Problem

agents.defaults.contextInjection is the only place this setting can be configured. There is no per-agent override (agents.list[].contextInjection).

This becomes load-bearing for setups that run multiple agents with different rule-strictness needs:

• A general-purpose main agent benefits from continuation-skip — most rules are advisory, multi-turn conversations are common, token cost matters.
• A specialized agent with prescriptive rules (anti-fallback guards, mandatory tool sequences, safety constraints) needs always so the rules are present in the system prompt on every turn — otherwise rules drift on continuation turns and the LLM falls back to its pre-trained defaults.

Today the only options are:

• Use continuation-skip globally and accept rule drift on the strict agent.
• Use always globally and pay the bootstrap re-injection cost on every turn for every agent.

Neither is great when the agents have genuinely different needs.

Concrete example (driving this request)

Krishna is an EM operating system agent with a strict capture flow:

1. MCP-only writes (no local file fallback for captures)
2. Mandatory tool order: read entities → list threads → run matcher → present match → wait for explicit ack → commit on ack
3. Anti-fallback rules in TOOLS.md ("never write captures to workspace memory files")

Under continuation-skip, turn 2+ of any session drops these rules from the system prompt. The LLM, no longer seeing the anti-fallback rules, drifts toward its pre-trained instinct: write a memory file, search local memory, etc. This was reproducible — turn 1 of a clean session followed the rules perfectly; turn 2 silently auto-logged to inbox without the required user prompt. Switching the default to always fixed it but now applies to all three agents on the host.

Proposal

Allow contextInjection in the per-agent config block, falling back to agents.defaults.contextInjection when unset:

{
  \"agents\": {
    \"defaults\": {
      \"contextInjection\": \"continuation-skip\"
    },
    \"list\": [
      { \"id\": \"main\" },
      { \"id\": \"krishna\", \"contextInjection\": \"always\" }
    ]
  }
}

Resolution order: per-agent override → defaults → "always" (code default).

Implementation pointers

The current resolver is single-source:

```ts // src/agents/bootstrap-files.ts export function resolveContextInjectionMode(config?: OpenClawConfig): AgentContextInjection { return config?.agents?.defaults?.contextInjection ?? "always"; } ```

Needs to take an `agentId` and look up the per-agent override first. Schema additions in `src/config/types.agents.ts` (AgentConfig) and `src/config/zod-schema.agent-runtime.ts` (AgentEntrySchema). Help text in `src/config/schema.help.ts`.

The same pattern would benefit `bootstrapMaxChars`, `bootstrapTotalMaxChars`, and `bootstrapPromptTruncationWarning` if the maintainers want to broaden the scope, since those are also in `agents.defaults` only.

Acceptance

• Per-agent `contextInjection` override is honored.
• Defaults remain backwards-compatible (no per-agent override = inherit from defaults = inherit "always").
• Documented at https://docs.openclaw.ai/configuration#contextInjection (or wherever the field lives).