[BHAWESHBHASKAR]

Summary

leep_commit_and_land and leep_resolve_and_land now run a pre-flight check. A call whose create answer carries a stopping warning the agent has not been shown creates the changeset but does not land it, and the result lists the warnings. A stopped changeset lands only with force: true.

The maintainer chose this option ("pre-flight check") to close the gap the twenty-agent demo found: the MCP tool created and landed in one call, so overlap warnings arrived only after the land.

Why

CLI agents run leep commit and see its warnings before they run leep land. MCP agents had no such pause. Now the agents most likely to collide, those with a duplicate task, a claim or the same file, are told before they land, and a call with nothing to warn about stays a single call.

Changes

The stop rule. It is decided in one place, Notice::stops_land, against the files the changeset changes, using the other side's path. A land stops on:

• a possible duplicate;
• a duplicate's land;
• a land that changed one of my files;
• a claim on one of my files or on a directory holding it;
• another session on the file itself.

near warnings, scope directories, related files and unknown levels never stop a land.

Only unseen warnings stop. The server says each warning once per session. It now remembers which files it said a warning as a stop for, and says it again when it would stop a file it did not stop before. For example: a non-stopping warning about ben's src/b.rs, followed later by ben editing my src/a.rs, stops my commit of src/a.rs. Per session, the server prefers the overlap row that stops for the request's files. Landed warnings name the request's file. Every stopping warning comes in one answer, beyond the usual cap of 3.

A stopped call:

• returns isError: false with structuredContent {status: "stopped", landed: false, changeset, target_ref, files_changed, stopped_by, guidance} and the text NOT LANDED: stopped before landing;
• leaves the target ref unchanged;
• guides the agent to land the same changeset with force: true (the guidance names the target ref), or to change its work.

Retry and force:

• A plain retry is stopped again. It lists the earlier warnings plus anything new from GET /v1/notices, so a host that automatically retries failed tool calls cannot land blind.
• changeset: <id> with force: true lands the remembered changeset on its own ref, with its own create_ref and land request id. No new changeset is made.
• The MCP server remembers up to 64 stopped changesets. One is forgotten only after it lands, so a failed forced land keeps its request id.

Other changes:

• When leep_task_start's context pack is shed, it is retried once if Retry-After is 2 s or less.
• No CLI change: leep commit already prints warnings before leep land.
• Demo claim 10: dee's commit into eli's duplicate task is stopped, main doesn't move, and dee then lands the same changeset with force.

Docs:

• the ADR 0096 amendment, MCP.md, ARCHITECTURE.md and DEMO-COORDINATION.md;
• tool descriptions and server instructions;
• changelog.d/mcp-preflight.md.

The documented limits are:

• create warnings that time out (150 ms bound) or a lost GET /v1/notices answer mean no stop, because warnings are advisory;
• several MCP servers on one LEEP_AGENT_SESSION don't share stops.

Testing

• Lint and format: fmt check, clippy -D warnings, check-version.sh and changelog-assemble.sh --check pass.
• Full suite: cargo test --workspace against Postgres 16: 2,012 passed, 1 failed, 9 ignored. The failure is a_hanging_coordination_store_changes_no_answer, an existing wall-clock assert: 314 ms against its 300 ms limit while another build ran on the machine. It passed 5 of 5 runs alone. This PR doesn't touch it. A proper fix (a paused-clock bound test) is in progress separately.
• MCP e2e: 43 of 43, with exact create and land counts through a counting middleware. The cases:
  • no overlap lands in one call;
  • a same-file overlap is stopped and not landed;
  • force lands without making a second changeset;
  • a stop on a non-main ref lands on that ref;
  • a retry is stopped again, with what is new since;
  • a warning already said does not stop;
  • near does not stop;
  • a failed forced land keeps its request id;
  • the resolve path;
  • the shed pack is retried (with Retry-After: 0, so it doesn't sleep).
• Derive: 40 of 40 notice tests, including say-again-when-it-comes-to-stop for same-file, claimed and landed warnings, and every stopping warning in one answer.
• Demo: coord_demo passes on SQLite and Postgres.
• No wall-clock asserts in the new tests.

Fable reviewed it ("ship after must-fixes"). The three must-fixes are included: wrong-ref guidance, said-once keyed by file, and the stop judged by the other side's path. All the should-fixes and nits are included too.

Compatibility

• API: additive only. Notice::stops_land is client logic; no wire field was removed.
• Visible changes:
  • MCP agents can now get a "stopped" result and need force: true to land past it.
  • A notices answer can carry more than 3 warnings when several of them stop a land.
• No migration.
• Branch history: the branch carries a no-op merge of its earlier history (coord: fix the friction the twenty-agent demo found #281, squashed into main as 582a33e).

---

Generated by Claude Code