[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AACP85PSKWWYB2PVRR92ZK":3,"skill-same:01M7AACP85PSKWWYB2PVRR92ZK":47},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":31,"updatedAt":32,"install":33,"resources":46},"01M7AACP85PSKWWYB2PVRR92ZK","add-harness",null,"Add a new agent harness (opencode, Gemini CLI, Codex, Amp, a local Ollama-backed runner, or any CLI that loads skills) to a shakedown config so its skill conformance can be measured, and diagnose a harness that `shakedown doctor` reports as not qualifying. Use this whenever someone wants shakedown to run against a harness it does not yet support, asks why a harness fails doctor, asks how to point an existing harness at a different model or provider, or is editing the `[harness.*]` block of a shakedown.toml. Use it even when they only name the tool (\"can this test opencode?\") without mentioning config or TOML.","JasperHG90\u002Fshakedown","JasperHG90","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F8079803?v=4","01M7AAAKN62DYZ9XRCXAN0HGEF","Jasper Ginn","01M7AAAKP8D78B8ANVKEV7MRK8",1,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FJasperHG90\u002Fshakedown\u002Fblob\u002Fmain\u002Fplugins\u002Fclaude-code\u002Fskills\u002Fadd-harness\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: add-harness\ndescription: >-\n  Add a new agent harness (opencode, Gemini CLI, Codex, Amp, a local\n  Ollama-backed runner, or any CLI that loads skills) to a shakedown config\n  so its skill conformance can be measured, and diagnose a harness that\n  `shakedown doctor` reports as not qualifying. Use this whenever someone\n  wants shakedown to run against a harness it does not yet support, asks why\n  a harness fails doctor, asks how to point an existing harness at a\n  different model or provider, or is editing the `[harness.*]` block of a\n  shakedown.toml. Use it even when they only name the tool (\"can this test\n  opencode?\") without mentioning config or TOML.\n---\n\n# Adding a harness to shakedown\n\nA harness is described entirely by config. No code is written to support\none, so this task is: find the flags, write the block, run `doctor`, and fix\nwhatever it names.\n\nThe reason `doctor` exists is that a harness can look configured and still\nbe measuring nothing. Every prerequisite below is verified by watching what\nthe agent actually did, never by reading a help string or an inventory\ncommand.\n\n## Step 0: Confirm the binary is here\n\nEverything below is read off the CLI in front of you, so start by\nestablishing there is one:\n\n```bash\ncommand -v \u003Ccli> && \u003Ccli> --version\n```\n\nIf it is not installed, stop and say so. shakedown runs harnesses; it does\nnot install them, and the block you would write without the binary is\nguesswork. Flags drift between versions, an agent's memory of a CLI is\nfrequently a version or two stale, and a plausible wrong flag produces a\nconfig that fails at `doctor` in a way that looks like the harness's fault\nrather than the block's. Ask the operator to install it, or work on a\nharness that is present.\n\nThe version matters beyond mere presence: record it, because it is part of\nwhat any measurement made with this block will mean.\n\n## Step 1: Check the binary can do the five things\n\nA harness qualifies if it can do these. Four are required.\n\n| # | prerequisite | required | if missing |\n|---|---|---|---|\n| 1 | run headless from a prompt and exit on its own | yes | unusable |\n| 2 | load a skill from a directory you control | yes | unusable |\n| 3 | emit machine-readable output with tool calls and text | yes | unusable, unless it can export the transcript afterwards (Step 3) |\n| 4 | continue a session | no | `inputs_resolved` reports `unsupported` |\n| 5 | run without a TTY | yes | unusable |\n| 6 | what else the model could see | reported only | nothing |\n\nRead the binary's `--help` and find the flag for each. What you are looking\nfor, in the usual naming:\n\n- A non-interactive or print flag (`-p`, `--prompt`, `run`).\n- An output format flag naming JSON or stream-json. Newline-delimited JSON\n  is what shakedown parses.\n- A session identifier flag, and a resume flag that takes it.\n- A model flag, so the same harness can be measured against several models.\n- A permission or approval flag, discussed in Step 4.\n\nPrerequisite 4 being optional is deliberate. A harness must never *fail* a\ndimension it cannot physically support, so a harness with no resume is\nreported as unsupported with a reason, and its other checks stay comparable\nagainst everything else.\n\n## Step 2: Write the block\n\n```toml\n[harness.my-harness]\nstart = [\n  \"my-agent\", \"run\", \"{prompt}\",\n  \"--json\",\n  \"--model\", \"{model}\",\n  \"--session\", \"{sid}\",\n]\nresume = [\n  \"my-agent\", \"run\", \"{reply}\",\n  \"--json\",\n  \"--model\", \"{model}\",\n  \"--resume\", \"{sid}\",\n]\nskills = \".my-agent\u002Fskills\"\nactivation_tool = \"Skill\"\n\n# Exactly one of these, and only for `--sandbox container`:\nimage      = \"ghcr.io\u002Fyou\u002Fmy-agent:1.2.3\"\n# dockerfile = \"docker\u002Fmy-agent.Dockerfile\"\n# context    = \".\"   # with `dockerfile`, to COPY from the repo root\n\n[harness.my-harness.env]\nMY_AGENT_TOKEN = \"${MY_AGENT_TOKEN}\"\n\n[[matrix]]\nharness = \"my-harness\"\nmodels  = [\"some-model\"]\n```\n\n**Commands are lists, not strings.** Each token is one argument, so a prompt\ncontaining a quote or a leading dash stays a prompt and can never become a\nflag. Substitution slots are `{prompt}`, `{reply}`, `{model}`, and `{sid}`.\n\n**`skills`** is the directory this harness scans for skills, relative to the\nworkspace. The skill under test is copied to `\u003Cskills>\u002F\u003Cname>\u002F`. Use the\nharness's own discovery path, the one it would find a skill in with no flags\nat all.\n\n**`activation_tool`** is the substring identifying a skill-activation call\nin this harness's output, `Skill` on Claude Code and `activate_skill` on\nGemini. It is what separates \"the skill never fired\" from \"it fired and\nfailed\". Without it a run where the model ignored the skill and improvised\nscores as an ordinary failure, and the number then describes the base model\nrather than the skill.\n\n**`env` is declared-only.** The sandbox gets these variables and nothing\nelse, which is what makes host contamination structurally impossible rather\nthan something a flag suppresses. Values come from the host through `${VAR}`\nso the TOML holds references, never secrets.\n\n**`image` or `dockerfile`** is needed only for `--sandbox container`, and\ndeclaring both is refused. Use `image` when you already build and push one,\nand `dockerfile` when you would rather keep the environment in the repo.\nEither way it holds the harness plus whatever the skills need at runtime,\nbecause a skill whose `bin\u002F` cannot execute fails checks for reasons that\nhave nothing to do with the skill. See `examples\u002Fdocker\u002F` for one per\nshipped harness.\n\nThe path is relative to `shakedown.toml`, and it is built once per run\nrather than per scenario.\n\n**`context`** is the directory a `dockerfile` build can `COPY` from,\ndefaulting to the dockerfile's own. Widen it with `context = \".\"` when the\nimage needs something this repo builds rather than something it installs\nfrom a registry — your own CLI, say. Without it the `COPY` is outside the\ncontext and Docker refuses it.\n\n## Step 3: Tell shakedown where the tool calls are\n\nHarnesses put tool calls at different depths. This is the one shape\nquestion, and it is config rather than code.\n\nRun the harness once by hand with its JSON flag and look at a line\ncontaining a tool call. Then choose:\n\n- **Nested**: the call sits inside a list under a dotted path, as Claude\n  Code's does in `message.content[]`. Set `container = \"message.content\"`.\n- **Flat**: the call is a top-level record, as Gemini CLI's is. Omit\n  `container` entirely.\n\nSince this is a genuine fork with a short list of answers, present it as a\ndecision rather than a prose question, and show the harness's own output as\nthe preview so the choice is made against evidence.\n\nThe remaining keys are renames, and the defaults match Claude Code:\n\n```toml\n[harness.my-harness.events]\ncontainer     = \"\"            # dotted path to a list; omit if flat\ndiscriminator = \"type\"        # the key naming a block's kind\ntool_marker   = \"tool_use\"    # that key's value on a tool call\nname_key      = \"name\"\nargs_key      = \"input\"\ntext_key      = \"text\"\ntext_marker   = \"text\"\n```\n\nThe four that name a key — `discriminator`, `name_key`, `args_key`,\n`text_key` — may each be a dotted path, which covers a flat record that\nstill buries what you need. opencode's tool call is top-level but carries\nits name and arguments under a `part` object, so it reads `name_key =\n\"part.tool\"` and `args_key = \"part.state.input\"`. (`tool_marker` and\n`text_marker` are values compared against, not keys, so a dot in one does\nnothing.)\n\nA harness matching neither shape fails at `doctor` step 3, loudly, instead\nof silently scoring zero everywhere.\n\n### When the harness prints prose and nothing else\n\nSome agents stream no events at all. Hermes prints only its final answer,\nso there is nothing to parse while it works. Before writing it off, check\nwhether it can hand over the transcript afterwards: `hermes sessions\nexport --format trace` writes Claude Code JSONL, and a start command may\nbe two commands.\n\n```toml\nstart = [\n  \"sh\", \"-c\",\n  \"hermes -z \\\"$1\\\" -m \\\"$2\\\" >&2 && hermes sessions export --format trace --no-redact -\",\n  \"hermes\", \"{prompt}\", \"{model}\",\n]\n```\n\nThe prompt stays a positional argument (`$1`) rather than being pasted\ninto the script, so it is still exactly one argument and still cannot\nbecome a flag. Check what the export does to the transcript on the way\nout, too: Hermes redacts tool arguments unless told not to, and those\narguments are what `tool_used` reads. Send the harness's own prose to\nstderr rather than `\u002Fdev\u002Fnull`: a run that fails for its own reasons then\nsays why, instead of scoring as a skill that never fired.\n\n## Step 4: Let the harness run the tool\n\nMost harnesses refuse tool calls by default, and a refusal still appears in\nthe transcript as a call. A config that omits the permission flag therefore\nproduces runs where the CLI was requested, denied, and never executed.\n\nshakedown reads the harness's denial records and fails `tool_used` with \"was\nrequested but denied\", so this shows up as a diagnosis rather than a silent\nwrong number. Fix it in `start` and `resume` by allowing the tools the skill\nneeds.\n\nWatch for a permission mode that covers file writes but not shell commands.\nClaude Code's `acceptEdits` allows `Write` while still prompting for `Bash`,\nso a skill that shells out to its CLI is denied under it.\n\n## Step 5: Run doctor until it qualifies\n\n```bash\nshakedown doctor --harness my-harness\n```\n\n`doctor` runs a canary skill that asks one question, then runs\n`echo shakedown-ok`. Seeing that shell call is only possible if the harness\nran headless, discovered the skill, surfaced it to the model, followed it,\nand emitted parseable output, so one cheap task settles prerequisites 1, 2,\n3 and 5 at once. The question is there to force a second turn, which\nsettles 4.\n\n```\n             my-harness\n  # prerequisite                    detail\n  1 headless run                 ok exit 0\n  2 skill surfaced at runtime  FAIL never activated\n  3 output parsed                ok 1 tool calls, 3 texts\n  4 session resume               ok 2 turns\n  5 no TTY required              ok ran without a terminal\n  6 environment visibility       ok 16 other skills visible\n\ndoes not qualify: blocked on skill surfaced at runtime\n```\n\nWork the failing number:\n\n- **1 fails**: the command needs a headless flag, or it is waiting on input.\n  A hang shows as a timeout rather than a non-zero exit.\n- **2 fails with \"never activated\"**: `skills` points somewhere the harness\n  does not scan. Check the canary landed there, in the workspace `doctor`\n  prints at the end.\n- **2 fails with \"activated but never ran\"**: discovery works and permission\n  does not. Go back to Step 4.\n- **3 fails**: the output is not newline-delimited JSON, or `events` does\n  not match its shape. Go back to Step 3.\n- **4 fails**: the resume flag did not carry the session. Some harnesses\n  resume by literal `latest` rather than an id, which is safe here because\n  the sandbox holds exactly one session. If the harness genuinely cannot\n  resume, delete the `resume` line and let the check report `unsupported`.\n- **6** never blocks. It reports how many other skills the model could see,\n  which on the `tmp` sandbox includes everything installed on the host. Use\n  `--sandbox container` when that contamination matters.\n\nIf `doctor` prints \"has no `[[matrix]]` entry, so there is no model to\ncheck it with\", the harness is described but nothing names a model for it.\nAdd a `[[matrix]]` entry, or pass `--model` for a one-off check.\n\n### Qualifying on `tmp` does not mean qualifying in a container\n\nThe sandbox is part of what you are configuring, and prerequisite 2 is a\ndifferent question in each. A harness can surface the skill perfectly on\n`tmp` and never surface it in a container, because the container has a\ndifferent filesystem, a different `HOME`, and none of the host's\nconfiguration. Every run then reports \"never activated\" and the skill\nlooks broken when the sandbox is at fault.\n\nSo if you intend to use `--sandbox container`, qualify it there too —\n**but only once the credential is declared**, because a container inherits\nnothing:\n\n```toml\n[harness.my-harness.env]\nMY_AGENT_TOKEN = \"${MY_AGENT_TOKEN}\"\n```\n\nWithout that, `doctor --sandbox container` fails on authentication, which\ntells you nothing about discovery and is easy to misread as a harness that\ndoes not qualify. Get the credential in first, confirm the harness runs at\nall in the image, and only then trust what prerequisite 2 says there. A\nsubscription login is the awkward case: on macOS it lives in the Keychain\nrather than in a file, so no amount of mounting carries it in and the\nharness needs a token it can read from the environment.\n\n## Pointing an existing harness at a different provider\n\nSame binary, different model behind it. This is what separates harness\nquality from model quality, so it is worth setting up.\n\n```toml\n[[matrix]]\nharness = \"claude-code\"\nlabel   = \"claude-code\u002Fglm-4.6\"\nmodels  = [\"glm-4.6\"]\nenv     = { ANTHROPIC_BASE_URL = \"https:\u002F\u002Fopen.bigmodel.cn\u002Fapi\u002Fanthropic\",\n            ANTHROPIC_AUTH_TOKEN = \"${GLM_TOKEN}\" }\n```\n\nGive it a `label`. An `env` override changes what is behind the harness, and\ncomparing two backends under one name would silently average them together.\n\n## Two traps worth naming\n\n**Do not reach for an isolation flag that disables skills.** Claude Code's\n`--safe-mode` lists skills among the customizations it turns off, so a run\nusing it never shows the skill to the model, and the result reads as \"this\nharness ignores instructions\". Isolation belongs to the sandbox. Use\n`--sandbox container`.\n\n**A static inventory does not prove runtime visibility.**\n`claude --plugin-dir X plugin details` reported `Skills (1)` for a run whose\ninit event showed the skill absent from the model's list. That is why the\nskill is copied into the harness's own discovery directory, and why `doctor`\nasserts on what the model did rather than on what a subcommand claims.\n","# Adding a harness to shakedown\n\nA harness is described entirely by config. No code is written to support\none, so this task is: find the flags, write the block, run `doctor`, and fix\nwhatever it names.\n\nThe reason `doctor` exists is that a harness can look configured and still\nbe measuring nothing. Every prerequisite below is verified by watching what\nthe agent actually did, never by reading a help string or an inventory\ncommand.\n\n## Step 0: Confirm the binary is here\n\nEverything below is read off the CLI in front of you, so start by\nestablishing there is one:\n\n```bash\ncommand -v \u003Ccli> && \u003Ccli> --version\n```\n\nIf it is not installed, stop and say so. shakedown runs harnesses; it does\nnot install them, and the block you would write without the binary is\nguesswork. Flags drift between versions, an agent's memory of a CLI is\nfrequently a version or two stale, and a plausible wrong flag produces a\nconfig that fails at `doctor` in a way that looks like the harness's fault\nrather than the block's. Ask the operator to install it, or work on a\nharness that is present.\n\nThe version matters beyond mere presence: record it, because it is part of\nwhat any measurement made with this block will mean.\n\n## Step 1: Check the binary can do the five things\n\nA harness qualifies if it can do these. Four are required.\n\n| # | prerequisite | required | if missing |\n|---|---|---|---|\n| 1 | run headless from a prompt and exit on its own | yes | unusable |\n| 2 | load a skill from a directory you control | yes | unusable |\n| 3 | emit machine-readable output with tool calls and text | yes | unusable, unless it can export the transcript afterwards (Step 3) |\n| 4 | continue a session | no | `inputs_resolved` reports `unsupported` |\n| 5 | run without a TTY | yes | unusable |\n| 6 | what else the model could see | reported only | nothing |\n\nRead the binary's `--help` and find the flag for each. What you are looking\nfor, in the usual naming:\n\n- A non-interactive or print flag (`-p`, `--prompt`, `run`).\n- An output format flag naming JSON or stream-json. Newline-delimited JSON\n  is what shakedown parses.\n- A session identifier flag, and a resume flag that takes it.\n- A model flag, so the same harness can be measured against several models.\n- A permission or approval flag, discussed in Step 4.\n\nPrerequisite 4 being optional is deliberate. A harness must never *fail* a\ndimension it cannot physically support, so a harness with no resume is\nreported as unsupported with a reason, and its other checks stay comparable\nagainst everything else.\n\n## Step 2: Write the block\n\n```toml\n[harness.my-harness]\nstart = [\n  \"my-agent\", \"run\", \"{prompt}\",\n  \"--json\",\n  \"--model\", \"{model}\",\n  \"--session\", \"{sid}\",\n]\nresume = [\n  \"my-agent\", \"run\", \"{reply}\",\n  \"--json\",\n  \"--model\", \"{model}\",\n  \"--resume\", \"{sid}\",\n]\nskills = \".my-agent\u002Fskills\"\nactivation_tool = \"Skill\"\n\n# Exactly one of these, and only for `--sandbox container`:\nimage      = \"ghcr.io\u002Fyou\u002Fmy-agent:1.2.3\"\n# dockerfile = \"docker\u002Fmy-agent.Dockerfile\"\n# context    = \".\"   # with `dockerfile`, to COPY from the repo root\n\n[harness.my-harness.env]\nMY_AGENT_TOKEN = \"${MY_AGENT_TOKEN}\"\n\n[[matrix]]\nharness = \"my-harness\"\nmodels  = [\"some-model\"]\n```\n\n**Commands are lists, not strings.** Each token is one argument, so a prompt\ncontaining a quote or a leading dash stays a prompt and can never become a\nflag. Substitution slots are `{prompt}`, `{reply}`, `{model}`, and `{sid}`.\n\n**`skills`** is the directory this harness scans for skills, relative to the\nworkspace. The skill under test is copied to `\u003Cskills>\u002F\u003Cname>\u002F`. Use the\nharness's own discovery path, the one it would find a skill in with no flags\nat all.\n\n**`activation_tool`** is the substring identifying a skill-activation call\nin this harness's output, `Skill` on Claude Code and `activate_skill` on\nGemini. It is what separates \"the skill never fired\" from \"it fired and\nfailed\". Without it a run where the model ignored the skill and improvised\nscores as an ordinary failure, and the number then describes the base model\nrather than the skill.\n\n**`env` is declared-only.** The sandbox gets these variables and nothing\nelse, which is what makes host contamination structurally impossible rather\nthan something a flag suppresses. Values come from the host through `${VAR}`\nso the TOML holds references, never secrets.\n\n**`image` or `dockerfile`** is needed only for `--sandbox container`, and\ndeclaring both is refused. Use `image` when you already build and push one,\nand `dockerfile` when you would rather keep the environment in the repo.\nEither way it holds the harness plus whatever the skills need at runtime,\nbecause a skill whose `bin\u002F` cannot execute fails checks for reasons that\nhave nothing to do with the skill. See `examples\u002Fdocker\u002F` for one per\nshipped harness.\n\nThe path is relative to `shakedown.toml`, and it is built once per run\nrather than per scenario.\n\n**`context`** is the directory a `dockerfile` build can `COPY` from,\ndefaulting to the dockerfile's own. Widen it with `context = \".\"` when the\nimage needs something this repo builds rather than something it installs\nfrom a registry — your own CLI, say. Without it the `COPY` is outside the\ncontext and Docker refuses it.\n\n## Step 3: Tell shakedown where the tool calls are\n\nHarnesses put tool calls at different depths. This is the one shape\nquestion, and it is config rather than code.\n\nRun the harness once by hand with its JSON flag and look at a line\ncontaining a tool call. Then choose:\n\n- **Nested**: the call sits inside a list under a dotted path, as Claude\n  Code's does in `message.content[]`. Set `container = \"message.content\"`.\n- **Flat**: the call is a top-level record, as Gemini CLI's is. Omit\n  `container` entirely.\n\nSince this is a genuine fork with a short list of answers, present it as a\ndecision rather than a prose question, and show the harness's own output as\nthe preview so the choice is made against evidence.\n\nThe remaining keys are renames, and the defaults match Claude Code:\n\n```toml\n[harness.my-harness.events]\ncontainer     = \"\"            # dotted path to a list; omit if flat\ndiscriminator = \"type\"        # the key naming a block's kind\ntool_marker   = \"tool_use\"    # that key's value on a tool call\nname_key      = \"name\"\nargs_key      = \"input\"\ntext_key      = \"text\"\ntext_marker   = \"text\"\n```\n\nThe four that name a key — `discriminator`, `name_key`, `args_key`,\n`text_key` — may each be a dotted path, which covers a flat record that\nstill buries what you need. opencode's tool call is top-level but carries\nits name and arguments under a `part` object, so it reads `name_key =\n\"part.tool\"` and `args_key = \"part.state.input\"`. (`tool_marker` and\n`text_marker` are values compared against, not keys, so a dot in one does\nnothing.)\n\nA harness matching neither shape fails at `doctor` step 3, loudly, instead\nof silently scoring zero everywhere.\n\n### When the harness prints prose and nothing else\n\nSome agents stream no events at all. Hermes prints only its final answer,\nso there is nothing to parse while it works. Before writing it off, check\nwhether it can hand over the transcript afterwards: `hermes sessions\nexport --format trace` writes Claude Code JSONL, and a start command may\nbe two commands.\n\n```toml\nstart = [\n  \"sh\", \"-c\",\n  \"hermes -z \\\"$1\\\" -m \\\"$2\\\" >&2 && hermes sessions export --format trace --no-redact -\",\n  \"hermes\", \"{prompt}\", \"{model}\",\n]\n```\n\nThe prompt stays a positional argument (`$1`) rather than being pasted\ninto the script, so it is still exactly one argument and still cannot\nbecome a flag. Check what the export does to the transcript on the way\nout, too: Hermes redacts tool arguments unless told not to, and those\narguments are what `tool_used` reads. Send the harness's own prose to\nstderr rather than `\u002Fdev\u002Fnull`: a run that fails for its own reasons then\nsays why, instead of scoring as a skill that never fired.\n\n## Step 4: Let the harness run the tool\n\nMost harnesses refuse tool calls by default, and a refusal still appears in\nthe transcript as a call. A config that omits the permission flag therefore\nproduces runs where the CLI was requested, denied, and never executed.\n\nshakedown reads the harness's denial records and fails `tool_used` with \"was\nrequested but denied\", so this shows up as a diagnosis rather than a silent\nwrong number. Fix it in `start` and `resume` by allowing the tools the skill\nneeds.\n\nWatch for a permission mode that covers file writes but not shell commands.\nClaude Code's `acceptEdits` allows `Write` while still prompting for `Bash`,\nso a skill that shells out to its CLI is denied under it.\n\n## Step 5: Run doctor until it qualifies\n\n```bash\nshakedown doctor --harness my-harness\n```\n\n`doctor` runs a canary skill that asks one question, then runs\n`echo shakedown-ok`. Seeing that shell call is only possible if the harness\nran headless, discovered the skill, surfaced it to the model, followed it,\nand emitted parseable output, so one cheap task settles prerequisites 1, 2,\n3 and 5 at once. The question is there to force a second turn, which\nsettles 4.\n\n```\n             my-harness\n  # prerequisite                    detail\n  1 headless run                 ok exit 0\n  2 skill surfaced at runtime  FAIL never activated\n  3 output parsed                ok 1 tool calls, 3 texts\n  4 session resume               ok 2 turns\n  5 no TTY required              ok ran without a terminal\n  6 environment visibility       ok 16 other skills visible\n\ndoes not qualify: blocked on skill surfaced at runtime\n```\n\nWork the failing number:\n\n- **1 fails**: the command needs a headless flag, or it is waiting on input.\n  A hang shows as a timeout rather than a non-zero exit.\n- **2 fails with \"never activated\"**: `skills` points somewhere the harness\n  does not scan. Check the canary landed there, in the workspace `doctor`\n  prints at the end.\n- **2 fails with \"activated but never ran\"**: discovery works and permission\n  does not. Go back to Step 4.\n- **3 fails**: the output is not newline-delimited JSON, or `events` does\n  not match its shape. Go back to Step 3.\n- **4 fails**: the resume flag did not carry the session. Some harnesses\n  resume by literal `latest` rather than an id, which is safe here because\n  the sandbox holds exactly one session. If the harness genuinely cannot\n  resume, delete the `resume` line and let the check report `unsupported`.\n- **6** never blocks. It reports how many other skills the model could see,\n  which on the `tmp` sandbox includes everything installed on the host. Use\n  `--sandbox container` when that contamination matters.\n\nIf `doctor` prints \"has no `[[matrix]]` entry, so there is no model to\ncheck it with\", the harness is described but nothing names a model for it.\nAdd a `[[matrix]]` entry, or pass `--model` for a one-off check.\n\n### Qualifying on `tmp` does not mean qualifying in a container\n\nThe sandbox is part of what you are configuring, and prerequisite 2 is a\ndifferent question in each. A harness can surface the skill perfectly on\n`tmp` and never surface it in a container, because the container has a\ndifferent filesystem, a different `HOME`, and none of the host's\nconfiguration. Every run then reports \"never activated\" and the skill\nlooks broken when the sandbox is at fault.\n\nSo if you intend to use `--sandbox container`, qualify it there too —\n**but only once the credential is declared**, because a container inherits\nnothing:\n\n```toml\n[harness.my-harness.env]\nMY_AGENT_TOKEN = \"${MY_AGENT_TOKEN}\"\n```\n\nWithout that, `doctor --sandbox container` fails on authentication, which\ntells you nothing about discovery and is easy to misread as a harness that\ndoes not qualify. Get the credential in first, confirm the harness runs at\nall in the image, and only then trust what prerequisite 2 says there. A\nsubscription login is the awkward case: on macOS it lives in the Keychain\nrather than in a file, so no amount of mounting carries it in and the\nharness needs a token it can read from the environment.\n\n## Pointing an existing harness at a different provider\n\nSame binary, different model behind it. This is what separates harness\nquality from model quality, so it is worth setting up.\n\n```toml\n[[matrix]]\nharness = \"claude-code\"\nlabel   = \"claude-code\u002Fglm-4.6\"\nmodels  = [\"glm-4.6\"]\nenv     = { ANTHROPIC_BASE_URL = \"https:\u002F\u002Fopen.bigmodel.cn\u002Fapi\u002Fanthropic\",\n            ANTHROPIC_AUTH_TOKEN = \"${GLM_TOKEN}\" }\n```\n\nGive it a `label`. An `env` override changes what is behind the harness, and\ncomparing two backends under one name would silently average them together.\n\n## Two traps worth naming\n\n**Do not reach for an isolation flag that disables skills.** Claude Code's\n`--safe-mode` lists skills among the customizations it turns off, so a run\nusing it never shows the skill to the model, and the result reads as \"this\nharness ignores instructions\". Isolation belongs to the sandbox. Use\n`--sandbox container`.\n\n**A static inventory does not prove runtime visibility.**\n`claude --plugin-dir X plugin details` reported `Skills (1)` for a run whose\ninit event showed the skill absent from the model's list. That is why the\nskill is copied into the harness's own discovery directory, and why `doctor`\nasserts on what the model did rather than on what a subcommand claims.\n",{"name":5,"description":7},"main","3ac632ac167bd480d8b07868b7001fc52e2a402c","plugins\u002Fclaude-code\u002Fskills\u002Fadd-harness",0,"https:\u002F\u002Fgithub.com\u002FJasperHG90\u002Fshakedown","https:\u002F\u002Fgithub.com\u002FJasperHG90","https:\u002F\u002Fjasperhg90.github.io\u002F","Data Engineer @ Xebia Data. I am interested in data analysis, history, politics, education, policy making and writing. ","2026-09-21 04:18:57",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add JasperHG90\u002Fshakedown --skill add-harness",{"tool":39,"command":40},"bunx","bunx skills add JasperHG90\u002Fshakedown --skill add-harness",{"tool":42,"command":43},"pnpm","pnpm dlx skills add JasperHG90\u002Fshakedown --skill add-harness","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FJasperHG90\u002Fshakedown\u002Fblob\u002Fmain\u002Fplugins\u002Fclaude-code\u002Fskills\u002Fadd-harness\u002FSKILL.md",true,[],[48,55,62,69,76,81],{"skillNo":49,"name":50,"nameZh":6,"description":51,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":52,"categories":53,"keywords":54,"publishedAt":19},"01M7AAFD8Q4CRZ5V34GB5655FX","fake-skill","A skill used only by shakedown's own offline tests.","https:\u002F\u002Fgithub.com\u002FJasperHG90\u002Fshakedown\u002Fblob\u002Fmain\u002Ftests\u002Ffake\u002Fskill\u002FSKILL.md",[],[],{"skillNo":56,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":59,"categories":60,"keywords":61,"publishedAt":19},"01M7AAFB03PDQMX0DY841ADDR9","shakedown-canary","Run the shakedown canary check. Use whenever someone asks to run the canary check, verify the harness, or confirm shakedown is wired up correctly.","https:\u002F\u002Fgithub.com\u002FJasperHG90\u002Fshakedown\u002Fblob\u002Fmain\u002Fsrc\u002Fshakedown\u002Fcanary\u002FSKILL.md",[],[],{"skillNo":63,"name":64,"nameZh":6,"description":65,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":66,"categories":67,"keywords":68,"publishedAt":19},"01M7AAF55DP6C2XSF7N23STFP3","create-cases","Author a shakedown cases file (`shakedowns\u002F\u003Cslug>.cases.toml`) for an existing agent skill: read the skill, turn what it promises into cases, supply fixtures for the commands it must not really run, and check the result with `shakedown case validate`. Use this whenever someone wants to measure, smoke-test, or add conformance cases for a skill, asks \"how do I test this skill\", points at a SKILL.md and wants cases written, or is editing a `.cases.toml`. Use it even when they only name the skill (\"write cases for connect-agent-to-hub\") without mentioning shakedown or TOML.","https:\u002F\u002Fgithub.com\u002FJasperHG90\u002Fshakedown\u002Fblob\u002Fmain\u002Fskills\u002Fcreate-cases\u002FSKILL.md",[],[],{"skillNo":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":19},"01M7AAEXCDG258QRVSSFQV0MV9","analyze-results","Read a shakedown report (`shakedown-report.json`) and say which file to open: the measured skill's own `SKILL.md`, the `shakedowns\u002F\u003Cslug>.cases.toml` holding its cases, the harness block in `shakedown.toml`, or a flag on the run itself. Use this whenever someone has run `shakedown case run` and wants to know what a number means or what to do about it, pastes a report or the PR comment rendered from one, asks why `skill_fired`, `tool_used`, `artifact_created` or `inputs_resolved` came out low, asks whether a rate is worth acting on, or says a conformance run failed and they cannot tell whose fault it is. Use it even when they only quote the percentage (\"my skill scores 60% on inputs_resolved, now what\") without naming shakedown or the report file.","https:\u002F\u002Fgithub.com\u002FJasperHG90\u002Fshakedown\u002Fblob\u002Fmain\u002Fskills\u002Fanalyze-results\u002FSKILL.md",[],[],{"skillNo":77,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M7AAEMM68JP8A128WHMTJJTX","https:\u002F\u002Fgithub.com\u002FJasperHG90\u002Fshakedown\u002Fblob\u002Fmain\u002Fskills\u002Fadd-harness\u002FSKILL.md",[],[],{"skillNo":82,"name":64,"nameZh":6,"description":65,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M7AAEEKJAS0KAJRM74K7FEFJ","https:\u002F\u002Fgithub.com\u002FJasperHG90\u002Fshakedown\u002Fblob\u002Fmain\u002Fplugins\u002Fopencode\u002Fskills\u002Fcreate-cases\u002FSKILL.md",[],[]]