Skip to content

estafeta 3.1.0: agent-relay for cross-provider work - #3

Merged
pashbel merged 1 commit into
mainfrom
agent-relay-3.1.0
Oct 15, 2026
Merged

pashbel merged 1 commit into
mainfrom
agent-relay-3.1.0

Conversation

@pashbel

@pashbel pashbel commented Oct 15, 2026

Copy link
Copy Markdown
Collaborator

Summary

Adds multi-provider work to estafeta. An orchestrator in any provider can:

  • consult another agent;

  • request a cross-review;

  • delegate a planned phase to a worker of another provider;

  • coordinate peer agents that the user opened.

  • Skill agent-relay (6th skill) with adapters for:

    • host-native tasks: T3 delegate_task, same-provider subagents;
    • headless CLIs: codex exec, claude -p, gemini -p.

    Flags that bypass the worker's sandbox are never used by default.

  • Runtime delegate (providers, start, status, collect, cancel).

    • The ledger lives beside the state, so state schema 2 is unchanged.
    • CLI workers run as supervised runtimes with a timeout.
    • A writer's claims are checked against its recorded checks, the DONE gate
      in its workspace, and its workspace changes against a baseline.
  • Golden path tests (scripts/test_relay.py). Five representative
    interaction sequences, with an expected result at every step:

    • a stub worker plays them on every change and in the new CI workflow;
    • live runs are verified with --live-verify.
  • Docs and release. Docs, guide and the 3.1.0 manifests are updated, and
    .github/workflows/ci.yml is added.

Plan and decisions: docs/plans/multi-provider.md. Evidence:
packs/session-workflow/VALIDATION.md.

Verification

  • macOS, Python 3.9.6:
    • runtime 70/70, installer 21/21, golden 5/5 scenarios (79 steps);
    • workflow.py check and claude plugin validate passed;
    • four deliberate regressions were each caught by a golden step.
  • Live, Codex gpt-6.1-sol through T3: a delegated write in a fixture, and
    two cross-review rounds. Round 1 found 12 issues and round 2 found 5; all
    are fixed with tests.
  • Live, claude -p CLI worker: consult on Haiku 5.5, and a write under
    the OS sandbox on Sonnet 5.5.

Not verified

  • live codex exec / gemini -p CLI transports;
  • the Gemini/Antigravity route;
  • native Linux and Windows (this CI run will be the first on Linux);
  • a third review round on the round 2 fixes.

🤖 Generated with Claude Code

…(3.1.0)

One project can now be worked on by several AI providers at once. An
orchestrator in any provider can consult another agent, get a cross-review,
or delegate a planned phase to a worker of another provider, and peer agents
the user opens can share the same records.

- New portable skill agent-relay with adapters for host-native tasks (T3
  delegate_task, same-provider subagents) and headless CLIs (codex exec,
  claude -p, gemini -p). Sandbox-bypass flags are never used by default; the
  Claude writer runs under Claude Code's OS sandbox.
- New runtime command delegate (providers, start, status, collect, cancel).
  A ledger beside the state keeps provider, model, packet and result, so
  state schema 2 is unchanged and 3.0 runtimes can still join as workers.
  CLI workers run as supervised runtimes with a timeout. collect reads only
  the final answer, and checks a writer's claims against recorded checks, the
  DONE gate in its workspace and its actual changes against a baseline.
- Golden path integration tests (scripts/test_relay.py): five representative
  interaction sequences with an expected result at every step, played with a
  deterministic stub worker on every change and in the new CI workflow, and
  verified against live runs with --live-verify.
- Live: Codex (T3) delegated write and two cross-review rounds (12 and 5
  findings, all fixed with tests); claude -p consult and sandboxed write.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pashbel
pashbel merged commit 96cdb22 into main Oct 15, 2026
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant