Timeline agent on the 2026-05-20 full run produced 0 phenomena: initial
round hit max_iterations=60 cap before recording, forced retry then hit
max_iterations=10 cap because every grounding-rejected call burns one
iteration in the new gateway. Two changes restore depth without re-
introducing the original "agent wanders off and never records" failure:
1. Raise retry cap 10 → 30. With grounding auto-rescue (prev commit)
most rejections heal on the first retry, but some still need 2-3
turns; 10 is empirically too tight, 30 leaves headroom.
2. Narrow the retry tool surface to RECORD + graph-write +
read-only-graph-query tools. Investigation tools (list_directory,
sqlite_query, parse_registry_key) are dropped on retry so the agent
can't restart its search loop — the retry is explicitly "record
what you already found, then stop".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First full-case run (runs/2026-05-20T20-15-04/) produced 83 GroundingError
rejections, almost all from a single failure mode: LLM cites a plausible-
looking inv-XXXXXXXX that doesn't exist, while the fact's value is in fact
present verbatim in one of its real tool outputs. The agent knew which
tool it read from, it just mis-typed the citation id.
Two-layer fix in evidence_graph.validate_fact_grounding:
Layer A (silent heal): when the cited inv-id misses, search the same
agent / task's invocations for one whose output contains the value
(strict or normalised substring). If exactly one matches, rewrite
fact.invocation_id in place and accept. Multi-match is NOT auto-
rescued — the candidate ids go back to the LLM so it picks deliberately.
Layer B (informative retry): GroundingError now appends the agent's
recent invocation ids and a brief tool-call summary, so the LLM has
the real ids in front of it for the next attempt rather than
fabricating again from memory.
Both layers preserve the design invariant: the fact's value must still
be present in a real tool output — nothing new can land grounded that
wasn't already verifiable. Cross-agent / cross-task isolation is also
preserved (rescue candidates filtered on agent + task_id).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Move _safe_json_loads from orchestrator.py to llm_client.py and have
_extract_tool_calls use it when parsing <tool_call> JSON blocks from
model output. orchestrator now imports it from llm_client.
Background: in the first full DeepSeek run (runs/2026-05-12T17-25-38),
~10 'Failed to parse tool call JSON' warnings appeared, all from regex
patterns where the LLM wrote \. or \* inside JSON string values:
Failed to parse tool call JSON: {..., "pattern": "Outlook Express|...|\.dbx"}
Failed to parse tool call JSON: {..., "pattern": "ethereal.*\.pcap"}
Failed to parse tool call JSON: {..., "pattern": "lookatlan.*\.txt|..."}
These are exactly the kind of stray-backslash errors stage-1 sanitize
already handles for orchestrator JSON calls — but tool-call extraction
was using bare json.loads. Result: each failed tool call silently dropped
on the floor, the LLM never got a result, and at least one network agent
burned 14m26s spinning before hitting max_iterations=40.
Now the sanitize/log-on-failure path is shared. Verified against the
three failure cases from yesterday's log: all three now parse cleanly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Note that the hard-coded HYPOTHESIS_EDGE_WEIGHTS table is a temporary
choice; an adaptive scheme should be explored once the full pipeline
is stable.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Issues found running the system end-to-end on the NIST CFReDS Hacking Case
disk image (SCHARDT.001, Mr. Evil). Four interconnected fixes:
1. HypothesisAgent boundary leak (two layers)
B.1 Tool set: BaseAgent._register_graph_tools was registering
add_phenomenon / add_lead / link_to_entity for every agent. With
an empty graph in Phase 2, HypothesisAgent "compensated" by
inventing phenomena, dispatching leads, and linking entities.
B.2 Prompt leak: BaseAgent's shared system prompt hard-coded "Call
investigation tools (list_directory, parse_registry_key, etc.)".
HypothesisAgent hallucinated list_directory and wasted 2 LLM
rounds on 'unknown tool' errors before backing off.
Fix:
- Split _register_graph_tools into _register_graph_read_tools +
_register_graph_write_tools.
- HypothesisAgent, ReportAgent, TimelineAgent override
_register_graph_tools to skip write tools.
- HypothesisAgent and TimelineAgent override _build_system_prompt
with focused, role-specific workflows (no Phase A-D investigation
boilerplate).
2. JSON parse failures in Phase 3 lead generation (5/6 hypotheses lost)
DeepSeek emits JSON with stray backslashes (Windows path references)
and occasional minor syntax slips. Old single-stage sanitize couldn't
recover; per-hypothesis fallback silently swallowed each failure.
Fix:
- _safe_json_loads: progressive — stage 0 as-is, stage 1 escape stray
\X (anything not in valid JSON escape set), log raw input on final
failure for diagnosis.
- New _call_llm_for_json helper: on parse failure, append the error
to the prompt and re-call LLM (self-correcting retry, up to 2).
- All 4 LLM-JSON callsites in orchestrator refactored to use it.
3. Phase 1 sometimes skipped add_phenomenon (LLM treated <answer> as deliverable)
Strengthen BaseAgent's RECORDING REQUIREMENT — explicit "your <answer>
is DISCARDED; only graph mutations propagate" plus a new rule:
negative findings (searched X, found nothing) MUST also be recorded
as phenomena, since they constrain the hypothesis space.
4. Phase 4 Timeline was a no-op
TimelineAgent inherited BaseAgent's Phase A-D prompt and never called
add_temporal_edge — produced 0 temporal edges. Override the prompt
with concrete workflow (build_filesystem_timeline ->
get_timestamped_phenomena -> 15-40 add_temporal_edge calls) and
restrict tool set to read-only + its 3 temporal tools.
Verified end-to-end: HypothesisAgent now 8 tools (no writes), ReportAgent
13 (no graph writes), TimelineAgent 10 (read + temporal + timeline).
All 60 unit tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous LLMClient used raw httpx + Claude Messages API (/v1/messages,
x-api-key, Anthropic SSE event types). Incompatible with DeepSeek.
Rewrite LLMClient.__init__/chat/close to use openai.AsyncOpenAI:
- /v1/chat/completions endpoint, OpenAI message format
- Bearer auth, native SDK error types
- Stream chunks via async for + chunk.choices[0].delta.content
Tool calling protocol (ReAct text-based tags) and all surrounding helpers
(_apply_progressive_decay, _fold_old_messages, _partition_tool_calls,
tool_call_loop, etc.) are unchanged — endpoint-agnostic by design.
New optional config params surfaced to config.yaml.agent:
- reasoning_effort: "high" | "medium" | "low" — DeepSeek/o1-style depth
- thinking_enabled: bool — DeepSeek extra_body.thinking switch
main.py and regenerate_report.py pass these through to LLMClient.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five interrelated cleanups:
1. Lead -> Phenomenon provenance
- Phenomenon.from_lead_id field on the dataclass
- BaseAgent.run(lead_id=...) writes self._current_lead_id
- _add_phenomenon auto-injects from agent state (LLM unaware)
- Orchestrator dispatch passes lead.id; Phase 1/2-auto/4/5 stay None
- Merge path preserves the first non-None lead_id on collision
2. Unified Phenomenon <-> Hypothesis link path
- HypothesisAgent only adds hypotheses, never links
- link_phenomenon_to_hypothesis tool + executor removed
- All links go through Orchestrator._judge_new_phenomena
- Phase 2 unconditionally judges after hypothesis generation
- Gap Analysis judges after each dispatch round
(Three previously-missing judge calls now in place.)
3. SSOT in agent subclasses
- Remove RoleTemplate dataclass, ROLE_TEMPLATES dict,
_instantiate_from_template method
- Each agent subclass owns name, role, and tool list
- agent_factory.py shrinks from 299 to 153 lines
- All 7 agents now route through _AGENT_CLASSES (filesystem,
registry, communication, network, timeline were previously dead
subclasses overridden by templates)
4. Configurable edge weights
- HYPOTHESIS_EDGE_WEIGHTS -> _DEFAULT_EDGE_WEIGHTS (private default)
- EvidenceGraph(edge_weights=...) override via config.yaml
- hypothesis_edge_weights section in config.yaml (commented example)
- main.py and regenerate_report.py read and pass through
5. regenerate_report.py auto-picks the latest run/*/graph_state.json
when no CLI arg is given (was a hardcoded date path)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Previous README described a Blackboard-based 4-phase, 6-agent system.
The actual code uses:
- EvidenceGraph with typed weighted edges (Phenomenon/Hypothesis/Entity)
- 5 phases (explicit Hypothesis Generation between survey and investigation)
- 7 agents (added HypothesisAgent)
Documents the confidence update formula, Phenomenon Jaccard merging,
Asset Library inode dedup, tool-result caching, Gap Analysis coverage
check, auto-persistence, and the resume mechanism.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
