Session modes
Magic Context has three effective runtime surfaces: primary sessions, subagents, and compaction-off mode. Primary sessions and subagents use the normal context-management pipeline with different feature sets. Compaction-off mode keeps the knowledge layer but does not manage the context window. The agent-facing reduce surface (ctx_reduce, §N§ prefixes, and reduce nudges) is gated by both the session mode and the session’s actual tool availability.
The modes
Section titled “The modes”Primary session
Section titled “Primary session”The full surface. Your agent gets the historian, compartments, memory/search/note guidance, prompt adjuncts, synthetic-todowrite, auto-search hints, and automatic safety valves. When ctx_reduce is available in the agent’s tool set, it also sees §N§ tags and receives nudges to reduce spent tool outputs as pressure builds.
This is the recommended mode for most users. The agent actively managing spent tool outputs usually beats fully automatic cleanup because the agent knows which results it has already used.
To hide the reduce surface for a particular agent, remove or deny ctx_reduce in that agent’s tool allow-list. The rest of Magic Context keeps running: historian compression, heuristic cleanup, compartments, memory, and search still work.
Caveman text compression (caveman_text_compression.enabled) is an orthogonal opt-in for primary sessions. It can run whether or not ctx_reduce is available, and it only rewrites old user/assistant prose; dropped tags still win.
Subagent
Section titled “Subagent”Subagent sessions (delegated workers and historian, sidekick, or dreamer child sessions) get a lightweight pass:
- Tagging and heuristic cleanup run normally
- No historian, no compartment injection, no prompt-adjunct blocks (
<project-docs>,<user-profile>) - No deferred-note nudges
- Heuristic drops run on every execute pass (not once-per-turn like primary sessions — subagents are effectively one parent turn)
- Overflow is handled via the overflow detection path without emergency-recovery state
- No caveman text compression
Subagents are driven by a parent agent, have bounded lifetimes, and often run in parallel. Turning on the full feature set in each subagent would create redundant work and per-agent cache churn.
Compaction-off mode
Section titled “Compaction-off mode”Set compaction.enabled: false in the user-level magic-context.jsonc and restart the harness. Memory, docs, user-profile, key-file, notes, search, expand, and raw-message indexing stay live through additive injection. The transform remains registered, but Magic Context stops managing context: it creates no new tags or compartments, writes no Magic Context boundary markers, performs no folds, drops, strips, splices, heuristic or emergency reclaim, synthetic context-management injection, temporal markers, nudges, or blocking. ctx_reduce is unavailable; ctx_expand remains available. The harness’s native compaction may run, but this setting does not enable it. fail_closed_blocking is inert: a failed transform passes input messages through.
The first turn after disabling may trigger one native compaction cycle on a long session. Only history hidden solely by Magic Context is exposed; a surviving native boundary remains authoritative, and Magic Context does not pre-trim the history. Marker cleanup is lazy per session, so an unresumed session is cleaned when it is next resumed. When switching back to compaction on, run /ctx-wrapup if the historian is runnable. OpenCode’s native compaction covers child sessions: subagents are managed by native compaction like any session, while Magic Context provides additive memory and docs injection without reclaim. Keep subagent tasks small or leave compaction on for projects that rely on long subagent runs.
Magic Context’s compaction.enabled in magic-context.jsonc is separate from OpenCode’s compaction.auto and compaction.prune in opencode.jsonc; they are different files and different owners. The coexistence contract covers OpenCode and Pi native compaction only. DCP and OMO keep their existing conflict policy. If transform_mode: "rust" is also configured, compaction-off mode selects the TypeScript transform and emits one frozen boot warning because Rust has no reduced-mode contract yet. The no-manager combination (native compaction disabled too) is allowed; Magic Context reports it rather than enabling native compaction. The sidebar shows raw usage with either native compaction or no active compaction, not a Magic Context execute-threshold percentage.
Feature comparison
Section titled “Feature comparison”| Feature | Primary session | Subagent | Compaction-off |
|---|---|---|---|
| Tag tracking | ✓ | ✓ | existing rows inert; no new rows |
§N§ tags in message text | when ctx_reduce is available | when ctx_reduce is available | |
ctx_reduce tool | tool allow-list controlled | tool allow-list controlled | |
ctx_expand tool | ✓ | ✓ | ✓ |
| Historian and compartments | ✓ | ||
<session-history> injection | ✓ | ||
<project-docs>, <user-profile> | ✓ | additive injection | |
| Channel 1 nudge (tool-output reminder) | when ctx_reduce is available | when ctx_reduce is available | |
| Channel 2 ceiling nudge | when ctx_reduce is available | when ctx_reduce is available | |
| Deferred-note nudges | ✓ | ||
| Synthetic-todowrite injection | ✓ | ||
| Auto-search hints | ✓ | ✓ | |
| Execute-pass cleanup and eligible age reclaim | ✓ | ✓ | |
| Force-band emergency drain (85% or threshold + 2) | ✓ | ||
| 95% block and recovery | ✓ | ||
| Caveman text compression | opt-in |
How it connects
Section titled “How it connects”Session modes are a lens on the overall architecture. Primary sessions get the context reduction surface when ctx_reduce is available, plus the historian and deterministic cleanup. Active context management uses cache-safe tagging and the cache architecture.