[sametcelikbicak]

Summary

Adds durable run and task checkpoints under .team/runs/ so the orchestrator can continue after compaction, a restart or an interrupted tool call without repeating completed work or skipping unfinished work.

• Append-only JSON: one run.json, one event file per task start/finish, one approval file per human gate. Writes are atomic (temp file + os.replace).
• Unique run ids and task ids, separate from reusable stage names. Attempt history is preserved.
• run_state.py commands: start-run, start-task, finish-task, retry, approve, status. status is read-only and is the resume entry point.
• Approval is tied to the artifact SHA-256; changing the artifact makes the approval stale. An artifact on disk is not approval.
• Existing projects start with no run and no approvals; old artifacts are reviewed, not assumed approved.
• The orchestrator reads status at session start and after compaction, records each dispatch and finish, retries explicitly, and records human approval only after the gate.
• Limits are documented: no automatic process restart, no exactly-once external side effects, one active run per project.

Files

• skills/team-memory/run_state.py — stdlib-only helper.
• scripts/test-run-state.py — 34-check self-test (also runs on Windows CI).
• agents/orchestrator.md, skills/team-memory/SKILL.md, commands/team.md, commands/team-status.md — protocol and resume behavior.
• docs/runs.md, docs/memory.md, docs/pipeline.md, README.md, README.tr.md, team/AGENTS.md — documentation.
• .github/workflows/checks.yml — checkpoint test on Ubuntu and Windows.
• TESTING-RUNS.md — verification results, manual scenarios and limits.

Verification

Check	Result
python3 scripts/test-run-state.py	pass, 34 checks
ruby scripts/validate.rb	pass, 47 files OK
./scripts/test-memory-compact.sh	pass
bash -n install.sh scripts/*.sh skills/team-memory/*.sh	pass
./scripts/smoke-install.sh	pass
npm run generate:reference	pass
npm run docs:build	pass

Manual OpenCode scenarios that need a live session (fresh /team, real context compaction) were not run in this environment; TESTING-RUNS.md marks them as pending instead of claiming a result.

Acceptance criteria

• Durable unique run and task attempt identity.
• Record before dispatch and terminal outcome after execution.
• Startup/resume reads checkpoints before dispatching through the orchestrator instructions and status; there is no runtime hook, and the orchestrator must run it.
• Completed tasks are not rerun after context loss.
• Running attempts are reported as uncertain; the orchestrator reconciles before retrying and does not mark them completed.
• Failed/interrupted attempts retry explicitly with history and a next action.
• Human approval is preserved; artifact existence alone is not approval.
• Atomic writes; concurrent tasks use distinct files; same-attempt races are reported as conflicts.
• Existing projects initialize safely without old approvals.
• Test continuation after context loss. Automated tests use a fresh process per scenario and cover state recovery, interruption, completed-task recovery, failed retry, approval invalidation and malformed state. A live OpenCode compaction was not run; it is accepted as equivalent to the fresh-process/fresh-session recovery tests (see TESTING-RUNS.md).
• Limits are documented.

Closes #48