Quick update since earlier today:

• Switched config from enabled: true to mode: "adaptive" | "aggressive".

  • Why: it makes intent explicit and avoids “magic values” (like setting ratios to 0) to get always-prune behavior.

• Replaced the earlier env-var-based wiring with a per-session WeakMap runtime registry.

  • Why: the env var approach was process-global and could leak settings across concurrent sessions in the gateway.

• Removed all context-pruning debug logging.

  • Why: it was adding custom session.jsonl entries + extra config knobs. Keeping the feature focused + low-noise should make it easier to review and upstream.

Behavior is still the same core promise: pruning happens only in-memory in the Pi context hook; the session *.jsonl on disk stays untouched/complete.

I took a look at #386 ("observation masking"). It’s aiming at the same root problem (tool results dominating context), but there are a few important differences vs this PR:

• Wiring / integration point

  • fix(agent): protect bootstrap prefix from context pruning #381: stays inside Pi’s normal extension system via additionalExtensionPaths + a per-session WeakMap runtime. Very little runner churn.
  • feat(agent): add simple observation masking extension #386: wires an ExtensionFactory directly via createAgentSession({ eventBus, preloadedExtensions, extensions }) and does its own discoverExtensions() inside the runner. That’s more moving parts + more runner surface area to maintain.

• What gets protected

  • fix(agent): protect bootstrap prefix from context pruning #381: protects the last keepLastAssistants assistant messages (and everything after that), so we don’t accidentally wipe the “current working set” even if the agent did many tool calls in one step.
  • feat(agent): add simple observation masking extension #386: keeps the last N tool results (keepLast). In practice that can be too aggressive: if an agent does 10 tool calls as part of one step, you can end up masking 9/10 of the tool results that are still relevant to the very next model call.

• Pruning behavior

  • fix(agent): protect bootstrap prefix from context pruning #381: has both adaptive (ratio-triggered soft-trim + hard-clear) and aggressive (always clear eligible tool results before cutoff).
  • feat(agent): add simple observation masking extension #386: is basically “always mask older tool results” (similar spirit to mode: "aggressive" here, but with different cutoff semantics).

• Safety / edge cases

  • fix(agent): protect bootstrap prefix from context pruning #381: explicitly skips tool results containing image blocks (we don’t try to partially prune them yet).
  • feat(agent): add simple observation masking extension #386: replaces tool result content wholesale, which would also drop any image blocks if they show up in tool results.

Net: #386 is simpler and very predictable, but it’s also easier to accidentally over-prune “still-needed” tool outputs and it adds more churn to the embedded runner. This PR’s aggressive mode gives you the simple/always-prune path while keeping the cutoff anchored to assistant messages and keeping the runner wiring minimal.

Landed on main.

Commits:

• 5ddf9b2 fix(agent): protect bootstrap prefix from pruning
• e0a30c4 docs: note bootstrap pruning guard (PR fix(agent): protect bootstrap prefix from context pruning #381)

Tests: pnpm lint && pnpm build && pnpm test