fix(ralph-loop): keep LLM in smart zone with fresh context per iteration

[edxeth]

The Problem: Context Window Overflow

The original ralph-loop implementation kept all iterations in a single session, causing the context window to fill up over time. This is exactly the problem described in "Why the Anthropic Ralph plugin sucks":

As LLMs receive more tokens, the relationships between tokens scale quadratically. Each additional token makes it harder for the model to process information and make good decisions.

Every LLM has a smart zone (first ~40% of context) and a dumb zone (remaining ~60%). The original implementation guaranteed you'd enter the dumb zone after just a few iterations:

Iteration	Context Usage	Zone
1	~20%	✅ Smart
2	~35%	✅ Smart
3	~50%	❌ Dumb
4+	60%+	❌ Dumb

The Solution: Fresh Context Per Iteration

This PR adds a default_strategy config option with "reset" as the new default:

Strategy	Behavior	Use Case
reset ⭐	Creates NEW session each iteration	Long-running loops, production work
continue	Keeps same session	Short tasks, debugging

With reset, each iteration starts with a fresh context window. The LLM always operates in the smart zone:

Iteration 1: Fresh context → Smart zone → Commit
Iteration 2: Fresh context → Smart zone → Commit  
Iteration 3: Fresh context → Smart zone → Commit
...

This matches how the original bash-loop Ralph works:

while :; do cat PROMPT.md | claude-code ; done

Each invocation = fresh context.

Changes

Core Feature

• New default_strategy config with reset (default) and continue options
• --strategy flag for per-loop override: /ralph-loop "task" --strategy=continue
• Strategy stored in loop state - persists across iterations

Bug Fixes

• Skip todo-continuation-enforcer when ralph-loop is active - prevents hook conflicts
• Conditional prompt prefix - reset allows keyword detection (new session needs mode injections), continue skips it (mode already applied)

Implementation

• TUI session switch via raw API call to /tui/select-session
• Strategy parsed from command arguments in both chat.message and PreToolUse handlers

Usage

Config (default for all loops):

{
  "ralph_loop": {
    "enabled": true,
    "default_strategy": "reset"
  }
}

Command (per-loop override):

/ralph-loop "Build feature X" --strategy=continue --max-iterations=10

Testing

• 42 ralph-loop tests pass ✅
• 37 todo-continuation-enforcer tests pass ✅
• Manually verified both strategies work correctly ✅

References

• Everything is a Ralph Loop - Original concept
• Why the Anthropic Ralph plugin sucks - The problem this PR fixes
• 11 Tips for AI Coding with Ralph Wiggum - Best practices

[github-actions]

All contributors have signed the CLA. Thank you! ✅
_{Posted by the CLA Assistant Lite bot.}

[edxeth]

I have read the CLA Document and I hereby sign the CLA

[cubic-dev-ai]

4 issues found across 14 files

Confidence score: 3/5

• Some regression risk: in src/hooks/ralph-loop/index.ts, a cancelled loop can be re-created after async session reset because writeState runs without re-checking active state, which could cause user-visible loop behavior.
• Several configuration/UX inconsistencies are flagged (docs and CLI/prompt strategy defaults) but have lower or uncertain confidence, so impact is less clear.
• Score 3 reflects a concrete runtime behavior risk in loop cancellation plus a few medium-severity inconsistencies that may confuse configuration.
• Pay close attention to src/hooks/ralph-loop/index.ts, docs/features.md, src/features/builtin-commands/templates/ralph-loop.ts, src/index.ts - loop reactivation risk and strategy configuration inconsistencies.

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="docs/features.md">

<violation number="1" location="docs/features.md:265">
P2: Docs use obsolete `context_strategy` key; config schema expects `default_strategy`, so users will configure an invalid option.</violation>
</file>

<file name="src/features/builtin-commands/templates/ralph-loop.ts">

<violation number="1" location="src/features/builtin-commands/templates/ralph-loop.ts:29">
P2: The prompt hard-codes the default strategy as "reset" even though `default_strategy` is configurable, so instructions become inaccurate when users set a different default.</violation>
</file>

<file name="src/hooks/ralph-loop/index.ts">

<violation number="1" location="src/hooks/ralph-loop/index.ts:350">
P2: Cancelled loop state can be recreated after async session reset because writeState runs without re-checking that the loop is still active.</violation>
</file>

<file name="src/index.ts">

<violation number="1" location="src/index.ts:490">
P2: Case-insensitive `--strategy` parsing passes through unnormalized values, so inputs like `--strategy=CONTINUE` won’t match downstream strict comparisons and can fall back to the default strategy.</violation>
</file>

---

Since this is your first cubic review, here's how it works:

• cubic automatically reviews your code and comments on bugs and improvements
• Teach cubic by replying to its comments. cubic learns from your replies and gets better over time
• Ask questions if you need clarification on any suggestion

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

1 issue found across 1 file (changes from recent commits).

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="src/index.ts">

<violation number="1" location="src/index.ts:663">
P2: Ralph Loop argument parsing is duplicated in three places (chat.message, /ralph-loop, /ulw-loop). This duplication is already diverging (only slash commands lowercase `strategy`), which risks inconsistent behavior and missed fixes. Consider extracting shared parsing into a helper to keep behavior consistent.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[edxeth]

@code-yeongyu this little feat really takes OMO to another level, it would be great if we could work out a planner for ralph loop-specific SPEC files :)

[cubic-dev-ai]

2 issues found across 4 files (changes from recent commits).

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="src/hooks/ralph-loop/index.ts">

<violation number="1" location="src/hooks/ralph-loop/index.ts:76">
P2: Reset strategy prompts do not inject the completion promise token, leaving {{COMPLETION_PROMISE}} unresolved and preventing custom completion promises from being honored.</violation>
</file>

<file name="src/index.ts">

<violation number="1" location="src/index.ts:520">
P2: Prompt sanitization was removed; raw CLI flags are now passed into ralphLoop.startLoop and become part of LLM prompts, regressing previous behavior that stripped flags from the task description.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}