The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes git bisect run in Phase 2c to automate regression localization (rules/reproduction.md). The command executed during bisection is derived from the 'Evidence Record', which is populated with data from external sources like Dash0 telemetry and Linear tickets. This presents a risk of command injection if an attacker can control the content of logs or ticket descriptions that are parsed into reproduction commands.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its heavy reliance on processing untrusted external data. Ingested content from Dash0 (spans, logs, web events) and Linear (issue descriptions) is merged into the 'Bug-Notes Ledger' (rules/bug-notes-ledger.md), which serves as the primary context for the 'holistic-analysis' and 'aw-executor' agents. Maliciously crafted log entries could attempt to influence the agent's logic to introduce vulnerabilities or perform unauthorized file operations.
Ingestion points: rules/evidence-resolution.md defines procedures for resolving Dash0 URLs, Linear tickets, and stack traces into agent context.
Boundary markers: The skill uses structured templates for its 'Evidence Record', but does not explicitly document the use of robust delimiters or 'ignore' instructions to prevent the agent from obeying instructions embedded within the ingested data.
Capability inventory: The agent can modify the local file system, create GitHub PRs using the gh CLI, and execute shell commands via test runners and git bisect.
Sanitization: Sanitization is limited to basic parsing (e.g., stripping interpolations from error messages), leaving the narrative content vulnerable to injection attacks.
[REMOTE_CODE_EXECUTION]: The skill automates the generation and execution of test code by delegating to sub-agents like /tdd and /e2e-testing (rules/reproduction.md). Because the requirements for these tests are derived from untrusted evidence, there is a risk that a generator could be tricked into including malicious code in the resulting test files, which are then executed on the host system.
[EXTERNAL_DOWNLOADS]: The skill interacts with external service providers via MCP servers for Dash0 and Linear. These operations involve fetching telemetry and ticket data from third-party infrastructures, which are inherent dependencies but represent a network-based data ingestion surface.

fix-bug