2dph/AGENTS.md

AGENTS.md — 2dph (deductionphile)

Evidence-first brain over the ops/eSlider stack. Facts need proof or they are (not confirmed).

Read first: PLAN → docs.

Method (detective, no fork)

≥2 independent sources of evidence, or the finding is (not confirmed). Link the lexicon yaml path that backs each claim.

• facts root = assertions backed by ≥2 independent sources (docker ps × compose × ssh-config × docs).
• info root = descriptive/narrative leafs, searchable, never asserted as fact.
• Search is deduction: facts → info → web-search (second independent source). An answer is confirmed only if it comes off the facts root.

Hard rules

1. Secrets. ~/.config/brain/, .env, .secrets/ never read into context, printed, or committed. The OnlyOffice password is obtained via the tunnel, never written to this repo.
2. Read-only data sources. Ladybug var/kb.lbug and Postgres are opened read-only for queries. Index rebuilds write to var/ (gitignored).
3. PII. brain-test, cs_brain client data is never read or quoted.
4. No main pushes. Feature branches + GitHub PR (gh); CI (Actions) must be green. Work board: Gitea issues.
5. TDD. Failing test before tool code. Unit tests run offline against fixtures; network/db calls are wrapped.
6. docs reflect behaviour. Any change updates docs/ + PLAN.md status.

Layout

PLAN.md       decisions + execution + open questions
docs/         published docs
skills/       in-project agent skills (vendored, no external links)
bin/          self-describing tools bin/{subject}/{method} (shebang)
bin/serve.go  async Go HTTP server entry (self-executing go run shebang)
bin/watch/    corpus watcher Go package (mtimes, no inotify deps)
bin/server/   async Go HTTP server (goroutines, bounded worker pool)
bin/mail/     mail pipeline: sync (Go), import (md), index_mail (rebuild)
bin/tools/    vendored python libs behind bin/* (kblib, yamlout, websearch)
bin/docker-entrypoint  container entrypoint (brain index|search|serve|watch)
compose.yaml  docker composition (root level, not docker/)
Dockerfile    multi-stage: python deps + static Go binaries
var/          kb.lbug, var/mail/*, caches (gitignored)
.venv/        ladybug + model2vec + mistune

Mail pipeline

bin/mail/sync.go --source onlyoffice,gmail --workers 8 --out var/mail   # raw message.json + attachments
bin/mail/sync.go --source gmail --query 'from:example.com' --out var/mail  # Gmail search (default in:inbox)
bin/mail/import --from-raw var/mail                                     # message.json → message.md (convert only)
bin/mail/index_mail                                                     # rebuild brain incl. all mail (fresh DB)

• sync (Go) downloads messages + attachments; Gmail uses paginated list + body.attachmentId (not partId) for attachments.
• import converts body + attachments to markdown. PDFs use poppler pdftotext -layout fast path (~15ms); textless/scanned PDFs fall back to docling (isolated subprocess — its native onnx can segfault the parent). Conversion never touches the brain DB (crash safety).
• index_mail always rebuilds from scratch (repo corpus + mail). Ladybug corrupts its WAL when brand-new leafs are bulk-inserted while FTS/vector indexes exist; a fresh DB with indexes created last is the only safe path. Keep conversion + indexing separate so a conversion crash can't leave the DB mid-transaction.

Tools

bin/facts/audit ["self"|"facts"|"info"|"stale"]   # 2-source + staleness gate
bin/facts/crm [--dry-run]                         # proof person↔company/company↔project (ooCRM × corpus SoT)
bin/kb/search "query" [--hop N] [--repo X]        # deduction search → YAML
bin/md/tables                                     # what the graph holds → YAML
bin/brain/deduce "question"                       # thinking wrapper

Never start a shell command with cd — use the tool working-directory parameter. Search before reading whole files.

GitHub safety rules (ABSOLUTE — never violate)

1. No absolute paths in committed files. Replace /mnt/, /home/<user>/, /Users/<user>/ with env vars ($HOME, $PROJECTS_ROOT, $DOCS_BASE).
2. No PII in commits. No real names, phones, emails of third parties. Test data must be synthetic (Alice, Bob, Charlie, Diana, example.com).
3. No credentials/secrets in commits. API keys, tokens, passwords, session strings, phone numbers only in gitignored .env files, referenced by path.
4. Curasoft, edelweiss — no files, no mentions. Remove all traces if found.
5. Check git history before push. If any commit contains leaks, rewrite history (rebase + force push) AND delete affected GitHub releases/tags.
6. docs/chat-import-plan.md — reference Gitea issue, never embed secrets.

Communication

Same tone as the corpus: plain, lists, no hype. Sign-off Andriy Oblivantsev. German C1 where useful. Caveman only for agent chat, never in committed docs.