[codex] deliver ambient roadmap features

[benclawbot]

Summary

• add a README product roadmap for not-yet-delivered final roadmap features, then mark this pass fully integrated after completing the roadmap
• deliver Action Flight Recorder as current behavior instead of leaving it on the roadmap
• add SafetySystem flight recorder read model plus first-party flight_recorder tool
• deliver Sensitive Context Firewall controls for redaction previews, capture pause/resume, forget-recent context, storage status, and capture enforcement
• expose firewall actions through the personal tool and update system prompt guidance
• deliver Recipe Catalog and Command Palette as built-in recipes plus a first-party recipe planning tool
• deliver Meeting and Voice Memory Mode with local transcript segments, source-linked notes, extracted decisions/questions/action items, and conversion into reminders/jobs/delegation drafts
• deliver Connector Packs With Permission Scopes with built-in Outlook/Gmail/Calendar, Slack/Teams, GitHub/Linear/Jira, Notion/Obsidian, and file-share connector definitions
• add first-party connector actions for scope grants, revocation, write preflight, write evidence recording, and write-audit lookup
• deliver Proactive Briefings and Next-Best Actions with morning briefings, end-task recaps, meeting prep cards, project resume cards, contextual recommendations, and never-suggest feedback
• add first-party briefing actions for proactive cards, saved recaps, low-noise recommendations, and feedback-rule lookup
• deliver Windows App Intent Manifests with iagent.intent.json discovery, validation, import persistence, action listing, and non-executing typed action plans
• add first-party intent actions that expose app/script parameters, examples, approval levels, rollback hints, and recipe-ready import hints without bypassing approval boundaries
• deliver Remote Dispatch and Watch Mode with authenticated clients, local/remote task submit, mobile-friendly status, scheduled jobs, approval-needed notifications, completion evidence, failure packets, and watch events
• add first-party dispatch actions for client creation/revocation, submit, approve, complete, fail, status, list, watch, and due-task lookup
• deliver Attention Budget and Interruption Control with quiet hours, hourly/daily caps, snooze/resume, critical-approval bypass, attention preflight, history, and digest summaries
• add first-party attention actions so proactive suggestions, reminders, briefing cards, and dispatch updates can be budgeted before interrupting the user

Validation

• rtk cargo test -q --test e2e test_action_flight_recorder
• rtk cargo test -q tool::tests::registry_exposes_action_flight_recorder_tool
• rtk cargo test -q --test e2e sensitive_context_firewall
• rtk cargo test -q tool::tests::personal_tool_schema_exposes_sensitive_context_firewall_actions
• rtk cargo test -q --test e2e recipe_catalog
• rtk cargo test -q tool::tests::registry_exposes_recipe_catalog_tool
• rtk cargo test -q --test e2e meeting_memory
• rtk cargo test -q tool::tests::registry_exposes_meeting_memory_tool
• rtk cargo test -q --test e2e connector_packs
• rtk cargo test -q tool::tests::registry_exposes_connector_pack_tool
• rtk cargo test -q --test e2e proactive_briefings
• rtk cargo test -q tool::tests::registry_exposes_proactive_briefing_tool
• rtk cargo test -q --test e2e intent_manifests
• rtk cargo test -q tool::tests::registry_exposes_intent_manifest_tool
• rtk cargo test -q --test e2e remote_dispatch
• rtk cargo test -q tool::tests::registry_exposes_remote_dispatch_tool
• rtk cargo test -q --test e2e attention_budget
• rtk cargo test -q tool::tests::registry_exposes_attention_budget_tool
• rtk cargo test -q prompt::prompt_tests
• rtk cargo check -q --lib
• git diff --check

cargo check still reports pre-existing unrelated warnings in ambient modules.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9e26b39d4a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Detect bearer tokens before storing sensitive text

redact_sensitive_text tokenizes input with split_whitespace(), so each token passed into classify_sensitive_token never contains spaces. The lower.contains("bearer ") check therefore never matches and headers like Authorization: Bearer abc123... are not flagged/redacted, which can leak auth tokens into preview output and stored context.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Use shared safety state when serving flight recorder views

Constructing a fresh SafetySystem here drops the in-memory actions log for the active runtime (the constructor initializes actions as empty), so flight_recorder cannot return rich kind:"action" entries with per-action summaries from the current session. In practice this degrades the tool to persisted audit-only data and misses the detailed action timeline the feature is meant to expose.

Useful? React with 👍 / 👎.