Bug: perseus memory compact hangs indefinitely when Ollama is configured but the model is slow

[tcconnally]

Running perseus memory compact hangs indefinitely when llm_provider: ollama is configured and the model (e.g. phi3) takes longer than the terminal timeout. There is no progress indicator and no timeout enforcement from the CLI side.

Steps to reproduce:

1. Configure memory.llm_provider: ollama with a slow model like phi3
2. Run perseus memory compact
3. Process hangs indefinitely

Expected

Progress indicator, or respect llm_timeout_s config value, or emit periodic 'still working...' messages.

Actual

Silent hang.

Workaround

Kill the process and disable llm_provider in config. Use deterministic mode.

[tcconnally]

Code review verification (v1.0.3)

Confirmed. _memory_do_compact in src/perseus/agora.py:77–106 has no overall wall-clock deadline:

if provider:
    new_body = _mneme_compact_llm(all_checkpoints, all_pythia, workspace, cfg, provider)
else:
    new_body = _deterministic_narrative(...)

Ollama with a large model streams tokens one-by-one. Each token may arrive within llm.timeout_s (30s default), but the total wall time is unbounded.

Suggested fix

Enforce total-call deadline at _memory_do_compact:

if provider:
    total_timeout = int(cfg.get('memory', {}).get('compact_total_timeout_s', 180))
    import concurrent.futures
    try:
        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as ex:
            fut = ex.submit(_mneme_compact_llm, all_checkpoints, all_pythia,
                            workspace, cfg, provider)
            new_body = fut.result(timeout=total_timeout)
    except concurrent.futures.TimeoutError:
        print(f'> ⚠ LLM compact exceeded {total_timeout}s; falling back to deterministic.',
              file=sys.stderr)
        new_body = _deterministic_narrative(all_checkpoints, all_pythia, '', workspace, cfg)
else:
    new_body = _deterministic_narrative(...)

Note: ThreadPoolExecutor cannot truly kill the worker; the in-flight HTTP request will continue until Ollama responds or the connection breaks. To actually interrupt, use requests with a (connect, read) tuple timeout. Related: #139 (same pattern in MCP).

Acceptance criteria

• Add memory.compact_total_timeout_s: 180 to DEFAULT_CONFIG.
• Test: mock _mneme_compact_llm to sleep 30s, set compact_total_timeout_s=2, assert deterministic fallback within 5s.

[tcconnally]

Fixed in PR #179 (integration branch targeting main).