bidirectional-bridge-claude-codex: Skill for Claude Code

Bridge Web Audit

Keep the application question and next missing witness central. This skill uses the existing bridge and tools; it does not connect PDX/BELIEF engines or add an orchestrator. Preserve the user's scope, tool preferences and read-only boundaries. Offline analysis stays offline.

Work from evidence

Read the program scope, saved campaign, available actors/sessions and relevant business rules. Distinguish actor identity from role, object owner from actor, and object tenant from actor tenant. Unknown facts stay explicit. Keep exact origin/port, serialized HTTP URL, method, route and ordered duplicate query shape beside each receipt. Recipes refuse URLs that Node would rewrite, such as unresolved dot segments or literal spaces/Unicode; prepare the encoded URL explicitly. A JS route is discovered until an actual capture exists.

Use the user's capture tools. The current executable MVP accepts its versioned JSON import format and has a Node HTTP runner; it does not import arbitrary HAR/Burp files or instrument a browser. Read formats-and-replay.md for the schema, working CLI commands and local fixture.

Keep retained bytes, derived JSON and interpretation separate. The hash binds the retained representation; it does not establish identity, policy, causality or impact. A 200 error, login redirect, timeout, copied source or fired hook is not a vulnerability. Credential headers resolve from environment references at execution and are redacted in the retained request; response data can still contain private information.

Resume shared work

Use memory_get_context on entry for the exact bug-bounty program, branch, application version and evidence epoch; load only useful working notes and source references. Inspect reference integrity and the campaign before relying on a note. Resume from the saved scope/change_cursor with memory_get_changes, paging without skipping to head_cursor. At handoff, record meaningful observations, missing witnesses and pending work with memory_upsert, using the current owned task and the revision read. On a conflict, reread and merge. Use the native manager for these MCP calls; delegated Codex workers receive the selected context through input artifacts because their project bridge MCP is disabled. Persist their proposed notes as manager with references to the worker output and the actual writer's task. Keep the campaign as the sole canonical hypothesis/evidence ledger. For abandoned work, preserve reason, tested conditions and reopening trigger; an active working note is never an oracle-confirmed vulnerability. Archive or supersede temporary notes when their context changes; earlier revisions remain.