[MarkusNeusinger]

Summary

• New optional input source_model on impl-generate.yml and bulk-generate.yml names the model that wrote an adopted implementation, for example gemini-4-argon. It must match ^[A-Za-z0-9][A-Za-z0-9._-]{0,39}$; anything else stops the run. Without source_ref it is dropped with a warning.
• Adopt-mode metadata now writes generated_by: <source_model> (still adopted:<ref> without one) and always adds adopted_from: <ref> (<sha>) right after generated_by, with the full SHA the adopt step resolved. Every other key is unchanged. The PR body line reads **Generation:** adopted from <ref> (<sha>) by <source_model>; the model below runs repair; the issue preview comment, the adopt commit, the step summaries and the ::notice:: lines name the model too.
• bulk-generate forwards source_model on both dispatch branches exactly like source_ref. impl-generate's dispatch_retry and its three manual rerun hints (infrastructure pause, retry cap, adoption failed) keep it as well.
• Hardening from review: neither workflow echoes a rejected source_ref or source_model any more (a newline in a dispatch value would start a workflow command), and bulk-generate's plan preview checks both with impl-generate's patterns before it prints anything, so a bad value stops the bulk run before its first dispatch.

Readers

• automation/scripts/sync_to_postgres.py copies a fixed key list, so adopted_from is ignored and gemini-4-argon syncs into the existing generated_by column (String(100)). No schema change.
• impl-review's metadata update and the regen write-back load the whole mapping and keep unknown keys. review_retest.py mirrors the key set of a fresh generation; adopted_from is inserted after the dict literal, so that mirror and its test stay as they are.
• No API model or frontend type parses metadata files; the site shows generated_by as a plain string (debug page) and has no model statistics that filter on it. Nothing in api/ or app/ changes.

Known gap

• The watchdog's generation rescues forward neither source_ref (as before) nor source_model: a pair the watchdog re-dispatches is generated by Claude. Left as it is on purpose.

Decisions

• adopted_from carries the full 40-character SHA, the value the adopt step resolves and the PR body already shows. The 20 MUI X files backfilled in chore(muix): name gemini-4-argon as author of 20 adopted files #12298 use the 9-character form 8cfea568c.
• A malformed source_model fails the run even without source_ref, so a typo never reaches a log line unchecked.
• The constant error for a rejected source_ref is new too: it had the same echo two lines above the new check.
• -f source_model= is always sent, empty or not, like -f source_ref=. Gemini Flash pointed out that GitHub answers 422 when a target workflow lacks the input; that only happens for a dispatch from this branch to a main without it, before merge.

Reviews

• Gemini 3.8 Flash (first review): sound with changes. Adopted: validate early in bulk-generate. Skipped: conditional -f source_model (see Decisions).
• Copilot round 1: two high findings (raw value echoed into ::error:: in impl-generate; raw value printed by bulk-generate before any check) and one low doc finding. All fixed in 1bd31a8, threads resolved.
• Copilot round 2 (requested because round 1 had high findings): no open issues. Two low doc notes about the metadata examples were fixed in 7a2ae73. Its headline about an unescaped malformed-model test input does not match the code: every malformed-value test passes the value through the environment. No further round.

Plan

N/A

Test plan

• tests/unit/workflows/test_impl_generate_adopt.py: input declarations on both workflows; only "Extract inputs" reads the raw input, every consumer reads steps.inputs.outputs.source_model; validation (valid names exported, malformed values including newline injections fail without being echoed, with and without source_ref; a model without source_ref warns and is dropped); the real adopt step names the model in its commit, notice and summary; the metadata writer emits generated_by and adopted_from in that order with all other keys equal to a fresh generation's; the failure handler's retry and its adoption-failed rerun command carry source_model; bulk-generate's dispatch loop, plan preview and its input checks; the generate job depends on the preview; both workflows use the same patterns.
• tests/unit/workflows/test_model_routing.py: the PR body with source_model round-trips through impl-review's and impl-repair's rescues (repair model opus survives).
• tests/unit/automation/scripts/test_sync_to_postgres.py: a muix metadata file with generated_by: gemini-4-argon and adopted_from scans cleanly; adopted_from is left out of the row.
• uv run ruff check ., ruff format --check ., mypy api core, uv run pytest tests/unit -n auto (13292 passed, 3 skipped before the review fixes; workflow and automation tests 2823 passed after them), tools.changelog check --base origin/main.
• Workflow YAML has no live verification loop. I reasoned through: set -u in bulk-generate's dispatch step (the env var is always defined there, possibly empty); ${SOURCE_MODEL:+…} expansions under bash -e; the inline Python in the metadata step receives only validated values (no quotes possible); the generate job's custom if keeps the implicit success() on needs: preview; a run dispatched by bulk-generate always targets the default branch, so both workflows need to be on main before source_model is used. No branch smoke was run, because one would open a real adopt PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_0175hA62juZoajMcqKxWFGza

[Copilot]

🟡 Changes recommended

Critical workflow-command injection risks remain in the bulk preview/dispatch path and invalid-input error path, with additional documentation inconsistencies.

3 open issues

• Validate SOURCE_MODEL before bulk preview and dispatch · New
• Prevent raw dispatch value from injecting workflow commands · New
• Document conditional generated_by and adopted_from metadata · New

What changed in this PR

This pull request adds adopt-mode attribution for the model that authored an externally supplied implementation, while preserving its source reference and resolved commit SHA.

Changes:

• Adds and validates optional source_model inputs across single and bulk generation workflows, including retries and reruns.
• Records generated_by and adopted_from metadata and surfaces the model in workflow and PR output.
• Adds workflow, routing, synchronization, documentation, and changelog coverage.
• Critical: Raw source_model values can be echoed before validation, enabling workflow-command/log injection in bulk dispatch and invalid-input paths.
• Nits: Two documentation examples describe metadata combinations that the workflow does not produce.

File	Summary
tests/​unit/​workflows/​test_model_routing.py	Verifies adopted PR bodies preserve repair-model routing.
tests/​unit/​workflows/​test_impl_generate_adopt.py	Tests validation, metadata output, propagation, retries, reruns, and bulk dispatch behavior.
tests/​unit/​automation/​scripts/​test_sync_to_postgres.py	Confirms generated_by synchronization while ignoring adopted_from.
docs/​workflows/​overview.md	Documents adopt-mode model attribution and metadata.
docs/​reference/​repository.md	Updates the repository metadata example.
changelog.d/​adopt-source-model.md	Adds the feature changelog entry.
agentic/​docs/​project-guide.md	Updates agent-facing adopt-mode and metadata guidance.
.github/​workflows/​impl-generate.yml	Validates and propagates source_model, records adoption metadata, and updates workflow outputs and retries.
.github/​workflows/​bulk-generate.yml	Adds the input, preview output, and dispatch forwarding.
.claude/​skills/​babysit-pipeline/​SKILL.md	Documents the new source_model dispatch input.

_{Review effort: Lite}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Copilot]

🔵 Needs a closer look

A malformed-model test case does not safely reach the validator because its shell input is unescaped.

No open issues 🎉

2 previously missed issues

In code that hasn't changed since last review

Show adopted provenance metadata in example

agentic/​docs/​project-guide.md:255

This example comments out adopted_from, so copying it produces no provenance key, and it labels the ordinary claude-opus-4-7 value as though it were the adopted author. Show the adopted variant as an actual generated_by/adopted_from pair (or clearly separate it as an alternate example) so the documented metadata format is not misleading.

Show adopted provenance metadata in example

docs/​reference/​repository.md:314

This example comments out adopted_from, so copying it produces no provenance key, and it labels the ordinary claude-opus-4-7 value as though it were the adopted author. Show the adopted variant as an actual generated_by/adopted_from pair (or clearly separate it as an alternate example) so the repository metadata format is not misleading.

3 resolved since last review

• Prevent raw dispatch value from injecting workflow commands
• Validate SOURCE_MODEL before bulk preview and dispatch
• Document conditional generated_by and adopted_from metadata

_{Review effort: Lite}