Jsingletonai Dejavu Dejavu

dejavu-test-integration

Verifies what /dejavu-integrate produced. Runs in the same workspace, on the same feature branch. Loose coupling — fast, catches compile and runtime bugs, does not catch logical errors.

Canonical sources (use these, not ambient knowledge)

All static checks and smoke-test shapes validate against these URLs. WebFetch each before running step 3.

• Scope-tagged docs index: https://docs.dejavu.ai/llms.txt
• OpenAPI (Platform REST): https://docs.dejavu.ai/openapi.json
• Published SDK skill (canonical call patterns): https://raw.githubusercontent.com/JSingletonAI/dejavu/main/skills/dejavu/SKILL.md
• Vercel AI SDK skill (if the target repo uses @ai-sdk/*): https://raw.githubusercontent.com/JSingletonAI/dejavu/main/skills/dejavu-vercel-ai-sdk/SKILL.md
• SDK source (cross-check version against frontmatter dejavu_tested_versions):
  • Repo root: https://github.com/JSingletonAI/dejavu
  • Python: https://github.com/JSingletonAI/dejavu/tree/main/dejavu
  • TypeScript: https://github.com/JSingletonAI/dejavu/tree/main/dejavu-ts

Read the Delegated skill: field in .dejavu-integration/plan.md — if it names a skill URL, fetch that skill and use its example blocks as the reference for both static checks (step 3) and the smoke test (step 5).

Non-invasiveness contract

Every check in this skill assumes the integration is additive and feature-flagged (see /dejavu-integrate "Integration principles"). Specifically:

• product.json must contain a feature_flag field.
• Steps 4–6 run in two passes:
  • Pass A — flag unset. All pre-existing tests must pass, smoke/E2E skip. The repo must behave like main. Any failure here is a hard fail — do not let the self-heal loop attempt a patch.
  • Pass B — flag set. New tests must pass, smoke and E2E run.
• If Pass A fails, the scorecard marks non_invasive: false and sets overall: fail with a distinct reason code the integrator's heal loop refuses to touch.

Preconditions

Refuse to start unless ALL of the following are true:

• .dejavu-integration/ directory exists in the repo root.
• .dejavu-integration/product.json, goal.md, and plan.md are readable and internally consistent (JSON parses, docs non-empty).
• Current branch name begins with dejavu-integrate/ (set by the companion skill). Prevents accidental runs on unrelated branches.
• Working tree is clean. The skill never modifies source files; any dirty state means the integration is mid-edit and not ready to verify.
• The same API key the integration used is available in the environment (VENICE_API_KEY for Platform, OPENAI_API_KEY for OSS — read which from product.json). Interactive mode asks if missing; CI mode exits 2.

Exit with a written rationale on any precondition failure. Never attempt to "fix up" state.

Pipeline

1. Read the contract

Load:

• product.json → which language, which product (Platform vs OSS), which dejavu version, write_site, read_site.
• plan.md → the mechanical contract (write pattern, read pattern, preserved behavior).
• goal.md → the intent (displayed in the scorecard only; not tested).

2. Install dependencies

Route by language from product.json:

If install fails → exit code 2 with stderr tail. Never move to testing if dependencies don't resolve.

3. Static sanity checks (fast, local, no API calls)

• Import check: does the write-site file import the expected Deja Vu surface? Authoritative list comes from ## Identify the User's Setup in https://docs.dejavu.ai/llms.txt:

  • Platform Python → from dejavu import MemoryClient
  • Platform TS → import MemoryClient from "dejavu-memory"
  • OSS Python → from dejavu import Memory
  • OSS TS → import { Memory } from "dejavu-memory/oss"

  If plan.md names a delegated skill (e.g., Vercel AI), use that skill's import signature instead of the list above. Mismatch → fail with line number.

• Version check: installed dejavu-memory version falls in the range from this skill's dejavu_tested_versions. Out of range → warn but continue.

• Type check (TS tracks only): run tsc --noEmit or tsup --dts. Non-zero → fail.

• Lint (if the repo has a linter configured): run the repo's own lint command. Lint failures from this skill's changes → fail; pre-existing lint failures → surface as a warning.

• Eager-init check: grep the write_site and read_site files (paths from product.json) for MemoryClient( or Memory( at module scope — i.e., not inside a function, method, or class body. MemoryClient() validates the API key in __init__ (network call) and OSS Memory() can eagerly initialize embedding/LLM providers — module-level instantiation hits the wire on import and breaks Pass A's test collection whenever the key is unset. Hit → fail with file:line and the lazy-init guidance from /dejavu-integrate step 8 constraint #7.

4. Run the repo's native test suite (two passes)

Pass A — feature_flag unset. Run the entire pre-existing suite (excluding the new test_dejavu_* files). Must be 100% green. Any failure here marks non_invasive: false in the scorecard and is a hard fail — the integrator's self-heal loop refuses to touch it.

Pass B — feature_flag set (value from product.json). Run the full suite including the new tests. All must pass.

Isolate integration-introduced failures using git diff main..HEAD --name-only. A test file that exists on main and fails only under the integration branch (flag set or unset) counts against the scorecard regardless of pass. A test file that already failed on main is surfaced as pre_existing_unrelated and does not count — but is still reported so the user can clean it up.

Capture output to .dejavu-integration/test-stdout-flag-off.log and .dejavu-integration/test-stdout-flag-on.log. Scorecard reports pass/fail per pass.

5. Smoke test (real API call, shortest round-trip)

Scripted end-to-end flow tailored to product.json. The call shapes below are the minimal ones; if plan.md names a delegated skill, use that skill's minimal example verbatim instead — it is the canonical shape for the detected stack.

Platform (Python):

from dejavu import MemoryClient
c = MemoryClient()                               # uses VENICE_API_KEY
uid = f"dejavu-test-integration-{os.urandom(4).hex()}"
c.add([{"role": "user", "content": "I prefer aisle seats"}], user_id=uid)
hits = c.search("seat preference", user_id=uid)
assert any("aisle" in h.get("memory", "") for h in hits), hits
c.delete_all(user_id=uid)                        # clean up

Platform (TS): same shape with MemoryClient from "dejavu-memory".

OSS (Python / TS): uses Memory() / new Memory() with default config (OpenAI LLM via OPENAI_API_KEY, local Qdrant). If the repo ships a docker-compose.yml with a Qdrant service, the skill starts it first and tears it down after. If no backing store is reachable → fail with a clear message naming the fix.

The smoke test always uses a disposable random user_id prefixed with dejavu-test-integration- so a failed cleanup doesn't pollute the user's real data. A background tidy step deletes any prefix-matching entries older than 24 hours on the next run.

Capture output to .dejavu-integration/smoke-stdout.log.

6. E2E integration test (run the app, exercise the flow)

Unit tests + smoke prove the SDK works in isolation. This step is the real signal: does memory actually appear in the app's user-visible output when the integration runs end-to-end?

Requires plan.md to contain an E2E recipe: section (authored by /dejavu-integrate step 5). If absent → status skipped (not fail), note in scorecard that the repo has no runnable entry point.

Recipe fields the skill reads:

• start — shell command to launch the app using $PORT for any network port. Run in background with stdout/stderr teed to .dejavu-integration/e2e-app.log.
• ready_probe — how to detect readiness. url=... status=... polls an HTTP endpoint; log="..." waits for a substring in e2e-app.log; sleep=N waits N seconds (last resort). 60-second hard timeout.
• compose_services — optional. If set, bring them up via docker compose up -d <services> before start, tear them down with docker compose down at the end.
• write_call — triggers the Deja Vu write path exactly once. Output is captured and surfaced on failure. 60-second hard timeout.
• write_async_wait_ms — pause after write_call to let async memory flushes land. Default 0.
• read_call — triggers the Deja Vu read path. Typically a fresh session or new request that should surface the stored memory.
• read_assert — substring, regex=..., or jsonpath=<expr>=<value> that must appear in read_call's stdout. This is the E2E pass gate.

Execution order:

1. Allocate an ephemeral TCP port; export as PORT.
2. Set DEJAVU_USER_ID to a disposable dejavu-test-integration-<rand> value and export it, so the app can use the same scoping the smoke test does if the recipe wants cleanup.
3. Bring up compose_services if named.
4. Run start in the background.
5. Poll ready_probe until success or 60s timeout. Timeout → fail.
6. Run write_call. Non-zero exit → fail (but continue to cleanup).
7. Sleep write_async_wait_ms.
8. Run read_call.
9. Evaluate read_assert against read_call's stdout. Miss → fail.
10. Cleanup (always, even on failure): SIGTERM the app, SIGKILL after 5s, docker compose down if services were started, delete_all memories matching dejavu-test-integration-* on Platform scenarios.

On any failure, the scorecard includes:

• Last 40 lines of e2e-app.log
• Full write_call output
• Full read_call output
• The expected vs actual for read_assert

7. Scorecard

Write .dejavu-integration/scorecard.md and .dejavu-integration/scorecard.json:

{
  "timestamp": "2026-04-20T14:03:11Z",
  "branch": "dejavu-integrate/remember-user-preferences",
  "product": "platform",
  "language": "python",
  "dejavu_version": "2.0.0",
  "non_invasive": true,
  "feature_flag": "DEJAVU_ENABLED",
  "results": {
    "install":      {"status": "pass", "duration_ms": 12043},
    "static_checks":{"status": "pass", "duration_ms": 812},
    "unit_tests_flag_off": {"status": "pass", "duration_ms": 3920, "count": 47,
                            "reason": "all pre-existing tests green with flag unset"},
    "unit_tests_flag_on":  {"status": "pass", "duration_ms": 4321, "count": 49},
    "smoke_test":   {"status": "pass", "duration_ms": 2890, "memory_id": "mem_..."},
    "e2e_test":     {"status": "pass", "duration_ms": 14200,
                     "ready_probe_ms": 3100, "write_exit": 0,
                     "read_assert_matched": true}
  },
  "friction": {
    "dependency_install_retries": 0,
    "pre_existing_test_failures": 0,
    "warnings": ["dejavu-memory 2.0.0 pinned; consider 2.0.1 for fix X"]
  },
  "overall": "pass"
}

The markdown version is human-readable and includes:

• Goal doc + plan doc reprinted at top (so reviewers don't have to hunt).
• Each check with pass/fail + log excerpt.
• Friction summary.
• Verbatim warnings from dejavu SDK (if any — e.g., deprecated field usage).
• Explicit "NOT checked" section listing what loose coupling misses: "Whether the stored data is what the user wants stored. Whether search runs at the right moment. Whether user_id matches the actual session scope. Human review required."

8. Report + exit

• Print the scorecard path + overall pass/fail to stdout.
• Do not commit the scorecard files. They live in .dejavu-integration/, which is gitignored. The user can inspect and optionally pin.
• On fail: print the first failing step's log tail (last 40 lines) and stop. Do not attempt to fix anything.

Artifacts (all under .dejavu-integration/)

Modes

Invocation

/dejavu-test-integration                       # interactive, all steps
/dejavu-test-integration --ci                  # non-interactive
/dejavu-test-integration --skip-smoke          # no API calls, no E2E
/dejavu-test-integration --skip-e2e            # unit + smoke only (faster CI)
/dejavu-test-integration --only-smoke          # just smoke
/dejavu-test-integration --only-e2e            # just E2E (assumes deps installed)

Composition: --skip-* can stack (--skip-smoke --skip-e2e = static + unit only, zero API cost). --only-* is mutually exclusive with all other flags.

Exit codes

Explicitly out of scope

• Modifying source files. The skill is read-only against the repo. If verification exposes a bug, re-run /dejavu-integrate on the same goal + plan; do not hand-patch.
• Fixing broken tests. Failing unit tests are a signal that the integration is wrong, not that the tests are wrong. The skill does not "try a different test."
• Deep logical correctness. The E2E step proves "something the user said earlier comes back later," which is a useful but shallow signal. It does NOT prove the integration picks the right facts to store, scopes user_id correctly across real users, or handles conflict resolution well. That's human review territory.
• Self-healing. This skill never modifies source files. The paired /dejavu-integrate skill in its default --heal mode consumes the scorecard produced here and drives its own remediation loop. Exit code 7 (non-invasiveness violation) is the explicit signal the heal loop must stop and surface to the user.
• Cross-branch comparisons. No main baseline diffing. The scorecard reflects this branch only.
• Running against production data. Every smoke test uses a disposable random user_id and cleans up after. Never touches any other user's data.

Source: JSingletonAI/dejavu — distributed by TomeVault.