The pre-task artifact producer. Collects multi-source context into a user-reviewed immutable bundle before any AI task runs.
An opensource opencode skill that solves context fragmentation — the 15-minute "where was that decision again — Slack? Jira? PR? Sheet?" tax that happens before every task.
| Continue / Cursor | Aider | Claude Code | Devin | Context Assembler | |
|---|---|---|---|---|---|
| Pre-task multi-source bundle | ❌ | ❌ | ❌ | ❌ | ✅ |
| User-curated (not AI-driven) | @ per-item | /add | Auto | Auto | ✅ explicit |
| Reviewable artifact before execute | ❌ | ❌ | ❌ | ❌ | ✅ mandatory gate |
| Immutable + replayable | ❌ | ❌ | jsonl log | ❌ | ✅ SHA-256 lock.json |
| Hard permission boundaries | ❌ | partial | ❌ | sandboxed | ✅ 5 middleware-enforced |
| Cross-source (Slack+Jira+Confluence) | via MCP | ❌ | ❌ | browser | ✅ first-class |
The decisive gap: No existing tool offers pre-task user-curated multi-source context bundles that the user can inspect, edit, save, and reuse before handing to an executor.
v0.0.0 — Scaffold (PR 1 of 25). NOT production-ready yet.
This repo implements the v1 plan over 25 PRs with Oracle review per PR.
| Stage | PRs | Status |
|---|---|---|
| 1: Foundation | 1-9 | In progress (PR 1) |
| 2: Connectors + Pilot | 10-16 | Planned |
| 3: Composition | 17-21 | Planned |
| 4: Production-Ready Gate | 22-25 | Planned |
- NO autonomous task decomposition
- NO auto-execute after assembly (mandatory review gate)
- NO smart source selection (sources explicit per task-type)
- NO proactive polling / background watch
- NO cross-task synthesis
- NO write-back to sources (read-only audit-proven)
- NO LLM-generated narrative in v1 (raw quotes with citations only)
- NO embedding service in v1 (deterministic content-hash ranking)
All enforced as middleware, not convention:
- Write boundary — No commit/push/merge/reset --hard to protected branches
- Artifact cap — ≤5 auto-bundles/day; HMAC-signed counter
- Surface boundary — No Slack/Jira/team broadcast without literal
APPROVE SENDper call - Audit boundary — Every action logged + replayable via
context-assembler replay <bundle_id> - Tool boundary — Destructive commands (rm -rf, DROP TABLE, force push) blocked at runtime
| Metric | Threshold |
|---|---|
| First-Pass Bundle Acceptance Rate (FPBAR) | ≥80% (≥70% per-category) |
| Replay determinism | ≥99% |
| Bundle coverage rate | ≥30% |
| Time-to-handoff median | ≤5min cold / ≤90s recall-hit |
20% of pilot corpus dual-rated by senior peer (defeats Hawthorne effect).
# After v1.0.0 ships
git clone https://github.com/hoainho/opencode-context-assembler-skill ~/.config/opencode/skills/context-assembler
context-assembler init # Templates user values into reference.yamlContext Assembler is upstream infrastructure. It produces bundle artifacts; downstream skills consume them via sidecar wrappers (zero modification to upstream skills):
bin/cba-wrap-ddd.sh→ diagnostic-driven-debuggingbin/cba-wrap-deep-design.sh→ deep-design (Metis/Oracle agents)bin/cba-wrap-pr-code-reviewer.sh→ pr-code-reviewer (4 parallel subagents)bin/cba-wrap-sdd.sh→ subagent-driven-development (implementer.md injection)bin/cba-wrap-review-work.sh→ review-work (5 parallel context-mining subagents)
MIT — see LICENSE
- opencode — the agent runtime
- diagnostic-driven-debugging-skill — previous skill by same author (proves the PR-gated + oracle-reviewed development cadence)
- Model Context Protocol — the MCP spec this skill builds on