Skip to content

v0.7.1

Latest

Choose a tag to compare

@github-actions github-actions released this 15 Nov 14:56
6f2d61e

On the built-in Node coordinator, a strict-mode session whose copy of a file went stale can read the file again after one deny and keep working, the stale-copy deny texts now say so, and an opted-in pre-edit with a malformed subagent id is refused with HTTP 400.

Upgrade note:

  • Restart the Node coordinator. A Node coordinator that an earlier plugin started keeps serving its workspace after the update, with the old behavior, until it stops. From the workspace root, check that ps -p "$(head -n 1 .coherence/server.pid)" -o command= names dist/coordinator.js, run kill "$(head -n 1 .coherence/server.pid)", then start a new session.
  • Python-backend workspaces. The recovery and the new deny texts come from agent-coherence releases after 0.16.0, which make the same change in the Python coordinator; 0.16.0 and earlier keep denying every Read. Upgrade the library with the plugin.
  • Deny texts. A script or prompt that matches the strict-mode deny texts byte for byte must accept the new wording (see Changed).

Fixed

  • On the built-in Node coordinator, a strict-mode session whose copy of a file went stale can read the file again and keep working (Cohexa-ai/agent-coherence#514). The session's first stale Read or Edit of the file is still denied. Every Read after it was denied too, for the rest of the session, whether a peer had committed, a peer had taken the file and stopped without committing, the session's grant had been reclaimed, or its own subagent had committed; only a Bash read run twice, a new session or a new subagent got the file back. Now the session's next Read of the file is allowed when the bytes it reads are the version the coordinator recorded, or the record holds no content: it is granted that version and answered {"status": "fresh", "version": <n>}, with any queued notices and no stale warning, and the session's Edits of the file are allowed after it. A commit that lands between the deny and that Read costs no second deny: the Read gets the newer version. An Edit or Write sent before that Read is still denied, so a write built from the old copy does not land. A Read of other bytes, or with no hash because the file is gone, stays denied (see Changed). The session's next allowed Read or Edit of the file uses the recovery up, so the next time the copy goes stale starts from a deny again. A Bash or Grep read that re-grants the file also uses up the record of the deny, so when the session's copy next goes stale, its first Read or Edit of the file is denied again. The hook clients do not change. The Python coordinator does the same in agent-coherence releases after 0.16.0; 0.16.0 and earlier keep denying every Read. A Node coordinator started by an earlier plugin keeps the old behavior until it restarts. Limits:
    • the coordinator keeps the record of the deny in memory, so after a restart the session is denied once more before its Read is allowed;
    • a Read and a Write sent in one message, or a Read of part of the file, can still let a whole-file write from the old copy through (Cohexa-ai/agent-coherence#624);
    • only the Claude Code hooks get the recovery. A pre-read that carries want_owner_generation, verify_only or name_kind, as a CoherentVolume, the MCP server, the substrate adapter, an effect fence's verification read and the raw HTTP capture send, keeps the deny and neither starts nor uses up the recovery; the Node coordinator now reads verify_only for this. A CoherentVolume from agent-coherence 0.16.0 or earlier marks none of its plain reads, so through this coordinator its second read heals a stale copy (Cohexa-ai/agent-coherence#629); use a release after 0.16.0;
    • the agent-coherence hook client's opt-in denial of an edit of a strict file another agent holds (echo refuse > .coherence/hook_on_held) stays off by default.

Changed

  • On the built-in Node coordinator, the strict-mode deny texts for a stale copy now say that reading the file again returns the current version (Cohexa-ai/agent-coherence#514). The text for a peer's commit and the text for a revoked grant keep their first sentence, and the rest now reads Read <path> again with the Read tool: the next Read returns the current version, and edits of <path> are denied until you have read it., where it told the agent to re-read the file and said that retrying the same operation would produce the same denial. Pre-read, pre-edit, pre-bash and pre-grep send these two texts. A stale Read whose bytes are not the recorded version, or that carries no hash, and a Read by a session that never read the file whose bytes differ from the record, now get one of two texts that never promise a Read: Stale read denied: <path> on disk does not match v<n>, the last version committed through the coordinator, and no session committed the change, so it was made outside the coordinator. ..., which keeps the same-denial sentence, or, while another session holds an uncommitted write of the file, Stale read denied: another agent is writing <path> and has not committed it yet, so <path> on disk does not match v<n>, the last version committed through the coordinator. Wait for that write to be committed, then Read <path> again with the Read tool. A session that never read the file got the commit text before, which named the last committer, or agent <unknown> when nobody had committed, as the writer of bytes no session committed. The out-of-band text for a reader that still holds the file is unchanged. Each text is the same bytes for a given file and event, and the Python coordinator sends the same bytes in agent-coherence releases after 0.16.0.

  • On the built-in Node coordinator, a pre-edit that opts in to acquire-or-fail with a malformed subagent id is now refused with HTTP 400 (Cohexa-ai/agent-coherence#519). A pre-edit that sends "if_unheld": true asks whether the caller's own agent identity may take the file. When its agent_id was present but not 1 to 64 characters of [A-Za-z0-9_-], such as "bad id!", a 65-character string, a number or a list, the Node coordinator read it as the parent session and answered for the parent, so a subagent whose parent held the file was admitted as the parent's own repeat. Now such a request answers HTTP 400 with {"error": "agent_id must be 1-64 chars of [A-Za-z0-9_-]"}, the wording the Python coordinator uses for the same fault, and changes nothing. The check runs before the tracked-path check, so an untracked path answers the same. An empty or null agent_id still means no subagent id, a well-formed one, including a CoherentVolume's 32-hex attempt id, is unchanged, and a pre-edit that does not opt in still reads a malformed id as the parent. The plugin's own hooks do not send if_unheld.

  • The plugin and marketplace manifests name a contact email, and the plugin ships a listing icon, .claude-plugin/icon.png (#276). claude plugin validate warned that no icon was present. Nothing changes at run time.