09-15 09:54  ran ~91 min   168 tool calls   (session latency_ms: 5447214)

The runs with ~200 tool calls are clearly hitting the global 200 cap. The session checkpoint for the last run shows runtime_checkpoint.iteration: 164, phase tools_completed, on the same read_file call. The reasoning_content shows the model re-deliberating where to route a single small fact (e.g. "python is not on PATH, use python3") over and over, re-reading the same 25 lines of memory/history.jsonl and the same 60 lines of one SKILL.md each time.

Steps to Reproduce

Run nanobot gateway (v0.3.0) with the default Dream config (intervalH: 2, maxIterations: 15, modelOverride: null) and a workspace with a few hundred memory/history.jsonl entries and ~50 skills.
Use a reasoning model as the default (here deepseek/deepseek-v4-flash-0731 via OpenRouter, temperature 0.1).
Wait for the scheduled Cron: executing job 'dream' (dream).
Watch the log: Tool call: read_file(...) for the same two paths repeats every ~30 s for an hour or more; the run only ends when the global maxToolIterations (200) is reached or the model eventually stops on its own.

Expected Behavior

A Dream-specific iteration cap that actually bounds the run (dream.maxIterations), or the key removed from the default config and the deprecation documented, so users aren't led to believe it does something.
Ideally a guard against repeated identical tool calls (same tool + same args N times in a row) during Dream, since consolidation is a narrow task and shouldn't need 200 iterations.

Relevant Logs

2026-09-15 09:54:17 | INFO  | - | Cron: executing job 'dream' (dream)
2026-09-15 09:54:17 | INFO  | - | Processing message from cli:user: You are a memory consolidation engine. Your sole task is to analyze conversation...
2026-09-15 09:55:09 | INFO  | - | Tool call: exec({"command": "tail -c 20000 /root/.nanobot/workspace/memory/history.jsonl 2>/dev/null | head -c 20000"})
...
2026-09-15 10:53:36 | INFO  | - | Tool call: read_file({"force": true, "limit": 25, "offset": 547, "path": "memory/history.jsonl"})
2026-09-15 10:53:48 | INFO  | - | Tool call: read_file({"limit": 60, "path": "skills/<skill-name>/SKILL.md"})
2026-09-15 10:53:54 | INFO  | - | Tool call: read_file({"force": true, "limit": 25, "offset": 547, "path": "memory/history.jsonl"})
2026-09-15 10:54:09 | INFO  | - | Tool call: read_file({"limit": 60, "path": "skills/<skill-name>/SKILL.md"})
2026-09-15 10:54:39 | INFO  | - | Tool call: read_file({"force": true, "limit": 25, "offset": 547, "path": "memory/history.jsonl"})
2026-09-15 10:55:15 | INFO  | - | Tool call: read_file({"limit": 60, "path": "skills/<skill-name>/SKILL.md"})
... (same pair repeats ~80 more times until 11:25)

Session checkpoint (sessions/dream:20260915-095417):

"runtime_checkpoint": {"phase": "tools_completed", "iteration": 164, "model": "deepseek/deepseek-v4-flash-0731",
 "assistant_message": {"role": "assistant", "content": "", "tool_calls": [{"function": {"name": "read_file",
 "arguments": "{\"force\": true, \"limit\": 25, \"offset\": 547, \"path\": \"memory/history.jsonl\"}"}}], ...

nanobot Version

0.3.0

Python Version

3.12

Operating System

Linux

Channel / Platform

Discord

LLM Provider

OpenRouter

Configuration (Optional)

agents:
  defaults:
    model: deepseek/deepseek-v4-flash-0731
    provider: openrouter
    maxTokens: 8192
    contextWindowTokens: 65536
    temperature: 0.1
    maxToolIterations: 200
    reasoningEffort: null
    dream:
      intervalH: 2
      modelOverride: null
      maxBatchSize: 20
      maxIterations: 15
      annotateLineAges: true

Additional Context

Each Dream burns up to ~200 LLM calls with a ~25k-token prompt, running 1–2 hours out of every 2-hour interval. On a 2-vCPU box it is the top CPU consumer and noticeably competes with live Discord replies.
Related but different from #5441 (cursor blocked by a tool error): here there is no error, the model just never converges.
Workarounds I'm considering: raising intervalH, setting dream.modelOverride to a faster/non-reasoning model, or disabling Dream. Happy to share the full Dream session jsonl (redacted) if useful.