fix(providers): discover workspace skills everywhere

[maria-rcks]

PR #8778 added cwd-aware skill discovery for Codex and OpenCode, and current main now covers Claude. This closes the remaining gaps: Cursor discovers project skills and translates selected $skill mentions to native /skill invocations, Grok probes its catalog in the project cwd, and mobile requests and resolves workspace snapshots.

before	after

Verified with 71 focused provider/client/mobile tests, targeted lint and formatting, and server/mobile typechecks.

Built with gpt-5.6-sol through the Codex harness.

---

Note

Medium Risk
Adds filesystem scans and CLI probes on provider refresh plus prompt rewriting in Cursor ACP; probe failures can block workspace snapshot caching until retry succeeds.

Overview
Closes workspace-aware skill gaps for Cursor and Grok and wires mobile/web composers to cwd-specific catalogs instead of machine-level provider skills.

Server: Cursor gets filesystem discovery of SKILL.md under standard skill roots (with scan budgets and frontmatter parsing), exposed via snapshotForCwd and used at send time to rewrite composer $skill mentions into Cursor’s /skill form. Grok’s grok inspect --json probe now fails explicitly on spawn/timeout/exit/decode errors so workspace snapshots are not cached as empty; machine-level Grok status still falls back to an empty list.

Clients:useComposerCommandMenu (and web ChatComposer) resolve skills with resolveProviderSkillsForCwd, export skills to editors, and call refreshProviders for the active cwd when no matching workspace snapshot exists, with a 10s retry cooldown. Thread detail passes thread cwd into the composer and feed skill lists.

^{Reviewed by Cursor Bugbot for commit 77ddc96. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add workspace-aware skill discovery for Cursor and Grok providers

• Mobile composers (NewTaskDraftScreen.tsx, ThreadComposer.tsx, ThreadDetailScreen.tsx) now source skills from the useComposerCommandMenu hook, which resolves provider skills against the active thread cwd and requests a provider refresh when no matching workspace snapshot exists.
• CursorDriver and GrokDriver gain workspace-specific snapshot functions that run skill discovery in the requested cwd when the instance is enabled; disabled instances reuse the machine snapshot.
• CursorSkills.ts adds recursive SKILL.md scanning with depth, entry, and byte budgets, frontmatter metadata parsing, project-over-user dedup across four root conventions, and typed CursorSkillsProbeError for incomplete scans.
• CursorAdapter now caches per-session discovered skill names and rewrites known dollar-prefixed mentions into Cursor slash invocations before sending to ACP.
• GrokSkills.ts changes discovery to a typed effect that fails explicitly on spawn, timeout, non-zero-exit, and decode errors; GrokProvider recovers these at the machine status boundary.
• Behavioral Change: discoverGrokSkills now fails instead of returning an empty list on probe errors; callers that expected empty-list fallbacks are updated but out-of-tree consumers may break. Strict probeCursorSkills fails when scan budgets are exhausted.

^{Macroscope summarized ee69cf3.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This change adds workspace skill discovery, provider-specific CLI/filesystem probing, and Cursor prompt rewriting across server, mobile, and web paths. Its broad new runtime behavior and multi-layer integration exceed the low-risk, self-contained change criteria.

Not approved because:

• Monthly spending limit reached (workspace setting). Approvability relies on correctness review in order to determine eligibility

^{Review your spending limits in Billing settings. You can add or adjust custom eligibility rules. Learn more.}

[archiekd]

Thank you @maria-rcks! I have been waiting for this for ages, hopefully this can get it asap!

[anirudhsama]

Thanks a ton for this @maria-rcks. This was a constant paper cut for me!

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ee69cf3. Configure here.}

[maria-rcks]

PR #8778 added cwd-aware skill discovery for Codex and OpenCode, and current main now covers Claude. This closes the remaining gaps: Cursor discovers project skills and translates selected $skill mentions to native /skill invocations, Grok probes its catalog in the project cwd, and mobile requests and resolves workspace snapshots.

before	after

Verified with 71 focused provider/client/mobile tests, targeted lint and formatting, and server/mobile typechecks.

Built with gpt-5.6-sol through the Codex harness.

---

Note

Medium Risk
Adds filesystem scans and CLI probes on provider refresh plus prompt rewriting in Cursor ACP; probe failures can block workspace snapshot caching until retry succeeds.

Overview
Closes workspace-aware skill gaps for Cursor and Grok and wires mobile/web composers to cwd-specific catalogs instead of machine-level provider skills.

Server: Cursor gets filesystem discovery of SKILL.md under standard skill roots (with scan budgets and frontmatter parsing), exposed via snapshotForCwd and used at send time to rewrite composer $skill mentions into Cursor’s /skill form. Grok’s grok inspect --json probe now fails explicitly on spawn/timeout/exit/decode errors so workspace snapshots are not cached as empty; machine-level Grok status still falls back to an empty list.

Clients:useComposerCommandMenu (and web ChatComposer) resolve skills with resolveProviderSkillsForCwd, export skills to editors, and call refreshProviders for the active cwd when no matching workspace snapshot exists, with a 10s retry cooldown. Thread detail passes thread cwd into the composer and feed skill lists.

^{Reviewed by Cursor Bugbot for commit 77ddc96. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add workspace-aware skill discovery for Cursor and Grok providers

• Mobile composers (NewTaskDraftScreen.tsx, ThreadComposer.tsx, ThreadDetailScreen.tsx) now source skills from the useComposerCommandMenu hook, which resolves provider skills against the active thread cwd and requests a provider refresh when no matching workspace snapshot exists.
• CursorDriver and GrokDriver gain workspace-specific snapshot functions that run skill discovery in the requested cwd when the instance is enabled; disabled instances reuse the machine snapshot.
• CursorSkills.ts adds recursive SKILL.md scanning with depth, entry, and byte budgets, frontmatter metadata parsing, project-over-user dedup across four root conventions, and typed CursorSkillsProbeError for incomplete scans.
• CursorAdapter now caches per-session discovered skill names and rewrites known dollar-prefixed mentions into Cursor slash invocations before sending to ACP.
• GrokSkills.ts changes discovery to a typed effect that fails explicitly on spawn, timeout, non-zero-exit, and decode errors; GrokProvider recovers these at the machine status boundary.
• Behavioral Change: discoverGrokSkills now fails instead of returning an empty list on probe errors; callers that expected empty-list fallbacks are updated but out-of-tree consumers may break. Strict probeCursorSkills fails when scan budgets are exhausted.

^{Macroscope summarized ee69cf3.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This change adds workspace skill discovery, provider-specific CLI/filesystem probing, and Cursor prompt rewriting across server, mobile, and web paths. Its broad new runtime behavior and multi-layer integration exceed the low-risk, self-contained change criteria.

Not approved because:

• Monthly spending limit reached (workspace setting). Approvability relies on correctness review in order to determine eligibility

^{Review your spending limits in Billing settings. You can add or adjust custom eligibility rules. Learn more.}

[archiekd]

Thank you @maria-rcks! I have been waiting for this for ages, hopefully this can get it asap!

[anirudhsama]

Thanks a ton for this @maria-rcks. This was a constant paper cut for me!

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ee69cf3. Configure here.}

[maria-rcks]

PR #8778 added cwd-aware skill discovery for Codex and OpenCode, and current main now covers Claude. This closes the remaining gaps: Cursor discovers project skills and translates selected $skill mentions to native /skill invocations, Grok probes its catalog in the project cwd, and mobile requests and resolves workspace snapshots.

before	after

Verified with 71 focused provider/client/mobile tests, targeted lint and formatting, and server/mobile typechecks.

Built with gpt-5.6-sol through the Codex harness.

---

Note

Medium Risk
Adds filesystem scans and CLI probes on provider refresh plus prompt rewriting in Cursor ACP; probe failures can block workspace snapshot caching until retry succeeds.

Overview
Closes workspace-aware skill gaps for Cursor and Grok and wires mobile/web composers to cwd-specific catalogs instead of machine-level provider skills.

Server: Cursor gets filesystem discovery of SKILL.md under standard skill roots (with scan budgets and frontmatter parsing), exposed via snapshotForCwd and used at send time to rewrite composer $skill mentions into Cursor’s /skill form. Grok’s grok inspect --json probe now fails explicitly on spawn/timeout/exit/decode errors so workspace snapshots are not cached as empty; machine-level Grok status still falls back to an empty list.

Clients:useComposerCommandMenu (and web ChatComposer) resolve skills with resolveProviderSkillsForCwd, export skills to editors, and call refreshProviders for the active cwd when no matching workspace snapshot exists, with a 10s retry cooldown. Thread detail passes thread cwd into the composer and feed skill lists.

^{Reviewed by Cursor Bugbot for commit 77ddc96. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add workspace-aware skill discovery for Cursor and Grok providers

• Mobile composers (NewTaskDraftScreen.tsx, ThreadComposer.tsx, ThreadDetailScreen.tsx) now source skills from the useComposerCommandMenu hook, which resolves provider skills against the active thread cwd and requests a provider refresh when no matching workspace snapshot exists.
• CursorDriver and GrokDriver gain workspace-specific snapshot functions that run skill discovery in the requested cwd when the instance is enabled; disabled instances reuse the machine snapshot.
• CursorSkills.ts adds recursive SKILL.md scanning with depth, entry, and byte budgets, frontmatter metadata parsing, project-over-user dedup across four root conventions, and typed CursorSkillsProbeError for incomplete scans.
• CursorAdapter now caches per-session discovered skill names and rewrites known dollar-prefixed mentions into Cursor slash invocations before sending to ACP.
• GrokSkills.ts changes discovery to a typed effect that fails explicitly on spawn, timeout, non-zero-exit, and decode errors; GrokProvider recovers these at the machine status boundary.
• Behavioral Change: discoverGrokSkills now fails instead of returning an empty list on probe errors; callers that expected empty-list fallbacks are updated but out-of-tree consumers may break. Strict probeCursorSkills fails when scan budgets are exhausted.

^{Macroscope summarized ee69cf3.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This change adds workspace skill discovery, provider-specific CLI/filesystem probing, and Cursor prompt rewriting across server, mobile, and web paths. Its broad new runtime behavior and multi-layer integration exceed the low-risk, self-contained change criteria.

Not approved because:

• Monthly spending limit reached (workspace setting). Approvability relies on correctness review in order to determine eligibility

^{Review your spending limits in Billing settings. You can add or adjust custom eligibility rules. Learn more.}

[archiekd]

Thank you @maria-rcks! I have been waiting for this for ages, hopefully this can get it asap!

[anirudhsama]

Thanks a ton for this @maria-rcks. This was a constant paper cut for me!

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ee69cf3. Configure here.}

[maria-rcks]

PR #8778 added cwd-aware skill discovery for Codex and OpenCode, and current main now covers Claude. This closes the remaining gaps: Cursor discovers project skills and translates selected $skill mentions to native /skill invocations, Grok probes its catalog in the project cwd, and mobile requests and resolves workspace snapshots.

before	after

Verified with 71 focused provider/client/mobile tests, targeted lint and formatting, and server/mobile typechecks.

Built with gpt-5.6-sol through the Codex harness.

---

Note

Medium Risk
Adds filesystem scans and CLI probes on provider refresh plus prompt rewriting in Cursor ACP; probe failures can block workspace snapshot caching until retry succeeds.

Overview
Closes workspace-aware skill gaps for Cursor and Grok and wires mobile/web composers to cwd-specific catalogs instead of machine-level provider skills.

Server: Cursor gets filesystem discovery of SKILL.md under standard skill roots (with scan budgets and frontmatter parsing), exposed via snapshotForCwd and used at send time to rewrite composer $skill mentions into Cursor’s /skill form. Grok’s grok inspect --json probe now fails explicitly on spawn/timeout/exit/decode errors so workspace snapshots are not cached as empty; machine-level Grok status still falls back to an empty list.

Clients:useComposerCommandMenu (and web ChatComposer) resolve skills with resolveProviderSkillsForCwd, export skills to editors, and call refreshProviders for the active cwd when no matching workspace snapshot exists, with a 10s retry cooldown. Thread detail passes thread cwd into the composer and feed skill lists.

^{Reviewed by Cursor Bugbot for commit 77ddc96. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add workspace-aware skill discovery for Cursor and Grok providers

• Mobile composers (NewTaskDraftScreen.tsx, ThreadComposer.tsx, ThreadDetailScreen.tsx) now source skills from the useComposerCommandMenu hook, which resolves provider skills against the active thread cwd and requests a provider refresh when no matching workspace snapshot exists.
• CursorDriver and GrokDriver gain workspace-specific snapshot functions that run skill discovery in the requested cwd when the instance is enabled; disabled instances reuse the machine snapshot.
• CursorSkills.ts adds recursive SKILL.md scanning with depth, entry, and byte budgets, frontmatter metadata parsing, project-over-user dedup across four root conventions, and typed CursorSkillsProbeError for incomplete scans.
• CursorAdapter now caches per-session discovered skill names and rewrites known dollar-prefixed mentions into Cursor slash invocations before sending to ACP.
• GrokSkills.ts changes discovery to a typed effect that fails explicitly on spawn, timeout, non-zero-exit, and decode errors; GrokProvider recovers these at the machine status boundary.
• Behavioral Change: discoverGrokSkills now fails instead of returning an empty list on probe errors; callers that expected empty-list fallbacks are updated but out-of-tree consumers may break. Strict probeCursorSkills fails when scan budgets are exhausted.

^{Macroscope summarized ee69cf3.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This change adds workspace skill discovery, provider-specific CLI/filesystem probing, and Cursor prompt rewriting across server, mobile, and web paths. Its broad new runtime behavior and multi-layer integration exceed the low-risk, self-contained change criteria.

Not approved because:

• Monthly spending limit reached (workspace setting). Approvability relies on correctness review in order to determine eligibility

^{Review your spending limits in Billing settings. You can add or adjust custom eligibility rules. Learn more.}

[archiekd]

Thank you @maria-rcks! I have been waiting for this for ages, hopefully this can get it asap!

[anirudhsama]

Thanks a ton for this @maria-rcks. This was a constant paper cut for me!

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ee69cf3. Configure here.}

[maria-rcks]

PR #8778 added cwd-aware skill discovery for Codex and OpenCode, and current main now covers Claude. This closes the remaining gaps: Cursor discovers project skills and translates selected $skill mentions to native /skill invocations, Grok probes its catalog in the project cwd, and mobile requests and resolves workspace snapshots.

before	after

Verified with 71 focused provider/client/mobile tests, targeted lint and formatting, and server/mobile typechecks.

Built with gpt-5.6-sol through the Codex harness.

---

Note

Medium Risk
Adds filesystem scans and CLI probes on provider refresh plus prompt rewriting in Cursor ACP; probe failures can block workspace snapshot caching until retry succeeds.

Overview
Closes workspace-aware skill gaps for Cursor and Grok and wires mobile/web composers to cwd-specific catalogs instead of machine-level provider skills.

Server: Cursor gets filesystem discovery of SKILL.md under standard skill roots (with scan budgets and frontmatter parsing), exposed via snapshotForCwd and used at send time to rewrite composer $skill mentions into Cursor’s /skill form. Grok’s grok inspect --json probe now fails explicitly on spawn/timeout/exit/decode errors so workspace snapshots are not cached as empty; machine-level Grok status still falls back to an empty list.

Clients:useComposerCommandMenu (and web ChatComposer) resolve skills with resolveProviderSkillsForCwd, export skills to editors, and call refreshProviders for the active cwd when no matching workspace snapshot exists, with a 10s retry cooldown. Thread detail passes thread cwd into the composer and feed skill lists.

^{Reviewed by Cursor Bugbot for commit 77ddc96. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add workspace-aware skill discovery for Cursor and Grok providers

• Mobile composers (NewTaskDraftScreen.tsx, ThreadComposer.tsx, ThreadDetailScreen.tsx) now source skills from the useComposerCommandMenu hook, which resolves provider skills against the active thread cwd and requests a provider refresh when no matching workspace snapshot exists.
• CursorDriver and GrokDriver gain workspace-specific snapshot functions that run skill discovery in the requested cwd when the instance is enabled; disabled instances reuse the machine snapshot.
• CursorSkills.ts adds recursive SKILL.md scanning with depth, entry, and byte budgets, frontmatter metadata parsing, project-over-user dedup across four root conventions, and typed CursorSkillsProbeError for incomplete scans.
• CursorAdapter now caches per-session discovered skill names and rewrites known dollar-prefixed mentions into Cursor slash invocations before sending to ACP.
• GrokSkills.ts changes discovery to a typed effect that fails explicitly on spawn, timeout, non-zero-exit, and decode errors; GrokProvider recovers these at the machine status boundary.
• Behavioral Change: discoverGrokSkills now fails instead of returning an empty list on probe errors; callers that expected empty-list fallbacks are updated but out-of-tree consumers may break. Strict probeCursorSkills fails when scan budgets are exhausted.

^{Macroscope summarized ee69cf3.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This change adds workspace skill discovery, provider-specific CLI/filesystem probing, and Cursor prompt rewriting across server, mobile, and web paths. Its broad new runtime behavior and multi-layer integration exceed the low-risk, self-contained change criteria.

Not approved because:

• Monthly spending limit reached (workspace setting). Approvability relies on correctness review in order to determine eligibility

^{Review your spending limits in Billing settings. You can add or adjust custom eligibility rules. Learn more.}

[archiekd]

Thank you @maria-rcks! I have been waiting for this for ages, hopefully this can get it asap!

[anirudhsama]

Thanks a ton for this @maria-rcks. This was a constant paper cut for me!

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ee69cf3. Configure here.}

[maria-rcks]

PR #8778 added cwd-aware skill discovery for Codex and OpenCode, and current main now covers Claude. This closes the remaining gaps: Cursor discovers project skills and translates selected $skill mentions to native /skill invocations, Grok probes its catalog in the project cwd, and mobile requests and resolves workspace snapshots.

before	after

Verified with 71 focused provider/client/mobile tests, targeted lint and formatting, and server/mobile typechecks.

Built with gpt-5.6-sol through the Codex harness.

---

Note

Medium Risk
Adds filesystem scans and CLI probes on provider refresh plus prompt rewriting in Cursor ACP; probe failures can block workspace snapshot caching until retry succeeds.

Overview
Closes workspace-aware skill gaps for Cursor and Grok and wires mobile/web composers to cwd-specific catalogs instead of machine-level provider skills.

Server: Cursor gets filesystem discovery of SKILL.md under standard skill roots (with scan budgets and frontmatter parsing), exposed via snapshotForCwd and used at send time to rewrite composer $skill mentions into Cursor’s /skill form. Grok’s grok inspect --json probe now fails explicitly on spawn/timeout/exit/decode errors so workspace snapshots are not cached as empty; machine-level Grok status still falls back to an empty list.

Clients:useComposerCommandMenu (and web ChatComposer) resolve skills with resolveProviderSkillsForCwd, export skills to editors, and call refreshProviders for the active cwd when no matching workspace snapshot exists, with a 10s retry cooldown. Thread detail passes thread cwd into the composer and feed skill lists.

^{Reviewed by Cursor Bugbot for commit 77ddc96. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add workspace-aware skill discovery for Cursor and Grok providers

• Mobile composers (NewTaskDraftScreen.tsx, ThreadComposer.tsx, ThreadDetailScreen.tsx) now source skills from the useComposerCommandMenu hook, which resolves provider skills against the active thread cwd and requests a provider refresh when no matching workspace snapshot exists.
• CursorDriver and GrokDriver gain workspace-specific snapshot functions that run skill discovery in the requested cwd when the instance is enabled; disabled instances reuse the machine snapshot.
• CursorSkills.ts adds recursive SKILL.md scanning with depth, entry, and byte budgets, frontmatter metadata parsing, project-over-user dedup across four root conventions, and typed CursorSkillsProbeError for incomplete scans.
• CursorAdapter now caches per-session discovered skill names and rewrites known dollar-prefixed mentions into Cursor slash invocations before sending to ACP.
• GrokSkills.ts changes discovery to a typed effect that fails explicitly on spawn, timeout, non-zero-exit, and decode errors; GrokProvider recovers these at the machine status boundary.
• Behavioral Change: discoverGrokSkills now fails instead of returning an empty list on probe errors; callers that expected empty-list fallbacks are updated but out-of-tree consumers may break. Strict probeCursorSkills fails when scan budgets are exhausted.

^{Macroscope summarized ee69cf3.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This change adds workspace skill discovery, provider-specific CLI/filesystem probing, and Cursor prompt rewriting across server, mobile, and web paths. Its broad new runtime behavior and multi-layer integration exceed the low-risk, self-contained change criteria.

Not approved because:

• Monthly spending limit reached (workspace setting). Approvability relies on correctness review in order to determine eligibility

^{Review your spending limits in Billing settings. You can add or adjust custom eligibility rules. Learn more.}

[archiekd]

Thank you @maria-rcks! I have been waiting for this for ages, hopefully this can get it asap!

[anirudhsama]

Thanks a ton for this @maria-rcks. This was a constant paper cut for me!

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ee69cf3. Configure here.}

[maria-rcks]

PR #8778 added cwd-aware skill discovery for Codex and OpenCode, and current main now covers Claude. This closes the remaining gaps: Cursor discovers project skills and translates selected $skill mentions to native /skill invocations, Grok probes its catalog in the project cwd, and mobile requests and resolves workspace snapshots.

before	after

Verified with 71 focused provider/client/mobile tests, targeted lint and formatting, and server/mobile typechecks.

Built with gpt-5.6-sol through the Codex harness.

---

Note

Medium Risk
Adds filesystem scans and CLI probes on provider refresh plus prompt rewriting in Cursor ACP; probe failures can block workspace snapshot caching until retry succeeds.

Overview
Closes workspace-aware skill gaps for Cursor and Grok and wires mobile/web composers to cwd-specific catalogs instead of machine-level provider skills.

Server: Cursor gets filesystem discovery of SKILL.md under standard skill roots (with scan budgets and frontmatter parsing), exposed via snapshotForCwd and used at send time to rewrite composer $skill mentions into Cursor’s /skill form. Grok’s grok inspect --json probe now fails explicitly on spawn/timeout/exit/decode errors so workspace snapshots are not cached as empty; machine-level Grok status still falls back to an empty list.

Clients:useComposerCommandMenu (and web ChatComposer) resolve skills with resolveProviderSkillsForCwd, export skills to editors, and call refreshProviders for the active cwd when no matching workspace snapshot exists, with a 10s retry cooldown. Thread detail passes thread cwd into the composer and feed skill lists.

^{Reviewed by Cursor Bugbot for commit 77ddc96. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add workspace-aware skill discovery for Cursor and Grok providers

• Mobile composers (NewTaskDraftScreen.tsx, ThreadComposer.tsx, ThreadDetailScreen.tsx) now source skills from the useComposerCommandMenu hook, which resolves provider skills against the active thread cwd and requests a provider refresh when no matching workspace snapshot exists.
• CursorDriver and GrokDriver gain workspace-specific snapshot functions that run skill discovery in the requested cwd when the instance is enabled; disabled instances reuse the machine snapshot.
• CursorSkills.ts adds recursive SKILL.md scanning with depth, entry, and byte budgets, frontmatter metadata parsing, project-over-user dedup across four root conventions, and typed CursorSkillsProbeError for incomplete scans.
• CursorAdapter now caches per-session discovered skill names and rewrites known dollar-prefixed mentions into Cursor slash invocations before sending to ACP.
• GrokSkills.ts changes discovery to a typed effect that fails explicitly on spawn, timeout, non-zero-exit, and decode errors; GrokProvider recovers these at the machine status boundary.
• Behavioral Change: discoverGrokSkills now fails instead of returning an empty list on probe errors; callers that expected empty-list fallbacks are updated but out-of-tree consumers may break. Strict probeCursorSkills fails when scan budgets are exhausted.

^{Macroscope summarized ee69cf3.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This change adds workspace skill discovery, provider-specific CLI/filesystem probing, and Cursor prompt rewriting across server, mobile, and web paths. Its broad new runtime behavior and multi-layer integration exceed the low-risk, self-contained change criteria.

Not approved because:

• Monthly spending limit reached (workspace setting). Approvability relies on correctness review in order to determine eligibility

^{Review your spending limits in Billing settings. You can add or adjust custom eligibility rules. Learn more.}

[archiekd]

Thank you @maria-rcks! I have been waiting for this for ages, hopefully this can get it asap!

[anirudhsama]

Thanks a ton for this @maria-rcks. This was a constant paper cut for me!

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ee69cf3. Configure here.}

[maria-rcks]

PR #8778 added cwd-aware skill discovery for Codex and OpenCode, and current main now covers Claude. This closes the remaining gaps: Cursor discovers project skills and translates selected $skill mentions to native /skill invocations, Grok probes its catalog in the project cwd, and mobile requests and resolves workspace snapshots.

before	after

Verified with 71 focused provider/client/mobile tests, targeted lint and formatting, and server/mobile typechecks.

Built with gpt-5.6-sol through the Codex harness.

---

Note

Medium Risk
Adds filesystem scans and CLI probes on provider refresh plus prompt rewriting in Cursor ACP; probe failures can block workspace snapshot caching until retry succeeds.

Overview
Closes workspace-aware skill gaps for Cursor and Grok and wires mobile/web composers to cwd-specific catalogs instead of machine-level provider skills.

Server: Cursor gets filesystem discovery of SKILL.md under standard skill roots (with scan budgets and frontmatter parsing), exposed via snapshotForCwd and used at send time to rewrite composer $skill mentions into Cursor’s /skill form. Grok’s grok inspect --json probe now fails explicitly on spawn/timeout/exit/decode errors so workspace snapshots are not cached as empty; machine-level Grok status still falls back to an empty list.

Clients:useComposerCommandMenu (and web ChatComposer) resolve skills with resolveProviderSkillsForCwd, export skills to editors, and call refreshProviders for the active cwd when no matching workspace snapshot exists, with a 10s retry cooldown. Thread detail passes thread cwd into the composer and feed skill lists.

^{Reviewed by Cursor Bugbot for commit 77ddc96. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add workspace-aware skill discovery for Cursor and Grok providers

• Mobile composers (NewTaskDraftScreen.tsx, ThreadComposer.tsx, ThreadDetailScreen.tsx) now source skills from the useComposerCommandMenu hook, which resolves provider skills against the active thread cwd and requests a provider refresh when no matching workspace snapshot exists.
• CursorDriver and GrokDriver gain workspace-specific snapshot functions that run skill discovery in the requested cwd when the instance is enabled; disabled instances reuse the machine snapshot.
• CursorSkills.ts adds recursive SKILL.md scanning with depth, entry, and byte budgets, frontmatter metadata parsing, project-over-user dedup across four root conventions, and typed CursorSkillsProbeError for incomplete scans.
• CursorAdapter now caches per-session discovered skill names and rewrites known dollar-prefixed mentions into Cursor slash invocations before sending to ACP.
• GrokSkills.ts changes discovery to a typed effect that fails explicitly on spawn, timeout, non-zero-exit, and decode errors; GrokProvider recovers these at the machine status boundary.
• Behavioral Change: discoverGrokSkills now fails instead of returning an empty list on probe errors; callers that expected empty-list fallbacks are updated but out-of-tree consumers may break. Strict probeCursorSkills fails when scan budgets are exhausted.

^{Macroscope summarized ee69cf3.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This change adds workspace skill discovery, provider-specific CLI/filesystem probing, and Cursor prompt rewriting across server, mobile, and web paths. Its broad new runtime behavior and multi-layer integration exceed the low-risk, self-contained change criteria.

Not approved because:

• Monthly spending limit reached (workspace setting). Approvability relies on correctness review in order to determine eligibility

^{Review your spending limits in Billing settings. You can add or adjust custom eligibility rules. Learn more.}

[archiekd]

Thank you @maria-rcks! I have been waiting for this for ages, hopefully this can get it asap!

[anirudhsama]

Thanks a ton for this @maria-rcks. This was a constant paper cut for me!

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ee69cf3. Configure here.}