feat: real agentic Propose via GitHub Copilot CLI, remove mock

.github/workflows/proposal-shared.yml

@@ -94,6 +94,10 @@ jobs:
               echo "::error::schedule/workflow_dispatch-triggered runs must receive capacity_query, queue_query, and review_limit"
               exit 1
             fi
+            if [ -z "${{ secrets.model_credentials }}" ]; then
+              echo "::error::schedule/workflow_dispatch-triggered runs must receive the model_credentials secret -- Propose calls a model to decide field values, it no longer mocks them"
+              exit 1
+            fi
           fi
       # Verifies the triggering comment is the proposal's actual first/
       # control comment, not just any comment containing matching checkbox
@@ -235,7 +239,33 @@ jobs:
           queue_query: ${{ inputs.queue_query }}
           review_limit: ${{ inputs.review_limit }}
           github_token: ${{ secrets.token }}
-      # Field values are mocked, not agent-decided yet — tracked in #11.
+      - name: Gather proposal context
+        id: context
+        if: steps.dequeue.outputs.result == 'dequeued'
+        uses: rubykatzen/starcast/actions/proposal@main
+        with:
+          mode: propose-context
+          parent_issue_id: ${{ steps.dequeue.outputs.issue_id }}
+          regulations_repo: ${{ inputs.regulations_repo }}
+          regulations_path: ${{ inputs.regulations_path }}
+          prefix: ${{ inputs.prefix }}
+          github_token: ${{ secrets.token }}
+      - name: Set up Node.js
+        if: steps.dequeue.outputs.result == 'dequeued'
+        uses: actions/setup-node@v4
+      - name: Install Copilot CLI
+        if: steps.dequeue.outputs.result == 'dequeued'
+        run: npm install -g @github/copilot
+      - name: Decide proposal content
+        id: inference
+        if: steps.dequeue.outputs.result == 'dequeued'
+        uses: actions/ai-inference@v1
+        with:
+          prompt: ${{ steps.context.outputs.prompt }}
+          system-prompt: 'You propose structured field values for a GitHub Issue, constrained by consumer-owned regulations text. Respond with exactly one JSON object and nothing else -- no prose, no Markdown code fence.'
+          model: claude-sonnet-4.5
+        env:
+          COPILOT_GITHUB_TOKEN: ${{ secrets.model_credentials }}
       - name: Create proposal
         id: propose
         if: steps.dequeue.outputs.result == 'dequeued'
@@ -245,6 +275,7 @@ jobs:
           parent_issue_id: ${{ steps.dequeue.outputs.issue_id }}
           prefix: ${{ inputs.prefix }}
           proposal_type_name: ${{ inputs.proposal_type_name }}
+          model_response: ${{ steps.inference.outputs.response }}
           entity_query: ${{ inputs.entity_query }}
           propose_transition_mutation: ${{ inputs.propose_transition_mutation }}
           github_token: ${{ secrets.token }}

.github/workflows/proposal.yml

@@ -18,7 +18,8 @@ jobs:
   # secrets.GITHUB_TOKEN cannot mutate Projects V2 (no such permission
   # scope exists for the auto-provisioned token) -- this caller needs a
   # real PAT with organization Projects read/write, stored as
-  # PROJECTS_TOKEN.
+  # PROJECTS_TOKEN. Propose's model call needs a separate PAT with an
+  # active GitHub Copilot seat, stored as COPILOT_PAT.
   handle:
     uses: ./.github/workflows/proposal-shared.yml
     with:
@@ -39,3 +40,4 @@ jobs:
         mutation($entityId: ID!) { updateProjectV2ItemFieldValue(input: {projectId: "PVT_kwDOD5LpBM4Bfjz_", itemId: $entityId, fieldId: "PVTSSF_lADOD5LpBM4Bfjz_zhZ13FI", value: {singleSelectOptionId: "98236657"}}) { clientMutationId } }
     secrets:
       token: ${{ secrets.PROJECTS_TOKEN }}
+      model_credentials: ${{ secrets.COPILOT_PAT }}

AGENTS.md

@@ -46,7 +46,8 @@ commit SHA. Do not recommend `@main` for stable consumers.
   enforces that `issue_comment` runs carry `issue_number`/`comment_id` and
   that `schedule`/`workflow_dispatch` runs do not, and that
   `schedule`/`workflow_dispatch` runs carry `capacity_query`/`queue_query`/
-  `review_limit`. On `issue_comment` runs, `validate` also verifies the
+  `review_limit`/the `model_credentials` secret. On `issue_comment` runs,
+  `validate` also verifies the
   triggering comment is actually the proposal's first/control comment
   (#63) via `mode: check-control-comment`, shared here rather than
   duplicated across `apply`/`reject`/`distill`/`rework` since all four
@@ -95,12 +96,22 @@ commit SHA. Do not recommend `@main` for stable consumers.
       starts with `prefix`, stripped, plus the always-available fixed set
       (`title`, `body`, `type`, `parent`, `labels`) — the menu a caller
       (in the future, an agent) picks from.
+    - `propose-context` gathers everything a model needs to decide a
+      proposal's content -- regulations text (fetched from
+      `regulations_repo`/`regulations_path`'s default branch via
+      `repository.object(expression: "HEAD:<path>")`), the candidate
+      issue's title/body, and the field catalog from `proposal_fields` --
+      into one self-contained prompt, output as `prompt`. Produces text
+      only; agnostic to whichever inference mechanism a caller wires in.
     - `propose` creates a proposal issue against a parent: resolves the
-      Issue Type, creates the sub-issue, fills every discovered custom
-      field with a trivial type-appropriate mock value (`propose` does
-      not read regulations or call a model yet — real agent judgment is
-      tracked in #11), and posts the control comment
-      (`CONTROL_COMMENT_BODY`) as the first comment.
+      Issue Type, creates the sub-issue, writes whichever fields
+      `--model-response` (a JSON blob shaped `{"title", "body",
+      "fields": {...}}`, produced from `propose-context`'s prompt by a
+      real model call) included, and posts the control comment
+      (`CONTROL_COMMENT_BODY`) as the first comment. Does not decide
+      content itself -- an unknown field name in the response is a
+      warning and a skip, not an error, since the model's output isn't
+      trusted to match the catalog exactly.
     - `propose`/`apply` optionally transition whatever "controlling
       entity" tracks an issue's state (e.g. a `ProjectV2Item`) as it
       moves through the lifecycle (#69): `entity_query` resolves the

CHANGELOG.md

@@ -22,15 +22,25 @@
     `content { ... on Issue { id } }`).
   - `list-fields`: discovers the prefixed proposal Issue Fields available
     in a repository.
+  - `propose-context`: gathers regulations text, the candidate issue's
+    title/body, and the field catalog into one self-contained prompt for
+    a model to decide proposal content from — output only, agnostic to
+    whichever inference mechanism a caller wires in.
   - `propose`: creates a proposal issue against a parent — Issue Type,
-    sub-issue relationship, field values, and the control comment. Field
-    *values* are mocked (a trivial type-appropriate placeholder) rather
-    than agent-decided; real regulation-constrained judgment is tracked
-    in #11.
+    sub-issue relationship, field values, and the control comment. Takes
+    the content to write as a `--model-response` JSON blob
+    (`{"title", "body", "fields": {...}}`) rather than deciding it
+    itself; an unknown field name in the response is a warning and a
+    skip, not an error.
   - `apply`/`reject`: unchanged behavior from the actions they replace
     (below), now reached via `mode: apply`/`mode: reject`.
-- `proposal-shared.yml`'s `propose` job now dequeues a candidate and
-  creates a (mocked) proposal for it, instead of a placeholder notice.
+- `proposal-shared.yml`'s `propose` job now dequeues a candidate, asks a
+  model (via `actions/ai-inference` wrapping GitHub Copilot CLI,
+  requiring a `model_credentials` secret backed by a PAT with an active
+  Copilot seat) what to propose, and creates the proposal from its
+  response — instead of a placeholder notice. `validate` hard-errors on
+  schedule/`workflow_dispatch` runs missing `model_credentials`, the
+  same way it already does for `capacity_query`/`queue_query`.
 - `entity_query`, `propose_transition_mutation`, and
   `apply_transition_mutation` inputs on `proposal-shared.yml`: optionally
   transition whatever "controlling entity" tracks an issue's state (e.g.

README.md

@@ -46,10 +46,9 @@ history remains available in Git.
 The current reusable workflows cover centralized Project intake for issues and
 pull requests, plus explicit label-based issue routing. The proposal-as-issue
 lifecycle described above (`Propose`/`Apply`/`Reject`/`Distill`/`Rework`) has a
-working contract, `proposal-shared.yml`: `Apply`/`Reject` are fully
-implemented, `Propose` dequeues and creates a proposal with mocked field
-values (real agent judgment not wired in yet), and `Distill`/`Rework` remain
-placeholders (tracked in #11).
+working contract, `proposal-shared.yml`: `Apply`/`Reject`/`Propose` are fully
+implemented (`Propose` decides field values via a real model call, not a
+mock), and `Distill`/`Rework` remain placeholders (tracked in #86/#87).
 
 ## Reusable workflows
 
@@ -223,21 +222,32 @@ jobs:
   empty/unset proposal field (title/body included) leaves the parent's field
   untouched. `Reject` just closes the proposal. Both are idempotent:
   re-running against an already-closed proposal is a no-op.
-- **`Propose` dequeues and creates, with mocked content** (#55, #11) — on
-  each scheduled run it runs the consumer-supplied `capacity_query`, and if
-  the result is below `review_limit`, runs `queue_query` and takes its first
+- **`Propose` dequeues, asks a model, and creates** (#55, #11) — on each
+  scheduled run it runs the consumer-supplied `capacity_query`, and if the
+  result is below `review_limit`, runs `queue_query` and takes its first
   candidate; at capacity or an empty queue are both clean no-ops. Both
   queries are consumer-owned GraphQL text: `capacity_query` must alias
   exactly one scalar numeric field as `capacity`, `queue_query` exactly one
   array-valued field as `queue` (each element at least `{ id }`), found by a
   recursive walk matching alias name and value type — zero or more than one
   match is a hard configuration error. Only one candidate is ever dequeued
-  per run, so no pagination/cursor state is needed between runs. When a
-  candidate is found, `Propose` creates the proposal issue against it —
-  Issue Type, sub-issue relationship, the control comment — but **fills every
-  field with a trivial mock value rather than a real agent decision**; wiring
-  in regulation-constrained judgment is tracked in #11. `Distill`/`Rework`
-  remain placeholders.
+  per run, so no pagination/cursor state is needed between runs.
+
+  Deciding what to propose and creating the proposal are two separate
+  steps, split the same way `Apply`/`Reject` are deterministic and
+  `Distill`/`Rework` are agent work: `propose-context` gathers the
+  regulations text, the candidate's title/body, and the catalog of
+  available fields into one self-contained prompt (schema-agnostic about
+  which inference mechanism reads it); the caller hands that prompt to a
+  model and gets back one JSON object (`{"title", "body", "fields": {...}}`);
+  `propose` takes that JSON as `--model-response` and does only the
+  mechanical part — Issue Type, sub-issue relationship, field writes, the
+  control comment. The dogfood caller (`.github/workflows/proposal.yml`)
+  wires this to GitHub Copilot CLI via `actions/ai-inference`, requiring a
+  PAT with an active Copilot seat (`model_credentials`/`COPILOT_PAT`) —
+  `validate` hard-errors on schedule/`workflow_dispatch` runs if it's
+  missing, the same way it does for `capacity_query`/`queue_query`.
+  `Distill`/`Rework` remain placeholders.
 - **The control comment has one fixed template**, posted by `Propose`
   immediately after creation: `- [ ] Apply`, `- [ ] Reject`,
   `- [ ] Distill`, `- [ ] Rework`, nothing else. Before `Apply`/`Reject`/

actions/proposal/action.yml

@@ -14,7 +14,16 @@ inputs:
     description: 'Id of the comment to verify is the proposal''s first/control comment (check-control-comment; see #63)'
     required: false
   parent_issue_id:
-    description: 'GraphQL node id of the issue to propose against (propose)'
+    description: 'GraphQL node id of the issue to propose against (propose, propose-context)'
+    required: false
+  regulations_repo:
+    description: '"owner/repo" that owns the regulations document (propose-context)'
+    required: false
+  regulations_path:
+    description: 'Path to the regulations document inside regulations_repo, read from its default branch (propose-context)'
+    required: false
+  model_response:
+    description: 'JSON {"title", "body", "fields": {...}} decided by a model from propose-context''s prompt (propose)'
     required: false
   capacity_query:
     description: 'GraphQL query text; must alias exactly one scalar numeric field as `capacity` (dequeue)'
@@ -55,6 +64,9 @@ outputs:
   fields:
     description: 'JSON field catalog (list-fields only)'
     value: ${{ steps.run.outputs.fields }}
+  prompt:
+    description: 'Self-contained prompt for a model to decide proposal content (propose-context only)'
+    value: ${{ steps.run.outputs.prompt }}
   summary:
     description: 'Human-readable summary of the outcome'
     value: ${{ steps.run.outputs.summary }}
@@ -70,6 +82,9 @@ runs:
         ISSUE_NUMBER: ${{ inputs.issue_number }}
         COMMENT_ID: ${{ inputs.comment_id }}
         PARENT_ISSUE_ID: ${{ inputs.parent_issue_id }}
+        REGULATIONS_REPO: ${{ inputs.regulations_repo }}
+        REGULATIONS_PATH: ${{ inputs.regulations_path }}
+        MODEL_RESPONSE: ${{ inputs.model_response }}
         CAPACITY_QUERY: ${{ inputs.capacity_query }}
         QUEUE_QUERY: ${{ inputs.queue_query }}
         REVIEW_LIMIT: ${{ inputs.review_limit }}
@@ -87,8 +102,11 @@ runs:
           list-fields)
             args+=(--repository "$REPOSITORY" --prefix "$PREFIX")
             ;;
+          propose-context)
+            args+=(--parent-issue-id "$PARENT_ISSUE_ID" --regulations-repo "$REGULATIONS_REPO" --regulations-path "$REGULATIONS_PATH" --prefix "$PREFIX")
+            ;;
           propose)
-            args+=(--parent-issue-id "$PARENT_ISSUE_ID" --prefix "$PREFIX" --type "$PROPOSAL_TYPE_NAME" --entity-query "$ENTITY_QUERY" --transition-mutation "$PROPOSE_TRANSITION_MUTATION")
+            args+=(--parent-issue-id "$PARENT_ISSUE_ID" --prefix "$PREFIX" --type "$PROPOSAL_TYPE_NAME" --model-response "$MODEL_RESPONSE" --entity-query "$ENTITY_QUERY" --transition-mutation "$PROPOSE_TRANSITION_MUTATION")
             ;;
           apply)
             args+=(--repository "$REPOSITORY" --issue-number "$ISSUE_NUMBER" --prefix "$PREFIX" --type "$PROPOSAL_TYPE_NAME" --entity-query "$ENTITY_QUERY" --transition-mutation "$APPLY_TRANSITION_MUTATION")