Track durable growth from nested instructions and skill activations

[opencode-agent]

Problem

V2 compaction bounds active model context, but dynamically loaded instruction-like conversation records continue accumulating in durable session history:

• Nested AGENTS.md discovery appends synthetic messages (intended once per path per session).
• Every skill invocation appends a skill message, including repeated activation of the same skill.

Completed compaction excludes older records from active context but does not prune or coalesce them in durable storage. A sufficiently long-lived session can therefore grow without bound on disk.

Scope

Track a later optimization for durable storage without changing current model-facing compaction semantics. Potential approaches include archival, checkpoint-aware pruning, or coalescing records whose relevant state is already represented elsewhere.

Any design should preserve:

• Correct post-compaction continuation.
• Nested-instruction deduplication across process or Location-layer restarts.
• Whatever replay, fork, revert, and audit guarantees V2 intends to support.
• Repeated skill activation when it is semantically meaningful.

Notes

This is not an immediate correctness blocker; active context is already bounded by compaction. It is a long-session storage-growth concern.

Source discussion: https://slack.com/archives/C0BE69AHCQP/p1784941847513000

Requested by: @rekram1-node (Aiden Cline via Slack)

[opencode-agent]

Closing after source verification: skill bodies and dynamically discovered nested AGENTS.md messages do not accumulate indefinitely in active model context; completed compaction bounds them through the normal checkpoint path. Their old records remain in durable history, but that is part of the broader event/session retention model rather than the context-growth problem this issue was intended to track. A separate restart-after-compaction nested-instruction dedup edge case was identified and should be tracked independently if we decide to address it.