القائمة
Use when reviewing code, diffs, PRs, implementation correctness, semantic drift, missing tests, runtime safety risks, generated context drift, or contract consistency under Harbor-spec.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
| name | harbor-code-review |
| description | Use when reviewing code, diffs, PRs, implementation correctness, semantic drift, missing tests, runtime safety risks, generated context drift, or contract consistency under Harbor-spec. |
Version: Harbor-spec v1.3.0
Purpose: Code review workflow for Harbor-managed repositories
Use this skill to review code changes under Harbor-spec.
This skill helps identify:
implementation bugs
semantic drift
contract gaps
schema gaps
test / DDT gaps
runtime safety risks
generated context gaps
Diary gaps
breaking change risks
This skill is an on-demand workflow entrypoint.
It is not source of truth.
It must follow:
AGENTS.md
.harbor/policy.yaml
.harbor/safety.yaml
.harbor/rules/agent-policy.md
.harbor/rules/contract-rules.md
.harbor/rules/ddt-rules.md
.harbor/rules/runtime-safety.md
.harbor/rules/diary-rules.md
.harbor/rules/project-rules.md, if present
Core principle:
A Harbor review checks not only whether code works, but whether code, contracts, tests, generated context, decision memory, and safety boundaries remain aligned.
Use this skill when reviewing:
code changes
diffs
PRs
patches
AI-generated code
refactors
bug fixes
feature changes
contract-impacting changes
test changes
DDT changes
schema changes
CLI output changes
JSON output changes
file write behavior changes
workspace layout changes
generated context changes
runtime safety changes
Also use when asked to check:
implementation correctness
semantic drift
contract consistency
schema consistency
missing tests
false green tests
runtime safety risk
generated context drift
Diary need
breaking change risk
AGENTS.md is the lightweight cross-tool entrypoint.
Use it for:
global role
workspace boundaries
default workflow
instruction priority
context loading order
completion expectations
Do not duplicate all AGENTS rules inside this skill.
Project-specific rules live in:
.harbor/rules/project-rules.md
Use Project Rules for:
technology stack
directory map
strictness map
verified test commands
project-specific contract sources
project-specific safety boundaries
local skill routing
If this skill conflicts with Project Rules, inspect .harbor/policy.yaml and .harbor/safety.yaml.
Machine policy wins over Markdown rules.
Use related skills when needed:
harbor-contract-change
Use when the review finds or involves Contract Impact that requires implementation or contract changes.
harbor-ddt-diary
Use when the review requires test updates, DDT binding changes, l3_version changes, Diary Drafts, changelog, or release notes.
harbor-safety-preflight
Use before risky operations, protected file changes, dependency changes, migrations, CI/CD, destructive commands, or .harbor/*.yaml changes.
harbor-context-refresh
Use when generated context under .harbor/views/** is stale or needs refresh.
harbor-workspace-migration-plan
Use for workspace diagnostics, migration dry-run, cleanup planning, or canonical workspace changes.
This skill may route to these skills, but should not replace them.
Before reviewing source files, load only necessary context.
Default order:
1. AGENTS.md
2. Project Rules, if present:
.harbor/rules/project-rules.md
3. Canonical project structure:
.harbor/views/project-structure.md
4. Relevant canonical L2 README:
.harbor/views/l2/<module>/README.md
5. Relevant canonical Module Capsule:
.harbor/views/modules/<module>/module-card.md
.harbor/views/modules/<module>/review-checklist.md
.harbor/views/modules/<module>/debug-playbook.md
6. Relevant contract rules when needed:
.harbor/rules/contract-rules.md
7. Relevant DDT rules when needed:
.harbor/rules/ddt-rules.md
8. Relevant runtime safety rules when needed:
.harbor/rules/runtime-safety.md
9. Relevant diary rules when needed:
.harbor/rules/diary-rules.md
10. Relevant source files
11. Relevant tests, fixtures, snapshots, schemas, or DDT targets
Do not read the whole repository unless the canonical project structure, L2 README, Module Capsule, source files, and tests are insufficient.
Generated context helps orientation.
Generated context is not source of truth.
Source code, tests, schemas, policy, and diary define the actual project state.
Check the change across these dimensions.
implementation correctness
implementation vs contract
schema / type consistency
args / returns / raises consistency
error paths
side effects
idempotency
runtime safety
tests / DDT coverage
semantic drift
breaking change risk
generated context update need
Diary need
tool honesty
Do not limit review to syntax or whether the code “looks fine”.
Classify the reviewed change.
Possible scopes:
feature
bugfix
refactor
test-only
docs-only
schema
CLI
JSON output
file write behavior
workspace layout
generated context
runtime safety
DDT / Diary
release / packaging
unknown
If scope is unclear, inspect the diff or changed files.
Do not assume low risk without evidence.
Return one of:
Contract Impact: yes
Contract Impact: no
Contract Impact: uncertain
Treat as Contract Impact when the change affects:
behavior
args
returns
raises
schema
side effects
state changes
idempotency
security
permission
persistence
event shape
database shape
export format
CLI output
JSON output
file write target
exit code
configuration behavior
migration behavior
external-visible result
user-visible result
If uncertain, inspect relevant contract sources.
Do not silently treat uncertainty as no impact.
Use .harbor/policy.yaml as the source of truth when available.
Default strictness:
strict
standard
light
Treat as strict when the change touches:
public API
public CLI
public JSON output
core schema
Pydantic model used externally
OpenAPI schema
parser
export
file writeback
review pipeline
workflow node
auth / permission / security
migration
CI/CD
critical path
user-visible result generation
external integration
tool schema
MCP tool schema
workspace layout
generated context behavior
If a helper appears light but is used by a strict path, upgrade the review strictness.
Check:
does the code implement the intended behavior?
does it handle expected inputs?
does it handle invalid inputs?
does it preserve invariants?
does it preserve idempotency if promised?
does it introduce hidden side effects?
does it change file writes?
does it change external calls?
does it change user-visible behavior?
does it change errors or exceptions?
does it change path handling?
does it change deterministic output?
Mark confirmed issues only when there is evidence.
Mark possible issues when evidence is incomplete but risk is real.
Inspect relevant contract sources:
docstring
type hints
Pydantic model
FastAPI route
OpenAPI schema
TypeScript type
Zod schema
database migration
event schema
CLI schema
MCP tool schema
JSON output contract
snapshot output
golden file
tests / fixtures
public behavior
user-visible behavior
file write behavior
Look for:
implementation differs from docstring
schema differs from implementation
tests verify old behavior
CLI output changes without contract update
JSON output changes without schema or test update
file write behavior not declared
side effects not documented
Raises path differs from contract
strictness metadata missing or wrong
Use labels:
Confirmed Semantic Drift
Possible Semantic Drift
Contract Gap
Schema Gap
Test / DDT Gap
Check:
are relevant tests present?
do tests verify the contract or only current implementation?
are strict targets bound to explicit l3_version?
is strategy="latest" used on strict targets?
were snapshots updated without contract reasoning?
were fixtures updated blindly?
are failure paths covered?
are side effects covered?
are CLI / JSON output contracts covered?
are file write behaviors covered?
For strict targets, this is forbidden:
@harbor_ddt_target("module.func", strategy="latest")
Use label:
DDT Violation
when found.
Use:
DDT False Green Risk
when tests may pass while the contract remains broken.
Check whether the change or suggested fix involves:
deleting files
batch-moving files
.env or secrets
migrations
CI/CD
Docker / deployment
dependencies
destructive commands
git push
git reset --hard
production config
auth / permission / billing
user data handling
.harbor/*.yaml
generated skills
external network access
If yes, classify safety:
ALLOW
ASK
DENY
Use .harbor/safety.yaml as source of truth when available.
Use harbor-safety-preflight before performing risky operations.
Do not silently approve or perform high-risk actions.
Generated context lives under:
.harbor/views/**
Check whether the change affects:
module boundaries
project structure
public contracts
docstrings
tests / DDT
CLI behavior
JSON output
workspace structure
module responsibilities
debug workflow
review workflow
If yes, generated context may need refresh.
Typical commands:
harbor project structure --write
harbor docs --changed --write
harbor module seal --changed --write
harbor finish --sync-context
harbor stale
harbor doctor
Do not manually edit .harbor/views/** as project truth.
Use label:
Generated Context Gap
when generated context is likely stale or missing.
Diary records why important changes happened.
Canonical Diary path:
.harbor/diary/YYYY-MM.jsonl
Diary Draft may be needed for:
Contract Change
Breaking Change
schema change
CLI behavior change
JSON output change
workspace layout change
migration behavior change
public API change
DDT strategy change
runtime safety policy change
non-obvious compatibility decision
important bugfix
architecture decision
release-relevant decision
Use label:
Diary Gap
when a decision should be recorded but no Diary Draft or entry exists.
Do not claim Diary was written unless it was actually written.
A breaking change occurs when existing users, callers, scripts, agents, tests, or downstream systems must change.
Check whether the change:
removes public field
renames JSON key
changes CLI output format
changes exit code
changes schema requirement
changes file path
changes migration behavior
changes public exception behavior
changes compatibility behavior
If risk exists, mark:
Breaking Change: yes / uncertain
and explain affected consumers.
Do not hide uncertain breaking risk.
The canonical Harbor workspace is:
.harbor/
Check whether the change affects:
.harbor/rules/**
.harbor/views/**
.harbor/diary/**
.harbor/reports/**
.harbor/cache/**
.harbor/state/**
.harbor/exports/**
.agents/skills/**
AGENTS.md managed block
Workspace layout changes are Contract Impact.
harbor workspace migrate --dry-run must remain read-only.
Do not assume harbor workspace migrate --write exists.
Use:
Workspace Contract Risk
when workspace behavior is changed without tests, docs, or safety reasoning.
Use these labels.
Confirmed Issue
Possible Issue
Possible Semantic Drift
Confirmed Semantic Drift
Contract Gap
Schema Gap
Test / DDT Gap
DDT Violation
DDT False Green Risk
Runtime Safety Risk
Generated Context Gap
Diary Gap
Breaking Change Risk
Workspace Contract Risk
Tool Honesty Issue
Suggested Improvement
Definitions:
Confirmed Issue:
Clear bug or violation with evidence.
Possible Issue:
Risk is plausible but evidence is incomplete.
Possible Semantic Drift:
Implementation may differ from contract, but more context is needed.
Confirmed Semantic Drift:
Implementation and contract clearly disagree.
Contract Gap:
Important behavior exists without clear contract.
Schema Gap:
Schema and implementation disagree or schema is missing.
Test / DDT Gap:
Contract behavior lacks adequate test coverage.
DDT Violation:
DDT rule is clearly violated, such as strategy="latest" on strict target.
DDT False Green Risk:
Tests may pass while the intended contract is broken.
Runtime Safety Risk:
Operation or change may be unsafe without confirmation.
Generated Context Gap:
.harbor/views/** likely needs refresh.
Diary Gap:
Important decision lacks Diary Draft or entry.
Breaking Change Risk:
Existing consumers may need migration.
Workspace Contract Risk:
.harbor workspace behavior changed without governance.
Tool Honesty Issue:
Result claims execution without evidence.
Suggested Improvement:
Non-blocking improvement.
Use:
high
medium
low
Use high severity for:
data loss risk
security risk
secret exposure
breaking public contract
strict DDT violation
migration risk
auth / permission / billing risk
destructive operation risk
false claim that tests passed
Use medium severity for:
contract gap in standard path
missing tests for meaningful behavior
possible semantic drift
generated context stale after meaningful change
Diary missing for important but non-critical decision
Use low severity for:
small clarity issue
minor missing edge case
non-blocking style issue
light helper test suggestion
documentation improvement
Suggest or run only when appropriate.
During review:
harbor checkpoint
harbor stale
harbor doctor
For generated context changes:
harbor finish --sync-context
For JSON / CI-style checks:
harbor stale --format json
harbor doctor --format json
For workspace diagnostics:
harbor workspace inspect
harbor workspace migrate --dry-run
harbor workspace migrate --dry-run is read-only and not part of the default coding workflow.
Do not run unless relevant.
Do not run the following unless the user explicitly requests it:
harbor log
harbor accept
harbor lock
harbor module promote-skill <module>
Never run harbor accept to silence unresolved drift.
Never claim Diary was written unless it was actually written.
Do not claim tests, lint, type checks, builds, CI, or Harbor commands were run unless actually run and observed.
Allowed wording:
已运行，结果是...
未运行，建议运行...
当前环境无法运行...
我只做了静态审查，未执行命令...
Forbidden wording when not executed:
测试已通过
harbor doctor 通过
stale check 已清理
lint 无问题
CI 会通过
If a command cannot be run, report:
Command not run:
Reason:
Risk:
Recommended next command:
Use Tool Honesty Issue when reviewing output that claims unobserved execution.
Start every review with:
Review Summary:
- Overall assessment:
- Contract Impact:
- Strictness:
- Tests / DDT:
- Runtime Safety:
- Generated Context:
- Diary:
- Breaking Change Risk:
Then list findings.
For each finding:
Finding <N>: <Label>
Location:
- <file>:<line or symbol>
Evidence:
- <specific evidence from code, contract, schema, test, generated view, or policy>
Risk:
- <why this matters>
Recommended fix:
- <specific fix>
Severity:
- high / medium / low
If no blocking Harbor-spec issue is found, say:
No blocking Harbor-spec issue found.
Still mention non-blocking suggestions if useful.
Use concise but explicit summaries.
Example:
Review Summary:
- Overall assessment: No blocking issue found, but one DDT gap should be addressed before release.
- Contract Impact: yes, JSON output shape changed.
- Strictness: strict, because public CLI JSON is affected.
- Tests / DDT: missing stable key test.
- Runtime Safety: no high-risk operation detected.
- Generated Context: refresh likely needed.
- Diary: Diary Draft recommended due to public JSON contract change.
- Breaking Change Risk: uncertain, downstream scripts may depend on previous key name.
Evidence may come from:
source code
tests
fixtures
schemas
docstrings
type hints
.harbor/policy.yaml
.harbor/safety.yaml
.harbor/rules/*.md
.harbor/views/**
.harbor/diary/**
command output actually observed
Generated context may support review orientation.
Generated context should not override source code, tests, schemas, policy, or diary.
Avoid:
reviewing only code style
ignoring contract impact
ignoring tests / DDT
ignoring runtime safety
ignoring generated context drift
ignoring Diary need
treating generated context as source of truth
treating skills as source of truth
marking Contract Impact: no without reasoning
missing strict target strategy="latest"
failing to flag JSON output changes
failing to flag file write behavior changes
failing to flag hidden breaking changes
claiming tests passed without execution evidence
A Harbor code review asks:
Does the change keep implementation, contracts, tests, generated context, decision memory, and safety boundaries aligned?
A change can be syntactically correct and still fail Harbor review if it creates drift.
Use before running risky commands, deleting or moving files, modifying secrets, changing migrations, installing dependencies, modifying CI/CD, changing production configuration, modifying .harbor/*.yaml, writing generated skills, or any operation with runtime safety risk under Harbor-spec.
Use when refreshing Harbor generated context under .harbor/views/**, including project structure, L2 README, Module Capsule, stale checks, doctor checks, or finish --sync-context after code, contract, schema, CLI, JSON, test, DDT, or module boundary changes.
Use when adding or modifying APIs, schemas, public functions, parser/export/writeback behavior, workflow nodes, CLI behavior, JSON output, file write behavior, or any change with possible Contract Impact under Harbor-spec.
Use when updating tests, DDT bindings, l3_version, contract versions, Diary Drafts, changelogs, release notes, or important architectural decisions under Harbor-spec.
Use when inspecting Harbor workspace layout, planning workspace cleanup, validating .harbor/ canonical workspace health, reviewing managed files, running workspace inspect, or previewing read-only workspace migration plans with migrate --dry-run.
Use when debugging, reviewing, or refactoring the harbor/core module. It routes the agent to the module capsule instead of loading the whole repository.
استنادا إلى تصنيف SOC المهني
