[nivSwisa1]

Source edits and subagent spawns are now held until the real verdict line is printed (max 3 holds per repo/session), so a prose grade or a file switch can't skip it; hook messages also name the SKILL.md as a fallback when the skill isn't listed.

🤖 Generated with Claude Code

[orca-security-us]

Orca Security Scan Summary

Status	Check	Issues by priority
Passed	Infrastructure as Code	0   0   0   0	View in Orca
Passed	OSS Licenses	0   0   0   0	View in Orca
Passed	SAST	0   0   0   0	View in Orca
Passed	Secrets	0   0   0   0	View in Orca
Passed	Vulnerabilities	0   0   0   0	View in Orca

[Copilot]

🟡 Changes recommended

Valid verdicts produced through the advertised SKILL.md fallback do not release the hold or reach reporting.

1 open issue

• Recognize fallback skill reads and validate subsequent verdicts · New

What changed in this PR

This PR updates agentic-advisor to v1.4.7, strengthening the verdict-before-work checks for source edits and subagent spawns.

Changes:

• Repeats holds while a verdict is missing, with an intended limit of three per repo/session.
• Adds SKILL.md fallback guidance and limits non-blocking reminders to read tools.
• Updates regression tests, documentation, and release metadata.

File	Description
tests/​agentic-advisor/​hooks_test.sh	Tests repeated holds, file switching, prose grades, and fallback guidance.
plugins/​agentic-advisor/​README.md	Documents repeated holds and read reminders.
plugins/​agentic-advisor/​hooks/​agentic-advisor-trigger.sh	Adds the hold counter and SKILL.md fallback messages.
plugins/​agentic-advisor/​CHANGELOG.md	Records the v1.4.7 behavior changes.
plugins/​agentic-advisor/​.claude-plugin/​plugin.json	Bumps the version to 1.4.7.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Fallback detection can bypass grading and misclassify telemetry, while reminder tracking and concurrent hold counting remain incorrect.

3 open issues

• Limit contamination checks to assistant Read invocations · New
• Restrict fallback verdict checks to assistant Read calls · New
• Count fallback Read runs in reminder tracking · New

1 previously missed issue

In code that hasn't changed since last review

Serialize hold counter updates to prevent concurrent overwrites

plugins/​agentic-advisor/​hooks/​agentic-advisor-trigger.sh:228

Concurrent Edit or Agent hooks can read the same hold count and overwrite each other's increments, allowing more than three denials per repo/session. Serialize the counter update with a per-repo/session lock and fail open if the lock cannot be acquired.

1 resolved since last review

• Recognize fallback skill reads and validate subsequent verdicts

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Verified fallback-detection failures on long transcripts can reapply holds, suppress telemetry, and misclassify baseline sessions.

1 open issue

• Avoid grep SIGPIPE causing valid fallback and telemetry failures · New

3 resolved since last review

• Count fallback Read runs in reminder tracking
• Restrict fallback verdict checks to assistant Read calls
• Limit contamination checks to assistant Read invocations

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Concurrent calls can exceed the hold cap, and fallback grading can reuse another repository’s verdict.

1 open issue

• Track grading completion per repository · New

1 previously missed issue

In code that hasn't changed since last review

Serialize hold-cap checks and increments

plugins/​agentic-advisor/​hooks/​agentic-advisor-trigger.sh:230

Parallel edits or subagent spawns can overwrite each other's counter increments: four overlapping calls can all be denied while recording only one hold. This exceeds the documented three-hold cap per repo/session. Serialize the check and increment with a per-marker lock that fails open on contention, and add coverage for concurrent calls.

1 resolved since last review

• Avoid grep SIGPIPE causing valid fallback and telemetry failures

_{Review effort: Balanced}