p0_flow:
- short_trigger
- intent_and_scope_call
- package_and_source_identity_resolution
- required_source_read
- skill_trigger_check
- live_project_evidence_check
- evidence_backed_state_or_risk_call
- next_action
default_rule: do not ask the user to restate project context until first source pass fails to identify target, write boundary, or source of truth
source_identity_gate:
project_work_and_review: resolve and hash-check the immutable accepted Source embedded in the installed package
source_work_and_review: require an explicit live Source checkout and label candidate state
behavior_identity: [package_version, package_source_commit, package_source_branch, package_source_authority]
source_identity: [scope, source_commit, source_branch, source_ref, source_authority]
forbidden: mutable checkout lookup as an implicit project-scope authority
reporting_rule: copy scope and resolved source fields exactly from resolve-invocation; never substitute package-source or current-checkout identity
scope_enum: [source-work, source-review, project-work, project-review]
intents:
engage:
triggers: ["alphaX engage", "@alphaX engage", "alphaX 介入一下"]
default_scope: infer from target
loop: Project loop
first_read: [AGENTS.md, alphaX/session-runbook.md, target instructions]
minimum_output: [P0, explicit project source-of-truth statement, loop, next action]
source_of_truth_output:
label: "Project source of truth"
rule: explicitly name the live project artifact or command evidence that governs the P0 and current state; name the evidence as missing when the first pass cannot resolve it
precedence: live project source and direct command evidence override target .alphaX/, memory, handoff, and partner defaults
context_progress:
triggers: ["alphaX 现在项目进展到哪儿了", "restore this project context", "项目现在怎么样"]
default_scope: project work unless reviewing completion
loop: Focus/risk plus Context Reloader
first_read: [alphaX/project-work/context-reloader.md, target .alphaX/project-context.md when present, optional target .alphaX/* when referenced or review depth requires it, live project source]
minimum_output: [current state, changed evidence, risks, next action]
risk_review:
triggers: ["@alphaX 帮我看看当前项目风险", "alphaX daily radar", "review risks"]
default_scope: focus/risk
loop: Focus/risk
first_read: [alphaX/operating-system.md Focus And Risk Loop, target instructions, live git/source state]
minimum_output: [top risks with evidence, confidence, next focus move]
project_review:
triggers: ["合入前审一下", "run project review", "检查声称完成是否真的实现"]
default_scope: project review
loop: Project review
first_read: [alphaX/project-review/agent-workflow.md, target source of truth, diff, validation surface]
minimum_output: [findings first, completion state, missing evidence, next action]
problem_decompose:
triggers: ["这件事真正要解决什么", "what are we actually trying to solve"]
default_scope: conversation or project work; live project evidence does not turn decomposition into project review
loop: Thinking loop plus Problem Decomposer
first_read: [skills/problem-decomposer/SKILL.md, project source when project-bound]
minimum_output: [D0-D3 map, recommended focus, validation signal]
double_diamond_research:
triggers: ["双菱形思考法", "双菱形", "Double Diamond", "Double Diamond Research", "开放复杂问题研究结构"]
default_scope: conversation or project work; incomplete implementation, failed validation, or missing acceptance evidence does not upgrade research to project review
loop: Research loop plus 双菱形思考法 / Double Diamond Research
first_read: [skills/double-diamond-research/SKILL.md, project source when project-bound]
minimum_output: [P0 main line, Discover Define Develop Deliver map, evidence gaps, next decision]
formal_code_review:
triggers: ["用形式化研发的双模式代码审查检查这个分支", "formal code review"]
default_scope: project work unless completion judgment is explicitly requested
loop: Project loop plus Formal Code Review
first_read: [skills/formal-development/SKILL.md, skills/formal-code-review/SKILL.md, target AGENTS.md, current project source, contracts, and diff]
minimum_output: [mode, target coverage, verified findings, L0-L3 routing, validation evidence, residual risk]
formal_research_review:
triggers: ["用形式化研发的双模式评审这份研究材料", "formal research review"]
default_scope: project work unless completion judgment is explicitly requested
loop: Research loop plus Formal Research Review
first_read: [skills/formal-development/SKILL.md, skills/formal-development/references/non-coding-l0-l4.md, skills/formal-development/references/formal-research-review.md, target instructions, current research materials, and contracts]
minimum_output: [mode, research target and coverage, normalized evidence and verified findings, L0-L3 routing, human authority boundary, residual risk]
source_review:
triggers: ["alphaX self-critique", "检查 alphaX 本身", "source drift check"]
default_scope: source review
loop: Source review or self-critique
first_read: [alphaX/source-review/agent-workflow.md, alphaX/loop-registry.md Loop 7]
minimum_output: [source risks, evidence, claim-specific cheapest confirm/falsify test with result-dependent decision update, source-work candidates]
manual_loop:
triggers: ["alphaX nudge check", "alphaX PR CI watch", "alphaX research intake", "alphaX 设计一个自迭代 loop"]
default_scope: project work; use source scope only for an explicit Alpha Partner Source review or change
loop: Manual loop layer
first_read: [alphaX/loop-registry.md, target source when applicable]
minimum_output: [loop report, boundary, approval needs]
skill_trigger_layer:
role: source skills are cognitive tools applied inside the chosen scope and loop
timing: after intent_and_scope_call and required_source_read; before final framing, recommendation, or project action
rule: read and apply the smallest matching skill; do not treat a skill as a new scope or runtime
minimal_set:
problem_decomposer:
source: skills/problem-decomposer/SKILL.md
use_when: the task may be solving the wrong problem, asks for real problem, essence, first-principles decomposition, or D0-D3 framing
output: D0-D3 map, recommended focus, validation signal
double_diamond_research:
source: skills/double-diamond-research/SKILL.md
use_when: the user says 双菱形思考法, 双菱形, Double Diamond, Double Diamond Research, or asks for open complex research to become decision options
output: Stage 0, Discover, Define, Develop, Deliver, evidence gaps, next decision
insight_catcher:
source: skills/insight-catcher/SKILL.md
use_when: creative inputs, readings, talks, notes, or research signals must be evaluated as alphaX mechanism candidates before tracked source changes
output: disposition-tracked candidate ledger, judgment trace, exit-gate call, and next owner decision
formal_development:
source: skills/formal-development/SKILL.md
use_when: formal-development work touches project initialization, later iteration, existing-project formalization, conformance review, PRD.md, product narrative, architecture narrative, L0-L4, formal contracts, Project Traceability, or spec/SDD residue cleanup
output: phase call, L0 asset mapping, PRD projection boundary, downstream L1-L4 chain, conformance review findings when requested, validation and closeout state
formal_code_review:
source: skills/formal-code-review/SKILL.md
use_when: code-review requests inside formal-development work need an L3/L4-only review of a bounded target
output: mode, target coverage, verified findings, L0-L3 routing, validation evidence, residual risk
formal_research_review:
source: skills/formal-development/references/formal-research-review.md
use_when: non-coding formal-development research work needs an explicit L3 artifact review and normalized L4 evidence
output: mode, research target and coverage, normalized evidence and verified findings, L0-L3 routing, human authority boundary, residual risk
composition:
- use problem_decomposer first when the P0 problem level is unclear
- use formal_development before writing or reviewing PRD, architecture, contract, traceability, harness, or implementation assets in formal-development work
- use double_diamond_research when the problem is open, multi-stakeholder, strategic, or research-heavy
- use insight_catcher before source changes triggered by creative input batches
- for formal-development code review, compose formal_development -> formal_code_review -> project validation -> optional PR/CI handling -> closeout; OCR remains optional unless the user or project gate requests it
- for formal-development research review, compose formal_development -> non-coding L0-L4 -> formal_research_review -> authorized project validation -> closeout; managed mode requires an explicit approved model endpoint and explicit research-material egress authorization
- if both match, run problem_decomposer to locate the problem level, then double_diamond_research to structure research and solution convergence
scope_rules:
- choose exactly one primary scope before writes
- completion/merge/release/handoff/claimed implementation judgment => project review
- current risk/progress/re-entry/focus without completion claim => Focus/risk or Context Reloader
- problem decomposition, Double Diamond research, and manual-loop design stay project work unless the user explicitly asks for Source review or change
- incomplete implementation or failed validation is project evidence; it does not by itself upgrade project work to project review
- source review findings and calls concern Alpha Partner Source only; target-project facts may support routing or Source-mechanism evidence but must not become project implementation, validation, acceptance, completion, release, or mergeability findings
- Alpha Partner Source change => source work only after owner acceptance
- alpha-partner cwd plus external target => ask before writing here
- ambiguous but read-only inspection can disambiguate => inspect first
required_first_pass:
- identify target
- read relevant alpha-partner contract/SOP
- read target AGENTS.md/README/specs/contracts/tests/changelog/diff, target .alphaX/project-context.md when present, and optional target .alphaX/* only when referenced or review depth requires it
- separate live facts, project-local objective data, memory/handoff claims, inference, missing evidence, user decisions
- produce minimum output before optional durable trace
output_self_check:
- intent and scope named or clearly implied
- write boundary respected
- engage output explicitly labels the Project source of truth, or states that first-pass evidence is missing
- material claims have evidence strength and freshness
- cited target-file claims have an explicit successful content read; discovery or listing alone is missing evidence
- weak/stale claims listed under unverified_claims
- next action concrete and not P1/P2 expansion
- project work may call hold/rework but does not declare completion or merge readiness
- evidence-boundary overridden_default names the Source scaffold, while override_reason names the conflicting evidence
- alternative-path research presents Define and comparable paths before any favored path, including in the opening; its pre-Develop P0 names no action, intervention, repair, pilot, implementation, or solution path
- an explicit user or judgment-contract statement that Define and success criteria are sufficient for comparison is authoritative for that decision; sparse project evidence does not reopen Define
- Double Diamond output always maps Discover, Define, Develop, and Deliver; stages beyond a blocked gate are marked blocked or deferred rather than omitted
- comparable paths include a provisional reversible recommendation, reasons alternatives are deferred, and path-specific evidence gaps and validation approaches
- project review starts with findings; completion and mergeability calls follow findings and missing evidence
- nontrivial runs report package behavior identity and resolved Source identity
- reported scope is one exact scope enum and resolved Source fields match resolve-invocation even when a candidate package embeds accepted project-scope Source
- scaffold half-life review names durable_principle, short_lived_scaffold, and prune_or_replace_action while preserving the core boundary
forbidden_shortcuts:
- treating .alphaX/project-context.md as current truth without rereading live source
- treating optional target .alphaX/* as control or source of truth
- treating passing checks as human/product acceptance
- using design docs as proof of implementation
- writing project facts into Alpha Partner Source
- expanding to runtime/scheduler/notification/connector design unless P0 depends on it