agentleFS
verivus-ossverivus-oss/llm-cli-gateway/.agents/skills/session-workflow/SKILL.md

session-workflow

Skill16 starsChanged 2 months agoMITView on GitHub

Manage gateway bookkeeping and provider-native conversation continuity across Claude, Codex, Gemini, Grok, Mistral, Devin, and Cursor. Use for multi-turn work, session inspection, and safe resume decisions.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

14 rules. About 0 are common, 4 are distinctive, and 10 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 10

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Before using a provider-specific session field, query:

    line 22
  2. Very specific:

    Use a native sessionId only when it came from that provider or was supplied by the caller as a verified native handle.

    line 34
  3. Very specific:

    All non-Claude providers use approvalStrategy:"legacy".

    line 51
  4. Very specific:

    When a user explicitly authorizes full provider permissions and native MCP access for review, follow multi-llm-review's full-access …

    line 57
  5. Very specific:

    Do not rely on continuity for the grant.

    line 64
  6. Distinctive:

    Create a gateway record when you need description, active-pointer, cache projection, or local workflow organization:

    line 76
  7. Distinctive:

    Treat the returned ID as gateway metadata.

    line 86
  8. Very specific:

    Use explicit workingDir for Claude, Grok, Mistral, and Devin.

    line 140
  9. Very specific:

    Use workingDir for a fresh Codex session.

    line 141
  10. Very specific:

    Use workingDir for Gemini's target checkout.

    line 145
  11. Very specific:

    Use workingDir for Cursor's process cwd.

    line 147
  12. Very specific:

    When async tools are registered, a sync call may return a deferred job.

    line 187
  13. Distinctive:

    Do not impose a review deadline or arbitrary review round count because a session is involved.

    line 192
  14. Very specific:

    Use session_delete({sessionId:"..."}) for an obsolete bookkeeping record and session_clear_all({cli:"..."}) only when the caller wants to …

    line 214

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Session Workflow

Separate gateway bookkeeping from native provider continuity. session_create, session_list, session_set_active, and session_get manage gateway metadata; they do not create or prove a provider-native conversation. Never pass a gateway-generated ID to a provider as though it were a native handle.

For a review session, dispatch requests through the local stdio gateway MCP surface. A session does not authorize a direct provider CLI, connector/shadow gateway, or shell fallback.

Discover before resuming

Before using a provider-specific session field, query:

provider_tool_capabilities({cli:"claude"})
provider_tool_capabilities({cli:"codex"})
provider_tool_capabilities({cli:"gemini"})
provider_tool_capabilities({cli:"grok"})
provider_tool_capabilities({cli:"mistral"})
provider_tool_capabilities({cli:"devin"})
provider_tool_capabilities({cli:"cursor"})

Use a native sessionId only when it came from that provider or was supplied by the caller as a verified native handle. A fresh gateway session often uses a gw-* bookkeeping ID. It is not resumable for Codex, Grok, Mistral, Devin, or Cursor, and must never be replayed to Gemini either.

Native continuity matrix

ProviderNative resumeImportant boundary
ClaudesessionId maps to --session-id; continuation maps to --continue.Managed native continuation is a high-risk posture input requiring an approval decision and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1.
CodexReal Codex UUID maps to codex exec resume <UUID>; resumeLatest:true resumes a previous session through --last.Do not rely on the resumed cwd or on which session --last selects (#258). Use a real UUID for a specific Codex session. Resume drops sandboxMode and fullAuto, though configOverrides can still set sandbox_mode.
GeminiA caller-owned Antigravity conversation ID maps to --conversation; resumeLatest:true maps to --continue.Fresh launches do not produce a reusable native ID. Check resumable:true; never replay gw-*.
GrokNative ID maps to --resume; resumeLatest:true maps to --continue.A fresh gw-* ID is bookkeeping only. Cwd affects --continue context.
MistralNative ID maps to --resume; resumeLatest:true maps to --continue.A fresh gw-* ID is bookkeeping only. Vibe session logging defaults on; doctor --json flags only explicit [session_logging] enabled = false.
DevinNative ID maps to --resume; resumeLatest:true maps to --continue.A fresh gw-* ID is bookkeeping only. Use workingDir or workspace to bind the selected cwd.
CursorNative chat/session ID maps to --resume; resumeLatest:true maps to --continue.A fresh gw-* ID is bookkeeping only. Use workingDir for its checkout, or a verified workspace.

All non-Claude providers use approvalStrategy:"legacy"; they reject Claude-only mcp_managed and ignore approvalPolicy. mcp_managed is usable only with Claude and its request-scoped gateway-owned MCP configuration.

Explicit full-access review sessions

When a user explicitly authorizes full provider permissions and native MCP access for review, follow multi-llm-review's full-access protocol instead of the safe examples below. Build and launch a fresh target-checkout stdio gateway with node dist/index.js --transport=stdio; do not use a globally installed gateway whose revision is unknown. Reapply the provider-native full-access mapping on every new job and verify it through the live capability surface.

Do not rely on continuity for the grant. In particular, the gateway drops sandboxMode on a Codex resume, so you cannot select a posture that way, and the resumed posture is not guaranteed to match the original either: configOverrides still passes through and Codex re-resolves configuration on a cold resume. A full-access Codex review therefore needs a fresh native session. Keep ambient provider MCP configuration available, do not turn request tool/MCP lists into an asserted allowlist, and record the exact base, diff or changed-file list, verification report, and durable job evidence with every review iteration. If the user requests 90-second progress checks, do not poll earlier than that cadence.

Gateway bookkeeping sessions

Create a gateway record when you need description, active-pointer, cache projection, or local workflow organization:

session_create({cli:"claude",description:"Refactor auth module",setAsActive:true})
session_list({cli:"claude"})
session_get({sessionId:"<gateway-session-id>"})
session_set_active({cli:"claude",sessionId:"<gateway-session-id>"})

Treat the returned ID as gateway metadata. Pair it with a separately verified native handle if you want provider continuation. Do not assume an active gateway session causes Codex, Gemini, Grok, Mistral, Devin, or Cursor to resume a provider conversation.

Safe request examples

Gateway tool calls:

claude_request({
  prompt:"Continue the approved task.",
  sessionId:"<verified-claude-id>",
  approvalStrategy:"legacy"
})

codex_request({
  prompt:"Continue the approved task.",
  resumeLatest:true,
  approvalStrategy:"legacy"
})

gemini_request({
  prompt:"Continue the approved task.",
  sessionId:"<caller-owned-antigravity-conversation>",
  workingDir:"<repo>",
  approvalStrategy:"legacy"
})

grok_request({prompt:"Continue.",resumeLatest:true,workingDir:"<repo>",approvalStrategy:"legacy"})
mistral_request({prompt:"Continue.",resumeLatest:true,workingDir:"<repo>",approvalStrategy:"legacy"})
devin_request({prompt:"Continue.",resumeLatest:true,workingDir:"<repo>",approvalStrategy:"legacy"})
cursor_request({prompt:"Continue.",resumeLatest:true,workingDir:"<repo>",approvalStrategy:"legacy"})

Codex resumeLatest is UNDER REVIEW (#258): --last is cwd-filtered upstream unless --all is passed, which the gateway does not emit, and the child is still spawned with the gateway-resolved cwd. Do not rely on which session it selects or on the resumed working directory. To target a specific native session, pass its verified real Codex UUID as sessionId.

For Codex, start source inspection with sandboxMode:"read-only"; use workspace-write only if the work needs write-producing commands. Do not use new fullAuto:true examples. On a Codex resume, sandboxMode and fullAuto are not emitted, so neither can select a posture there; configOverrides can still set sandbox_mode. Start a fresh native session when the posture must be known.

Target and concurrency discipline

Native session state is contextual. Verify the repository/workspace and cwd before every resume, especially when multiple workstations or repositories run simultaneously:

  • Use explicit workingDir for Claude, Grok, Mistral, and Devin.
  • Use workingDir for a fresh Codex session. Do not assume it leaves resumeLatest unaffected: --last is cwd-filtered upstream and the child is still spawned with the gateway-resolved cwd, so it can influence which session is selected. Use a verified real Codex UUID to target a specific session.
  • Use workingDir for Gemini's target checkout. includeDirs remains an extra read path and still does not select cwd.
  • Use workingDir for Cursor's process cwd. Its workspace is a separate Cursor-native selector (a saved-workspace name, a .code-workspace file, or a directory), so passing both an absolute workspace path and a different workingDir is rejected rather than silently ranked.

For providers whose latest-session selection is cwd-scoped, an unscoped local CLI child runs in a fresh neutral temporary directory, not the gateway repository. Their cwd-scoped resumeLatest therefore fails closed unless workingDir, workspace, or a configured default workspace supplies a stable target. Codex was documented as the exception, on the basis that resumeLatest selects a globally latest session. Do not rely on that either (#258): --last is cwd-filtered upstream and the child is still spawned with the gateway-resolved cwd. Verify it, or start a fresh session when the target must be certain.

The gateway's cache/session metadata does not prove that a provider resumed the intended repository. Record the native handle, selected target, provider, and commit/diff evidence together.

Cache-aware turns

Claude, Codex, Gemini, Grok, and Mistral accept either prompt or structured promptParts. Devin and Cursor accept only flat prompt. Keep the stable system/tools/context prefix byte-identical only where promptParts is supported, and retain a canonical flat equivalent for the other two providers.

claude_request({
  promptParts:{system:"<stable>",context:"<stable spec>",task:"Follow-up task"},
  sessionId:"<verified-claude-id>",
  approvalStrategy:"legacy"
})

prompt and promptParts are mutually exclusive. cache-state://global, cache-state://session/{sessionId}, and cache-state://prefix/{hash} contain token/hash aggregates, not prompt/response text. A shared prefix hash is not evidence that every provider hit an equivalent cache.

Async and review continuity

When async tools are registered, a sync call may return a deferred job. Poll with llm_job_status and read llm_job_result once terminal. SQLite/PostgreSQL job results survive restarts; acknowledged memory is ephemeral; backend none does not register async/job tools.

Do not impose a review deadline or arbitrary review round count because a session is involved. A mandatory review remains incomplete until every required reviewer returns an evidence-backed unconditional approval. A provider failure must be repaired/retried or reported INCOMPLETE/BLOCKED, never skipped.

Personal Agent Config Kit

Kit mode is local-only, supports only Claude/Codex/Mistral, requires healthy SQLite/PostgreSQL durable admission, and retires native continuation after a gateway restart. It disables validation and least-cost routing. Use explain_effective_config before consequential Kit work. If effective Kit caps constrain an exhaustive review, do not claim an unconditional/complete review without an approved uncapped profile or explicit user direction.

explain_effective_config({workingDir:"<repo>"}) can inspect a candidate Kit scope. Do not pass that workingDir to a Claude or Mistral Kit request: Kit rejects it before compiling context. Select that Kit target through an already configured registered workspace alias or the configured default workspace. It never inherits the gateway process cwd.

Cleanup

Use session_delete({sessionId:"..."}) for an obsolete bookkeeping record and session_clear_all({cli:"..."}) only when the caller wants to clear that gateway metadata. Neither command deletes a provider-native conversation.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.