project-skill-specialization

name	project-skill-specialization
execution_scope	state_machine
execution_owner	T5-SPECIALIZE-EXECUTOR-SKILLS
description	从 T5 reboost 或 legacy handoff 及其权威来源构建、预览或校验项目专属 external-executor Skill Suite。适用于 Suite 缺失、过期、恢复前核验或发布失败诊断；不会补造项目事实、执行实验或替换运行中的 Suite。
tools	["read_file","list_files","glob_files","grep_search","bash_run","finish_task"]
allowed_read_prefixes	["","project.yaml","external_executor/","ideation/","literature/","novelty/","user_seeds/"]
allowed_write_prefixes	["external_executor/"]

Project Skill Specialization

Act as the operational controller for project Skill specialization. Use the bundled wrappers to invoke the repository's deterministic specializer; do not perform context extraction, Guidance rendering, or template editing yourself.

Establish the operating roots

1. Locate the nearest repository root containing both:
   • researchos/;
   • skills/external_executor_skills/skill_specialization.yaml.
2. Locate the project workspace containing both:
   • project.yaml;
   • external_executor/.
3. Treat the directory containing this file as <skill-dir>.
4. Read before execution:
   • references/compiler-interface.md;
   • references/source-authority-and-uncertainty.md;
   • references/specialization-contract.md;
   • references/report-and-return-contract.md.
5. Read references/recovery-and-safety.md before retrying, replacing an existing Suite, or responding to a failed preflight or build.

Do not broaden repository roots, workspace roots, source authority, writable paths, or compiler modes by inference.

Own only invocation and result interpretation

This Skill owns:

• resolving the repository and workspace passed to the compiler;
• choosing build, dry-run, or validate-only mode from the request;
• running deterministic preflight and specialization wrappers;
• reading external_executor/report/skill_specialization_report.json;
• returning a compact status and actionable failure surface.

The Project Skill Specializer owns:

• reading handoff and source artifacts;
• building project_skill_context.yaml;
• applying source-over-handoff rules;
• preserving uncertain fields;
• validating the Context Schema and specialization mapping;
• copying and rendering all 13 executor Skills;
• verifying marker and template integrity;
• staging, publishing, and rollback;
• writing the specialization report.

Never duplicate those responsibilities in prose edits, ad hoc Python, or direct file manipulation.

Choose one mode

Use exactly one mode:

• Build: generate and atomically publish the project Context and all 13 specialized Skills. This is the default after T5-REBOOST or legacy T5-HANDOFF.
• Dry run: execute the complete build and validation path in staging without publishing. Use when the user asks to preview, debug, or test specialization.
• Validate only: validate an existing specialized Suite without rebuilding it. Use when the output already exists and only integrity or readiness must be checked.

Do not use validate-only to repair stale or missing outputs. Do not use build to overwrite a Suite whose executor status is running.

Run deterministic preflight

Run:

python <skill-dir>/scripts/preflight_specialization.py \
  --workspace <workspace> \
  --repo-root <repo-root> \
  --mode build|dry-run|validate-only \
  --json

The preflight checks repository assets, required control artifacts, the 13 template directories and markers, compiler availability, and mode-specific active-executor safety. Use the same mode that will be passed to the specializer.

Proceed when status is pass or warning. Stop when it is fail. Treat warnings about optional source artifacts as potential uncertain fields, not as permission to invent values.

Invoke the specializer

Build and publish

python <skill-dir>/scripts/run_specialization.py \
  --workspace <workspace> \
  --repo-root <repo-root> \
  --mode build \
  --json

Dry run

python <skill-dir>/scripts/run_specialization.py \
  --workspace <workspace> \
  --repo-root <repo-root> \
  --mode dry-run \
  --json

Validate existing output

python <skill-dir>/scripts/run_specialization.py \
  --workspace <workspace> \
  --repo-root <repo-root> \
  --mode validate-only \
  --json

Use the wrapper rather than calling internal builder, renderer, validation, or publication modules separately. All entrypoints must converge on the repository's public specialize_project_skills(...) service.

Inspect the durable report

For a published build or validation, read:

<workspace>/external_executor/report/skill_specialization_report.json

Produce a compact summary with:

python <skill-dir>/scripts/summarize_specialization_report.py \
  --workspace <workspace> \
  --json

Do not determine readiness by checking only whether external_executor/skills/ exists. The report, Context Schema validation, 13-Skill count, marker checks, and template-integrity results must agree.

Handle the result

ready

Confirm that:

• project_skill_context.yaml exists and validates;
• all 13 Skills were specialized;
• required uncertain fields are empty;
• template integrity passed;
• the published Suite path is present.

Return control to the T5-SPECIALIZE-EXECUTOR-SKILLS task or caller.

incomplete

The Suite was generated successfully, but one or more fields required by an injection mapping remain uncertain.

• Preserve the generated Suite and report.
• List the required uncertain paths and their source notes.
• Do not fill them from general knowledge, repository defaults, or conversational assumptions.
• Route their resolution to runtime context-alignment, an upstream source correction, or the existing human gate.
• Do not reinterpret incomplete as a failed compiler run or a failed scientific audit.

failed

• Confirm that the previous published Suite, if any, was preserved.
• Report the exact error codes, paths, Skill names, and field paths from the wrapper/report.
• Use references/recovery-and-safety.md to choose the repair owner.
• Do not bypass Schema validation, change mapping paths, edit generated Guidance manually, or use a force overwrite to make the report pass.

Non-negotiable boundaries

Never:

• create or edit project_skill_context.yaml manually;
• directly write text inside the 13 Guidance marker regions;
• modify text outside those marker regions;
• use an LLM to resolve source conflicts or uncertain values;
• add a per-Skill project-context node or per-Skill configuration file;
• treat handoff as authoritative when a designated source artifact provides a different value;
• perform context alignment, novelty audit, protocol review, resource readiness, or claim approval during specialization;
• copy legacy skills_customization into the 13-Skill output;
• publish a partial Suite;
• overwrite a Suite while executor_status.json says running;
• delete failed, stale, or previous evidence as part of specialization.

Return contract

Return a compact result containing:

skill=project-skill-specialization
mode=build|dry-run|validate-only
status=ready|incomplete|failed
workspace=<workspace>
context=external_executor/project_skill_context.yaml|not_published
skills=13/13|<count>/13
report=external_executor/report/skill_specialization_report.json|not_published
required_uncertain_fields=<paths>
errors=<codes>
next_action=continue_to_gate|resolve_uncertain_context|repair_specializer|stop

Do not include the full project Context, private research content, or long report payload in the return message.