feat: Shared context & self-knowledge (env, sessions, compaction)

[buihongduc132]

Summary

• Design and implement a consistent “shared context” model across turns: env/git metadata, self‑knowledge, session resume, and smart compaction.

Motivation

• Improve continuity, reduce prompt size, and give the model actionable context every turn.

Scope

• Env context per turn (repo path/branch/diffs); self‑knowledge store; session list/resume.
• Compaction: auto-compact heuristics; apply-patch context improvements; Optimize Your Input (ephemeral turn).
• Respect environment_context knobs (e.g., network access) when generating prompts.

References

• Related base PRs:
  • Env context each turn: ``openai/codex PR 2483
  • Self knowledge: ``openai/codex PR 3036
  • Sessions list/resume: ``openai/codex PR 2898
  • Apply-patch context lines: ``openai/codex PR 2596
  • Refresh context-left after /compact: ``openai/codex PR 2105
  • Auto compact: ``openai/codex PR 2207
  • Optimize Your Input: ``openai/codex PR 2426
  • Respect environment_context for network access: ``openai/codex PR 3153

Acceptance Criteria

• Shared context appears in every turn with correct gating; session resume works; compaction reduces tokens without losing fidelity; tests and docs included.\n\nNote: Links to base PRs are defanged to avoid cross-repo noise. Open locally via:\n- gh pr view openai/codex#<PR_NUMBER> -w\n