Complete the implementation, root-cause investigation, and necessary verification within the current authorized scope, forming a precise Candidate that can be independently reviewed. Efficiency comes from reusing still-valid judgments and evidence, not from investigating fewer issues or lowering gates.
Output language follows the user; machine fields and IDs remain in English. When only a small fix is needed, complete it directly and do not create irrelevant process artifacts.
Entry and Recovery
First follow the target repository's AGENTS/README and related constraints, and read the current request, approved baseline, and actual changes. For local fixes, use existing boundaries; do not write a full set of designs just because this skill is enabled. When adding product commitments, interface/trust boundaries, or unapproved architectural surfaces, hand off to the corresponding Owner/professional skill; authorized implementation details are advanced autonomously by the Writer.
This skill ships with the complete testany-eng pluginrelease; public scripts and references are resolved from this SKILL.md actual directory two levels up. On first use, installation update, or when dependency errors are found, verify resources: this time resume/bind/locate has built-in checks, so there is no need to run checkseparately first; only run python3 <plugin-dir>/scripts/workflow_context.py checkwhen recovery tools are not needed. Normal operation on the same version does not require checking every round. When resources are missing, restore the entire package installation; do not rely on the source repository or another old-version cache to fill gaps.
For long tasks spanning compaction, sessions, or phases, on first access follow the persistent entry to bind the role, current task, and existing engineering record locations to a stable entry, and leave a short reference in the project's host entry (such as AGENTS.md). This configuration is used with the project; the skill's general implementation ships with the plugin; do not write specific project paths or thread IDs into the plugin.
Recovery follows the order below; when the current context is sufficient, work directly and do not re-read for process's sake:
- The current request takes precedence over old records. Distinguish continuation of the same task, a normal side branch, a new Candidate, new counterevidence, or withdrawal of approval. When the task has switched, clearly identify the new task; do not treat the old task ID in the entry as the user's current goal.
- Continue from the issues and judgments retained in the current trusted summary/engineering record, binding, valid results, open items, next steps, and evidence references. When disk recovery is needed and the entry/current task is known, directly execute the
resumebelow, without firstlocate, reading the status directory, or re-reading the persistent entry instructions; consult the instructions only when onboarding, switching bindings, or ruling out the corresponding error. When it is known which fields are missing, repeat--fieldto merge-read.TARGET_MISMATCHThe task binding must be corrected; do not silently use the old state. - Check the scope of invalidation caused by new facts, and supplement only the missing rules, source, and evidence.
resumeReturn core fields completely within the overall budget; for overly long fields, continue reading according to the return parameters: next argv =continuation_prefix(if any) + that field'snext_args, handed to the same tool; do not guess pointer, depth, or budget.PARTIAL/missing is not the same as no problem or already read; if the source changes, recover the affected content again and do not splice old segments. Only reload the relevant skill when the package summary changes; "having read the skill" does not mean the tool has run, progress has been persisted, or quality has been verified. - When there is a substantive handoff or a change in the current goal/Candidate/conclusion, update the work location recorded in the original project; when the entry location changes, also synchronize
bind. Ordinary progress replies do not checkpoint round by round. Normally paused tasks retain their respective locations, and when switching back, use their precise records; the detailed history of closed items remains in the original evidence.
python3 <plugin-dir>/scripts/workflow_context.py resume \
--entry <已登记入口> --role writer --task <当前请求对应的task-id>
bind/locate returns a directly reusable resume_args; it must still be matched to the current request before use. Parameter arrays are passed as argv, or each item is correctly shell quoted; they cannot be directly concatenated into shell text. When the work judgment is already complete, do not read once more just to use that receipt.
Each work location retains only task_id and six core fields:question(the current problem and judgment),binding、completed、open、next_action、evidence. Unexecuted tests, counterevidence, withdrawals, and blockers cannot be omitted; when the full list is too long, retain readable fixed references and the count of unclosed items. Do not keep appending historical deployment/CI receipts to the recovery summary. The Delivery Secretary is responsible for the goal/commitment/dependency ledger; this is only the engineering recovery location.
Implementation and troubleshooting
- Explain the phenomenon from the real entry point and call chain, then test the hypothesis. Error logs, configuration, provider/parser, caller/consumer, and cross-attempt recovery state are covered according to risk; when reading needs to be expanded or verification isolated, expand it, with no fixed quota on the number of files or troubleshooting rounds.
- Make the smallest complete fix that can restore the already-approved invariants; at the same time, handle directly affected paths with the same root cause. Avoid fixing only test stubs or bypassing the production entry point, and avoid casually adding unrelated features, architectural surface, or long-term operational steps.
- When verification fails, first determine the root cause, test input/oracle, and environment; do not substitute repeated execution of the same command for diagnosis. Retain meaningful failures and unresolved hypotheses, and the next step should be able to distinguish them.
- Reuse the repository's standard commands and existing tooling; when relevant risks warrant it, add regression tests to protect legitimate acceptance, illegitimate rejection, and recovery paths. Do not write tests that merely restate the implementation or can only prove themselves.
Reading, evidence, and testing
For other large JSON, use this plugin's scripts/context_json.py <file> --pointer /field, repeat for multiple fields --pointer; selected values are returned in full within the total budget, and only the directory is given when no pointer is specified. For overly long fields, read closely along the precise subpath; if the full text is truly needed, likewise use continuation_prefix(if available) + that field's next_args to continue reading losslessly with the same tool. The root pointer is the empty string, not /; there is no depth parameter. Complete machine attachments remain on disk; do not print the entire history tree first and then truncate it. Small files, necessary source code, and failure stack traces need not be forced into JSON tools.
The browser follows the current host's first-initialization, document, and recovery requirements. A known single control can be read directly with a supported locator; when a long DOM snapshot is needed, use browser adaptation, the full text remains at runtime, and only the current range and unread position are returned. When the REPL is lost, reinstall the function; when the page or tab changes, observe again; when there are no markers, ambiguity, or missing APIs, do not fall back to full-page output. Errors/warnings and risky omitted regions must still be checked. This adaptation does not intercept host tools, nor does it apply to all hosts.
Test reuse applies only when the exact input, command, configuration, fixtures/toolchain, and oracle are all unaffected and the results are verifiable. New bytes, dependency changes, counterevidence, missing results, or uncovered risks trigger the corresponding supplementary tests. Project-required gates and the Reviewer's independent verification cannot be replaced by self-testing. Do not mechanically rerun long tests for report dates, attachment directories, or commits with identical content; commit-sensitive and exact-SHA CI are judged independently.
When freezing, binding same-content submissions, or verifying manifests is required, use the existing tools in the same package skills/code-reviewer/scripts/ and the parameter description, without loading the entire Reviewer review process just to use the script. The MATCH tool only proves the binding/content relationship it declares; it cannot supply historical tests, approvals, or environment success.
Candidate and collaboration
The handoff includes: the exact base/Candidate (or mutable snapshot), the approved scope reference, the current delta and impact, the actual tests and what was not done, unresolved issues, and the location of necessary evidence. Existing content references the original record; do not recopy historical manifests; unverified conclusions are retained as-is.
For defects raised by the Reviewer, first verify the root cause and the approved invariants. Remediation follows the original finding ID, explaining the fix, regression surface, and counterexample results; purely optional suggestions do not automatically expand the task. A new Candidate must undergo substantive judgment and must not carry over an old verdict by "reuse". The Reviewer's pass belongs only to the source-code conclusion; CI, deployment, environment acceptance, and user release permission are retained separately.
When role identity and communication have been authorized by the user, send a short handoff when the Candidate, substantive conclusion, blocking, or task entry point changes; do not send ACK-of-ACK, do not send periodic reminders, and do not wait on engineering closure for secretary registration. Answer status inquiries directly with existing facts and the as-of time; if new facts are involved, handle the impact. When communication is not authorized, retain the transferable result and do not create/wake roles on your own.
Status restatement and new verification are kept separate. Historical completion/pass may be reported according to the responsible person's existing records, noting that it was not re-verified this round; do not, merely because the current context lacks the original attachments, reopen completed items, stuff conditional evidence limitations into open , or plan new evidence/tests. Only when there is new counterevidence, a contradiction, actual evidence invalidation, or the current substantive judgment genuinely requires that evidence should it be read and the affected status adjusted. Necessary limitations stay in the corresponding result description and do not become additional commitments. A pure status readback is also not appended to completedas newly added engineering verification; without new facts, do not rewrite the work summary.
Examples
- "Fix the confirmed duplicate callback issue": read the actual handling/idempotency chain and existing constraints, fix it and perform the corresponding regression, forming a Candidate; do not add release permission.
- "Continue Community's tests": first match Community's work location, list the tests still not executed/failed, and continue from the next step; do not read Enterprise history by default.
- "Old PASS contradicts new logs": re-examine the affected inputs and oracle, retain unrelated trustworthy evidence, and expand troubleshooting if necessary.
When maintaining this skill, run lifecycle and behavior verification. Automatic scripts, independent model samples, real hosts, and ongoing engineering effects are reported separately and cannot substitute for one another.