gitea-admin/obsidian-wiki
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(mcp): add verified ingest job phases

Browse Source 
Expose authoritative ingest phases and supervisor-backed process liveness without semantic activity inference. Require deterministic registry, manifest, archive, file-index, and declared-output verification before an agent-successful job can finish succeeded.\n\nPreserve legacy state values, derive defaults for older job records, keep verification outside the manager lock, and retain partial outputs on verification failure.\n\nValidated with the full unittest suite (205 tests).
This commit is contained in:
Codex
2026-07-16 14:26:58 +00:00
parent a2b4bc393b
commit 2183404966
8 changed files with 888 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
AGENTS.md
+1
View File
@@ -131,6 +131,7 @@ See `wiki-query` and `wiki-export` skills for how the filter is applied.
- **Merge, do not overwrite.** Automatic identifier extraction may add terms, but it must not discard curated registry metadata. Registry values are the durable source for future index rebuilds and remote MCP lookup.
- **Prefer precise identifiers in search ranking.** Exact or prefix matches in structured part-number and series fields should rank above incidental filename substrings such as timestamps or archive hashes.
- **Keep remote ingest gated.** MCP ingest tools are write-capable and must remain disabled unless the service config explicitly enables the `ingest` tool group via `OBSIDIAN_MCP_ENABLED_TOOLS=file-assets,ingest`, `mcp_service.sh tools enable ingest`, or `mcp_service.sh start --enable-ingest`. Uploaded files and direct text notes must stage outside the vault under the MCP runtime root; `wiki-ingest` is responsible for accepting, archiving, and indexing sources inside the vault. Use `ingest_text_note` only for short remote-agent thoughts/notes; it must wrap the text as a temporary markdown upload and then use the fixed `/wiki-ingest` job path. Remote agents should use `create_ingest_upload_ticket` and the returned `upload_url` plus `X-Obsidian-Wiki-Upload-Token` for HTTP multipart uploads; bearer-token uploads remain only for local/admin compatibility. Prefer `start_job=false`, then `preview_upload_metadata`, then `start_ingest_job` after the preview looks correct. `validate_upload` and `preview_upload_metadata` may report exact sha256 duplicates, but must not claim fuzzy dedupe. Upload tickets are consumed only by authorized `POST` uploads, the upload endpoint must enforce configured host/origin restrictions, and oversized request bodies must fail with HTTP `413`. Do not expose base64 or chunked upload tools in the MCP tool surface. They may launch local OpenCode only through the fixed `/wiki-ingest` prompt; never expose arbitrary prompt or shell execution through MCP. Remote ingest responses must not expose local runtime paths; internal job records may keep paths for the fixed local prompt. Remote ingest must keep the OpenCode readiness preflight in front of job launch, and `check_ingest_opencode_readiness(run_smoke_test=true)` should stay available for operators to verify that OpenCode can load the project `wiki-ingest` skill before accepting uploads.
- **Keep remote ingest status authoritative.** Preserve the `queued`, `preflight`, `agent_running`, `verifying`, and `complete` phases. Compute `process_alive` only from process-supervisor ownership while `phase=agent_running`; do not infer semantic activity from logs or elapsed time. Treat `MCP_INGEST_OK` as permission to begin deterministic verification, not as final success. Final `state=succeeded` requires registry, manifest, file-index, configured archive, and declared output-path checks to pass. A failed verification must finish with `state=failed`, preserve partial outputs for diagnosis, and report compact path-redacted failures.
- **Keep remote query gated.** MCP query tools must remain disabled unless the service config explicitly enables the `query` tool group via `OBSIDIAN_MCP_ENABLED_TOOLS=file-assets,query`, `mcp_service.sh tools enable query`, or `mcp_service.sh start --enable-query`. They may launch local OpenCode only through the fixed `/wiki-query` prompt, with no arbitrary prompt or shell execution. `ask_wiki` is the preferred remote-agent query entrypoint; the async `start_wiki_query`/status/result tools remain available for long jobs. `public_only` defaults to false, meaning remote answers may use the full local vault visibility; enable the tool group only for trusted clients and set `public_only=true` per call for filtered user-facing answers. Remote query responses may expose log tails but must not expose local log file paths, and structured fields such as `answer_markdown`, `citations`, `warnings`, and `confidence` must remain backward-compatible with the legacy `answer` field.
- **Keep OpenCode model choice server-side.** MCP clients must not pass arbitrary model names through ingest/query tools. Configure local defaults in `.runtime/obsidian-wiki-mcp.env`: ingest uses `OBSIDIAN_MCP_INGEST_OPENCODE_MODEL=openai/gpt-5.5` with `OBSIDIAN_MCP_INGEST_OPENCODE_VARIANT=high`; query uses `OBSIDIAN_MCP_QUERY_OPENCODE_MODEL=openai/gpt-5.5` with `OBSIDIAN_MCP_QUERY_OPENCODE_VARIANT=minimal`.