Target-info mode (menu 12) asked for company, industry, and location as three
blank prompts. Once the company name is known the local Ollama model can often
supply the other two, so ask it and offer the answers as editable defaults.
- llm.research_target(): TargetResearchInput/TargetResearchOutput schemas plus a
research SystemPromptGenerator that tells the model to return empty strings
when it does not genuinely recognize the organization, so an unknown small
client yields blank prompts instead of a confident hallucination.
APITimeoutError is translated to LLMTimeoutError like generate_candidates.
- clean_research_field(): strips, collapses whitespace, caps at 80 chars, and
turns anything non-string into "" so model output cannot be pasted unbounded
into a prompt default.
- main.hcatOllamaResearchTarget(): runs the call inside the existing spinner and
degrades to blank suggestions on timeout or any other failure, so research can
never block the attack.
- attacks.ollama_attack(): shows suggestions as "Industry (freight rail): "
defaults, labelled explicitly as the model's GUESSES rather than OSINT.
- New ollamaAutoResearch config toggle (default true) in both config examples.
Research uses only the configured local Ollama server; the client name is never
sent to a search engine or third-party company-data API.
Co-Authored-By: Claude <noreply@anthropic.com>
Adds a third LLM Attack generation mode ("cracked") that samples the
plaintexts already recovered this session from <hashfile>.out and asks the
model to infer the target organization's password conventions and emit new
candidates in the same style.
- llm.py: new _CRACKED_PROMPT (offensive candidate generation, explicitly
not the denylist-oriented _WORDLIST_PROMPT), a _PROMPTS mode->prompt map,
and a "cracked" branch in _build_request that tells the model not to
repeat already-cracked passwords.
- main.py: extract the wordlist sampling logic into the shared module-level
helper _sample_plaintext_file(path, cap, source_label) and call it from
both the wordlist and cracked branches of hcatOllama; guard missing and
empty .out files.
- attacks.py: offer mode 3 only when <hashfile>.out exists and is non-empty
(matching _markov_pick_training_source), with a clear message otherwise.
- README: document the three modes and the widened ollamaMaxSampleLines
scope.
Co-Authored-By: Claude <noreply@anthropic.com>
The Atomic Agents client was built with no timeout, so an Ollama server that
accepted the TCP connection but never replied (most commonly a large model still
loading into VRAM) left the CLI blocked in agent.run() forever. The caller's
`except Exception` handler only ever fired on connection-refused.
- llm.generate_candidates() takes a `timeout` parameter (default
DEFAULT_TIMEOUT_SECONDS = 300.0) and forwards it to the OpenAI client.
- openai.APITimeoutError is translated into a new domain-level
llm.LLMTimeoutError so main.py need not import openai itself, keeping the
atomic-agents/instructor dependency isolated to llm.py as documented.
- hcatOllama passes the new `ollamaTimeout` config value and prints
timeout-specific guidance (elapsed seconds, VRAM-loading hint, the setting to
raise) instead of the misleading "ensure Ollama is running" message.
- Documented `ollamaTimeout` in config.json.example and README.
Also closes two test gaps: the defensive `except ValueError` handler in
hcatOllama now has coverage, and the misleadingly-named `test_unknown_mode_raises`
is split into explicit unit tests of _build_request's mode validation. Ticked the
completed steps in the LLM Atomic Agents plan doc.
Co-Authored-By: Claude <noreply@anthropic.com>
