…wslabs#15)

… E2E tests

Gemini CLI supervisor agents were not receiving the agent profile system
prompt, causing them to respond as a generic agent instead of knowing
their role and available MCP tools. Fixed by writing the system prompt
to a GEMINI.md file in the working directory (Gemini CLI reads this for
project-level instructions), with backup/restore of existing files.

Added test_supervisor_orchestration.py with 10 E2E tests (2 per provider)
that verify real supervisor→worker delegation via handoff and assign MCP
tools using the analysis_supervisor profile. This fills the gap where
existing E2E tests only tested worker lifecycle, not the orchestration
flow. Also fixed send_message test polling (single check → 60s loop)
and updated all provider docs, skills, and lessons learned (#14, #15).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

… E2E tests

Gemini CLI supervisor agents were not receiving the agent profile system
prompt, causing them to respond as a generic agent instead of knowing
their role and available MCP tools. Fixed by writing the system prompt
to a GEMINI.md file in the working directory (Gemini CLI reads this for
project-level instructions), with backup/restore of existing files.

Added test_supervisor_orchestration.py with 10 E2E tests (2 per provider)
that verify real supervisor→worker delegation via handoff and assign MCP
tools using the analysis_supervisor profile. This fills the gap where
existing E2E tests only tested worker lifecycle, not the orchestration
flow. Also fixed send_message test polling (single check → 60s loop)
and updated all provider docs, skills, and lessons learned (#14, #15).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Update analysis_supervisor.md agent profile with "How Message Delivery
  Works" section explaining that messages arrive when the agent is idle,
  and agents must NOT run shell commands to wait (fixes Kimi CLI
  busy-wait blocking inbox delivery)
- Add lesson #17: Proactive Models Block Inbox Delivery by Busy-Waiting
- Fix 10 issues found in skills review:
  - Add Kiro CLI to provider lists (contributed lesson #9)
  - Fix lesson cross-references (#15→#11/#12 for supervisor E2E, #9→#8
    for Ralph)
  - Fix outdated "literal=True" → "bracketed paste" in lesson #10
  - Fix "prompt_toolkit (Kimi CLI, Codex)" → Codex is Rust-native TUI
  - Fix Kimi timeout "restore on cleanup" → NOT restored (concurrent
    instances)
  - Fix init timeout 60s→120s in implementation checklist
  - Add missing supervisor E2E command in Ralph fix_plan.md

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…dation, and README docs

- Rename lessons-learned.md to lessons-learnt.md
- Add 7 new lessons (#14-#20): shell warm-up, TERM compatibility, alt-screen
  vs scrollback, startup prompt loop, stale permission detection, per-directory
  locks, assign+handoff e2e validation
- Add docs/tool-restrictions.md as essential reading reference
- Add assign+handoff orchestration step (Step 9) to SKILL.md with
  troubleshooting guide referencing examples/assign/
- Add alt-screen vs scrollback to "Before You Start" checklist
- Expand Step 6 with cross-references to lessons #14, #15, #17
- Fix stale lessons-learned.md reference in provider-template.md
- Add Skills section to README with tool compatibility table (Claude Code,
  Kiro CLI, Gemini CLI, Strands Agents SDK, Microsoft Agent Framework, etc.)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat: add cao-provider skill for scaffolding new CLI agent providers

Comprehensive skill with annotated template, 10 lessons learned from
building 7 providers, test guide, and step-by-step implementation checklist.
Placed in skills/ (not .claude/) so it's tool-agnostic for any AI assistant.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat(skill): enhance cao-provider skill with lessons learnt, e2e validation, and README docs

- Rename lessons-learned.md to lessons-learnt.md
- Add 7 new lessons (#14-#20): shell warm-up, TERM compatibility, alt-screen
  vs scrollback, startup prompt loop, stale permission detection, per-directory
  locks, assign+handoff e2e validation
- Add docs/tool-restrictions.md as essential reading reference
- Add assign+handoff orchestration step (Step 9) to SKILL.md with
  troubleshooting guide referencing examples/assign/
- Add alt-screen vs scrollback to "Before You Start" checklist
- Expand Step 6 with cross-references to lessons #14, #15, #17
- Fix stale lessons-learned.md reference in provider-template.md
- Add Skills section to README with tool compatibility table (Claude Code,
  Kiro CLI, Gemini CLI, Strands Agents SDK, Microsoft Agent Framework, etc.)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(skill): add --provider and --auto-approve to assign e2e examples

The launch commands in lesson #20 and SKILL.md Step 9 were missing
--provider (defaults to kiro_cli otherwise) and --auto-approve/--yolo
flags needed for new provider testing.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* docs(skill): expand lesson #20 with all 11 e2e tests as minimum success criteria

Detail each test across 5 suites (handoff, assign, send_message,
allowed_tools, supervisor_orchestration) with what it validates and
why it matters. Emphasize that all 11 must pass — a supervisor that
silently does workers' jobs or ignores tool restrictions is only
caught by the full suite.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Three small fixes flagged by the diff-verification pass against plan.md:

- Task 12: list the per-suite pytest invocations the plan calls out as
  required backend checks, and elevate the criteria-catalog evaluation
  from a "completion report" item to a blocking acceptance gate (matches
  plan DoD awslabs#15, which the original task wording softened).
- Task 07: spell out the Linear non-regression verification as an
  explicit acceptance bullet so the implementer knows to check the
  Linear resolver and adapter registration remain intact.
- Task 08: note "latest arm wins" semantics on activate_plan
  re-arming, so behavior is explicit rather than implicit.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>