OpenCode v1.18.15 Fixes AI Coding Agent Memory

OpenCode’s v1.18.15 release landed on August 7, 2026, with a deceptively practical theme: make long-running coding-agent sessions more trustworthy across chronology, compaction, attachments, and desktop use. The open-source coding agent’s repository shows roughly 195,000 stars, 25,000 forks, and an MIT license, making it one of the most visible open alternatives in the AI coding-agent market.

What changed in v1.18.15

The release fixes a class of bugs that developers usually discover only after an agent has been running for a while. Message chronology now remains correct even when imported or legacy message IDs arrive out of order. That matters because coding agents often combine persisted sessions, tool results, imported transcripts, and newly streamed messages. If an application sorts by identifiers rather than creation order, a revert or fork can operate on the wrong state.

OpenCode also changed revert and fork actions to use real message chronology instead of message-ID ordering. In practical terms, the agent’s branch point should correspond to what the user actually saw and what the model actually did, not to an incidental database key. This is the kind of maintenance work that rarely appears in launch headlines but directly affects confidence in autonomous development.

The release also improves repeated context compaction. Earlier tool-call history is now retained in summaries instead of losing orphaned results. For an agent, a tool call without its result is more than an aesthetic defect: it can make the model believe an operation was never completed, repeat a command, or draw a conclusion without the evidence that produced it.

A second fix makes blob-based attachments load correctly in the web UI. That expands the reliability of workflows that include screenshots, design references, logs, or other binary inputs alongside source code.

Desktop improvements are part of the architecture

OpenCode v1.18.15 is not only a core-agent release. The desktop app received broader locale coverage, JSON export for complete session transcripts, and more reliable session timeline behavior. Stored messages now stay ordered by creation time, and session lists sort by persisted activity time more reliably.

These features point to a broader product direction. A coding agent is becoming a durable work surface rather than a disposable terminal conversation. Exportable transcripts support audits, bug reports, handoffs, and internal documentation. Correct timelines make those transcripts useful. Localization lowers the friction of putting the tool in front of a wider engineering team.

OpenCode’s repository describes two built-in agents: a full-access build agent and a read-only plan agent. The plan mode denies file edits by default and asks permission before shell commands, which creates a useful boundary between exploration and mutation. The new chronology and compaction fixes strengthen that boundary because planning decisions and execution history remain easier to reconstruct.

Why compaction deserves engineering attention

Context windows are not a free substitute for memory. When a coding agent works across many files, the conversation fills with shell output, diffs, compiler errors, test logs, and tool metadata. Eventually, the system must compress or prune history. A poor compaction strategy can preserve the prose while dropping the causal chain.

The v1.18.15 change to repeated compaction is therefore important. Retaining earlier tool-call history in summaries helps preserve the relationship between an action and its result. That relationship is essential for safe autonomy: the model needs to know not only that a test failed, but which command produced the failure and what changed afterward.

Teams evaluating coding agents should test this directly. Start a session, make several edits, run tests, compact the context, and ask the agent to explain the current state. Then fork or revert the session and check whether the resulting branch matches the visible timeline. These tests are more informative than a single benchmark score because they measure operational memory.

A practical upgrade checklist

Before upgrading a shared installation, export a representative session as JSON and record the current version. Test the build and plan agents separately. Include a workflow with a binary attachment, a workflow with imported or resumed history, and a workflow that performs multiple rounds of compaction.

For teams running OpenCode through remote workspaces, verify that the project root and persisted session data are resolved correctly. For desktop users, check locale selection, transcript export, and timeline ordering. In CI or internal evaluation, retain the release artifact’s checksum and record the exact model/provider configuration used during testing.

The release page lists platform assets for Linux, macOS, and desktop packages, including a release attestation JSON file. That makes it possible to treat the upgrade as a supply-chain event rather than an untracked binary replacement.

The larger developer-tooling signal

OpenCode’s latest update is a reminder that the competitive frontier for coding agents is moving below the model layer. Tool-call recovery, session state, compaction, permissions, attachments, and reproducible exports determine whether an agent can be trusted inside a real repository.

The lesson for builders is straightforward: an agent that writes impressive code once is a demo; an agent that preserves chronology, recovers from partial failures, and exposes its history is infrastructure. OpenCode v1.18.15 spends its release budget on that infrastructure, and that is exactly why developers should pay attention.

Sources: OpenCode repository and release notes: https://github.com/anomalyco/opencode and https://github.com/anomalyco/opencode/releases. Accessed August 9, 2026.

Related reading

• Grok Build 1.0: xAI Terminal Coding Agent Guide
• GitHub Copilot SDK GA: Build Production AI Agents
• Go for AI-Assisted Software Engineering: A Verification-First Guide