Improve skill quality: phase 0 fix, failure modes, tmp isolation, layered triggers, session cleanup

[opensouth21]

Quality improvement pass on 2 core SKILL.md files and 1 reference template. 3 files changed, 191 insertions, 13 deletions.

Changes

browser-act/SKILL.md

• Split description into layered triggers: 500-char wall of text -> description + 11 triggers entries
• Added Pre-flight Check: auto-run on load, verifies CLI, version, and core guide integrity
• Added 4 Failure Modes: FM-1 (CLI not installed), FM-2 (output truncated), FM-3 (session leak), FM-4 (auth failure)

browser-act-skill-forge/SKILL.md

• Fixed Phase 0: Removed 'Already completed -> skip'. Replaced with real 3-level detection
• Added 6 Failure Modes: FM-F1 to FM-F6
• Per-instance tmp isolation: shared tmp/ -> tmp/{skill-name}-{datetime}/

output_template.md

• Added Session Cleanup section to generated SKILL.md

Design note

Install command left flexible because PyPI version (1.0.5) differs from schema version (2.0.2). Runtime compatibility via CLI built-in handshake.

[browseract-ai]

Thank you for the contribution. The PR contains several useful ideas, particularly session
cleanup, temporary-directory isolation, and clearer failure-mode documentation.

However, we cannot merge it as-is because some changes introduce compatibility and safety
concerns, including:

• Treating Skill version 2.0.2 as the CLI package version.
• Moving discovery keywords from the standard description field to a non-standard triggers
  field.
• Closing sessions without respecting cross-conversation ownership.
• Automatically installing the CLI without explicit user confirmation.
• Using undefined or non-portable shell variables and commands.

For future contributions, could you please split independent changes into several smaller,
focused PRs? This will make each change easier to evaluate, test, and merge separately.