feat(canvas): give agents the learnings a repository accepted

[alliscode]

Why

Agents working in a repository should have the guidance its administrators accepted in Hub, without someone repeating those decisions in every session. This change gives agents that context through an approved Hub connection, without sending the person's prompt to Hub.

What changes

When someone submits a prompt, the Replica plugin asks Hub for the repository's accepted learnings and adds them as session context. Hub checks the connection's consent and the person's cached repository read permission. The learning store records the recall as a use by that person.

sequenceDiagram
  actor Person
  participant CLI as Copilot CLI with the Replica plugin
  participant Hub
  participant Store as Repository's learning store
  Person->>CLI: Submit a prompt
  CLI->>Hub: Ask for accepted learnings, without the prompt
  Hub->>Store: Recall learnings as that person
  Store-->>Hub: Text, learning IDs, and content digests
  Note over Store: Record use by the person
  Hub-->>CLI: Return accepted learnings
  CLI->>CLI: Check the format and add session context

Loading

The connection approval page now explains that Copilot can add accepted learnings to prompts. Hub records consent only when someone approves the current disclosure. Existing connections and older approval forms still connect without learnings; those users must disconnect and connect again to enable them.

The new POST /api/canvas/learnings/recall endpoint returns the learning text, IDs, and content digests, which identify changes to the wording. The plugin accepts only the bounded format produced by replica-learning, including its safety guidance that learnings are data, not commands. The first successful response replaces earlier learning context, including after a restart or session resume. Later unchanged responses add nothing; changed or empty responses replace earlier guidance. Turning Learnings off explicitly clears it.

The hook uses an existing connection and saved Hub selection; submitting a prompt never starts pairing or Hub discovery. Token renewal happens in the background for a later prompt. Hub limits each connection to 30 recalls per minute and stops slow recalls early enough for the plugin's three-second deadline. Errors, invalid responses, and timeouts add no context and never fail the prompt; older responses cannot overwrite a newer one.

Copilot CLI 1.0.96 requires --experimental or /experimental on for extensions. Loading a repository's project copy in prompt mode (copilot -p) also requires GITHUB_COPILOT_PROMPT_MODE_EXTENSIONS=true. Its ACP server does not load these hooks, so ACP clients do not receive learnings through this plugin.

The live integration check uses an isolated local-mode Hub and synthetic data; GitHub-backed OAuth and cross-account authorization remain unverified by that check. Its off/restore case exercises Hub's learnings_off response, not disconnecting a real Copilot account.

[Copilot]

🔵 Needs a closer look

Consent and prompt-context delivery cross security boundaries, and GitHub-backed cross-account authorization remains unverified.

2 open findings

• Unbounded recall audit history can exhaust memory · New
• Stale session objects can restore obsolete guidance · New

What changed in this PR

Adds repository-approved learnings to Copilot sessions through existing Hub connections, without sending prompts to Hub.

Changes:

• Adds consent-gated, rate-limited learning recall.
• Registers a prompt hook with response validation and session-context replacement.
• Updates approval disclosures, persistence, tests, and guides.

File	Description
hub/​test/​integration/​canvas_pairings_test.rb	Tests versioned consent and older approvals.
hub/​test/​integration/​canvas_learning_recall_test.rb	Tests recall access, auditing, limits, and timeouts.
hub/​test/​fixtures/​files/​slow-replica-learning.mjs	Simulates delayed recall and lingering output.
hub/​test/​fixtures/​files/​learning.txt	Supplies accepted learning text.
hub/​docs/​security.md	Documents recall request limits.
hub/​docs/​security-rate-limits.md	Defines the per-connection recall budget.
hub/​docs/​prototype-canvas.md	Documents recall delivery and validation.
hub/​docs/​learnings.md	Explains agent access to accepted learnings.
hub/​docs/​authentication-and-repository-authorization.md	Explains learning consent during approval.
hub/​db/​schema.rb	Records consent fields and constraint.
hub/​db/​migrate/​20261013120000_add_learnings_consent_to_canvas_connections.rb	Adds consent persistence.
hub/​config/​routes.rb	Routes learning recall requests.
hub/​config/​locales/​canvas_pairings.en.yml	Adds the learning disclosure.
hub/​app/​views/​canvas_pairings/​show.json.props	Exposes the disclosure version.
hub/​app/​models/​canvas_pairing.rb	Records consent and transfers it to grants.
hub/​app/​models/​canvas_grant.rb	Updates schema annotations for consent.
hub/​app/​javascript/​pages/​canvas-pairings-page/​types.ts	Types the disclosure version.
hub/​app/​javascript/​pages/​canvas-pairings-page/​stories.tsx	Updates approval story data.
hub/​app/​javascript/​pages/​canvas-pairings-page/​index.tsx	Submits the disclosure version.
hub/​app/​javascript/​pages/​canvas-pairings-page/​index.test.tsx	Tests disclosure display and submission.
hub/​app/​controllers/​concerns/​security_rate_limiting.rb	Adds the recall budget.
hub/​app/​controllers/​canvas_pairings_controller.rb	Passes the submitted consent version.
hub/​app/​controllers/​canvas_api/​learning_recalls_controller.rb	Serves consent-gated recall responses.
hub/​app/​controllers/​canvas_api/​base_controller.rb	Makes permission freshness overridable.
.github/​extensions/​replica-prototype/​register-prototype-canvas/​index.test.mjs	Tests prompt-hook registration and delivery.
.github/​extensions/​replica-prototype/​register-prototype-canvas/​index.mjs	Registers the learning hook.
.github/​extensions/​replica-prototype/​recall-learnings/​index.test.mjs	Tests validation and session delivery behavior.
.github/​extensions/​replica-prototype/​recall-learnings/​index.mjs	Validates and replaces learning context.
.github/​extensions/​replica-prototype/​README.md	Explains delivery and CLI requirements.
.github/​extensions/​replica-prototype/​hub-client/​index.mjs	Adds recall requests and background renewal.

🧠 Review effort: Max

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🔵 Needs a closer look

The consent and private-data delivery boundary needs human review because GitHub-backed cross-account authorization remains unverified.

0 open findings

2 resolved since last review

• Unbounded recall audit history can exhaust memory
• Stale session objects can restore obsolete guidance

🧠 Review effort: Max

[nshki]

Thanks @alliscode! Appreciate you and @thegovind working on the learning loop features here. I believe this closes the core loop.