Docs: setup gaps found during end-to-end OpenCode + Docker + Ollama install #233

Activity

1. cl0ckt0wer

   added a commit that references this issue on May 8, 2026

   fix(plugin): add session instruction injection and consolidation pipe…

   ...

   This issue will close once commit 2f35f1d is merged into the 'main' branch.

   2f35f1d
2. rohitg00 commented on May 8, 2026

   

   rohitg00

   on May 8, 2026

   Owner

   More actions

   Thanks @cl0ckt0wer — six concrete documentation gaps with reproducible context is the kind of issue every project wants. Triaging each one:

   1. AGENTMEMORY_TOOLS=all silently fails on the MCP shim — the behavior gets fixed by #238 (the shim becomes proxy-aware and exposes all 51 tools when a server is reachable). The README still needs a paragraph clarifying "shim talks to a running server for the full toolset; fall back is 7 tools when no server is reachable." Will land alongside #238.

   2. No stdio→HTTP MCP bridge documented — your 22-line bridge script is exactly what we should be shipping. Two paths:

   • After fix(mcp): proxy all 51 tools to server, not just 7 hardcoded IMPLEMENTED_TOOLS (closes #234) #238 lands, the shim is the bridge — no custom script needed (point command: npx, args: [-y, @agentmemory/mcp] + AGENTMEMORY_URL=http://server:3111 and you get all 51 tools).
   • For users who want a hand-rolled bridge anyway (custom auth, header injection, etc.), I'll add a docs/mcp-bridge.md page with your script as the canonical example. Mind if we credit you in the doc?

   3. AGENTMEMORY_URL missing from README config table — pure docs. Easy fix, will batch.

   4. Self-hosted Ollama embeddings end-to-end guide — agreed, the current "Any model" entry is too sparse. Filing the install + pull + env block + Docker host.docker.internal networking gotcha as a docs/ollama-self-hosted.md page. The Docker networking bit is the non-obvious one and deserves its own callout.

   5. Windows Docker Compose env_file absolute-path silent failure — this is a real Compose-on-Windows issue, not unique to agentmemory, but absolutely worth documenting in the Windows install section. Will add: "use environment: inline or place .env next to docker-compose.yml — absolute Windows paths in env_file: silently no-op."

   6. npm v11 + onnx-proto@4.0.4 malformed version — upstream bug in onnx-proto. We can document the --force workaround under Windows install while we wait for them to publish a fixed semver. Will also pin away from the broken transitive if practical.

   I'll bundle these as a docs PR (docs/setup-gaps.md + README patches) after #237 / #238 / #239 land so the OpenCode setup story is documented against the post-fix state, not the current broken one.

   Leaving open until that docs PR ships. Good list.
3. rohitg00

   added a commit that references this issue on May 15, 2026

   docs(readme): clarify MCP shim vs full server tool count (#233 gap 1) (…

   ...

   ff3e024
4. rohitg00

   closed this as completedin 68fddd4on May 18, 2026
5. Tanmay-008

   added a commit that references this issue on May 19, 2026

   feat: OpenCode plugin with 22 auto-capture hooks (closes rohitg00#236, …

   ...

   This issue will close once commit 68fddd4 is merged into the 'main' branch.

   68fddd4
6. gitrishiom commented on May 26, 2026

   

   gitrishiom

   on May 26, 2026

   Last edited by gitrishiom

   More actions

   Absolutely yes on the credit.

   After re-reading this, I think the lowest-friction addition would be a tiny memory_session_receipt fixture in the OpenCode / Docker / Ollama path. agentmemory already has most of the underlying primitives (AGENT_ID, AGENTMEMORY_AGENT_SCOPE, Session Replay, /agentmemory/export, /agentmemory/audit), so my hunch is the docs could assemble them as one reviewable record after setup.

   I think this can stay docs-only at first, with no new runtime feature. For example:

   {
     "project": "hyper-agents",
     "agent_host": "OpenCode via stdio MCP",
     "agent_id": "software_ops_founder",
     "agent_scope": "shared",
     "hook_coverage": ["SessionStart", "PreToolUse", "PostToolUse", "Stop"],
     "mcp_tools_visible": 51,
     "server_url": "http://localhost:3111",
     "memory_export_ref": "/agentmemory/export",
     "delete_audit_ref": "/agentmemory/audit?event=forget",
     "replay_session_ref": "viewer:session/<session-id>",
     "redaction_policy": "local sqlite + configured provider env only; no external sync",
     "known_gaps": [
       "fallback MCP shim exposes 7 tools when no upstream server is reachable",
       "scope filtering depends on AGENT_ID + AGENTMEMORY_AGENT_SCOPE being set before session start"
     ]
   }

   Why I think this is worth documenting:

   • setup becomes reviewable instead of relying on "I think the memory layer is wired correctly"
   • the new multi-agent isolation story becomes easier to explain with concrete evidence
   • downstream teams get a stable place to point when they want replay / export / delete / audit proof for one session

   If this fits your planned docs PR, I think it could sit near the OpenCode/Ollama walkthrough with a short note that the record is assembled from existing features, not a separate API contract.
7. healdigital

   added a commit that references this issue on Jul 27, 2026

   feat: OpenCode plugin with 22 auto-capture hooks (closes rohitg00#236, …

   ...

   This issue will close once commit 23a979e is merged into the 'main' branch.

   23a979e