Loop ticks should use evolution, judges, and memory - not bypass them

[electronicBlacksmith]

Problem

phantom_loop ticks bypass the entire intelligence stack - no evolved config, no memory recall, no post-session evolution, no cross-model judges. Each tick is a naive LLM call with no benefit from anything Phantom has learned.

The loop is a context-window management strategy, not a reason to skip the intelligence layer. Interactive sessions get the full stack; loops - the primary mechanism for long-running autonomous work - get none of it.

Proposed fix (3 phases)

Phase 1: Inject evolved config + memory into tick prompts

Each tick gets persona, domain knowledge, error recovery strategies, and recalled memories. Zero extra LLM calls (one local embedding query per tick for memory recall).

• Extend RunnerDeps with optional memoryContextBuilder, evolvedConfig, roleTemplate
• Wire them from src/index.ts (same instances already in the router scope)
• Extend buildTickPrompt() to inject evolved config sections before the goal and memory context before the state file
• Call contextBuilder.build(loop.goal) per tick for memory recall

Modified: src/loop/runner.ts, src/loop/prompt.ts, src/index.ts

Phase 2: Post-loop evolution and memory consolidation

After a loop finishes, synthesize a SessionSummary from accumulated tick transcripts and feed it through afterSession() and consolidateSessionWithLLM(). Phantom learns from autonomous work, not just interactive conversations.

• Accumulate tick prompt/response pairs in-memory during the run
• On finalize, build a SessionSummary (loop status maps to outcome: done->success, stopped->abandoned, budget_exceeded/failed->failure)
• Call evolution.afterSession(summary) - the pipeline is completely channel-agnostic, no changes needed
• Call consolidateSessionWithLLM() to store the run as a vector-backed episode

Modified: src/loop/runner.ts, src/index.ts

Phase 3: Mid-loop critique checkpoints

For long loops (10+ ticks), run a Sonnet 4.6 judge every N ticks to detect drift before the budget burns out. This is the only phase that adds a new LLM call.

• New src/loop/critique.ts module: reads state file + tick history, asks quality judge if the loop is making progress or stuck
• Configurable checkpoint_interval (default 5 ticks, disabled for short loops)
• Critique injected into next tick prompt as "Reviewer feedback" section

Modified: src/loop/runner.ts, src/loop/prompt.ts
New: src/loop/critique.ts

Key notes

• All existing evolution/judge/memory code is channel-agnostic and reusable without modification
• Primary integration point is buildTickPrompt() in src/loop/prompt.ts
• 300-line file cap applies - Phase 3 critique logic lives in a separate module
• Builds on Loop progress feedback in Slack is silent after start message #5 / fix(loop): restore Slack feedback for phantom_loop runs #7 (loop Slack feedback pipeline)