You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
GPT models don't output thinking blocks so I'm using a separate plug-in that hooks into Opencode's built in compaction to route it to openai's API for /compact.
Thus it would be nice to just use magic context for it's memory/dreamer agents and disable the tool output pruning.
Proposed solution
a compaction=false config setting that would not inject the nudges, not install the tools for compaction, and not auto compact at all. Let Opencode handle it internally.
Shipped in v0.34.0: compaction.enabled: false keeps the knowledge surfaces (project memory, docs injection, ctx_search, ctx_expand, notes) while native compaction owns history. Switching requires a restart; turning compaction back on suggests /ctx-wrapup as the catch-up path. Details in the release notes: https://github.com/cortexkit/magic-context/releases/tag/v0.34.0
Activity
mason: issue #266 compaction-off investigation report
docs: issue #266 compaction-off mode investigation report
docs: compaction-off (#266) final spec — folded from 6 adversarial ro…
docs: fold OC peer verifications into #266 spec (newest-marker confir…
feat(config): compaction-off mode config substrate (#266 S1)
feat(compaction-off): S1 config substrate — schema block, isCompactio…
feat(compaction-off): S4 tool registration + guidance variant (#266)
feat(compaction-off): S4 tool registry + guidance — ctx_reduce unregi…
feat(compaction-off): S2 conflict detector + native-config writers (#266
feat(compaction-off): S2 conflict detector + native writers — mode-aw…
feat(compaction-off): S3 OpenCode transform gating + mode transitions (…
feat(compaction-off): S3 transform gating + transitions — reduced mod…
feat(compaction-off): command and sidebar mode surfaces (#266 S5)
feat(compaction-off): S5 commands + sidebar — honest surfaces (refusa…
mason: add Pi compaction-off parity (#266 S6)
feat(compaction-off): S6 Pi parity — native compaction proceeds, redu…
mason: add #266 S8 schema fence probes (#1187)
feat(fence): S8 child-spawn fence probe + loud stale-server surfaces …
feat(compaction-off): add rust downgrade and docs (#266 S7)
feat(compaction-off): S7 rust downgrade + docs + integrated e2e sweep (…
mason: remediate #266 compaction-off audit
fix(compaction-off): audit remediation — verified cleanup gating, not…
mason: persist #266 compaction notice retries
fix(compaction-off): durable notice-pending + cleanup-pending record …
alfonso-magic-context commentedon Aug 7, 2026
Shipped in v0.34.0:
compaction.enabled: falsekeeps the knowledge surfaces (project memory, docs injection, ctx_search, ctx_expand, notes) while native compaction owns history. Switching requires a restart; turning compaction back on suggests /ctx-wrapup as the catch-up path. Details in the release notes: https://github.com/cortexkit/magic-context/releases/tag/v0.34.0