@freibergergarcia/phone-a-friend 2.7.1 → 2.8.0

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "phone-a-friend",
   "description": "CLI relay that lets AI coding agents collaborate by sending prompts and repository context to backend agents.",
-  "version": "2.7.1",
+  "version": "2.8.0",
   "author": {
     "name": "Bruno Freiberger"
   }

package/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "phone-a-friend",
-  "version": "2.7.1",
+  "version": "2.8.0",
   "description": "CLI relay that lets AI coding agents collaborate by sending prompts and repository context to backend agents.",
   "author": {
     "name": "Bruno Freiberger"

package/README.md

@@ -194,7 +194,7 @@ phone-a-friend --to codex --prompt "Review the auth module" --session auth-revie
 phone-a-friend --to codex --prompt "Now fix those issues" --session auth-review
 ```
 
-Sessions work reliably with Claude, Codex, and OpenCode. Ollama replays history (may hit token limits on long conversations). Gemini sessions are currently unsupported.
+Sessions work reliably with Claude, Codex, Gemini, and OpenCode. Ollama replays history (may hit token limits on long conversations). (Gemini resumes natively via `--session-id`/`--resume`; resume depends on Gemini's session retention.)
 
 ### Job tracking
 

package/commands/phone-a-friend.md

@@ -186,8 +186,8 @@ PAF_CONTEXT_EOF
 
    "$RELAY_BIN" --to codex --repo "$PWD" --prompt "$(cat "$PROMPT_FILE")" --context-file "$CONTEXT_FILE" $PAF_NO_DIFF [--fast] [--session <id>]
    # For gemini, omit --model by default (let auto-routing pick); see "Gemini model selection" below.
-   # Do NOT pass --session to gemini — it will error (see "Session continuity" below):
-   "$RELAY_BIN" --to gemini --repo "$PWD" --prompt "$(cat "$PROMPT_FILE")" --context-file "$CONTEXT_FILE" $PAF_NO_DIFF [--fast]
+   # Gemini supports --session via native resume (see "Session continuity" below):
+   "$RELAY_BIN" --to gemini --repo "$PWD" --prompt "$(cat "$PROMPT_FILE")" --context-file "$CONTEXT_FILE" $PAF_NO_DIFF [--fast] [--session <id>]
    ```
 
    Use delimiter names that do not appear in the payload. The quoted heredoc
@@ -264,14 +264,10 @@ Benefits: the backend keeps full conversation history, so follow-up prompts
 can be shorter (no need to re-send context from previous turns).
 
 **Backend-specific behavior:**
-- **Codex, Claude, OpenCode**: native session resume. Follow-up prompts
-  can send deltas only.
+- **Codex, Claude, Gemini, OpenCode**: native session resume. Follow-up
+  prompts can send deltas only.
 - **Ollama**: replays full history each call. Sessions work but prompt
   size grows with each turn. Keep follow-ups concise.
-- **Gemini**: `--session` is **not supported**. PaF rejects it with a
-  RelayError (`--session is not supported by the gemini backend ...`).
-  Each Gemini relay call must be self-contained. Do not pass `--session`
-  with `--to gemini`.
 
 On the FIRST relay under a new session label, PaF prints an informational
 stderr line: `[phone-a-friend] Session label "..." not found in store.

package/commands/phone-a-team.md

@@ -320,9 +320,8 @@ command:
 
   **Generate session IDs for every backend that supports session resume.**
   PaF declares a resume strategy per backend (`native-session` for
-  codex, claude, opencode; `transcript-replay` for ollama; `unsupported`
-  for gemini). Generate a session ID for any backend whose strategy is
-  NOT `unsupported`:
+  codex, claude, gemini, opencode; `transcript-replay` for ollama).
+  Generate a session ID for every backend:
 
   | Backend | resumeStrategy | Generate SESSION_ID? |
   |---|---|---|
@@ -330,11 +329,11 @@ command:
   | claude | native-session | yes |
   | opencode | native-session | yes |
   | ollama | transcript-replay | yes |
-  | gemini | unsupported | NO, omit `--session` entirely |
+  | gemini | native-session | YES, generate a SESSION_ID |
 
-  For `--backend both` (codex + gemini), generate a SESSION_ID for codex
-  only. For `--backend all`, generate SESSION_IDs for codex, claude,
-  opencode, ollama (every backend that runs), but never gemini.
+  For `--backend both` (codex + gemini), generate a SESSION_ID for both
+  codex and gemini. For `--backend all`, generate SESSION_IDs for codex,
+  claude, opencode, ollama, and gemini (every backend that runs).
 
 ### Algorithm
 
@@ -404,9 +403,7 @@ command:
    asked for a diff/branch/staged review.
 
    Include `--session <SESSION_ID>` for every session-capable backend
-   (`codex`, `claude`, `opencode`, `ollama`). Omit `--session` only for
-   `gemini` because PaF rejects it (Gemini's resume strategy is
-   `unsupported`).
+   (`codex`, `claude`, `gemini`, `opencode`, `ollama`).
 
    On the FIRST relay under a new session label, PaF prints an
    informational stderr line: `[phone-a-friend] Session label "..." not
@@ -605,15 +602,12 @@ PAF_TEAM_CONTEXT_EOF
 
   Always include `--fast` (relay prompts are self-contained). For
   `--to claude`, `--fast` has no effect. Include `--session` for every
-  session-capable backend: `codex`, `claude`, `opencode`, `ollama`. Pass
-  the backend-specific ID from `SESSION_IDS`. For `gemini`, omit
-  `--session` entirely; PaF rejects `--session` against Gemini (resume
-  strategy declared `unsupported`).
+  session-capable backend: `codex`, `claude`, `gemini`, `opencode`,
+  `ollama`. Pass the backend-specific ID from `SESSION_IDS`.
 
   When `--session` is used, the session lets the backend remember
   previous rounds, so follow-up prompts can focus on feedback deltas
-  rather than re-sending full context. For Gemini (no session), each
-  round is stateless — see "Per-round relay rules" below.
+  rather than re-sending full context.
 
   **Direct mode** (`RELAY_MODE = direct`):
   ```bash
@@ -774,11 +768,6 @@ The backend already has them in its session history.
 call (prompt size grows per turn). Sessions work but follow-up prompts
 must stay concise to avoid hitting size limits.
 
-**Exception: Gemini (no `--session`)**: Gemini runs stateless every round.
-Each Gemini relay call must include enough context for the backend to
-answer independently — include a brief task recap and the latest output
-alongside the feedback in every round.
-
 **Per-round relay rules (direct mode, no session)**:
 - Each relay call sends ONLY:
   - The original TASK_DESCRIPTION

package/dist/index.js

@@ -6979,6 +6979,9 @@ function looksLikeCodexAlready(output) {
     "already registered"
   ].some((token) => text.includes(token));
 }
+function codexMarketplaceSource(resolvedRepo) {
+  return resolvedRepo.includes("@") ? GITHUB_REPO : resolvedRepo;
+}
 function syncCodexPluginRegistration(source) {
   const lines = [];
   try {
@@ -7248,7 +7251,7 @@ function installHosts(opts) {
     }
   }
   if (shouldInstallCodex && syncCodexCli) {
-    lines.push(...syncCodexPluginRegistration(resolvedRepo));
+    lines.push(...syncCodexPluginRegistration(codexMarketplaceSource(resolvedRepo)));
   }
   return lines;
 }
@@ -77825,15 +77828,16 @@ var GeminiBackend = class {
     "workspace-write",
     "danger-full-access"
   ]);
-  // Session resume is declared 'unsupported' rather than 'transcript-replay':
-  // run() never reads opts.sessionHistory, and the --resume code path below
-  // depends on a session ID that the upstream extractor cannot reliably
-  // produce (see extractGeminiSessionId). Until the Gemini CLI's session
-  // surface is verified, --session against this backend is rejected at the
-  // relay layer instead of silently no-opping.
+  // Session resume mirrors Claude's native-session model: PaF generates the
+  // session UUID client-side (requiresClientSessionId), pins it on the first
+  // call with `--session-id <uuid>`, and resumes later calls with
+  // `--resume <uuid>`. Because the ID is client-generated and deterministic,
+  // PaF never relies on extracting an ID from Gemini's output and never uses
+  // `--resume latest`. History is not replayed (server-side session state),
+  // so opts.sessionHistory is intentionally unused.
   capabilities = {
-    resumeStrategy: "unsupported",
-    requiresClientSessionId: false
+    resumeStrategy: "native-session",
+    requiresClientSessionId: true
   };
   async run(opts) {
     if (!isInPath("gemini")) {
@@ -77902,22 +77906,17 @@ var GeminiBackend = class {
     }
   }
   async runOnce(opts, model) {
-    const args = [];
     const useJsonOutput = Boolean(opts.schema);
     const prompt = opts.schema ? injectSchemaPrompt(opts.prompt, opts.schema) : opts.prompt;
-    if (opts.sandbox !== "danger-full-access") {
-      args.push("--sandbox");
-    }
-    args.push("--yolo");
-    args.push("--include-directories", opts.repoPath);
-    args.push("--output-format", useJsonOutput ? "json" : "text");
-    if (opts.resumeSession && opts.sessionId) {
-      args.push("--resume", opts.sessionId);
-    }
-    if (model) {
-      args.push("-m", model);
-    }
-    args.push("--prompt", prompt);
+    const args = buildGeminiArgs({
+      prompt,
+      repoPath: opts.repoPath,
+      sandbox: opts.sandbox,
+      model,
+      useJsonOutput,
+      sessionId: opts.sessionId ?? null,
+      resumeSession: Boolean(opts.resumeSession)
+    });
     try {
       const result = await spawnCli("gemini", args, {
         timeoutMs: opts.timeoutSeconds * 1e3,
@@ -77945,10 +77944,42 @@ var GeminiBackend = class {
         }
         throw new GeminiBackendError("Gemini reached turn limit, response may be incomplete");
       }
+      if (err instanceof SpawnCliError && opts.sessionId && isUnknownSessionFlagError(err.stderr)) {
+        const flag = opts.resumeSession ? "--resume" : "--session-id";
+        throw new GeminiBackendError(
+          `The installed Gemini CLI does not support the \`${flag}\` flag required for --session resume. Upgrade it (\`${INSTALL_HINTS.gemini}\`), or drop --session to run a one-shot relay.`
+        );
+      }
       throw err;
     }
   }
 };
+function buildGeminiArgs(opts) {
+  const args = [];
+  if (opts.sandbox !== "danger-full-access") {
+    args.push("--sandbox");
+  }
+  args.push("--yolo");
+  args.push("--include-directories", opts.repoPath);
+  args.push("--output-format", opts.useJsonOutput ? "json" : "text");
+  if (opts.sessionId) {
+    if (opts.resumeSession) {
+      args.push("--resume", opts.sessionId);
+    } else {
+      args.push("--session-id", opts.sessionId);
+    }
+  }
+  if (opts.model) {
+    args.push("-m", opts.model);
+  }
+  args.push("--prompt", opts.prompt);
+  return args;
+}
+function isUnknownSessionFlagError(stderr) {
+  const text = stderr.toLowerCase();
+  if (!/unknown argument|unknown option|unrecognized/.test(text)) return false;
+  return text.includes("session-id") || text.includes("resume");
+}
 function formatDeadModelError(model, entry, cachePath) {
   return `Model \`${model}\` returned 404 from Gemini (ModelNotFoundError). Cached as unavailable until ${entry.expiresAt} at ${cachePath}. Run without \`--model\` to use Gemini's auto-routing, set \`PHONE_A_FRIEND_GEMINI_DEAD_CACHE=false\` to bypass the cache, or delete the cache file to clear it.`;
 }
@@ -78071,9 +78102,27 @@ async function* parseNDJSONStream(body, signal) {
     throw new Error("Stream ended unexpectedly");
   }
 }
+function extractOpenCodeErrorMessage(event) {
+  const error3 = event.error;
+  if (error3 && typeof error3 === "object") {
+    const data = error3.data;
+    if (data && typeof data.message === "string" && data.message.trim()) {
+      return data.message;
+    }
+    if (typeof error3.name === "string" && error3.name.trim()) {
+      return error3.name;
+    }
+    try {
+      return JSON.stringify(error3);
+    } catch {
+    }
+  }
+  return "opencode reported an error";
+}
 async function* parseOpenCodeStreamJSON(stdout, opts) {
   let buffer = "";
   let sessionReported = false;
+  let errorReported = false;
   function* processLines(lines) {
     for (const line of lines) {
       const trimmed = line.trim();
@@ -78088,6 +78137,10 @@ async function* parseOpenCodeStreamJSON(stdout, opts) {
         opts.onSessionCreated(parsed.sessionID);
         sessionReported = true;
       }
+      if (!errorReported && parsed.type === "error" && opts?.onError) {
+        opts.onError(extractOpenCodeErrorMessage(parsed));
+        errorReported = true;
+      }
       if (parsed.type === "text") {
         const part = parsed.part;
         if (part?.text && typeof part.text === "string") {
@@ -78381,6 +78434,18 @@ function isClaudeAuthError(msg) {
   const text = msg.toLowerCase();
   return text.includes("not logged in") || text.includes("please run /login");
 }
+function extractClaudeSchemaOutput(stdout) {
+  let parsed;
+  try {
+    parsed = JSON.parse(stdout);
+  } catch {
+    return stdout;
+  }
+  if (typeof parsed === "object" && parsed !== null && !Array.isArray(parsed) && Object.prototype.hasOwnProperty.call(parsed, "structured_output")) {
+    return JSON.stringify(parsed.structured_output);
+  }
+  return stdout;
+}
 var ClaudeBackend = class {
   name = "claude";
   localFileAccess = true;
@@ -78466,7 +78531,7 @@ var ClaudeBackend = class {
         label: "claude"
       });
       if (result.stdout) {
-        return result.stdout;
+        return opts.schema ? extractClaudeSchemaOutput(result.stdout) : result.stdout;
       }
       throw new ClaudeBackendError("claude completed without producing output");
     } catch (err) {
@@ -78716,42 +78781,48 @@ var OpenCodeBackend = class {
     process.on("SIGINT", onSigint);
     const stderrChunks = [];
     child.stderr?.on("data", (chunk) => stderrChunks.push(chunk));
-    const closePromise = new Promise((resolve5, reject) => {
-      child.on("close", (code, signal) => {
-        if (timedOut) {
-          reject(new OpenCodeBackendError(
-            `opencode timed out after ${opts.timeoutSeconds}s`
-          ));
-        } else if (signal) {
-          reject(new OpenCodeBackendError(
-            `opencode killed by signal ${signal}`
-          ));
-        } else if (code !== 0 && code !== null) {
-          const stderr = Buffer.concat(stderrChunks).toString().trim();
-          reject(new OpenCodeBackendError(
-            stderr || `opencode exited with code ${code}`
-          ));
-        } else {
-          resolve5();
-        }
-      });
-    });
+    let streamError = null;
+    const closePromise = new Promise(
+      (resolve5) => {
+        child.on("close", (code, signal) => resolve5({ code, signal }));
+      }
+    );
     let chunkCount = 0;
     try {
       for await (const chunk of parseOpenCodeStreamJSON(
         child.stdout,
-        { onSessionCreated: opts.onSessionCreated }
+        {
+          onSessionCreated: opts.onSessionCreated,
+          onError: (msg) => {
+            if (streamError === null) streamError = msg;
+          }
+        }
       )) {
         chunkCount++;
         yield chunk;
       }
-      await closePromise;
+      const { code, signal } = await closePromise;
+      if (timedOut) {
+        throw new OpenCodeBackendError(
+          `opencode timed out after ${opts.timeoutSeconds}s`
+        );
+      }
+      if (signal) {
+        throw new OpenCodeBackendError(`opencode killed by signal ${signal}`);
+      }
+      if (code !== 0 && code !== null) {
+        const stderr = Buffer.concat(stderrChunks).toString().trim();
+        throw new OpenCodeBackendError(
+          stderr || streamError || `opencode exited with code ${code}`
+        );
+      }
+      if (streamError) {
+        throw new OpenCodeBackendError(streamError);
+      }
       if (chunkCount === 0) {
         throw new OpenCodeBackendError(OPENCODE_NO_OUTPUT_MESSAGE);
       }
     } catch (err) {
-      closePromise.catch(() => {
-      });
       if (err instanceof OpenCodeBackendError) throw err;
       const msg = err instanceof Error ? err.message : String(err);
       if (timedOut) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@freibergergarcia/phone-a-friend",
-  "version": "2.7.1",
+  "version": "2.8.0",
   "description": "CLI relay that lets AI coding agents collaborate",
   "keywords": [
     "ai-agent",

package/plugins/phone-a-friend/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "phone-a-friend",
-  "version": "2.7.1",
+  "version": "2.8.0",
   "description": "CLI relay that lets AI coding agents collaborate by sending prompts and repository context to backend agents.",
   "author": {
     "name": "Bruno Freiberger"

package/plugins/phone-a-friend/skills/phone-a-friend/SKILL.md

@@ -260,8 +260,8 @@ PAF_CONTEXT_EOF
 
    "$RELAY_BIN" --to codex --repo "$PWD" --prompt "$(cat "$PROMPT_FILE")" --context-file "$CONTEXT_FILE" $PAF_NO_DIFF [--fast] [--session <id>]
    # For gemini, omit --model by default (let auto-routing pick); see "Gemini model selection" below.
-   # Do NOT pass --session to gemini — it will error (see "Session continuity" below):
-   "$RELAY_BIN" --to gemini --repo "$PWD" --prompt "$(cat "$PROMPT_FILE")" --context-file "$CONTEXT_FILE" $PAF_NO_DIFF [--fast]
+   # Gemini supports --session via native resume (see "Session continuity" below):
+   "$RELAY_BIN" --to gemini --repo "$PWD" --prompt "$(cat "$PROMPT_FILE")" --context-file "$CONTEXT_FILE" $PAF_NO_DIFF [--fast] [--session <id>]
    ```
 
    Use delimiter names that do not appear in the payload. The quoted heredoc
@@ -389,10 +389,10 @@ can be shorter (no need to re-send context from previous turns).
   can send deltas only.
 - **Ollama**: replays full history each call. Sessions work but prompt
   size grows with each turn. Keep follow-ups concise.
-- **Gemini**: `--session` is **not supported**. PaF rejects it with a
-  RelayError (`--session is not supported by the gemini backend ...`).
-  Each Gemini relay call must be self-contained. Do not pass `--session`
-  with `--to gemini`.
+- **Gemini**: native session resume (same as Codex/Claude/OpenCode).
+  PaF generates the session UUID client-side, pins it with `--session-id`
+  on the first call, and resumes with `--resume` later. Follow-up prompts
+  can send deltas only.
 
 On the FIRST relay under a new session label, PaF prints an informational
 stderr line: `[phone-a-friend] Session label "..." not found in store.

package/plugins/phone-a-friend/skills/phone-a-team/SKILL.md

@@ -65,18 +65,14 @@ For each backend in the comma-separated list, the session label is `${TEAM_ID}-<
 | codex | native-session | yes (never used here; recursion guard) |
 | opencode | native-session | yes |
 | ollama | transcript-replay | yes |
-| gemini | unsupported | **no, omit the flag entirely** |
+| gemini | native-session | yes |
 
 Build the relay flag accordingly per backend:
 
 ```bash
 session_flag_for() {
   local b="$1"
-  if [ "$b" = "gemini" ]; then
-    echo ""
-  else
-    echo "--session ${TEAM_ID}-${b}"
-  fi
+  echo "--session ${TEAM_ID}-${b}"
 }
 ```
 

package/skills/phone-a-friend/SKILL.md

@@ -260,8 +260,8 @@ PAF_CONTEXT_EOF
 
    "$RELAY_BIN" --to codex --repo "$PWD" --prompt "$(cat "$PROMPT_FILE")" --context-file "$CONTEXT_FILE" $PAF_NO_DIFF [--fast] [--session <id>]
    # For gemini, omit --model by default (let auto-routing pick); see "Gemini model selection" below.
-   # Do NOT pass --session to gemini — it will error (see "Session continuity" below):
-   "$RELAY_BIN" --to gemini --repo "$PWD" --prompt "$(cat "$PROMPT_FILE")" --context-file "$CONTEXT_FILE" $PAF_NO_DIFF [--fast]
+   # Gemini supports --session via native resume (see "Session continuity" below):
+   "$RELAY_BIN" --to gemini --repo "$PWD" --prompt "$(cat "$PROMPT_FILE")" --context-file "$CONTEXT_FILE" $PAF_NO_DIFF [--fast] [--session <id>]
    ```
 
    Use delimiter names that do not appear in the payload. The quoted heredoc
@@ -389,10 +389,10 @@ can be shorter (no need to re-send context from previous turns).
   can send deltas only.
 - **Ollama**: replays full history each call. Sessions work but prompt
   size grows with each turn. Keep follow-ups concise.
-- **Gemini**: `--session` is **not supported**. PaF rejects it with a
-  RelayError (`--session is not supported by the gemini backend ...`).
-  Each Gemini relay call must be self-contained. Do not pass `--session`
-  with `--to gemini`.
+- **Gemini**: native session resume (same as Codex/Claude/OpenCode).
+  PaF generates the session UUID client-side, pins it with `--session-id`
+  on the first call, and resumes with `--resume` later. Follow-up prompts
+  can send deltas only.
 
 On the FIRST relay under a new session label, PaF prints an informational
 stderr line: `[phone-a-friend] Session label "..." not found in store.

package/skills/phone-a-team/.codex/SKILL.md

@@ -65,18 +65,14 @@ For each backend in the comma-separated list, the session label is `${TEAM_ID}-<
 | codex | native-session | yes (never used here; recursion guard) |
 | opencode | native-session | yes |
 | ollama | transcript-replay | yes |
-| gemini | unsupported | **no, omit the flag entirely** |
+| gemini | native-session | yes |
 
 Build the relay flag accordingly per backend:
 
 ```bash
 session_flag_for() {
   local b="$1"
-  if [ "$b" = "gemini" ]; then
-    echo ""
-  else
-    echo "--session ${TEAM_ID}-${b}"
-  fi
+  echo "--session ${TEAM_ID}-${b}"
 }
 ```