Skip to content

reference.md

Latest commit

 

History

History
104 lines (76 loc) · 3.84 KB

reference.md

File metadata and controls

104 lines (76 loc) · 3.84 KB

Reference

Commands

/delve <topic>

Run the full research pipeline on the given topic.

Flags:

Flag Default Description
--quick off Skip DIVE and VERIFY stages (scan + synthesize only)
--depth <level> medium shallow (2 agents) / medium (3-4) / deep (5-6)
--providers claude off Single-model mode, no external AI calls
--output <path> auto Custom output path for synthesis report

Conflicts: --quick + --depth -> depth is ignored (quick always skips dive/verify).

/delve resume [run_id]

Resume an interrupted run from its last checkpoint. Without run_id, lists recent runs for selection.

/delve status

List recent runs with their status, topic, depth, and output path.

Constants

Name Value Notes
CACHE_DIR ~/.cache/delve/runs Run storage
WORKER_TIMEOUT_MS 120000 Per-agent timeout
COVERAGE_THRESHOLD 0.7 Min weighted coverage to proceed
CONTESTED_THRESHOLD 0.3 Above -> mark as draft
P_WEIGHTS P0=3, P1=2, P2=1 For coverage calculation
MAX_CLAIMS shallow=20, medium=35, deep=60 Depth-dependent claim limit

Output format

Reports are saved to docs/research/YYYY-MM-DD-<topic-slug>-<run_id_short>.md with YAML frontmatter:

---
topic: <topic>
run_id: <run_id>
date: <YYYY-MM-DD>
depth: <shallow | medium | deep>
verification: <verified | partially-verified | unverified>
completion: <complete | incomplete | synthesis_only | draft | cancelled | no_evidence>
sources: <count>
claims_verified: <count>
claims_contested: <count>
---

Report sections: Key Findings, Contested Points, Gaps & Open Questions, Sources, Methodology.

Quality labels

verification_status (uses verified_ratio, not coverage):

  • verified: verified_ratio >= 0.8, 0 rejected among P0 claims, >50% verified claims backed by T1/T2 sources, full pipeline
  • partially-verified: 0.5-0.79, or degraded verify, or --providers claude, or ≤50% T1/T2 backing
  • unverified: < 0.5, or verify skipped/failed

completion_status:

  • complete: all stages ran, all P0 sub-questions covered
  • incomplete: DIVE had partial success
  • synthesis_only: no new research, used existing docs
  • draft: contested_ratio > 0.3
  • cancelled: user interrupted
  • no_evidence: 0 sources found

Decision gates

Gate Condition Action
no_evidence 0 sources + 0 existing research Warn user, offer abort or first-principles
reuse Comprehensive existing research < 30 days Skip to SYNTHESIZE
refresh Existing research > 30 days DIVE focuses on deltas
extend Partial existing research DIVE focuses on uncovered aspects
full-run No prior research Full pipeline

Error handling

Error Stage Action
WebSearch unavailable SCAN Fallback to existing research, warn user
0 sources found SCAN no_evidence gate
Worker timeout DIVE Mark timeout, retry P0 once
P0 failed after retry DIVE ABORT
Coverage < 0.7 DIVE Proceed with incomplete + warn
Claim extraction fails VERIFY Skip verify, mark unverified
Prompt injection detected DIVE/VERIFY Strip content, log event, continue

Troubleshooting

"Run directory was cleaned up" — The run completed or the cache was cleared. Start a new run.

Resume re-dispatches all DIVE workers — A prompt template changed since the original run, triggering cascade invalidation. This is expected behavior.

Verification always returns partially-verified — Check if --providers claude is active. Same-model verification is capped at partially-verified by design.

HITL checkpoint not appearing — Only shown in non-quick mode during DECOMPOSE. Quick mode skips directly to SYNTHESIZE.