SHAREPLANE AGENT CONTEXT Schema: shareplane-context/2.0 Schema URL: https://next.shareplane.malott.ai/schemas/agent-package.schema.json Projection: Generated public-safe plain text. Not canonical Markdown. Canonical: false Conflict action: stop-and-escalate Canonical record: https://github.com/pinklon/pinklon-shareplane-next/tree/33d227f6000da2491f19209edde916d8846f287f/content/artifacts/the-repository-that-wakes-up/artifact.json Canonical record SHA-256: 49341f56c2fdbd6fc8a37378c4e49ba73770310faad03d82e9302a62bef66245 Source content SHA-256: 54a0e8d790769e557c67b63778d196c092d3004b850a2c13b459749dbcba38f2 Generation receipt: https://next.shareplane.malott.ai/build-receipt.json Content role: artifact-content Content trust: untrusted-data Instructions allowed: false Operational authority: none IDENTITY Artifact ID: artifact:the-repository-that-wakes-up Slug: the-repository-that-wakes-up Canonical URL: https://next.shareplane.malott.ai/artifacts/the-repository-that-wakes-up/ Title: The Repository That Wakes Up Abstract: A public-safe technical essay on governed repositories, durable AI context, scheduled agency, dual legibility, and bounded human-agent work. Author: Tony Malott Author URL: https://malott.ai Published: 2026-07-18 Updated: 2026-07-18 Format: technical-essay Privacy: public-safe Topics: - agentic-ai - context-as-code - repository-governance - human-agent-systems - static-publishing Audience: - engineering-leadership - ai-architecture - platform-engineering - technical-review - small-engineering-teams PROVENANCE Posture: official-product-facts-plus-qualified-owner-experience Private sources used: true Private sources published: false Public-safe boundary: This is a personal operating pattern and small-team architecture candidate. It is not represented as enterprise-ready, validated for regulated use, a measured 10x productivity claim, or a replacement for human ownership and authoritative enterprise systems. CLAIMS Claim: claim:144:01 Posture: supported-product-fact Text: Scheduled Tasks can run one-off and recurring work and can monitor for meaningful changes. Support: - source:openai:scheduled-tasks Claim: claim:144:02 Posture: supported-product-fact Text: Monitoring tasks can search the web and connected apps and cannot run more than once per hour. Support: - source:openai:scheduled-tasks Claim: claim:144:03 Posture: supported-product-fact Text: Active task limits vary by plan and tasks created inside projects cannot access project files. Support: - source:openai:scheduled-tasks Claim: claim:144:04 Posture: supported-product-fact Text: Codex automations support scheduled focused workflows and local automations depend on the local machine and Codex client being available. Support: - source:openai:codex-automations Claim: claim:144:05 Posture: supported-product-fact Text: ChatGPT apps can search and reference external sources, and some apps can take actions. Support: - source:openai:apps Claim: claim:144:06 Posture: supported-product-fact Text: The GitHub app can pull live repository content, subject to authorization, plan, administrator, and indexing constraints. Support: - source:openai:github Claim: claim:144:07 Posture: supported-product-fact Text: Agent action surfaces create sensitive-data and prompt-injection risks and use confirmation or supervision safeguards for higher-impact actions. Support: - source:openai:agent Claim: claim:144:08 Posture: qualified-owner-thesis Text: A governed repository can provide durable context, explicit authority, version history, validation, and recoverable work state for Tony's operating pattern. Support: - source:owner:experience - source:github:issue-144 Claim: claim:144:09 Posture: qualified-owner-thesis Text: Dual legibility allows humans and agents to work from the same explicit, versioned representation. Support: - source:owner:experience - source:github:issue-144 Claim: claim:144:10 Posture: qualified-owner-experience Text: In Tony's experience, reducing coordination and consistency work can feel like doing the job of ten people, but this is not a measured universal productivity claim. Support: - source:owner:experience - source:github:issue-144 Claim: claim:144:11 Posture: qualified-applicability-claim Text: The pattern is strongest for individuals or bounded engineering teams already comfortable with Git and explicit repository discipline. Support: - source:owner:experience - source:github:issue-144 Claim: claim:144:12 Posture: required-boundary Text: The pattern is not established as enterprise-ready or validated for regulated use. Support: - source:github:issue-144 PUBLIC SOURCES Source: source:openai:scheduled-tasks Title: Scheduled Tasks in ChatGPT Type: public-source Role: official-product-fact Description: Official product facts; behavior and availability may change. Locator: https://help.openai.com/en/articles/10291617-tasks-in-chatgpt Source: source:openai:codex-automations Title: Codex automations Type: public-source Role: official-product-fact Description: Scheduled engineering workflows; local execution depends on the client and machine. Locator: https://openai.com/academy/codex-automations/ Source: source:openai:apps Title: Apps in ChatGPT Type: public-source Role: official-product-fact Description: Connected-data and action surface; availability varies. Locator: https://help.openai.com/en/articles/11487775-connectors-in Source: source:openai:github Title: Connecting GitHub to ChatGPT Type: public-source Role: official-product-fact Description: Repository access depends on authorization, plan, administration, and indexing. Locator: https://help.openai.com/en/articles/11145903-connecting-github-to-chatgpt Source: source:openai:agent Title: ChatGPT agent Type: public-source Role: official-product-fact Description: Sensitive-data and prompt-injection risk remain material. Locator: https://help.openai.com/en/articles/11752874-chatgpt-agent RELATIONSHIPS Relationship: relatedTo Target: artifact:your-work-is-evaporating Label: Your Work Is Evaporating Description: A SharePlane article defining work evaporation as the loss of recoverable context around AI-assisted work and presenting a governed continuity system that extracts decisions, binds evidence, records provenance, and makes work reusable beyond the originating session. Posture: declared-owner-approved-related-work Evidence: Issue #144 owner approval and PR #172 related-work correction record. Relationship: relatedTo Target: artifact:how-shareplane-works Label: How SharePlane Works Description: SharePlane is a static publishing library and a governed context-as-code operating model. This artifact explains how versioned authority, human judgment, bounded agents, deterministic generation, validation, receipts, and exact-head publication turn rough ideas into durable public work. Posture: declared-owner-approved-related-work Evidence: Issue #144 owner approval and PR #172 related-work correction record. Relationship: relatedTo Target: artifact:your-second-brain-is-not-a-production-architecture Label: Your Second Brain Is Not a Production Architecture Description: A source-derived architecture explainer separating personal second-brain markdown workflows from production agent context-plane requirements. Posture: declared-owner-approved-related-work Evidence: Issue #144 owner approval and PR #172 related-work correction record. ARTIFACT CONTENT [HEADING 2] The Repository That Wakes Up [HEADING 2] TL;DR [PARAGRAPH] A frontier model can reason brilliantly and still forget the work the moment the conversation ends. A governed repository changes that. It gives the model durable context, explicit source authority, version history, relationships, validation, and an exact place to resume. Scheduled tasks and Codex automations can return to that state, reconcile what changed, and advance bounded work without forcing the human to reconstruct the project from chat history. [PARAGRAPH] This is not a universal architecture. It is a method I use because I already work comfortably with Git, issues, pull requests, CI, and explicit review gates. The strongest fit is an individual or a small engineering team with the same habits. At enterprise scale, different systems may need to remain authoritative for identity, records, access control, workflow, and deployment. [PARAGRAPH] The point is simple: the chat is the workspace. The repository is the memory. Commit on the way in. Commit on the way out. [PARAGRAPH] Continue for the operating pattern, the leverage it creates, and the ways it can fail. [HEADING 2] 1. A brilliant stranger is still a stranger [PARAGRAPH] A frontier model without maintained context is a brilliant stranger. It may understand the technology, recognize the pattern, and produce a strong answer. It does not automatically know what I decided last week, which source governs, which branch is safe, why a design was rejected, or what must not change. [PARAGRAPH] That gap is usually disguised as a prompting problem. It is more often a state-management problem. [PARAGRAPH] Chats are excellent working surfaces. They are weak system-of-records. They accumulate half-decisions, provisional language, abandoned alternatives, and critical details trapped behind phrases such as “we already settled this.” A human can sometimes reconstruct the thread because memory fills the gaps. An agent sees fragments unless the important result leaves the conversation. [PARAGRAPH] My operating rule is therefore blunt: commit on the way in, commit on the way out. [PARAGRAPH] On the way in, establish the current baseline, the governing issue, the exact authority, the allowed paths, the validation contract, and the stop conditions. On the way out, preserve the meaningful result, the evidence, the unresolved questions, the receipt, and the next executable action. The repository does not need every stray thought. It needs the significant state required to resume the work without folklore. [PARAGRAPH] This turns version control into something broader than code history. It becomes a history of maintained understanding. [HEADING 2] 2. Dual legibility is the leverage mechanism [PARAGRAPH] The repository works only when it is legible to both humans and agents. [PARAGRAPH] Human legibility means an engineer, reviewer, owner, or successor can inspect the material, challenge it, and understand why the system behaves as it does. Agent legibility means a model can locate authority, relationships, constraints, open work, and required actions without ingesting an entire archive or guessing from anonymous chunks. [PARAGRAPH] I call this dual legibility. [PARAGRAPH] It is not about making every file tiny or writing robotic prose for machines. It is about stable identity, bounded documents, explicit metadata, clear authority ranking, progressive disclosure, and predictable paths. A short task should not require reading a hundred files. A consequential task should be able to follow the graph from a compact entry point into deeper evidence. [PARAGRAPH] The compact architecture looks like this: [LIST ITEM] canonical source carries meaning, ownership, and authority; [LIST ITEM] issues carry intent, state, decisions, and unresolved work; [LIST ITEM] branches and pull requests create bounded change surfaces; [LIST ITEM] validators and receipts prove what happened; [LIST ITEM] generated pages, indexes, packages, and feeds project the canonical source; [LIST ITEM] scheduled agents return to the same governed state rather than starting from zero. [PARAGRAPH] Retrieval still matters. Broad search and vector retrieval are useful when the corpus is large, weakly structured, or exploratory. But retrieval is not the same as maintained understanding. Relevance does not prove authority. Recency does not prove approval. A vector match does not tell an agent whether it may act. [PARAGRAPH] The repository gives retrieval something better to retrieve: explicit, versioned, governed context. [HEADING 2] 3. Scheduled agency closes the loop [PARAGRAPH] Durable context becomes far more useful when something can return to it. [PARAGRAPH] OpenAI’s current Scheduled Tasks support one-off and recurring work, plus monitoring that checks for meaningful changes and notifies the user when something is worth reporting. Tasks can search the web and connected apps, but they also have practical limits: they cannot run more than hourly, active-task limits vary by plan, and tasks created inside projects cannot access those project files. Codex automations are a separate, more focused mechanism for scheduled engineering workflows. OpenAI notes that local Codex automations work best when the machine is awake and Codex is running. [PARAGRAPH] Those details matter because “autonomous” is too often used as a decorative adjective. A useful agent loop needs a reliable state source, an execution surface, bounded authority, and a review destination. [PARAGRAPH] A typical flow in my own work starts with a raw thought. I capture it as an issue rather than trusting the chat to remain discoverable. The issue records the thesis, public-safe boundary, evidence needs, and what is not yet authorized. A scheduled reconciliation later reads the issue and repository state, detects what changed, and identifies the next safe action. An implementation agent creates an isolated branch, runs validations, and opens a draft pull request. The publishing compiler generates the page, graph, feed, context, package, and receipts. I review the exact immutable preview. Only then do I authorize the exact-head merge. [PARAGRAPH] The schedule is not the intelligence. The model is not the memory. The repository is not the agent. The value comes from connecting all three without letting any one of them impersonate the others. [HEADING 2] 4. Coherent scale changes the economics of one person [PARAGRAPH] The most surprising benefit is not faster writing or faster file editing. It is the ability to treat a corpus as a connected system. [PARAGRAPH] When ownership, semantics, relationships, validation, and change boundaries are explicit, one human working with capable agents can coordinate changes across documents, schemas, code, indexes, receipts, and generated projections without manually touching every surface. Change the canonical source once, trace the affected paths, regenerate the derived outputs, and verify the result. [PARAGRAPH] In my own work, this can feel like doing the job of ten people, not because the model replaces ten specialists, but because it collapses the coordination, search, rewriting, consistency checking, and mechanical update work that normally sits between them. [PARAGRAPH] That is an experiential statement, not a universal productivity benchmark. The expert judgment still has to exist. The authority still has to be real. The model can accelerate a bad architecture just as efficiently as a good one. [PARAGRAPH] The same structure that lets you update a hundred connected artifacts coherently can also let you damage a hundred connected artifacts coherently. Stale authority, weak validation, excessive standing context, or an incorrectly scoped agent can propagate coherent nonsense at extraordinary speed. Exact change boundaries, review, negative tests, and human judgment are what make the leverage usable. [PARAGRAPH] This is why I care so much about generated-output ownership. Generated files are not hand-merged as if they were competing opinions. Canonical source wins. Derived output is discarded and regenerated. If the source itself is ambiguous, the lane stops. A green validator cannot invent missing authority. [HEADING 2] 5. A team brain is a handoff system, not a replacement for people [PARAGRAPH] The pattern becomes more valuable when work must survive a handoff. [PARAGRAPH] Most team transitions are embarrassingly primitive. The successor receives a folder, a calendar invitation, several stale diagrams, and a series of conversations with people who each remember a different system. The real operating model remains trapped in habit and oral history. [PARAGRAPH] A well-structured repository can encode what the team owns, how work enters, which sources govern, why key decisions were made, what is active, what is blocked, which controls must pass, and how another engineer or agent resumes. It can preserve architecture, standards, procedures, risks, evidence, and exact continuation state in one reviewable history. [PARAGRAPH] That is what I mean by a team brain. The phrase is a metaphor for legible operating context, not a claim that people are disposable or that tacit expertise can be eliminated. [PARAGRAPH] Dual legibility is what makes the handoff practical. Humans and agents work from the same explicit representation. A successor can read it. An agent can traverse it. A reviewer can challenge it. A validator can test it. Git can show how it changed. [PARAGRAPH] This does not eliminate communication, leadership, peer review, or ownership. It reduces key-person dependency and makes the unavoidable conversations more productive because the baseline is inspectable. [HEADING 2] 6. Where this pattern fits, and where it does not [PARAGRAPH] I can assemble this personally because I control the repositories, subscriptions, permissions, tools, and operating rules. A small engineering team may be able to do something similar when its members already understand Git and agree on information architecture, ownership, maintenance, and review. [PARAGRAPH] That does not make the pattern enterprise-ready. [PARAGRAPH] Enterprises have different requirements for identity, record-level access, privacy, retention, deletion, data residency, segregation of duties, validated systems, transaction volume, concurrency, continuity, and platform ownership. Git may remain one authoritative surface while EDMS, ITSM, identity platforms, databases, deployment systems, security tooling, or formal approval systems govern other facts. The repository can reference and project those systems without pretending to replace them. [PARAGRAPH] Connected apps also widen the risk boundary. OpenAI’s current apps model can search and reference external data, and some apps can take actions. The GitHub app can pull live repository content, subject to plan availability, repository authorization, administrator approval, and GitHub indexing behavior. ChatGPT agent can act on websites, but OpenAI explicitly warns about sensitive data and prompt-injection risk and uses confirmations and supervision for higher-impact actions. Those are not minor footnotes. They are part of the architecture. [PARAGRAPH] This method is one operating pattern among several. Many people are converging on context-as-code, repository-native knowledge, agentic engineering, and compiled-memory ideas. I am documenting the combination I use, the discipline it requires, and the failures I keep finding. [PARAGRAPH] The system is still imperfect. Cloud deployments fail. Credentials expire. workflows race. Packages go missing. Authorities become stale when main advances. The answer is not to hide that friction. The answer is to make the state and failure visible enough that the next human or agent can recover without starting over. [PARAGRAPH] A frontier model gives me reasoning power. Maintained context gives that power direction. Versioned authority gives it boundaries. Scheduled agency gives it recurrence. Review and validation keep the loop honest. [PARAGRAPH] This article is not the declaration of a finished system. It is a commit. [HEADING 2] Operating loop [LIST ITEM] Preserve authority before action. [LIST ITEM] Make context legible to people and agents. [LIST ITEM] Let scheduled agency return to versioned state. [LIST ITEM] Scale coordination without scaling ambiguity. [LIST ITEM] Treat handoff as a system property. [LIST ITEM] Keep enterprise boundaries explicit. PUBLIC SURFACES Human page: https://next.shareplane.malott.ai/artifacts/the-repository-that-wakes-up/ Metadata JSON: https://next.shareplane.malott.ai/artifacts/the-repository-that-wakes-up/artifact.json Receipt: https://next.shareplane.malott.ai/artifacts/the-repository-that-wakes-up/receipt.json Context: https://next.shareplane.malott.ai/artifacts/the-repository-that-wakes-up/context.txt Agent-package manifest: https://next.shareplane.malott.ai/artifacts/the-repository-that-wakes-up/agent-package.json Agent-package ZIP: https://next.shareplane.malott.ai/artifacts/the-repository-that-wakes-up/agent-package.zip Collection catalog: https://next.shareplane.malott.ai/catalog.json Graph: https://next.shareplane.malott.ai/graph.json Agent index: https://next.shareplane.malott.ai/llms.txt