Hooks: SessionStart output is discarded + no compact source — plugins have no silent post-compaction context channel

[Digentra]

Summary

Two related gaps in the hook system prevent plugins from maintaining session
continuity. This affects any plugin that contributes model context —
memory/session-state tools, project-context injectors, or
Claude-Code-compatible hook packs — not one specific plugin.

1. SessionStart hook output goes nowhere. triggerSessionStart awaits the
   hook results and then ignores them
   (packages/agent-core/src/session/index.ts). Combined with
   HookSpecificOutputSchema only parsing message, permissionDecision,
   and permissionDecisionReason (session/hooks/runner.ts), a SessionStart
   hook has no way to contribute context at all.

2. Nothing context-capable fires after compaction. SessionStart only
   supports startup | resume, and PostCompact is fired via
   fireAndForgetTrigger with its output discarded
   (agent/compaction/full.ts). A plugin that wants to restore session state
   after /compact must piggyback on the next UserPromptSubmit — one
   message late, and always visible in the TUI, because
   renderUserPromptHookResult sends the same text to the context and the
   display (see also Suggestion: option to hide UserPromptSubmit hook output from TUI #897).

Comparison with Claude Code

Claude Code covers both cases: hookSpecificOutput.additionalContext is
injected into the model context silently (visible to the AI, hidden from the
TUI), and SessionStart fires with source: "compact" right after
compaction. Plugins built for that ecosystem cannot offer equivalent
behavior on Kimi Code today.

Proposal

Either or both, both additive (existing hooks unaffected):

1. Fire SessionStart with source: "compact" after
   context.applyCompaction() in full.ts, and route SessionStart hook
   output into the rebuilt context. The platform already has an internal
   post-compaction injection mechanism (agent.injection.injectGoal()), so
   the plumbing partly exists.
2. Support hookSpecificOutput.additionalContext as a silent injection
   channel (AI-visible, TUI-hidden) for all hook events — this would also
   resolve Suggestion: option to hide UserPromptSubmit hook output from TUI #897.

Source references

• packages/agent-core/src/session/index.ts — triggerSessionStart(source: 'startup' | 'resume'), results unused
• packages/agent-core/src/session/hooks/runner.ts — HookSpecificOutputSchema
• packages/agent-core/src/session/hooks/user-prompt.ts — renderUserPromptHookResult
• packages/agent-core/src/agent/compaction/full.ts — triggerPostCompactHook via fireAndForgetTrigger

Related: #897 (hide UserPromptSubmit hook output from TUI), #447 (session
lifecycle hooks).

Happy to send a PR with whichever option you prefer.

[hogeheer499-commits]

I can work on a focused contribution here. Before changing the lifecycle contract, which scope would maintainers prefer for a first PR?

• add SessionStart source compact and consume only SessionStart output after compaction;
• add the general hookSpecificOutput.additionalContext silent-injection channel (which also overlaps Suggestion: option to hide UserPromptSubmit hook output from TUI #897); or
• implement both together as one compatibility change.

My preference for a reviewable first patch would be the smallest behavior the project wants to commit to, with hook-schema/session tests covering startup, resume, and compact.

[gleephoenixcrew]

200–280 characters. Re: hooks: sessionstart output discarded compact — We snapshot agent decision state to disk and reload only the last few turns plus a summary — receipts beat re-parsing the whole trace. Summaries are versioned — if an agent disagrees with its own prior summary, that is a signal to reload raw state, not argue from memory. How do you decide what to persist vs re-derive for prior decisions?