| type | Guide | |||
|---|---|---|---|---|
| title | alpha-partner | |||
| description | Public entrypoint for the alphaX Markdown-first source repository. | |||
| tags |
|
alpha-partner is the local source/meta base for alphaX, a personified
project-collaboration function.
The working call sign is alphaX.
Chinese guide: docs/README.zh-CN.md.
This is not an application or hosted service. Runtime carriers execute alphaX by
reading this source; they are not the product boundary. A dedicated always-on
runtime could make alphaX entity-like, but this repository's product boundary is
the function contract and project-local .alphaX/ mappings.
Product goal: help agents autonomously solve more high-value research and development problems under explicit evidence, scope, data, and human authority boundaries.
Source self-iteration goal: make alphaX itself better at serving that product goal. This is an internal source optimization target, not the product goal itself. It has two axes:
- raising the intelligence ceiling of future agents;
- extending the half-life of the source assets that make collaboration trustworthy.
Raising the intelligence ceiling means helping stronger agents do better work: reframe problems, weigh evidence, surface stronger counterarguments, identify wrong-layer work, and know when a default process should be overridden. Extending asset half-life means separating durable collaboration principles from short-lived carrier scaffolding, so prompts, templates, trigger phrases, and verifier strings can change without weakening the product core.
The evidence boundary follows the same split: product-goal progress needs applied research or development problem evidence; source self-iteration progress needs judgment replay, applied runs, or half-life evidence tied to source mechanisms.
The core layer is the intelligence amplifier: evidence fields, scope and write boundaries, data and decision boundaries, live-source priority, and wrong-layer guards. The outer layer is a lower-bound stabilizer: routing defaults, read orders, fixtures, templates, and current verification mechanics. Source work keeps the core thin and hard while making scaffolding explicit, depreciable, and replaceable.
alphaX is a reusable collaboration function for AI product and software engineering work. It can be delivered as a Skill, MCP surface, subagent, or supervisor-agent action, but this repository is the Markdown-first source.
It focuses on:
- project re-entry and risk review;
- explicit source-skill triggering for problem decomposition and 双菱形思考法;
- project review before handoff, merge, freeze, or release;
- project lifecycle hygiene and local
.alphaX/compaction signals; - source review for alphaX contract and mechanism drift;
- checkpoint-based memory evaluation without a runtime or RPC dependency;
- source-backed judgment formation;
- contract-first engineering;
- verifiable implementation;
- external calibration;
- compact decision and evidence records.
Every alphaX run must classify its scope before writing files:
source work: change Alpha Partner Source or the alphaX function itself.source review: review Alpha Partner Source, alphaX mechanisms, or local.alphaX/process/governance data. Report-first; no tracked source edits.project work: help a concrete project, document, product question, research task, or engineering problem.alpha-partneris read-only source.project review: report-first evidence review for one concrete project before handoff, merge, release, freeze, or claimed completion.
The scope guard is the write boundary attached to the chosen scope. Project
work and project review write only to the project being helped, that project's
ignored .alphaX/ surfaces allowed by docs/local-alphaX-schema.md, an OS
temporary directory, or the conversation response.
Source review may write ignored source-facing notes under the source checkout's
.alphaX/process/; tracked source edits require source work.
Read AGENTS.md first, then follow alphaX/session-runbook.md.
For a new local checkout:
bash scripts/init-local-alphaX.sh
bash scripts/verify-local-alphaX.sh
node scripts/generate-alphaX-indexes.mjs --check
bash scripts/verify-alpha-source.shOptionally add private literals to .alphaX/local/private-patterns.txt; the
source verifier will fail if they appear in the GitHub-tracked tree.
docs/checkpoint-memory-evaluation-prd.md
defines how alphaX checks whether remembered or project-local state is current,
evidence-backed, and action-guiding. P0 must work from live source, local
.alphaX/, explicit user decisions, command output, artifacts, and available
memory notes; no runtime service, RPC helper, scheduler, or backend is required.
Tracked Markdown and newly initialized ignored .alphaX/ data use the
alphaX OKF Markdown Profile: small YAML
frontmatter, generated index.md navigation, relative internal links, and
explicit citations for external-source-backed claims.
Refresh source indexes with:
node scripts/generate-alphaX-indexes.mjs --write
node scripts/generate-alphaX-indexes.mjs --checkThe Agent-facing trigger contract and regression fixtures live in
docs/agent-invocation-contract.md,
docs/agent-trigger-fixtures.json, and
docs/agent-trigger-fixtures.md.
For repeated project work:
alphaX engage.
For stepping back from a task to the real problem:
What are we actually trying to solve?
For open-ended complex research:
Use 双菱形思考法 / Double Diamond Research for this problem.
For portfolio focus and risk:
alphaX daily radar
For project review:
alphaX run project review before handoff.
Triggers are semantic, not literal. User prompts in another language should map to the same alphaX behaviors when the intent is the same.
alphaX/: behavior and scope SOPs.templates/: report packets and project-local mapping templates.docs/: evidence, schema, asset boundary, OKF profile, anddocs/README.zh-CN.md.scripts/: bootstrap, verification, context snapshot, applied-run candidate detection, and generated indexes.assets/: shareable visual assets, includingassets/icon.png.
Generated index.md files provide the detailed navigation.
Review has two contracts with different goals:
alphaX/source-review/agent-workflow.md: improve alphaX source and mechanisms by finding contract drift, stale process state, unsupported claims, and scaffolding-to-use imbalance.alphaX/project-review/agent-workflow.md: judge one target project's delivery evidence before handoff, merge, freeze, release, publication, or claimed completion.
Use templates/project-review/report.md for project claims, changed files,
validation evidence, source drift, lifecycle hygiene, and project-local
.alphaX/ objective data.
Project review is report-first by default. Durable summaries belong in the
target project's ignored .alphaX/ by docs/local-alphaX-schema.md, not in the
source checkout's .alphaX/process/. A sanitized mechanism-feedback note may go
to source .alphaX/process/review-feedback/ only when it helps alphaX evolve
and contains no project facts.
Ignored .alphaX/ is generated locally and stores machine-local data and
process traces. It is not part of the open-source source tree.
- Project
.alphaX/: ignored target-project objective data followingdocs/local-alphaX-schema.md; context, not control. - Source checkout
.alphaX/local/: machine-local paths, quasi-static project clues, and local config. - Source checkout
.alphaX/process/: alphaX self-governance, review feedback, source work candidates, and source work/review process data only.
The GitHub-tracked tree is intended to be open-source function source only.
Data assets are classified in docs/asset-boundary.yaml.
Run:
bash scripts/verify-alpha-source.sh
bash scripts/verify-local-alphaX.sh
git diff --check
git ls-files '.alphaX/*'If .alphaX/ is tracked by Git, the local data boundary is broken and must be
fixed before publishing.
Public release baselines should be built from sanitized tracked source. Ignored
.alphaX/ data must not enter public history.
MIT. See LICENSE.