Agentsop LLM Artifact Versioning

Versioned, Reproducible LLM Artifact — Overlay SOP

"Prompts are effectively the weights of an LLM application." — DSPy core philosophy [arxiv.org/abs/2310.03714] (R1 §1)

"Treat the compiled program as a (program × LM) pair. Changing the LM invalidates the artifact — recompile." — dspy-sop SKILL, Dilemma Case B (R1 §2)

This is an enhancement overlay, not a framework SOP. It sits on top of whatever stack you use (DSPy, LangChain, raw API) and adds one discipline: define, pin, and version the entire deployable bundle as a unit. It is the broad sibling of [[agentsop-per-model-artifacts]] — that skill versions one compiled prompt; this one versions everything that ships together.

1. 何时激活 (When to activate)

Activate when any of these appears in the user's intent, codebase, or workflow:

Do NOT activate when:

• One-off prompt edit with no deploy and no reproduction requirement — just edit and run.
• A single-component proof-of-concept where there is exactly one prompt, one model, no retrieval, and it will never be reproduced. Flat v1.json is fine.
• A vendor owns the whole lifecycle (OpenAI Prompt Library, Anthropic Workbench managed prompts, fully vendor-managed RAG) — let them version it.
• The team rewrites the app daily during exploration — bundle versioning has no shelf life until the app shape stabilizes (same boundary as dspy-sop "signature still changing").

2. 核心心智模型 (Core mental model)

An LLM app artifact ≠ an ML model

The single most common mistake is reasoning about an LLM app the way you reason about a trained model. They are not the same shape.

The deployable artifact is:

            ┌──────────────── DEPLOYABLE BUNDLE (one tag) ────────────────┐
            │                                                             │
  prompts       compiled        model pins        retrieval       eval-set
  (text +       programs        (snapshot id      config          version
   hashes)      (program.json)   per call site)   (index ptr,     (sha256)
            │                    embed model,                             │
            │                    top_k, reranker)                        │
            │                                                             │
            └── version them TOGETHER, or you can't reproduce a deploy ──┘

The load-bearing claim: the deployable artifact = prompts + compiled programs + model pins

• retrieval config + eval-set version. Version them together or you cannot reproduce a deploy — and you cannot roll back without producing a never-tested combination.

Why "together" is non-negotiable

Each part can change without touching the others (R1 §3, §4, §6):

• A model alias rolls forward server-side (gpt-4o → new snapshot) — prompt unchanged, behavior changed.
• A model snapshot gets a silent server-side patch (R1 §4) — pin unchanged, behavior changed.
• A retriever knob moves in a dashboard (top_k 5→10) — no commit, behavior changed.
• The eval set gains 50 examples — the "same" score now means something different (R1 §5).

If these are versioned separately, "the deploy" is not a thing you can name. If they are versioned as one bundle with one id, the deploy is reproducible and rollback is atomic (R1 §8).

The registry analogy and its limit (R1 §7)

MLflow's Model Registry gives the right primitives: versioning, stage transitions (Staging/Production/Archived), reproduce-from-config, compare-versions [~/.claude/skills/mlflow/SKILL.md]. Borrow those primitives. But the analogy breaks on heterogeneity: a registry versions one model + signature + run; the LLM bundle is a composite of many models, prompts, configs, and an eval set. You can log the manifest into MLflow as one "model", but the registry was built for the single-weights case. The overlay exists to make the composite explicit.

3. SOP 工作流 (SOP workflow)

Five steps. Each has an exit criterion. The output is one versioned, reproducible bundle.

Step 1 — Enumerate artifact components

List every component that affects behavior at runtime. Nothing implicit:

• Every prompt / system-prompt file (path + content hash).
• Every compiled program (program.json or save_program dir + sha).
• Every model call site (router, synthesizer, reranker, judge) and its model.
• Retrieval config: which index, its embed model + dim, top_k, reranker on/off, chunking.
• The eval set that produced the bundle's scores.
• Framework versions (dspy, langchain, llama-index, python).

Exit: a written component inventory — no "and whatever the dashboard says" gaps.

Step 2 — Pin each component

• Models: dated snapshot, never an alias, never latest, at every call site (OP-2, R1 §3). Cross-link [[agentsop-per-model-artifacts]] for per-prompt swap-test detail.
• Prompts: content sha256, committed to VCS.
• Compiled programs: path + sha; prefer save_program=True for portability (R1 §2).
• Config: move every knob into version-controlled config (OP-3, config-as-code). Dashboards are not artifacts.
• Eval set: sha256 over canonical eval bytes (R1 §5).

Exit: every component has an immutable identifier. Grep for aliases finds none.

Step 3 — Bundle + version

Write a single manifest.<version>.json (OP-1) referencing every pinned component, and assign one monotonic bundle id (OP-5). Components keep their own internal versions; the bundle has one id production deploys atomically.

Exit: deploy/manifest.v<n>.json exists and a git tag deploy/v<n> points to it.

Step 4 — Tie to eval-set version

Record eval_set_sha + scores in the manifest (OP-4). A bundle without a pinned eval set has no reproducible score. Cross-link [[agentsop-regression-gate]]: the gate re-runs this exact eval set on the new bundle vs the parent bundle at PR time, and fails the PR on regression.

Exit: manifest shows eval_set_sha and the scores measured on it; the regression gate is wired to that eval-set version.

Step 5 — Enable rollback

Production deploy reads the bundle tag, not HEAD (OP-6, R1 §8). Keep prior bundles tagged and reachable. Rollback = re-point the deploy to the previous tag, which restores prompts + model + config + index pointer together — never a partial revert.

Exit: one-command rollback (deploy deploy/v<n-1>) restores a known-good combination.

Maintenance loop

• Canary re-run the eval set against pinned snapshots periodically — catches silent server-side drift the snapshot pin cannot (R1 §4).
• On any provider deprecation, re-bundle (the per-prompt swap-test lives in [[agentsop-per-model-artifacts]]; the bundle re-version lives here).

4. 操作模型 (Trigger / Action / Output / Evidence)

OP-1 — Artifact manifest

OP-2 — Model-snapshot pinning

OP-3 — Config-as-code

OP-4 — Eval-set linkage

OP-5 — Bundle-and-version

OP-6 — Rollback recipe

OP-7 — Component-drift audit (supporting)

OP-8 — Retrieval-config binding (supporting)

5. 困境决策案例 (Dilemma cases)

Case A — "Model provider silently updates: output drift with no version bump"

困境 (Dilemma): A RAG app has shipped on gpt-4o-2024-08-06 for the synthesizer. No code, prompt, config, or model pin has changed in three months — every component's id in the manifest is identical. Yet the quarterly canary on the pinned eval set drops from 78% to 73%. Support escalations are rising. What changed, and what is the fix when nothing in the bundle moved?

约束 (Constraints):

• The snapshot string is unchanged — the pin did its job.
• OpenAI's policy is "snapshots are stable," but server-side safety/throughput/hallucination patches ship without a snapshot bump (R1 §4).
• Reverting the model is impossible: the canary IS the latest behavior on the same snapshot.
• The eval-set version is pinned, so the score drop is a true behavior shift, not a measurement artifact (OP-4) — apples vs apples.

决策步骤 (Decision steps):

1. Confirm it is drift, not measurement. Because the manifest pins eval_set_sha, the 73% and the old 78% were measured on the same eval set. Rule out apples-vs-oranges first (OP-7) — a changed eval_set_sha would be the more boring explanation.
2. Compare sample-output hashes. Store sample-output hashes per bundle at build time (R1 §4); re-run the same sample inputs and diff. Divergence with an unchanged snapshot confirms silent server-side drift.
3. Re-bundle, do not edit in place. Recompile/re-tune the prompt component against the new server behavior (per-prompt mechanics live in [[agentsop-per-model-artifacts]]), then mint a new bundle deploy/v<n+1> even though the model snapshot string is unchanged. The built_at field carries the temporal lineage the snapshot string cannot.
4. Gate the new bundle. [[agentsop-regression-gate]] re-runs the pinned eval set on v(n+1) vs the v(n) parent; merge only if it recovers above the gate threshold.
5. Tighten the canary cadence to weekly for 30 days post-incident; revert to quarterly when stable. The canary against the pinned eval set is the only true contract with the provider (R1 §4) — the pin alone is necessary but not sufficient.

结果 (Outcome): Re-bundling recovers to ~79%. The pin protected against alias rollover and most drift; the eval-set-linked canary caught what the pin missed. Crucially, the bundle id incremented even though no human-authored component "changed" — the behavior changed, so the bundle changed.

可提取的操作 (Extractable operation): A snapshot pin is necessary but not sufficient. Tie every bundle to a pinned eval set and canary against it; when behavior drifts on a stable snapshot, mint a new bundle version rather than mutating the live one.

Case B — "Prompt change shipped without an eval — and now we can't reproduce or roll back cleanly"

困境: Under deadline pressure, an engineer edits the synthesizer system prompt directly in the running config to fix one bad answer, redeploys, and moves on. No new bundle, no eval run, no tag. A week later a different regression appears in prod, and a model-snapshot rollover also landed in the same window. Now: which prompt is live? Was it ever evaluated? Can we roll back to before the prompt edit without also reverting unrelated changes?

约束:

• The live prompt no longer matches any committed bundle manifest — the manifest is stale.
• The model alias rolled forward in the same window, so even "revert the prompt" leaves an untested (old-prompt × new-model) combination.
• No eval was run on the edited prompt, so there is no score to compare against — the regression gate was bypassed entirely.
• Customers are affected; time pressure is real again.

决策步骤:

1. Stop the bleeding with a bundle rollback, not a prompt rollback. Deploy the last tagged bundle deploy/v<n> (OP-6, R1 §8). Because the bundle pins prompt and model and config together, this restores a known-good combination — it does not produce the untested (old-prompt × new-model) Frankenstein that a prompt-only revert would.
2. Reconstruct what was live. The ad-hoc prompt edit is the root cause of the reproducibility loss: behavior was changed outside the bundle. Capture the edited prompt's content hash now so it is at least recoverable, then delete the out-of-band path.
3. Re-introduce the fix the right way. Put the prompt fix into VCS (config-as-code, OP-3), bump the model pin to the dated snapshot it will actually run on (OP-2), and build a new bundle deploy/v<n+1>.
4. Run the gate. [[agentsop-regression-gate]] runs the pinned eval set on v(n+1) vs the parent. The original one-bad-answer case should be in the eval set now (turn the incident into a test). Merge only on pass.
5. Add a pre-deploy guard. Refuse to deploy if the live prompt hashes do not match the tagged bundle's manifest — converts "someone edited prod directly" from a silent reproducibility hole into a deploy-time error.

结果: Rollback restores service immediately because the bundle is atomic. The fix re-lands behind the gate. The lesson the team internalizes: a prompt change shipped without an eval-linked bundle is invisible to rollback and reproduction — it is exactly the failure mode this overlay exists to prevent.

可提取的操作: No behavior-affecting change ships outside a versioned, eval-linked bundle. Roll back by bundle tag (atomic) — never by reverting one component, which yields untested combinations.

6. 反模式与边界 (Anti-patterns & boundaries)

Anti-patterns

1. Versioning code but not prompts/configs. Git tracks the Python; the prompt text lives in a dashboard and the model id in an env var. The "deploy" is then unreproducible — the committed code does not determine behavior. Prompts are weights (R1 §1); version them.
2. "latest" (or any alias) model pin. gpt-4o, claude-3-5-sonnet, latest are moving targets that change behavior with no version bump (R1 §3). Pin dated snapshots at every call site (OP-2).
3. No eval-set linkage. A bundle that records a score but not the eval set that produced it cannot be compared to its successor — "score went up" across two eval sets is meaningless (R1 §5, OP-4). Pin eval_set_sha.
4. Per-component tags instead of one bundle tag. Tagging prompts, models, and configs separately makes rollback non-atomic — you revert one part and ship a never-tested combination (R1 §8, OP-6). One bundle, one tag.
5. Editing live prod components out of band. A "quick" dashboard prompt edit (Case B) breaks reproducibility and rollback silently. All behavior changes flow through a new bundle.
6. Treating the bundle like a single ML model. Logging just the compiled program and calling it "the artifact" drops the model pins, retrieval config, and eval set. The artifact is a composite (R1 §7) — save_program captures one component, not the bundle (R1 §2).
7. Retrieval config left implicit. Index pointer, embed model, and top_k unbundled means a re-embed or a knob change silently alters answers (OP-8, R1 §6). Pin them in the manifest.
8. Deleting old bundles to "save space." Disk is cheap; rollback targets are precious. Keep prior bundle tags reachable.

Boundaries (when this overlay is overkill)

• Single-component PoC. One prompt, one model, no retrieval, never reproduced — a flat v1.json is enough. Adopt this overlay at the second component or the first real deploy.
• Vendor owns the lifecycle. OpenAI Prompt Library / Anthropic Workbench / managed RAG — let the vendor version it.
• Pre-stabilization exploration. If the app shape changes daily, bundle versioning has no shelf life yet — stabilize first (same boundary as dspy-sop).
• Versioning ONE compiled prompt. Use [[agentsop-per-model-artifacts]] directly; this overlay only adds value once there are multiple components to bundle.
• The CI comparison mechanism itself. This overlay requires a regression gate but does not implement it — see [[agentsop-regression-gate]].

7. 跨框架对照 (Cross-framework comparison)

How four mechanisms handle (or fail to handle) the whole-bundle versioning problem. None of them natively version the composite; each covers a slice.

DSPy save_program — one component, not the bundle

• Stores: per-predictor instructions, bootstrapped demos, signature shapes (R1 §2) [dspy.ai/tutorials/saving/].
• Does NOT store: model identity, dataset, metric, retrieval config, eval-set version.
• Role in the bundle: program.json (or a save_program=True dir) is one component the manifest references. The overlay pins the rest. Per-prompt lifecycle of this component → [[agentsop-per-model-artifacts]].

LangChain Hub — versioned text templates, model as a hint

• Stores: prompt template text, partial variables, model name as a non-enforced hint.
• Does NOT store: eval scores, dataset binding, retrieval config, recompile lineage.
• Role in the bundle: a viewer/diff tool for the prompt component. It does not pin a model, bind an eval set, or run a gate. Use Hub as the prompt viewer; the bundle manifest is the source of truth. hub.pull("user/rag-qa:v3") resolves a prompt version, not a deploy.

MLflow Model Registry — the closest analogy, built for one model

• Stores: pickled model/program, param + metric history per run, signature, versions with explicit Stage labels (None/Staging/Production/Archived) [~/.claude/skills/mlflow/SKILL.md, mlflow.org/docs/latest/model-registry.html].
• Does NOT store natively: the heterogeneity — many model snapshots, multiple prompts, retrieval config, eval-set hash (add via params/tags).
• Role in the bundle: borrow the primitives — versioning, stage promotion, reproduce, compare (R1 §7). Press it into service by logging the whole manifest as one "model" with components as params/artifacts. The gap: it was built for the single-weights case; the LLM bundle is a composite. The analogy holds for versioning + promotion + rollback, not for the artifact's shape.

Plain git + manifest — the lightweight default

• Stores: whatever you put in deploy/manifest.<version>.json — prompt hashes, compiled program shas, model snapshots per call site, retrieval config, eval_set_sha, framework versions.
• Strengths: zero infra, fully grep-able and git-blame-able; the bundle tag (deploy/v<n>) gives atomic deploy + rollback (R1 §8); the regression gate ([[agentsop-regression-gate]]) wires to the pinned eval set in CI.
• Limitations: no UI, no RBAC, no built-in compare. Scales to small/medium teams.
• Decision rule: start here. Migrate the registry layer to MLflow when you exceed team scale; keep the manifest discipline regardless.

Summary

Companion skills

• [[agentsop-per-model-artifacts]] — versions ONE compiled prompt as a (program × LM × dataset) triple. This overlay wraps that as one component of the whole bundle. Use it for per-prompt swap-tests, snapshot churn, and dataset-hash discipline.
• [[agentsop-regression-gate]] — the CI mechanism that re-runs the pinned eval set on new bundle vs parent and fails the PR on regression. This overlay requires it (Step 4, OP-4) but does not implement it.
• dspy-sop — the compile loop that produces the compiled-program component.

Quick-reference appendix

Manifest skeleton

{
  "bundle_version": "v7", "built_at": "...", "commit": "<git-sha>", "parent_bundle": "v6",
  "prompts": [{"name": "synth_system", "path": "prompts/synth_system.txt", "sha256": "..."}],
  "compiled_programs": [{"name": "rag_synth", "path": "artifacts/rag_synth/.../v3.json", "sha256": "..."}],
  "models": [
    {"call_site": "router",      "provider": "openai",    "snapshot": "gpt-4o-mini-2024-07-18"},
    {"call_site": "synthesizer", "provider": "openai",    "snapshot": "gpt-4o-2024-08-06"},
    {"call_site": "judge",       "provider": "anthropic", "snapshot": "claude-3-7-sonnet-20250219"}
  ],
  "retrieval": {"index_path": "indices/kb/.../<corpus-sha8>/", "embed_model": "text-embedding-3-small",
                "dim": 1536, "top_k": 8, "reranker": "bge-reranker-v2", "corpus_sha8": "a8f1c2e9"},
  "eval_set_sha": "9c3f...", "scores": {"dev": 0.81, "test": 0.79},
  "frameworks": {"dspy": "2.6.1", "python": "3.11.8"}
}

Decision tree

Preparing to deploy / "what's in prod?" ─► Step 1-3 (enumerate, pin, bundle)
   │
Score regressed, nothing committed changed? ─► Case A: confirm eval_set unchanged,
   │                                            check sample-hashes, re-bundle, gate
Prompt edited live without eval? ──────────► Case B: roll back by BUNDLE tag (atomic),
   │                                            re-land fix behind the gate
Need rollback? ─────────────────────────────► OP-6: deploy previous bundle tag
   │
Only ONE compiled prompt to version? ───────► use [[agentsop-per-model-artifacts]] instead
Need the CI comparison itself? ─────────────► use [[agentsop-regression-gate]]

Pre-deploy guards (recommended)

• no-alias-pin: fail if any manifest model id lacks a date suffix or is latest (OP-2).
• manifest-matches-live: fail deploy if live component hashes ≠ tagged bundle (Case B).
• eval-set-pinned: fail bundle if eval_set_sha missing or scores recorded without it (OP-4).
• regression-gate: re-run pinned eval set on new bundle vs parent — see [[agentsop-regression-gate]].

Key references

• DSPy save_program / (program × LM × data) triple: dspy-sop-skill/SKILL.md (R1 §1, §2).
• Per-prompt artifact lifecycle, snapshot pinning, silent drift, config-as-artifact: d-per-model-artifacts-skill/SKILL.md (R1 §3–§6).
• Registry analogy (versioning, stage transitions, reproduce): ~/.claude/skills/mlflow/SKILL.md (R1 §7).
• OpenAI deprecations: platform.openai.com/docs/deprecations
• Anthropic deprecations: docs.anthropic.com/en/docs/about-claude/model-deprecations
• This skill's evidence file: references/R1-source-evidence.md.