Skip to content

opencode.md

Latest commit

 

History

History
66 lines (52 loc) · 3.65 KB

opencode.md

File metadata and controls

66 lines (52 loc) · 3.65 KB
Raw

OpenCode — Tier 1: real injection via the shipped plugin

Last verified: 2026-08-31, against opencode-ai 1.18.25 (npm), plugin SDK @opencode-ai/plugin 1.18.25, probes on google/gemini-2.5-flash via opencode run. Repo note: the OpenCode project moved — sst/opencode now redirects to anomalyco/opencode.

The mechanism (why this host is special)

OpenCode loads JS plugins in-process (Bun) and exposes a typed Hooks interface. There is no session-start hook; our shim (.opencode/plugins/memory-kit.js) rides experimental.chat.system.transform, which fires on every model call and lets a plugin append strings to the system prompt. That is deliberately stronger than a one-shot injection:

  • memory is re-read from disk per call — always current;
  • it travels in the SYSTEM prompt, not the transcript — compaction cannot drop it, which makes Claude Code's PreCompact block unnecessary here rather than missing.

The shim also implements the caps check (180 / 32 KB / 3000, same numbers as the hook), the unadopted-repo pointer (never scaffolds), skills registration, and a compaction nudge via experimental.session.compacting (append-to-summary-prompt: "capture unsaved observations into MEMORY.md right after compaction").

Install (verified via the local-path form)

// opencode.json in the user's repo
{ "plugin": ["memory-kit@git+https://github.com/awrshift/agent-memory-kit.git"] }
// or a local checkout path: { "plugin": ["/path/to/agent-memory-kit"] }

The root package.json (name memory-kit, main → the shim) exists exactly for this resolution and is covered by the CI version-agreement check. Both forms verified 2026-08-31: the probes below ran on the local-path form, and the git+ form was then probed against live GitHub post-push — a fresh scratch repo with only opencode.json + a canary MEMORY.md answered YES without reading a file.

What works (verified, canary probes on 2026-08-31)

  • Injection: in an adopted scratch repo, the MEMORY.md canary was present in the system context WITHOUT the agent reading any file, and the identity working agreement was there too. The core "wakes up already knowing" property holds — per call, not just per session.
  • All 8 skills available — registered via config.skills.paths pointing at the plugin's skills/ dir (the mechanism compound-engineering's own spec distrusted "until tested" — now tested). OpenCode also natively reads .claude/skills/ and ~/.claude/skills/, so user skills appear alongside.
  • The unadopted-repo guarantee: a virgin repo got the one-line setup pointer in context and ZERO files created.
  • OpenCode auto-loads AGENTS.md (and even ~/.claude/CLAUDE.md) by itself (documented-only, its rules docs) — the T2 protocol block works here as a fallback even without the plugin.

Degradation vs Claude Code (verified by API shape)

  • No compaction block. experimental.session.compacting can only append context to the summarization prompt, not refuse it. Mitigated by the standing system-prompt injection (the cache itself cannot be lost) plus the capture-after-compaction instruction.
  • No test-edit guard yet. permission.ask (allow/ask/deny) exists and could carry a protect-tests port — manual-check-needed, deferred until someone wants it.
  • The relevant hooks are experimental.* — no API stability guarantee; re-verify on OpenCode upgrades. Pin: probes ran on 1.18.25.

Notes

  • Auth: any provider works; probes used Google via GOOGLE_GENERATIVE_AI_API_KEY.
  • Skill invocation is by name (flat namespace) — identity.md's /memory-kit:<name> spelling maps to plain <name> here.