Skip to content

Claude Code host: remaining work before 2.0 (#2438) #2469

Backlog

Description

@bradygaster

Everything left before Claude Code support in Squad is done, from epic #2438. Phases 0, 2, 3, 4, 5 and 6 are merged (#2445, #2450, #2463). Phase 1 is merged in #2451 except for the live golden trace.

Each item is marked Must (blocks calling Claude support done for 2.0), Should (strongly wanted for 2.0) or Later (can ship after 2.0).

1. Finish Phase 1: execution harness (#2441)

2. Bugs from the end-to-end run

3. Plugin and dashboard hardening (after #2463)

4. Phase 7: host-neutral canonical coordinator

5. Review, test and docs loose ends

Also found while working these

6. Release gate (prepared in #2506, not yet dispatched)

🤖 Generated with Claude Code

2026-11-12: demo run findings

A real claude -p run on a fresh repository, recorded for the README:

2026-11-16: status

All code work for Claude Code support is merged. Merged 2026-11-13/14: #2496 (narrower Squad Bash allowlist), #2500 (linked team repositories), #2501 (ADR-003 option C accepted), #2502 (squad_decide says where a git-branch backend stored it), #2503 (live scenario evidence), #2504 (no Copilot Init Mode skills for Claude), #2505 (a user's team decision reaches decisions.md in the same turn, so a teammate on Copilot sees it; one team with one person on Claude Code and another on Copilot CLI checked live both ways).

What's left for this issue:

Activity

  1. bradygaster commented on Nov 12, 2026

    @bradygaster
    CollaboratorAuthor

    ADR-003 evidence: does the Claude coordinator read .github/agents/squad.agent.md?

    ADR-003 option B rests on one cost: a non-Copilot coordinator may read the 80 KB Copilot coordinator on demand, because its digest says "if .github/agents/squad.agent.md exists in this workspace, read it as the authoritative coordinator contract". This comment checks whether that happens. The analysis is in #2486.

    Method. I went through every local Claude Code session transcript for a Squad test repository (~/.claude/projects/<repo>/*.jsonl, Claude Code 2.1.317–2.1.318, 2026-11-10 to 2026-11-12) and checked every tool call for .github/agents/squad.agent.md as a Read file_path, a Bash command, or a Grep or Glob pattern/path. The script is below.

    Result: 0 of 43 sessions and 0 of 223 tool calls touched the file.

    Most of those sessions are short probes, so here are the ones that did real team work. In both repositories the 82–83 KB Copilot file was present, and the coordinator's digest pointed at it:

    Repository Sessions Tool calls squad_state / .squad/ accesses Member spawns Reads of the Copilot file
    Demo, 2026-11-12: fresh squad init --client claude, cast a team, build an MVP (22 tests pass) 4 116 72 6 0
    End-to-end arcade test, 2026-11-10 3 61 24 26 0
    Team work total 7 177 96 32 0
    Plugin labs, harness probes and other short sessions 36 46 4 7 0

    The coordinators worked from their digest, the squad-coordinator-* skills and squad_state. The same pattern shows in the demo transcripts: the coordinator reads team.md, config.json and routing.md through squad_state, then casts or dispatches.

    Caveats.

    • These are 7 substantive sessions on one model family and one Claude Code version, so treat this as a strong signal, not proof. A longer or stranger session (a protocol question the digest doesn't cover, a ceremony, Ralph mode) could still follow the pointer.
    • All the sessions are mine (lab, test and demo runs), not user sessions.

    What it means for ADR-003. The pointer is close to dead text in practice, so B's stated gain is mostly theoretical. Option C, rewording the digest so it doesn't point at the Copilot file, matches how the coordinator already behaves.

    Script
    // node evidence.mjs ~/.claude/projects/<repo>/*.jsonl
    import fs from 'node:fs';
    let sessions = 0, tools = 0, hits = 0;
    for (const file of process.argv.slice(2)) {
      sessions++;
      for (const line of fs.readFileSync(file, 'utf8').split('\n')) {
        let m; try { m = JSON.parse(line); } catch { continue; }
        if (m.type !== 'assistant') continue;
        for (const c of m.message?.content ?? []) {
          if (c.type !== 'tool_use') continue;
          tools++;
          const i = c.input ?? {};
          if (/\.github\/agents\/squad\.agent\.md/.test(String(i.file_path ?? i.command ?? i.pattern ?? i.path ?? ''))) hits++;
        }
      }
    }
    console.log({ sessions, tools, hits });
  2. bradygaster commented on Nov 13, 2026

    @bradygaster
    CollaboratorAuthor

    ADR-003: team feedback and proposed direction

    Reviewed the Teams discussion on 2026-11-12. All three substantive replies (Q, Bubu, and Davidbot) favor option C for 2.0, with option B explored separately as an isolated, unshipped spike. This matches Brady's earlier direction in the thread. To make the sequencing explicit: the replies recommend C now; consider shipping B later only if the experiment earns it, not delaying C until B is complete.

    Related: analysis PR #2486; Claude Code epic #2438; Teams discussion.

    C: deliver a complete Claude experience for 2.0

    Make non-Copilot coordinator instructions self-sufficient without depending on .github/agents/squad.agent.md. The user-visible goal is correct specialist delegation, preserved project policy and safeguards, and useful results through Claude-native tools—not merely a shorter prompt or a different file layout.

    Shared conditions from the replies:

    • Keep mandatory routing, authorization/approval, state-ownership, and recovery rules immediately available. Load detailed procedures through tested host-native mechanisms.
    • Generate host instructions from the shared source contract where practical; avoid separately maintained rulebooks. Behavioral parity does not require identical wording or tool sequences.
    • Test with the Copilot coordinator file deliberately absent. Cover specialist routing, project-policy overrides (including ADO rather than GitHub), review rejection and approval stops, configured backend selection/isolation and outages, and restart/context-compaction recovery without duplicate effects or invented approvals.
    • Preserve team-owned customizations and explicit policy precedence through setup and upgrade.

    The reported 43 sessions establish file-consumption observations, not behavioral completeness. Successful task completion alone is insufficient evidence that approval gates and backend rules were respected.

    B: bounded experiment, not a shipping commitment

    Prototype a host-neutral canonical workspace contract with Claude and Copilot projections in disposable, explicitly opted-in workspaces. No default installer migration or live-team migration as part of the spike. Give it a fixed experiment budget; C must proceed independently.

    Reuse C's behavioral scenarios, then test fresh and customized legacy workspaces, monorepo/remote placement, repeatable upgrades, customization preservation and precedence, rollback that restores prior bytes, and no Copilot regression. Assess whether B measurably improves correctness or maintainability enough to justify migration risk—not just whether the directory layout is cleaner.

    Outcome: pursue C for the 2.0 Claude experience; investigate B separately and return with evidence before any shipment decision. The replies are design recommendations, explicitly not independent implementation reviews. This comment records the discussion; it does not mark ADR-003 Accepted, approve an implementation, or authorize shipping B.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions