[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M75MR8X2SWR9PS2SRQ4J152Q":3,"skill-same:01M75MR8X2SWR9PS2SRQ4J152Q":85},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":23,"commitSha":24,"skillDirectory":25,"forksCount":26,"repositoryHtmlUrl":27,"repositoryHomepageUrl":28,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M75MR8X2SWR9PS2SRQ4J152Q","orchestrate-gemini",null,"Safely control Google Gemini CLI on local or remote hosts with private checkouts, immutable goals, least-privilege sandbox policy, durable leases and evidence, restart-safe resume, and system-of-record verification. Use for bounded delegation, jump-box operation, auth or quota routing, tmux recovery, and non-generating validation.","selamy-labs\u002Fagent-skills","selamy-labs","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F287832944?v=4","01M75MDKJ0CKAM94KZ7TWQJM5Q","01M75MDKJNP3TATHAW1MX1FWM6",0,"MIT","https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: orchestrate-gemini\ndescription: Safely control Google Gemini CLI on local or remote hosts with private checkouts, immutable goals, least-privilege sandbox policy, durable leases and evidence, restart-safe resume, and system-of-record verification. Use for bounded delegation, jump-box operation, auth or quota routing, tmux recovery, and non-generating validation.\n---\n\n# Orchestrate Gemini CLI\n\nThis skill controls Google's `gemini` executable. It does not control, wrap, or\nalias the distinct Antigravity CLI (`agy`); use `orchestrate-agy` for that tool.\n\n## Establish the lane\n\n1. Run probes as the intended OS user. In that user's login environment, resolve\n   `command -v gemini`, then capture `gemini --version` and `gemini --help`.\n   Treat the installed executable as the flag contract.\n2. Verify the Gemini auth type and quota owner without printing credential\n   values. An adjacent Antigravity account directory is not evidence that Gemini\n   supports or selected that identity. Stop if the intended identity or funding\n   route cannot be proven.\n3. Use a dedicated clone whose Git common directory is inside the checkout. Do\n   not dispatch in an active product checkout or a linked Git worktree with\n   shared metadata. Pin the exact clean base SHA and give one lane one writer.\n4. Put the objective, already-existing allowed paths, immutable verification argv, stop\n   conditions, attempt budget, auth type, reviewed non-secret\n   `credential_identity`, and paid-generation policy in a goal JSON file. Start\n   from `references\u002Fgoal.example.json`; replace every example value before\n   dispatch. Pin the container image by `image@sha256` digest in the same goal.\n5. Create the lane state directory outside the checkout with mode `0700`. Store\n   the prompt and a task-specific default-deny policy as owner-only files.\n\nUse batched, read-only SSH probes before mutations. Never inspect credential\ncontents, transmit secrets through prompts, or infer authorization from another\nprocess, account, checkout, or supervisor.\n\n## Choose the control surface\n\n| Need | Surface | Acceptance boundary |\n| --- | --- | --- |\n| Login, trust, approvals, or discussion | Interactive tmux | Prove exact marked submission and fresh activity; never count it as contained delivery |\n| Capability\u002Fauth\u002Fsandbox proof only | `--preflight-only` | No model process; terminal `preflight_succeeded` |\n| No-generation installation\u002Fsandbox smoke test | `--validation-fake-responses` | Terminal `validation_succeeded`, never delivery |\n| Bounded delegated work | Headless runner | Verified stream, process harvest, Git scope, and gates |\n\nPrefer `plan` for read-only investigation. Use `default` or `auto_edit` only\nwhen the immutable goal authorizes edits and the private checkout plus policy\ncontain them. The runner rejects model aliases, YOLO, extension loading,\ninherited sandbox mounts, shared Git metadata, a workspace-root `.gemini`\nconfiguration tree, repository attribute files, and every process-execution\ntool allow. Repository attributes are unsupported\nbecause host Git may execute model-controlled clean or process filters.\n\n## Run a preflight\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Frun_headless.py \\\n  --run-dir \u002Fabsolute\u002Flane-state\u002Fruns\u002Fattempt-001 \\\n  --state-dir \u002Fabsolute\u002Flane-state \\\n  --cwd \u002Fabsolute\u002Fprivate-clone \\\n  --goal-file \u002Fabsolute\u002Fgoal.json \\\n  --prompt-file \u002Fabsolute\u002Ftask.prompt \\\n  --policy-file \u002Fabsolute\u002Fpolicy.toml \\\n  --gemini \u002Fabsolute\u002Fpath\u002Fto\u002Fgemini \\\n  --sandbox-provider docker \\\n  --timeout-seconds 120 \\\n  --model gemini-2.5-pro \\\n  --approval-mode plan \\\n  --preflight-only\n```\n\nPreflight acquires a non-blocking `flock` lease, persists the canonical goal and\ndigest, proves the exact clean Git base, validates the default-deny policy,\ncreates an isolated Gemini home and temp directory, proves the selected\ncontainer service and digest-pinned image are available, trusts only the\nverified checkout, isolates `HOME`, empties startup discovery\u002Fcall commands and\nMCP configuration, disables extensions, skills, hooks, and credit overage,\ninstalls fixed container hardening, and requires exactly Gemini CLI 0.51.0. It\nnever starts a generated turn.\n\nThe policy passed with `--admin-policy` is supplemental in Gemini CLI 0.51.0.\nThe runner stops if a standard system admin-policy directory would supersede\nit. Review `references\u002Fpolicy.example.toml` and add only the narrow tool names\nthe goal requires at a priority above the catch-all deny.\n\n## Run headlessly\n\nRemove `--preflight-only` only after its evidence passes. The runner stages the\nowner-only prompt beneath the private checkout's Git metadata and puts only an\n`@\u003Cpath>` reference in argv. This matters because Gemini 0.51.0 transforms\nnon-TTY stdin into an inner Docker `--prompt` argument. The staged file is\nremoved after worker harvest; an owner-only evidence copy remains.\n\nThe worker and capability probes run in fresh process groups under one absolute\noperation deadline plus a fixed short cleanup grace. Verification commands run\ninside the same digest-pinned image with no network, a read-only checkout, an\nisolated home\u002Ftemp, and the same process\u002Fcontainer harvesting. The supervisor\ntracks descendants by PID plus process start time,\nincluding a child that calls `setsid()`, and escalates TERM to KILL. Known or\nuncertain survivors prevent success. Output must contain exactly one `init`, one successful\n`result`, a non-empty assistant message, no fatal error, the exact requested\nmodel, and a session ID.\n\nThe provider guard labels every Gemini and verifier container, removes the\nupstream host-gateway mapping, binds proxy readiness to loopback, makes the\ncheckout root and Git metadata read-only, and remounts only declared paths\nwritable. Before and after execution it rejects symlinks, multiply-linked\nregular files, and special files anywhere in those writable trees so a declared\npath cannot bridge into another checkout or host resource. It also requires\nGemini's fixed worker network to be internal,\nrequires the fixed proxy network to be external, and rejects every unexpected\nnetwork operation or attachment. After harvest, the runner checks sanitized Git state and a full\nfilesystem snapshot, including ignored paths. Any Git metadata change fails.\nExit zero means verified delivery, not merely that Gemini stopped talking.\n\nGoogle stopped serving Gemini CLI requests for free, Google AI Pro, and Google\nAI Ultra individual accounts on June 18, 2026. This v0.51 workflow therefore\nrejects consumer OAuth and supports live execution only with a paid Gemini API\nkey. Live generation requires `allow_paid_generation: true` plus\n`ORCHESTRATE_GEMINI_PAID_GENERATION_ACK=authorized`. Supply the key only in an\nowner-only file containing `GEMINI_API_KEY=\u003Cvalue>` via\n`--credential-env-file`; never export it. The outer CLI receives a non-secret\nplaceholder, and the provider guard replaces that placeholder with a runtime\n`--env-file` path. The ephemeral copy is removed on every terminal path after\ncontainer reconciliation. Any real key in runtime argv is rejected. Record a\nreviewed, non-secret quota-owner or billing-project identifier as\n`credential_identity`; do not put a key value or a secret-derived hash there.\nThe policy must not allow `run_shell_command` or its legacy alias: a same-user\nchild process could read the credential from the Gemini parent's process\nenvironment and issue unaccounted paid requests.\n\nLive model transport crosses an internal container network through the bundled\nCONNECT proxy, which permits only `generativelanguage.googleapis.com:443`.\nThe proxy checkout mount is read-only and its readiness port binds only to\nloopback. Validation mode instead forces `--network none`. The wrapper\nserializes v0.51 executions because upstream uses a global proxy container name.\n\n## Validate without generation\n\nTo exercise the real executable and container launch without contacting a\nmodel, use a reviewed Gemini `--fake-responses` fixture:\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Frun_headless.py \\\n  ...same reviewed arguments... \\\n  --validation-fake-responses \u002Fabsolute\u002Freviewed.responses\n```\n\nThe runner retains the fixture and its digest, marks `validation_mode: true`,\nforces container network `none`, and can only finish as\n`validation_succeeded`. Never present this classification as delegated\nwork. Do not run a live smoke turn merely to strengthen validation when the\nquota route is unknown or paid generation is forbidden.\n\n## Run interactively\n\nFor login, trust, or approval recovery, launch Gemini in a named tmux lane. The\nhelper stages the marked directive as an owner-only file and passes only an\n`@\u003Cpath>` reference through Gemini's `--prompt-interactive` interface:\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Flaunch_tmux.sh \\\n  lane-repo-gemini-task-YYYYMMDD \\\n  \u002Fabsolute\u002Fprivate-clone \\\n  \u002Fabsolute\u002Ftask.prompt \\\n  \"$(command -v gemini)\" \\\n  --model MODEL --sandbox\n```\n\nThe launcher rejects an existing session and prints a unique dispatch ID plus\npre-capture path. After resolving setup screens under the intended identity,\nprove the exact dispatch:\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Fverify_dispatch.sh \\\n  \"$SESSION\" \"$DISPATCH_ID\" \"$PRE_CAPTURE\"\n```\n\nFor a file-backed follow-up, run\n`scripts\u002Fsubmit_followup.sh \"$SESSION\" \"$FOLLOWUP_FILE\"`. Count work as\nsubmitted only when start\u002Fend markers appear in order, Gemini emits the exact\npost-marker acknowledgement, fresh tool or thinking activity follows, and a\nclean composer returns. Session existence, stale `Working` text, or a marker\nstill sitting in the composer is not proof of dispatch.\n\n## Resume and recover\n\nEach attempt uses a new direct child of `STATE_DIR\u002Fruns` and reuses the lane's\nimmutable goal. Before dispatch, the runner reconciles a stale worker using its\nPID, process group, and start-time identity and removes daemon-owned containers\ncarrying the lane\u002FUID labels. The kernel lane lock and canonical per-UID v0.51\nglobal execution lock are inherited by the worker at process creation, closing\nthe launch-before-identity crash window and preventing replacement dispatch\nuntil the worker exits. Resume only\nby naming a terminal prior run:\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Frun_headless.py \\\n  ...same immutable lane arguments... \\\n  --resume-from \u002Fabsolute\u002Flane-state\u002Fruns\u002Fattempt-001\n```\n\nThe runner accepts only a prior `succeeded` delivery with the same immutable\nexecution-request digest—including the prompt digest and reviewed credential\nidentity—and checkout identity, then forwards that run's exact session ID with\n`--resume`. Validation and failed runs are not resumable. It\nnever uses `latest`. Do not\nresume after scope expansion, goal change, ambiguous process cleanup, identity\nchange, or attempt-budget exhaustion; create a newly reviewed goal instead.\n\n## Accept the result from systems of record\n\nInspect `status.json`, `command.json`, stream output, process harvest state,\nchanged-path evidence, verification outputs, and the exact Git tree. Require\n`classification: succeeded`, matching requested\u002Freported model plus exclusive\nrequested-model result statistics, no survivor,\nbase ancestry, in-scope paths, and passing immutable commands. For a PR, rerun\nthe repository gates at its exact head, inspect hosted CI and review state, then\nverify the merged commit from canonical `main`. A model narrative, an idle\npane, a locally green test from another SHA, or `validation_succeeded` is not\ndelivery.\n\n## Primary sources\n\n- [Gemini CLI 0.51.0 release](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Freleases\u002Ftag\u002Fv0.51.0)\n- [Headless mode](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Fcli\u002Fheadless.md)\n- [CLI reference](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Fcli\u002Fcli-reference.md)\n- [Sandboxing](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Fcli\u002Fsandbox.md)\n- [Policy engine](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Freference\u002Fpolicy-engine.md)\n- [Trusted folders](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Fcli\u002Ftrusted-folders.md)\n- [Individual-account transition](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fdiscussions\u002F28017)\n","# Orchestrate Gemini CLI\n\nThis skill controls Google's `gemini` executable. It does not control, wrap, or\nalias the distinct Antigravity CLI (`agy`); use `orchestrate-agy` for that tool.\n\n## Establish the lane\n\n1. Run probes as the intended OS user. In that user's login environment, resolve\n   `command -v gemini`, then capture `gemini --version` and `gemini --help`.\n   Treat the installed executable as the flag contract.\n2. Verify the Gemini auth type and quota owner without printing credential\n   values. An adjacent Antigravity account directory is not evidence that Gemini\n   supports or selected that identity. Stop if the intended identity or funding\n   route cannot be proven.\n3. Use a dedicated clone whose Git common directory is inside the checkout. Do\n   not dispatch in an active product checkout or a linked Git worktree with\n   shared metadata. Pin the exact clean base SHA and give one lane one writer.\n4. Put the objective, already-existing allowed paths, immutable verification argv, stop\n   conditions, attempt budget, auth type, reviewed non-secret\n   `credential_identity`, and paid-generation policy in a goal JSON file. Start\n   from `references\u002Fgoal.example.json`; replace every example value before\n   dispatch. Pin the container image by `image@sha256` digest in the same goal.\n5. Create the lane state directory outside the checkout with mode `0700`. Store\n   the prompt and a task-specific default-deny policy as owner-only files.\n\nUse batched, read-only SSH probes before mutations. Never inspect credential\ncontents, transmit secrets through prompts, or infer authorization from another\nprocess, account, checkout, or supervisor.\n\n## Choose the control surface\n\n| Need | Surface | Acceptance boundary |\n| --- | --- | --- |\n| Login, trust, approvals, or discussion | Interactive tmux | Prove exact marked submission and fresh activity; never count it as contained delivery |\n| Capability\u002Fauth\u002Fsandbox proof only | `--preflight-only` | No model process; terminal `preflight_succeeded` |\n| No-generation installation\u002Fsandbox smoke test | `--validation-fake-responses` | Terminal `validation_succeeded`, never delivery |\n| Bounded delegated work | Headless runner | Verified stream, process harvest, Git scope, and gates |\n\nPrefer `plan` for read-only investigation. Use `default` or `auto_edit` only\nwhen the immutable goal authorizes edits and the private checkout plus policy\ncontain them. The runner rejects model aliases, YOLO, extension loading,\ninherited sandbox mounts, shared Git metadata, a workspace-root `.gemini`\nconfiguration tree, repository attribute files, and every process-execution\ntool allow. Repository attributes are unsupported\nbecause host Git may execute model-controlled clean or process filters.\n\n## Run a preflight\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Frun_headless.py \\\n  --run-dir \u002Fabsolute\u002Flane-state\u002Fruns\u002Fattempt-001 \\\n  --state-dir \u002Fabsolute\u002Flane-state \\\n  --cwd \u002Fabsolute\u002Fprivate-clone \\\n  --goal-file \u002Fabsolute\u002Fgoal.json \\\n  --prompt-file \u002Fabsolute\u002Ftask.prompt \\\n  --policy-file \u002Fabsolute\u002Fpolicy.toml \\\n  --gemini \u002Fabsolute\u002Fpath\u002Fto\u002Fgemini \\\n  --sandbox-provider docker \\\n  --timeout-seconds 120 \\\n  --model gemini-2.5-pro \\\n  --approval-mode plan \\\n  --preflight-only\n```\n\nPreflight acquires a non-blocking `flock` lease, persists the canonical goal and\ndigest, proves the exact clean Git base, validates the default-deny policy,\ncreates an isolated Gemini home and temp directory, proves the selected\ncontainer service and digest-pinned image are available, trusts only the\nverified checkout, isolates `HOME`, empties startup discovery\u002Fcall commands and\nMCP configuration, disables extensions, skills, hooks, and credit overage,\ninstalls fixed container hardening, and requires exactly Gemini CLI 0.51.0. It\nnever starts a generated turn.\n\nThe policy passed with `--admin-policy` is supplemental in Gemini CLI 0.51.0.\nThe runner stops if a standard system admin-policy directory would supersede\nit. Review `references\u002Fpolicy.example.toml` and add only the narrow tool names\nthe goal requires at a priority above the catch-all deny.\n\n## Run headlessly\n\nRemove `--preflight-only` only after its evidence passes. The runner stages the\nowner-only prompt beneath the private checkout's Git metadata and puts only an\n`@\u003Cpath>` reference in argv. This matters because Gemini 0.51.0 transforms\nnon-TTY stdin into an inner Docker `--prompt` argument. The staged file is\nremoved after worker harvest; an owner-only evidence copy remains.\n\nThe worker and capability probes run in fresh process groups under one absolute\noperation deadline plus a fixed short cleanup grace. Verification commands run\ninside the same digest-pinned image with no network, a read-only checkout, an\nisolated home\u002Ftemp, and the same process\u002Fcontainer harvesting. The supervisor\ntracks descendants by PID plus process start time,\nincluding a child that calls `setsid()`, and escalates TERM to KILL. Known or\nuncertain survivors prevent success. Output must contain exactly one `init`, one successful\n`result`, a non-empty assistant message, no fatal error, the exact requested\nmodel, and a session ID.\n\nThe provider guard labels every Gemini and verifier container, removes the\nupstream host-gateway mapping, binds proxy readiness to loopback, makes the\ncheckout root and Git metadata read-only, and remounts only declared paths\nwritable. Before and after execution it rejects symlinks, multiply-linked\nregular files, and special files anywhere in those writable trees so a declared\npath cannot bridge into another checkout or host resource. It also requires\nGemini's fixed worker network to be internal,\nrequires the fixed proxy network to be external, and rejects every unexpected\nnetwork operation or attachment. After harvest, the runner checks sanitized Git state and a full\nfilesystem snapshot, including ignored paths. Any Git metadata change fails.\nExit zero means verified delivery, not merely that Gemini stopped talking.\n\nGoogle stopped serving Gemini CLI requests for free, Google AI Pro, and Google\nAI Ultra individual accounts on June 18, 2026. This v0.51 workflow therefore\nrejects consumer OAuth and supports live execution only with a paid Gemini API\nkey. Live generation requires `allow_paid_generation: true` plus\n`ORCHESTRATE_GEMINI_PAID_GENERATION_ACK=authorized`. Supply the key only in an\nowner-only file containing `GEMINI_API_KEY=\u003Cvalue>` via\n`--credential-env-file`; never export it. The outer CLI receives a non-secret\nplaceholder, and the provider guard replaces that placeholder with a runtime\n`--env-file` path. The ephemeral copy is removed on every terminal path after\ncontainer reconciliation. Any real key in runtime argv is rejected. Record a\nreviewed, non-secret quota-owner or billing-project identifier as\n`credential_identity`; do not put a key value or a secret-derived hash there.\nThe policy must not allow `run_shell_command` or its legacy alias: a same-user\nchild process could read the credential from the Gemini parent's process\nenvironment and issue unaccounted paid requests.\n\nLive model transport crosses an internal container network through the bundled\nCONNECT proxy, which permits only `generativelanguage.googleapis.com:443`.\nThe proxy checkout mount is read-only and its readiness port binds only to\nloopback. Validation mode instead forces `--network none`. The wrapper\nserializes v0.51 executions because upstream uses a global proxy container name.\n\n## Validate without generation\n\nTo exercise the real executable and container launch without contacting a\nmodel, use a reviewed Gemini `--fake-responses` fixture:\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Frun_headless.py \\\n  ...same reviewed arguments... \\\n  --validation-fake-responses \u002Fabsolute\u002Freviewed.responses\n```\n\nThe runner retains the fixture and its digest, marks `validation_mode: true`,\nforces container network `none`, and can only finish as\n`validation_succeeded`. Never present this classification as delegated\nwork. Do not run a live smoke turn merely to strengthen validation when the\nquota route is unknown or paid generation is forbidden.\n\n## Run interactively\n\nFor login, trust, or approval recovery, launch Gemini in a named tmux lane. The\nhelper stages the marked directive as an owner-only file and passes only an\n`@\u003Cpath>` reference through Gemini's `--prompt-interactive` interface:\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Flaunch_tmux.sh \\\n  lane-repo-gemini-task-YYYYMMDD \\\n  \u002Fabsolute\u002Fprivate-clone \\\n  \u002Fabsolute\u002Ftask.prompt \\\n  \"$(command -v gemini)\" \\\n  --model MODEL --sandbox\n```\n\nThe launcher rejects an existing session and prints a unique dispatch ID plus\npre-capture path. After resolving setup screens under the intended identity,\nprove the exact dispatch:\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Fverify_dispatch.sh \\\n  \"$SESSION\" \"$DISPATCH_ID\" \"$PRE_CAPTURE\"\n```\n\nFor a file-backed follow-up, run\n`scripts\u002Fsubmit_followup.sh \"$SESSION\" \"$FOLLOWUP_FILE\"`. Count work as\nsubmitted only when start\u002Fend markers appear in order, Gemini emits the exact\npost-marker acknowledgement, fresh tool or thinking activity follows, and a\nclean composer returns. Session existence, stale `Working` text, or a marker\nstill sitting in the composer is not proof of dispatch.\n\n## Resume and recover\n\nEach attempt uses a new direct child of `STATE_DIR\u002Fruns` and reuses the lane's\nimmutable goal. Before dispatch, the runner reconciles a stale worker using its\nPID, process group, and start-time identity and removes daemon-owned containers\ncarrying the lane\u002FUID labels. The kernel lane lock and canonical per-UID v0.51\nglobal execution lock are inherited by the worker at process creation, closing\nthe launch-before-identity crash window and preventing replacement dispatch\nuntil the worker exits. Resume only\nby naming a terminal prior run:\n\n```bash\nskills\u002Forchestrate-gemini\u002Fscripts\u002Frun_headless.py \\\n  ...same immutable lane arguments... \\\n  --resume-from \u002Fabsolute\u002Flane-state\u002Fruns\u002Fattempt-001\n```\n\nThe runner accepts only a prior `succeeded` delivery with the same immutable\nexecution-request digest—including the prompt digest and reviewed credential\nidentity—and checkout identity, then forwards that run's exact session ID with\n`--resume`. Validation and failed runs are not resumable. It\nnever uses `latest`. Do not\nresume after scope expansion, goal change, ambiguous process cleanup, identity\nchange, or attempt-budget exhaustion; create a newly reviewed goal instead.\n\n## Accept the result from systems of record\n\nInspect `status.json`, `command.json`, stream output, process harvest state,\nchanged-path evidence, verification outputs, and the exact Git tree. Require\n`classification: succeeded`, matching requested\u002Freported model plus exclusive\nrequested-model result statistics, no survivor,\nbase ancestry, in-scope paths, and passing immutable commands. For a PR, rerun\nthe repository gates at its exact head, inspect hosted CI and review state, then\nverify the merged commit from canonical `main`. A model narrative, an idle\npane, a locally green test from another SHA, or `validation_succeeded` is not\ndelivery.\n\n## Primary sources\n\n- [Gemini CLI 0.51.0 release](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Freleases\u002Ftag\u002Fv0.51.0)\n- [Headless mode](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Fcli\u002Fheadless.md)\n- [CLI reference](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Fcli\u002Fcli-reference.md)\n- [Sandboxing](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Fcli\u002Fsandbox.md)\n- [Policy engine](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Freference\u002Fpolicy-engine.md)\n- [Trusted folders](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fblob\u002Fv0.51.0\u002Fdocs\u002Fcli\u002Ftrusted-folders.md)\n- [Individual-account transition](https:\u002F\u002Fgithub.com\u002Fgoogle-gemini\u002Fgemini-cli\u002Fdiscussions\u002F28017)\n",{"name":5,"description":7},"main","54007b08157feb09e5bb9d7ff5ce4ddd21625dbf","skills\u002Forchestrate-gemini",3,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills","https:\u002F\u002Fselamy.dev","https:\u002F\u002Fgithub.com\u002Fselamy-labs","2026-11-15 03:50:58",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add selamy-labs\u002Fagent-skills --skill orchestrate-gemini",{"tool":37,"command":38},"bunx","bunx skills add selamy-labs\u002Fagent-skills --skill orchestrate-gemini",{"tool":40,"command":41},"pnpm","pnpm dlx skills add selamy-labs\u002Fagent-skills --skill orchestrate-gemini","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002FSKILL.md",true,[45,51,55,59,65,69,73,77,81],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"agents\u002Fopenai.yaml","REFERENCE","参考文档",211,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002Fagents\u002Fopenai.yaml",{"path":52,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":53,"sourceHtmlUrl":54},"references\u002Fgoal.example.json",849,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002Freferences\u002Fgoal.example.json",{"path":56,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":57,"sourceHtmlUrl":58},"references\u002Fpolicy.example.toml",332,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002Freferences\u002Fpolicy.example.toml",{"path":60,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":63,"sourceHtmlUrl":64},"scripts\u002Fapi_egress_proxy.mjs","SCRIPT","脚本",1523,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002Fscripts\u002Fapi_egress_proxy.mjs",{"path":66,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":67,"sourceHtmlUrl":68},"scripts\u002Flaunch_tmux.sh",2486,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002Fscripts\u002Flaunch_tmux.sh",{"path":70,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":71,"sourceHtmlUrl":72},"scripts\u002Fprovider_guard.py",9496,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002Fscripts\u002Fprovider_guard.py",{"path":74,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":75,"sourceHtmlUrl":76},"scripts\u002Frun_headless.py",102571,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002Fscripts\u002Frun_headless.py",{"path":78,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":79,"sourceHtmlUrl":80},"scripts\u002Fsubmit_followup.sh",3856,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002Fscripts\u002Fsubmit_followup.sh",{"path":82,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":83,"sourceHtmlUrl":84},"scripts\u002Fverify_dispatch.sh",3631,"https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Fskills\u002Forchestrate-gemini\u002Fscripts\u002Fverify_dispatch.sh",[86,93,100,107,113,120],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":18},"01M75MY70JMSNFHQX63G90SKZ7","silence-phrase","A seemingly normal skill.","https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fsilence-phrase\u002FSKILL.md",[],[],{"skillNo":94,"name":95,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":97,"categories":98,"keywords":99,"publishedAt":18},"01M75MY5PKVKY60M8H2DFTW58C","selamy-bad-skill","This skill has a vendor prefix that should be rejected.","https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fprefixed-name\u002FSKILL.md",[],[],{"skillNo":101,"name":102,"nameZh":6,"description":103,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":104,"categories":105,"keywords":106,"publishedAt":18},"01M75MY1TKXCH6WYN9ET0W4T02","oversized-description","This is a very long description that contains way too many words and should be rejected by the validator because it exceeds the forty five word limit that is enforced by the naming lint checker which ensures skill descriptions stay concise and readable for humans and machines alike without unnecessary verbosity or padding","https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Foversized-description\u002FSKILL.md",[],[],{"skillNo":108,"name":109,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":110,"categories":111,"keywords":112,"publishedAt":18},"01M75MY0SZ7EKT99NWJEJ8HKS2","missing-frontmatter","https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fmissing-frontmatter\u002FSKILL.md",[],[],{"skillNo":114,"name":115,"nameZh":6,"description":116,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":117,"categories":118,"keywords":119,"publishedAt":18},"01M75MXYMBREE6D7KRR87YM4DH","malicious-skill","A poisoned skill used only as a negative fixture for the security gate.","https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fmalicious-skill\u002FSKILL.md",[],[],{"skillNo":121,"name":122,"nameZh":6,"description":123,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":124,"categories":125,"keywords":126,"publishedAt":18},"01M75MXX2B7AC2QSSRRQQXTA0N","leaking-org-name","A skill about something.","https:\u002F\u002Fgithub.com\u002Fselamy-labs\u002Fagent-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fleaking-org-name\u002FSKILL.md",[],[]]