Agents, collaboration and optional hooks

Qiongli 2 keeps the model in your chosen Host. Qiongli supplies research contracts, source-bound handoffs and project tools. Installing the CLI-exported Plugin gives you Skills and Full MCP; it does not create another agent or change your model.

Reply only

Qiongli 2.0 includes the independent no-qiongli Skill. In Codex, select it as $no-qiongli; in other Hosts, select no-qiongli from the installed Skills. You can also say “reply only” or “no tools” naturally. The entry answers directly without loading the main Skill or research workflow:

Reply only for this conversation: explain the paragraph below. Do not use tools or read or save files.

Codex and Claude Plugin packages place it at skills/no-qiongli/SKILL.md, beside the research entry. Update the Plugin and reload its Skills to discover it. Skills-only exports include workflow/no-qiongli/SKILL.md; that self-contained directory can be installed as a Skill through your Host. The export itself does not register a Host or connect MCP. There is no qiongli no-qiongli CLI command.

The model answers from visible conversation content, skipping resource reads, search, MCP, agents, project checks and Graph updates. If material is missing, it asks for an excerpt or explains the limit. The choice stays active until you explicitly resume execution; “for this reply only” limits it to one response. It also applies when using a workflow shortcut. Ordinary requests keep their existing routing, and “no” in a quoted passage does not select this mode.

This is a Skill instruction, not a CLI flag or a Host permission lock. It does not change higher-priority Host rules, cancel running work or disable automatic Host hooks. To remove this Plugin's context hooks, separately use qiongli install plugin --hooks off and complete its installation confirmations.

Choose the smallest useful collaboration

Need	Use	Completion evidence
Independent methodological judgment	A fresh native subagent reviewing the same bounded sources	Actual reviewer result, execution identity and sources read
Draft plus review	A writer, then a reviewer of the exact candidate	Candidate identity, findings and coordinator's reconciliation
Independent components	Native subagents with separate output/file scopes	Returned components and an integration check
Another Host's review or edits	A portable handoff and returned review/candidate	Matching task, source version and candidate; unresolved conflicts remain visible

Ask naturally, for example: “Use an independent agent to check whether this results section respects our study design. Return findings before changing the manuscript.” Small edits can stay in one conversation. There is no required agent count or discussion round count.

Qiongli's model-collaborator guidance uses the Host's actual delegation tools. The coordinator gives each participant an outcome, source scope, file ownership and return format, then collects real results. A role label or a queued task is not an independent completed review. If no reviewer can run, that requirement stays open; same-conversation self-review is labelled accordingly.

Native tool availability depends on the running Host, configuration and model integration. Preserve the user's model settings. See the official Codex subagent guide and Claude Code subagent guide.

Work with another Host

Use the bundled templates/agent-handoff.md to send the requested outcome, permitted source excerpts/files, observed source hashes or revision, exact candidate and allowed actions. The other Host returns templates/agent-review-packet.md or candidate edits. Use an available, authorized communication tool, or transfer the packet manually. Prepared packets remain awaiting external review until a result returns.

The coordinator compares the returned source/candidate identity with the current project, resolves disagreements against evidence and previews any accepted changes. Stale reviews need reconciliation or a new review. One coordinator owns canonical research writes; separate candidate files or worktrees avoid concurrent overwrites. Keep claim/decision IDs, citekeys, source anchors, design limitations and previous stage-summary links across the exchange. A review is not new primary evidence for the Graph.

Do not transfer private conversations, credentials or approval tokens. Existing Full MCP runs are bound to their starting Host and authenticated reads; a packet does not transfer that checkpoint or its authority. Automatic cross-Host task claiming, concurrent canonical writes and crash recovery remain later work under ADR 0218. Portable review/edit exchange is available without those features.

Optional context hooks

A hook can remind the model what to recheck after compaction, resume or child-agent startup. The CLI provides qiongli hooks context: a native command that reads an event JSON object from stdin and returns a brief context reminder. It needs no Python, Node, MCP connection or model call. It does not read project files or transcripts, save summaries, approve writes or keep a completed task running. Hooks are optional; ordinary Skill routing remains sufficient.

qiongli install offers context hooks when you choose Plugin. They default to off on first install; later installs and upgrades preserve your choice. You can also choose explicitly:

qiongli install plugin --hooks context
qiongli install plugin --hooks off

The preview shows the events and command before you confirm file changes. The configuration lives in the Plugin manifest and calls its bundled native binary; no separate script runtime or MCP service is needed. Each Host has its own choice and installation. --hooks off removes this Plugin's reminder entries. It leaves manual Host settings intact. Avoid enabling the same reminder through both the Plugin and a manual configuration.

After registration, Codex still requires you to review and trust the commands in /hooks or its Hook settings. A changed definition can require trust again. Claude Code needs 2.1.139 or newer for command hooks with args; the guide checks this before writing files. Check the Plugin entries in /hooks. Reload the Plugin or start a new session, then verify that resume/compaction or child startup delivers a reminder. Export and registration checks do not prove event delivery. To inspect exported configuration, use qiongli app plugin-source-status --target codex --destination /absolute/path/qiongli-next: source.context_hooks reports the receipt-verified choice; host_state remains not-verified. Use --target claude for Claude Code.

If you prefer manual Host configuration, the following example uses an absolute native binary path; replace it with yours, preserving shell quoting. On Windows use the absolute qiongli.exe path and quote/escape it for the Host's command shell and JSON.

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "resume|compact",
        "hooks": [
          {
            "type": "command",
            "command": "\"/absolute/path/qiongli\" hooks context",
            "timeout": 5
          }
        ]
      }
    ],
    "SubagentStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "\"/absolute/path/qiongli\" hooks context",
            "timeout": 5
          }
        ]
      }
    ]
  }
}

For Codex, merge these entries into the appropriate .codex/hooks.json and review/trust them in /hooks; project-local configuration also needs project trust. For Claude Code, merge the hooks entries into the appropriate .claude/settings.json. Keep other entries intact and avoid registering this same reminder at multiple scopes. Configuration and event support follow the Host's own Codex hooks documentation or Claude Code hooks documentation.

Check qiongli hooks context --help first: an older CLI may not contain this command. To inspect the protocol locally, pipe a SessionStart object with source: "resume" into the command and close stdin. It returns hookSpecificOutput.additionalContext; unsupported events return {}. Invalid or over-64-KiB input exits 1 without printing its contents. The Hook result is only a reminder, never evidence of an independent review or saved research.

This implementation uses command hooks only. Do not assume all Hosts implement the same prompt/agent/MCP hook handlers. Confirm delivery in the actual Host; protocol tests alone do not qualify a live installation. The historical observed capability matrix keeps its original evidence scope.