[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7F0VVFSW6MV4T7CTPV9FVSW":3,"skill-same:01M7F0VVFSW6MV4T7CTPV9FVSW":52},{"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":6,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M7F0VVFSW6MV4T7CTPV9FVSW","implement-in-worktree",null,"Orchestrate one implementation ticket in one isolated Git worktree with one Herdr-managed coding agent, one goal, verification, and a pull request. Use only when explicitly invoked from inside Herdr.","lza6\u002FArmor-piercing-system-prompts","lza6","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F128880206?v=4","01M7EB3K7WPX5T28EJ2NVHE6EG","Chinese-tingfeng","01M7EB3K97E1KQ16QWKJMFNGX8",11,"MIT","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F05-技能与代理开发\u002Fimplement-in-worktree\u002FSKILL.md",[],[],"2026-11-16 10:05:34",false,"---\nname: implement-in-worktree\ndescription: Orchestrate one implementation ticket in one isolated Git worktree with one Herdr-managed coding agent, one goal, verification, and a pull request. Use only when explicitly invoked from inside Herdr.\n---\n\n# Implement in Worktree\n\nRun one autonomous implementation lane without losing the primary session's\nownership of scope, review, or merge authorization.\n\n## Hard gates\n\n1. Before any other action, run `test \"${HERDR_ENV:-}\" = 1`. If it fails, stop\n   and say this skill requires execution inside Herdr. Do not substitute plain\n   Git worktrees, background shells, internal subagents, or another multiplexer.\n2. Confirm the invocation describes exactly one bounded ticket or story with a\n   testable outcome. Ask the user only for product choices that cannot be\n   discovered or safely inferred.\n3. Treat the user's invocation as permission to create a worktree, worker,\n   branch, commits, and exactly one draft pull request for that ticket. The\n   worker may choose and update its evidence-based title and body, apply\n   relevant existing labels, monitor and repair CI, and mark the pull request\n   ready once it is green, mergeable, and has no known defect. None of those\n   actions require another approval. This is not permission to comment, review,\n   merge, publish a release or package, change repository settings, create or\n   change labels or rules, create another pull request, or broaden scope unless\n   the user says so explicitly.\n4. Read every applicable `AGENTS.md`. Read repository product or architecture\n   context when those instructions require it.\n\n## Learn and capture live state\n\n- Use `herdr --help` and the relevant `herdr worktree`, `herdr agent`, `herdr\n  pane`, and `herdr notification` help before controlling Herdr. The installed\n  CLI is authoritative.\n- Capture the invoking workspace, tab, pane ID, repository root, current branch,\n  tracked worktree state, remote, and authenticated GitHub identity. Use explicit\n  IDs and `--no-focus`; never target another client's focused pane.\n- Fetch remote state when authorized by the requested PR workflow. Resolve and\n  record the exact base SHA. By default, require local `main` and `origin\u002Fmain`\n  to agree. If they diverge or tracked changes would contaminate the lane, stop\n  and ask the invoking session to resolve the base. Do not disturb untracked\n  user files.\n- Derive a short unique topic branch, worktree label, and worker name from the\n  ticket. Prefer `feature\u002F`, `fix\u002F`, or `docs\u002F` according to the work.\n\n## Base drift and final integration policy\n\n- Treat the recorded base SHA as the ticket's implementation base. Do not\n  repeatedly merge or rebase `main` during exploration and implementation each\n  time `origin\u002Fmain` advances. Finish the coherent topic work first.\n- Immediately before final canonical qualification, fetch `origin\u002Fmain`. When\n  it advanced beyond the recorded or previously integrated base, merge the\n  exact current `main` into the topic with ordinary non-force history even when\n  Git reports no textual conflict. A clean merge is not proof that the combined\n  behavior remains semantically compatible.\n- Run the required focused, canonical, milestone, live, and CI qualification on\n  that integrated head. `CLEAN`, `MERGEABLE`, and a green result against an\n  older base do not satisfy the final gate.\n- Strict required-status freshness remains authoritative. If `main` advances\n  after the green run and GitHub requires another update, integrate the new tip\n  and requalify rather than bypassing or weakening the requirement.\n- Record the original base, each integrated main SHA, resulting parents,\n  conflict resolutions, semantic compatibility repairs, and final\n  requalification in the pull request and handoff.\n\n## Create exactly one lane\n\n1. Create one Herdr-managed Git worktree\u002Fworkspace from the recorded base with\n   `--no-focus`. Parse its workspace and root-pane IDs from returned JSON.\n2. Start one agent in that root pane. Use Codex by default. Use Claude only when\n   the user explicitly requests it or Codex is unavailable and Claude can honor\n   the same persistent-goal contract.\n3. Give the worker exactly one persistent goal covering the complete ticket.\n   Require the worker to create it through its supported goal mechanism before\n   editing and to keep it active until the pull request is ready. If the chosen\n   agent cannot maintain a goal, stop rather than silently replacing it with a\n   plain prompt.\n4. Include the invoking pane ID in the worker brief so the worker can report\n   results and blockers through Herdr.\n\nThe brief must contain:\n\n- the task, acceptance criteria, out-of-scope boundaries, branch, base SHA, and\n  worktree path;\n- the applicable repository instructions and required product or architecture\n  reading;\n- an exploration-first phase covering relevant code, tests, contracts, history,\n  and useful upstream or open-source implementations;\n- permission to use read-only research subagents when valuable, while keeping a\n  single implementation owner and avoiding overlapping edits;\n- a requirement that every Codex review use `$codex-review` and every Claude\n  review use `$claude-code-review`; reviewers are non-interactive background\n  workers and must never receive a Herdr pane, tab, workspace, worktree, or\n  `herdr agent` process; a Codex implementation owner must use a native Codex\n  subagent and must never launch `codex exec`, while only a non-Codex harness may\n  use the Codex CLI; apply the equivalent native-versus-CLI rule to Claude Code;\n- focused and canonical test expectations, snapshot review, and the mandatory\n  live Herdr stress checkpoint below, including real API, PTY, or visual paths\n  that the feature exposes, plus the prohibition against weakening gates;\n- permission boundaries for credentials and external actions: use only supplied\n  credentials, never copy secrets into prompts, files, logs, commits, PRs, or\n  comments, use the user's configured Git\u002FGitHub identity, and state that this\n  invocation authorizes one draft pull request plus its later ready transition\n  without another text or action approval;\n- the pull-request and structured-handoff requirements below.\n\nSubmit the brief through `herdr agent prompt ...` without `--wait`. Dispatching\nmust return after Herdr accepts the prompt so independent lanes can start in\nparallel; a long-running goal must not turn prompt submission into a timeout.\nUse `agent get`, `agent read`, and a separate `agent wait` only for deliberate\nmonitoring after every intended lane has been dispatched. Inspect the agent\nbefore responding to `blocked`, `unknown`, a monitoring timeout, or a stalled\nprompt. Do not blindly resend a prompt.\n\n## Exploration checkpoint\n\nThe worker must inspect before editing. It may proceed autonomously when the\nticket fits the existing design and can satisfy its acceptance criteria.\n\nStop the lane when evidence shows any of the following:\n\n- a substantial prerequisite refactor should land on `main` first;\n- an upstream API, hook, stable identity, platform capability, or permission is\n  missing;\n- the only available approach would be misleading, fragile, unsafe, or knowingly\n  deficient;\n- the requested behavior conflicts with repository contracts or needs a product\n  choice outside the ticket.\n\nDo not jump through compatibility hoops merely to produce code. Preserve the\nworktree, do not open a deficient pull request, and report to the invoking pane\nthrough Herdr with the evidence, affected acceptance criteria, attempted safe\nalternatives, and recommended next action. Ask the missing question when user\ninput can unblock the work. Use the worker's supported blocked or stale goal\nstate truthfully; never mark the goal complete. If its goal system delays a\nblocked transition, keep the goal active and still report immediately.\n\n## Implement and qualify\n\n- Implement only the ticket and necessary supporting changes. Preserve unrelated\n  user work.\n- Run focused tests during development and every repository-required canonical\n  gate before committing. Review all snapshot or golden-file diffs explicitly.\n- For live TUI, Herdr, harness, API, PTY, or visual qualification, the worker may\n  create disposable test tabs and panes inside its assigned Herdr workspace.\n  Address them by explicit IDs with `--no-focus`; never reuse the invoking\n  coordinator's pane, another workstream's workspace, or a user's unrelated\n  pane. Temporary test panes do not authorize another implementation agent,\n  reviewer, worktree, or parallel implementation owner.\n- When the feature integrates with Herdr or a harness, exercise the real\n  structured API and relevant UI\u002FPTY behavior in the worker's Herdr workspace.\n  A mock-only result is insufficient when live qualification is feasible.\n- Treat `herdr pane send-text` as literal text injection, not as a terminal key\n  event transport. Do not use it to qualify escape sequences, modifier chords,\n  BackTab, bracketed paste framing, or another exact terminal protocol. When\n  `pane send-keys` does not expose the required key distinctly, drive the exact\n  bytes through a real PTY tool such as Expect inside the disposable pane.\n- A successful pane send proves only that Herdr enqueued the input. Likewise,\n  `pane wait-output` searches immediately and may match text that was already\n  present. Prove each live step with a unique new screen postcondition, terminal\n  transcript event, durable revision, or external-state inspection; a generic\n  label such as `saved`, `ready`, or `complete` is not transition evidence.\n- Define the live oracle before combining stress dimensions. Seed fixtures\n  through a stable CLI or API when fixture creation is not the behavior under\n  test, exercise one invariant first, and inspect its exact durable result before\n  adding load, resize, repetition, Unicode, or another neighboring state. A\n  compound script whose timeout could mean transport, setup, rendering,\n  persistence, eligibility, or teardown is not useful completion evidence.\n- Derive changed and deliberately unchanged cases from the feature contract.\n  Keep supported actions separate from conservative no-ops so a correct refusal\n  is not misdiagnosed as dropped input. Increase content size and repetition\n  only after the minimal case proves the same path with an unambiguous oracle.\n- When a PTY driver echoes child terminal traffic into a Herdr pane, terminal\n  capability queries can elicit replies that remain buffered as later shell\n  input. Keep child terminal traffic contained (for example, Expect\n  `log_user 0`), emit only explicit text sentinels, and verify the pane is back\n  at an uncontaminated shell before reusing it.\n- Before handoff, run the implemented feature from the exact topic-branch build\n  in a disposable live pane inside the assigned Herdr workspace and actively try\n  to break it. This is a completion gate, not an optional walkthrough or a\n  substitute for automated tests.\n- Derive an adversarial matrix from the feature's actual risks. At minimum,\n  exercise applicable boundary and empty inputs, unusually large content or\n  collections, overflow and narrow\u002Fshallow layouts, Unicode and control-heavy\n  text, rapid or repeated input, repeated activation\u002Fdeactivation, cancellation,\n  resize\u002Freflow, and restart or recovery. Repeat idempotent actions enough to\n  prove that they converge on the same result without duplicate durable writes,\n  deliveries, receipts, or resources. For a nonvisual feature, drive the real\n  command, API, persistence, or integration path from that live pane rather than\n  inventing a cosmetic TUI scenario.\n- Exercise meaningful combinations with neighboring state such as editing,\n  selection, collapse, scrolling, pending persistence, failure, retry, and undo\n  when the ticket can interact with them. Use both keyboard and mouse paths when\n  the behavior has both. Do not use private user content or perform an\n  irreversible external action merely to manufacture stress evidence.\n- Treat every failure found during the stress pass as implementation work:\n  reproduce it with a focused automated regression where practical, fix it, and\n  rerun the relevant stress case and canonical gate. An unresolved defect means\n  the lane is blocked or incomplete and must be reported; it cannot be hidden as\n  a residual risk while declaring the goal complete.\n- Record every temporary test tab and pane as it is created. Close those\n  disposable resources before handoff, remove only test state created by the\n  worker, and report the live scenarios exercised plus any resource that could\n  not be cleaned up. Preserve the implementation pane and worktree for primary\n  review.\n- Inspect the final diff, status, new files, generated artifacts, and commit\n  identity. Commit coherent changes without secrets or machine-specific paths.\n- Push only the topic branch. Open exactly one draft pull request against `main`\n  with a concise evidence-based title and a body covering summary, acceptance\n  criteria, tests, live\u002Fmanual verification, risks, limitations, and follow-ups.\n  Keep that body accurate as branch evidence changes. The invocation is already\n  approval for the title, body, creation, and updates, so do not stop to request\n  another outbound-text approval.\n- Inspect the repository's existing labels and apply only relevant existing\n  labels. Common Proqi choices are `enhancement`, `bug`, `documentation`,\n  `rust`, and `accessibility`. Never create labels or repository rules unless\n  separately authorized.\n\nMonitor the pull request's complete CI result. Diagnose failures before retrying.\nFix failures caused by the branch and rerun the required gates. One evidence-\nbased retry is reasonable for an infrastructure flake; repeated unexplained\nfailure is a blocker, not permission for indefinite retries. Never weaken a\ngate, threshold, test, or snapshot to obtain green status.\n\nWhen `main` advances after a green run, apply the final integration policy above\nand obtain a fresh complete CI result for the new integrated head. Do not treat\nan older green result or a conflict-free merge as current integration evidence.\n\nWhen the complete pull request is green and mergeable and no known defect or\nreview issue remains, mark it ready for review without requesting another\napproval. Do not post comments or reviews while doing so. A ready pull request\nis still only a review artifact; the worker never merges it.\n\n## Finish and hand back\n\nThe worker is done only when the pull request is green, mergeable, and ready for\nreview, the full acceptance criteria are met, and no known defect or review\nissue remains. It must then mark its goal complete and send a structured Herdr\nhandoff to the invoking pane. The original skill invocation already authorizes\nthis internal handoff, so the worker must not ask Oliver for separate approval\nbefore submitting it. The handoff must contain:\n\n- pull-request URL, branch, base and head SHAs, and worktree\u002Fworkspace IDs;\n- implementation summary and important design decisions;\n- focused and canonical verification plus a concrete stress-test matrix listing\n  the live inputs, repetitions, boundary conditions, state combinations, and\n  observed results; never summarize this only as \"manual testing passed\";\n- CI status and any residual risks or intentionally deferred work;\n- exact cleanup identifiers.\n\nThe worker never merges its pull request. The invoking primary session reviews\nthe complete diff and validation itself, and sends revisions back to the same\nworker when needed. The primary merges only when the user's original request\nexplicitly authorized merging; otherwise it reports readiness and waits.\n\nAfter a successful merge, remove only the Herdr worktree\u002Fworkspace created by\nthis invocation, then safely delete its merged local branch with `git branch\n-d`. Never use forced deletion for routine cleanup. Preserve blocked or unmerged\nlanes for inspection and report their exact identifiers.\n","# Implement in Worktree\n\nRun one autonomous implementation lane without losing the primary session's\nownership of scope, review, or merge authorization.\n\n## Hard gates\n\n1. Before any other action, run `test \"${HERDR_ENV:-}\" = 1`. If it fails, stop\n   and say this skill requires execution inside Herdr. Do not substitute plain\n   Git worktrees, background shells, internal subagents, or another multiplexer.\n2. Confirm the invocation describes exactly one bounded ticket or story with a\n   testable outcome. Ask the user only for product choices that cannot be\n   discovered or safely inferred.\n3. Treat the user's invocation as permission to create a worktree, worker,\n   branch, commits, and exactly one draft pull request for that ticket. The\n   worker may choose and update its evidence-based title and body, apply\n   relevant existing labels, monitor and repair CI, and mark the pull request\n   ready once it is green, mergeable, and has no known defect. None of those\n   actions require another approval. This is not permission to comment, review,\n   merge, publish a release or package, change repository settings, create or\n   change labels or rules, create another pull request, or broaden scope unless\n   the user says so explicitly.\n4. Read every applicable `AGENTS.md`. Read repository product or architecture\n   context when those instructions require it.\n\n## Learn and capture live state\n\n- Use `herdr --help` and the relevant `herdr worktree`, `herdr agent`, `herdr\n  pane`, and `herdr notification` help before controlling Herdr. The installed\n  CLI is authoritative.\n- Capture the invoking workspace, tab, pane ID, repository root, current branch,\n  tracked worktree state, remote, and authenticated GitHub identity. Use explicit\n  IDs and `--no-focus`; never target another client's focused pane.\n- Fetch remote state when authorized by the requested PR workflow. Resolve and\n  record the exact base SHA. By default, require local `main` and `origin\u002Fmain`\n  to agree. If they diverge or tracked changes would contaminate the lane, stop\n  and ask the invoking session to resolve the base. Do not disturb untracked\n  user files.\n- Derive a short unique topic branch, worktree label, and worker name from the\n  ticket. Prefer `feature\u002F`, `fix\u002F`, or `docs\u002F` according to the work.\n\n## Base drift and final integration policy\n\n- Treat the recorded base SHA as the ticket's implementation base. Do not\n  repeatedly merge or rebase `main` during exploration and implementation each\n  time `origin\u002Fmain` advances. Finish the coherent topic work first.\n- Immediately before final canonical qualification, fetch `origin\u002Fmain`. When\n  it advanced beyond the recorded or previously integrated base, merge the\n  exact current `main` into the topic with ordinary non-force history even when\n  Git reports no textual conflict. A clean merge is not proof that the combined\n  behavior remains semantically compatible.\n- Run the required focused, canonical, milestone, live, and CI qualification on\n  that integrated head. `CLEAN`, `MERGEABLE`, and a green result against an\n  older base do not satisfy the final gate.\n- Strict required-status freshness remains authoritative. If `main` advances\n  after the green run and GitHub requires another update, integrate the new tip\n  and requalify rather than bypassing or weakening the requirement.\n- Record the original base, each integrated main SHA, resulting parents,\n  conflict resolutions, semantic compatibility repairs, and final\n  requalification in the pull request and handoff.\n\n## Create exactly one lane\n\n1. Create one Herdr-managed Git worktree\u002Fworkspace from the recorded base with\n   `--no-focus`. Parse its workspace and root-pane IDs from returned JSON.\n2. Start one agent in that root pane. Use Codex by default. Use Claude only when\n   the user explicitly requests it or Codex is unavailable and Claude can honor\n   the same persistent-goal contract.\n3. Give the worker exactly one persistent goal covering the complete ticket.\n   Require the worker to create it through its supported goal mechanism before\n   editing and to keep it active until the pull request is ready. If the chosen\n   agent cannot maintain a goal, stop rather than silently replacing it with a\n   plain prompt.\n4. Include the invoking pane ID in the worker brief so the worker can report\n   results and blockers through Herdr.\n\nThe brief must contain:\n\n- the task, acceptance criteria, out-of-scope boundaries, branch, base SHA, and\n  worktree path;\n- the applicable repository instructions and required product or architecture\n  reading;\n- an exploration-first phase covering relevant code, tests, contracts, history,\n  and useful upstream or open-source implementations;\n- permission to use read-only research subagents when valuable, while keeping a\n  single implementation owner and avoiding overlapping edits;\n- a requirement that every Codex review use `$codex-review` and every Claude\n  review use `$claude-code-review`; reviewers are non-interactive background\n  workers and must never receive a Herdr pane, tab, workspace, worktree, or\n  `herdr agent` process; a Codex implementation owner must use a native Codex\n  subagent and must never launch `codex exec`, while only a non-Codex harness may\n  use the Codex CLI; apply the equivalent native-versus-CLI rule to Claude Code;\n- focused and canonical test expectations, snapshot review, and the mandatory\n  live Herdr stress checkpoint below, including real API, PTY, or visual paths\n  that the feature exposes, plus the prohibition against weakening gates;\n- permission boundaries for credentials and external actions: use only supplied\n  credentials, never copy secrets into prompts, files, logs, commits, PRs, or\n  comments, use the user's configured Git\u002FGitHub identity, and state that this\n  invocation authorizes one draft pull request plus its later ready transition\n  without another text or action approval;\n- the pull-request and structured-handoff requirements below.\n\nSubmit the brief through `herdr agent prompt ...` without `--wait`. Dispatching\nmust return after Herdr accepts the prompt so independent lanes can start in\nparallel; a long-running goal must not turn prompt submission into a timeout.\nUse `agent get`, `agent read`, and a separate `agent wait` only for deliberate\nmonitoring after every intended lane has been dispatched. Inspect the agent\nbefore responding to `blocked`, `unknown`, a monitoring timeout, or a stalled\nprompt. Do not blindly resend a prompt.\n\n## Exploration checkpoint\n\nThe worker must inspect before editing. It may proceed autonomously when the\nticket fits the existing design and can satisfy its acceptance criteria.\n\nStop the lane when evidence shows any of the following:\n\n- a substantial prerequisite refactor should land on `main` first;\n- an upstream API, hook, stable identity, platform capability, or permission is\n  missing;\n- the only available approach would be misleading, fragile, unsafe, or knowingly\n  deficient;\n- the requested behavior conflicts with repository contracts or needs a product\n  choice outside the ticket.\n\nDo not jump through compatibility hoops merely to produce code. Preserve the\nworktree, do not open a deficient pull request, and report to the invoking pane\nthrough Herdr with the evidence, affected acceptance criteria, attempted safe\nalternatives, and recommended next action. Ask the missing question when user\ninput can unblock the work. Use the worker's supported blocked or stale goal\nstate truthfully; never mark the goal complete. If its goal system delays a\nblocked transition, keep the goal active and still report immediately.\n\n## Implement and qualify\n\n- Implement only the ticket and necessary supporting changes. Preserve unrelated\n  user work.\n- Run focused tests during development and every repository-required canonical\n  gate before committing. Review all snapshot or golden-file diffs explicitly.\n- For live TUI, Herdr, harness, API, PTY, or visual qualification, the worker may\n  create disposable test tabs and panes inside its assigned Herdr workspace.\n  Address them by explicit IDs with `--no-focus`; never reuse the invoking\n  coordinator's pane, another workstream's workspace, or a user's unrelated\n  pane. Temporary test panes do not authorize another implementation agent,\n  reviewer, worktree, or parallel implementation owner.\n- When the feature integrates with Herdr or a harness, exercise the real\n  structured API and relevant UI\u002FPTY behavior in the worker's Herdr workspace.\n  A mock-only result is insufficient when live qualification is feasible.\n- Treat `herdr pane send-text` as literal text injection, not as a terminal key\n  event transport. Do not use it to qualify escape sequences, modifier chords,\n  BackTab, bracketed paste framing, or another exact terminal protocol. When\n  `pane send-keys` does not expose the required key distinctly, drive the exact\n  bytes through a real PTY tool such as Expect inside the disposable pane.\n- A successful pane send proves only that Herdr enqueued the input. Likewise,\n  `pane wait-output` searches immediately and may match text that was already\n  present. Prove each live step with a unique new screen postcondition, terminal\n  transcript event, durable revision, or external-state inspection; a generic\n  label such as `saved`, `ready`, or `complete` is not transition evidence.\n- Define the live oracle before combining stress dimensions. Seed fixtures\n  through a stable CLI or API when fixture creation is not the behavior under\n  test, exercise one invariant first, and inspect its exact durable result before\n  adding load, resize, repetition, Unicode, or another neighboring state. A\n  compound script whose timeout could mean transport, setup, rendering,\n  persistence, eligibility, or teardown is not useful completion evidence.\n- Derive changed and deliberately unchanged cases from the feature contract.\n  Keep supported actions separate from conservative no-ops so a correct refusal\n  is not misdiagnosed as dropped input. Increase content size and repetition\n  only after the minimal case proves the same path with an unambiguous oracle.\n- When a PTY driver echoes child terminal traffic into a Herdr pane, terminal\n  capability queries can elicit replies that remain buffered as later shell\n  input. Keep child terminal traffic contained (for example, Expect\n  `log_user 0`), emit only explicit text sentinels, and verify the pane is back\n  at an uncontaminated shell before reusing it.\n- Before handoff, run the implemented feature from the exact topic-branch build\n  in a disposable live pane inside the assigned Herdr workspace and actively try\n  to break it. This is a completion gate, not an optional walkthrough or a\n  substitute for automated tests.\n- Derive an adversarial matrix from the feature's actual risks. At minimum,\n  exercise applicable boundary and empty inputs, unusually large content or\n  collections, overflow and narrow\u002Fshallow layouts, Unicode and control-heavy\n  text, rapid or repeated input, repeated activation\u002Fdeactivation, cancellation,\n  resize\u002Freflow, and restart or recovery. Repeat idempotent actions enough to\n  prove that they converge on the same result without duplicate durable writes,\n  deliveries, receipts, or resources. For a nonvisual feature, drive the real\n  command, API, persistence, or integration path from that live pane rather than\n  inventing a cosmetic TUI scenario.\n- Exercise meaningful combinations with neighboring state such as editing,\n  selection, collapse, scrolling, pending persistence, failure, retry, and undo\n  when the ticket can interact with them. Use both keyboard and mouse paths when\n  the behavior has both. Do not use private user content or perform an\n  irreversible external action merely to manufacture stress evidence.\n- Treat every failure found during the stress pass as implementation work:\n  reproduce it with a focused automated regression where practical, fix it, and\n  rerun the relevant stress case and canonical gate. An unresolved defect means\n  the lane is blocked or incomplete and must be reported; it cannot be hidden as\n  a residual risk while declaring the goal complete.\n- Record every temporary test tab and pane as it is created. Close those\n  disposable resources before handoff, remove only test state created by the\n  worker, and report the live scenarios exercised plus any resource that could\n  not be cleaned up. Preserve the implementation pane and worktree for primary\n  review.\n- Inspect the final diff, status, new files, generated artifacts, and commit\n  identity. Commit coherent changes without secrets or machine-specific paths.\n- Push only the topic branch. Open exactly one draft pull request against `main`\n  with a concise evidence-based title and a body covering summary, acceptance\n  criteria, tests, live\u002Fmanual verification, risks, limitations, and follow-ups.\n  Keep that body accurate as branch evidence changes. The invocation is already\n  approval for the title, body, creation, and updates, so do not stop to request\n  another outbound-text approval.\n- Inspect the repository's existing labels and apply only relevant existing\n  labels. Common Proqi choices are `enhancement`, `bug`, `documentation`,\n  `rust`, and `accessibility`. Never create labels or repository rules unless\n  separately authorized.\n\nMonitor the pull request's complete CI result. Diagnose failures before retrying.\nFix failures caused by the branch and rerun the required gates. One evidence-\nbased retry is reasonable for an infrastructure flake; repeated unexplained\nfailure is a blocker, not permission for indefinite retries. Never weaken a\ngate, threshold, test, or snapshot to obtain green status.\n\nWhen `main` advances after a green run, apply the final integration policy above\nand obtain a fresh complete CI result for the new integrated head. Do not treat\nan older green result or a conflict-free merge as current integration evidence.\n\nWhen the complete pull request is green and mergeable and no known defect or\nreview issue remains, mark it ready for review without requesting another\napproval. Do not post comments or reviews while doing so. A ready pull request\nis still only a review artifact; the worker never merges it.\n\n## Finish and hand back\n\nThe worker is done only when the pull request is green, mergeable, and ready for\nreview, the full acceptance criteria are met, and no known defect or review\nissue remains. It must then mark its goal complete and send a structured Herdr\nhandoff to the invoking pane. The original skill invocation already authorizes\nthis internal handoff, so the worker must not ask Oliver for separate approval\nbefore submitting it. The handoff must contain:\n\n- pull-request URL, branch, base and head SHAs, and worktree\u002Fworkspace IDs;\n- implementation summary and important design decisions;\n- focused and canonical verification plus a concrete stress-test matrix listing\n  the live inputs, repetitions, boundary conditions, state combinations, and\n  observed results; never summarize this only as \"manual testing passed\";\n- CI status and any residual risks or intentionally deferred work;\n- exact cleanup identifiers.\n\nThe worker never merges its pull request. The invoking primary session reviews\nthe complete diff and validation itself, and sends revisions back to the same\nworker when needed. The primary merges only when the user's original request\nexplicitly authorized merging; otherwise it reports readiness and waits.\n\nAfter a successful merge, remove only the Herdr worktree\u002Fworkspace created by\nthis invocation, then safely delete its merged local branch with `git branch\n-d`. Never use forced deletion for routine cleanup. Preserve blocked or unmerged\nlanes for inspection and report their exact identifiers.\n",{"name":5,"description":7},"main","d69096e239f2a89aabc2cd078c47664a9d3b1e11","05-技能与代理开发\u002Fimplement-in-worktree",3,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts","https:\u002F\u002Fgithub.com\u002Flza6","个人vx：Tf00798\r\n个人✈：@TFGY999","2026-10-08 14:29:22",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add lza6\u002FArmor-piercing-system-prompts --skill implement-in-worktree",{"tool":38,"command":39},"bunx","bunx skills add lza6\u002FArmor-piercing-system-prompts --skill implement-in-worktree",{"tool":41,"command":42},"pnpm","pnpm dlx skills add lza6\u002FArmor-piercing-system-prompts --skill implement-in-worktree","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F05-技能与代理开发\u002Fimplement-in-worktree\u002FSKILL.md",true,[46],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"agents\u002Fopenai.yaml","REFERENCE","参考文档",297,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F05-技能与代理开发\u002Fimplement-in-worktree\u002Fagents\u002Fopenai.yaml",[53,60,65,72,79,86],{"skillNo":54,"name":55,"nameZh":6,"description":56,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":19},"01M7FFG485KY83TD6HSC4YPCMC","obsidian-clipper-template-creator","Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator\u002FSKILL.md",[],[],{"skillNo":61,"name":55,"nameZh":6,"description":56,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":62,"categories":63,"keywords":64,"publishedAt":19},"01M7FFG0T6AZ0STEJPSB3CX84K","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator-antigravity-awesome-skills-main\u002FSKILL.md",[],[],{"skillNo":66,"name":67,"nameZh":6,"description":68,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":19},"01M7FFFY0QEM14SKWPDYGMXM2P","python-api","Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP\u002FmAP semantics, and source-style import diagnostics.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Fpython-api\u002FSKILL.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":78,"publishedAt":19},"01M7FFFVZQZ2GJ0C0TGP7JHD20","file-evaluation","Folder-based PASCAL VOC AP\u002FmAP evaluation from ground-truth and detection text files.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Ffile-evaluation\u002FSKILL.md",[],[],{"skillNo":80,"name":81,"nameZh":6,"description":82,"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},"01M7FFFSFNYSE77SJ6PE1CK2DA","object-detection-metrics","Guides legacy Object-Detection-Metrics PASCAL VOC AP\u002FmAP evaluation from detection text folders and direct Python API classes.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002FSKILL.md",[],[],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":19},"01M7FFFP0GCK1RVXBCPQ8APBMT","nx-workspace-patterns","Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fnx-workspace-patterns\u002FSKILL.md",[],[]]