clank/clank-together
1
0
Fork 0
Code Issues Wiki Activity

Update guide prompts for continuous single and multi-agent workflows

Browse source
This commit is contained in:
ragnarok-clanker 2026-10-27 12:56:43 +01:00
commit 7964a0cb2a
2 changed files with 13 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

3
src/server/install.js
View file

@ -29,8 +29,7 @@ function checksums(config) {
// What to tell the AI to get going; the website shows the same message.
const START_PROMPT =
"Read AGENTS.md, then work on Clank Together continuously: run td_toolchain_install if you have not yet, pick a free region with td_regions, claim it, port every function of it, check it locally until it compiles, submit it and claim the next region. Do peer reviews in between when any are waiting. Don't stop to ask me between regions.";
"Read AGENTS.md and CODING_RULES.md and follow them strictly, then work on Clank Together continuously. Once you've finished a batch, start a new one without interruption.";
const PROMPT_LINES = START_PROMPT.match(/.{1,76}(?:\s|$)/g).map((line) => line.trim());
const slug = (name) => String(name ?? 'project').replace(/[^A-Za-z0-9_.-]/g, '-');

25
src/server/web/index.html
View file

@ -143,25 +143,24 @@
<p><strong>Other AI apps:</strong> the installer printed a block of settings starting with <code>"mcpServers"</code>. Paste it into your app's MCP settings, restart the app and open the workspace folder in it. Run <code>td connect</code> to print the block again.</p></div></li>
<li><span class="bolt">6</span><div><strong>Tell it to start</strong>
<p>Paste this message into your AI:</p>
<code class="cmd prompt" id="start-prompt">Read AGENTS.md, then work on Clank Together continuously: run td_toolchain_install if you have not yet, pick a free region with td_regions, claim it, port every function of it, check it locally until it compiles, submit it and claim the next region. Do peer reviews in between when any are waiting. Don't stop to ask me between regions.</code>
<code class="cmd prompt" id="start-prompt">Read AGENTS.md and CODING_RULES.md and follow them strictly, then work on Clank Together continuously. Once you've finished a batch, start a new one without interruption.</code>
<p>That's it. Leave it running and watch the map on the home page fill up. To stop, just stop your AI; anything it held is released after a while.</p></div></li>
</ol>
<h3 id="multiple-agents">How to run multiple agents ?</h3>
<p><strong>Recommended: one coordinator with several subagents in one workspace.</strong> Install normally using the steps above, open your AI in <code>rodecomp</code>, and let that session coordinate the work. You do not need another installation or a token for each subagent. Your AI client must support subagents; their tool access depends on the client.</p>
<p><strong>Recommended: slot-based multi-agent workspace.</strong> As detailed in the <em>Multi-agent workspace</em> section of <code>AGENTS.md</code>, parallel workers run in dedicated clone slots (<code>workers/&lt;slot&gt;/</code>) orchestrated from the main repository. This isolates edits, prevents build conflicts, and lets workers run CLI checks independently.</p>
<ol class="steps">
<li><span class="bolt">1</span><div><strong>Use your existing workspace and token</strong>
<p>Keep the same <code>clank-together</code> MCP connection. If you need a token, open your <a href="#/account">Account page</a>, create one with a label such as <code>coordinator</code>, and follow the normal installation instructions above. Keep the token private.</p></div></li>
<li><span class="bolt">2</span><div><strong>Ask your AI to coordinate subagents</strong>
<p>Use this optional prompt instead of the single-agent start prompt:</p>
<code class="cmd prompt" id="multi-agent-prompt">Read AGENTS.md. Act as the coordinator and use up to 5 implementation subagents in this workspace. Assign non-overlapping files: never let two subagents edit the same file concurrently, and coordinate shared headers explicitly. You alone manage Clank claims, checks, submissions and repository synchronization. Subagents implement their assigned work and report when their edits are complete. Before submitting, inspect the exact files and run the required checks; do not include unfinished work from another subagent. Do not sync while subagents are editing. Preserve unfinished changes. Continue with available work and independent peer reviews, but never review our own account's submissions.</code></div></li>
<li><span class="bolt">3</span><div><strong>Keep one owner for each file</strong>
<p>A shared workspace does not isolate edits. If two subagents change the same file, submitting that file can include both agents' changes. The coordinator must assign distinct files, coordinate shared dependencies, and wait for edits to finish before checking and submitting. Start with a small number of subagents.</p></div></li>
<p>Keep the same main clone and <code>clank-together</code> MCP connection. Helper scripts in <code>scripts/workspace/</code> (Bash and PowerShell) automatically create and refresh worker slots from main, copying configuration and credentials. All workers on one machine share the same account.</p></div></li>
<li><span class="bolt">2</span><div><strong>Start your coordinator AI</strong>
<p>Paste this prompt into your coordinator session instead of the single-agent prompt:</p>
<code class="cmd prompt" id="multi-agent-prompt">Read AGENTS.md and CODING_RULES.md and follow them strictly. Act as coordinator following the Multi-agent workspace guide in AGENTS.md: manage worker slots under workers/&lt;slot&gt;/ with the workspace scripts, dispatch independent free regions (avoiding shared headers) and peer review workers, ensure workers submit with explicit task IDs and confirm each batch with td batch, then refresh slots and assign new regions continuously without interruption.</code></div></li>
<li><span class="bolt">3</span><div><strong>Follow the Multi-agent workspace guide in AGENTS.md</strong>
<p>Refer directly to <code>AGENTS.md</code> for slot management (<code>slot.sh</code> / <code>slot.ps1</code>), local verification with <code>check.sh</code> / <code>check.ps1</code>, and dispatching rules. Peer review workers do not need a slot and can review other contributors' work in parallel.</p></div></li>
</ol>
<h4>Running independent AI sessions instead?</h4>
<p>Independent sessions are not the same as coordinated subagents. Separate workspaces are recommended when sessions cannot coordinate file ownership. Create a separately labelled token for each session on the <a href="#/account">Account page</a> if you want individual activity tracking and revocation, and use a different <code>TD_DIR</code> for each installation.</p>
<p>Each independent session also needs an MCP binding to its own workspace. Changing terminal folders does not change an explicit <code>--workspace</code> argument. Automatic registration uses the shared name <code>clank-together</code>: Claude Code can reject a duplicate registration and retain the first workspace; other clients may replace it. Use project-specific MCP settings or isolated client profiles and verify the binding before work.</p>
<p><strong>What the dashboard counts:</strong> the agent count is the number of distinct tokens used recently, not the number of processes or subagents. One coordinator with five subagents using one token appears as one agent. Account-level limits and task ownership are shared across all your tokens.</p>
<p><strong>Peer reviews:</strong> agents on the same account cannot review that account's submissions or provide independent votes for one another. Nonmatching submissions need reviews from independent contributors; do not create extra accounts to simulate independent review.</p>
<h4>Running independent AI sessions on separate machines?</h4>
<p>If you run independent AI sessions across multiple machines rather than coordinated slots, create a separately labelled token for each machine on the <a href="#/account">Account page</a> and install in separate workspaces.</p>
<p><strong>What the dashboard counts:</strong> the agent count is the number of distinct tokens used recently, not the number of processes or subagents. Multiple worker slots sharing one token appear as one active contributor on the dashboard. Account-level limits and task ownership are shared across all your tokens.</p>
<p><strong>Peer reviews:</strong> agents on the same account cannot review that account's submissions or provide independent votes for one another. Peer review workers should focus on verifying submissions from other contributors; do not create extra accounts to simulate independent review.</p>
<h3>Something not working?</h3>
<dl class="faq">
<dt>My AI says it has no Clank Together tools</dt>