Skip to content

Implementation verification

This page retains dated implementation/audit evidence. Statements about local-only work describe those runs, not the repository's current publication or push status. Repository-alignment verification subsequently passed 508 existing tests with 89.27% combined coverage, plus three new distribution-contract tests separately. See the alignment report for its exact commands, security scope and remaining boundaries.

The following historical checks ran locally before publication on 10 November 2026 for 0.1.0, including the optional navigation projection and its concurrency, recovery, FSM and keyboard tests.

Historical implementation results

Check Result
Python 3.12 / aiogram 3.31.0 / redis-py 7.4.1 / Redis 7.4.11, frozen source 471 passed; 89.19% combined line/branch coverage
Python 3.14 / aiogram 3.31.0 / redis-py 7.4.1, frozen source 471 passed
Python 3.10 / aiogram 3.22.0 / redis-py 5.0.8 389 core cases plus all 82 optional cases passed across the core/optional/boundary runs
Fresh wheel with nav/redis, redis-py 5.0.8 91 optional/component/ownership cases passed outside checkout
Fresh sdist with nav/redis, redis-py 5.0.8 Same 91 cases passed independently
Standalone wheel/sdist, nav absent 389 core cases passed for each; optional modules skipped explicitly
Base wheel/sdist, Redis and nav absent 232 cases passed for each; infrastructure/optional cases skipped explicitly
Final base and Redis-only artifact import boundaries Passed; missing nav import gives installation guidance
Frozen standalone checkout sync without peer checkout Passed
Ruff lint and formatting Passed
Strict core and optional source/example typing Zero errors
Wheel/sdist metadata through Twine Passed
Example compilation, CLI help and four-style Dispatcher example tests Passed
Installed dependency audit No known vulnerabilities; unpublished local packages skipped
Kit strict documentation build Passed
Whitespace and Python file-size checks Passed; files below 500 lines

The 471 unique cases comprise 389 core tests and 82 optional navigation tests. Earlier full endpoint runs passed 443 cases; subsequent optional runs cover the added keyboard/schema/business-recovery cases, and the final minimum-version boundary run covers shutdown ordering. The frozen Python 3.12/3.14 runs execute the entire final suite in one command. Artifact core suites were run before the last optional-only changes; final core files were verified byte-for-byte equal and the rebuilt artifacts reran the optional suite plus component/ownership tests.

The audit adds 69 regression cases to the earlier 311-test implementation. See the findings and fixes for the corrected contracts. The final full runs used frozen source; clean-artifact runs imported installed packages outside the checkout rather than editable source.

Earlier memory-only compatibility runs also passed on Python 3.11 and 3.13. CI covers Python 3.10–3.14, minimum/latest aiogram endpoints and Redis 7/8 with minimum/latest supported redis-py clients. Python 3.14 with aiogram 3.22 is excluded because that release pins dependencies predating Python 3.14 support. The updated CI workflow has been checked locally but has not run remotely.

Behavioral coverage

Shared contracts cover guarded sessions/activity generations, immutable snapshots, deduplicated operations, fingerprint conflicts, expiry, configuration, controls, capacity and retention. Redis tests use actual isolated servers and independent clients/processes rather than an emulated Redis.

Recovery tests cover competing workers, expired claims, killed claiming processes, stale acknowledgements, lost commit responses, reconnection, script-cache loss, confirmed receipts, uncertain sends, exhausted retries and Continue/Cancel/timeout races. Atomic tests interpose between reads and commits to verify namespace and configuration-loss fences, immutable profiles, parked-state fencing, validated claim snapshots, heartbeat replay ordering and delegated effect capabilities.

Corruption tests reject invalid identities, policies, jobs, receipts, tokens, unsupported schemas and binary data without overwriting recovery evidence. Queue tests cover missing jobs, orphan due entries, reconciliation continuing past corrupt work, bounded scan overflow, index reconstruction and explicit release after repair. Retention admission is checked with 128 live claims ahead of a prunable terminal record.

Dispatcher.feed_update tests verify full-handler protection, enabled/disabled FSM isolation, consistent local/distributed lock ordering, actual borrowed-lock identity, unchanged middleware configuration and lossless Redis FSM keys. They also verify separate dispatchers sharing Redis FSM state and raw-state reload after acquiring ownership. Revoked scopes cannot reacquire, delegated effects cannot use another conversation's capability, and timeout/lease cancellation consumes only its own request. These cancellation cases run on CPython 3.10 as well as modern Python.

All four prompt styles use the Redis runtime with fake Telegram sessions. No real bot messages were sent.

Limits and deferred capabilities

  • Memory remains single-process and loses all state on restart.
  • Redis infrastructure persistence, replication, rollback and Sentinel failover were not qualified. Process-crash recovery assumes authoritative Redis data survives; lost history cannot reconstruct external effects.
  • Redis Cluster, external session-store integrations and a durable analytics outbox remain deferred. The optional navigation adapter is implemented separately. Core and Redis need no aiogram-nav.
  • Custom FSM isolation requires a supported adapter and is rejected. Redis FSM builders must include bot identity; selecting idle Redis storage does not persist application drafts or a memory FSM.
  • Unknown business outcomes require evidence-based recovery. Leases cannot retract in-flight Telegram/database requests; application writes still need generation guards and idempotency.
  • Real Telegram-client UX on iOS, Android and Desktop was not manually tested.
  • Existing older Redis layouts require explicit migration, never automatic history/configuration replacement.
  • Changes remain local and uncommitted. No publication, push or deployment occurred.

The dedicated Redis test containers were removed after verification.

Optional navigation integration

The integration runs a separate four-style/two-backend matrix. Tests assert atomic nav/activity/deadline updates, mapped counters and exact touch receipts, inherited policies, explicit restarts, no acceptance from background renders, prompt-control precedence, shared memory/Redis FSM coordination and shutdown ordering. They also cover process exit after confirmed rendering, failed final cancellation commits, unknown deliveries, malformed component parking and lost/replaced claims.

Reply/hybrid tests inspect actual Telegram method arguments after Continue, Cancel and ordinary accepted input. They verify application keyboard restoration, explicit and implicit removal, silent restoration, durable background recovery, no extra accepted activity and no repeated business cancellation. The canonical presenter receives temporary-keyboard provenance through an optional wrapper; standalone hybrid delivery behavior is unchanged.

See the integration guide. The manual peer-repository CI workflow covers Python 3.10–3.14 and an additional Redis 8 job; no remote CI execution is claimed. Publication, Kit dependency extras, external session-store transactions, Redis Cluster and durable analytics outboxes remain separate milestones.

Follow-up integration ownership review (10 November 2026)

The subsequent review added guards against changing reserved adapter/hooks or coordination after binding, including changes while an FSM handler awaits its lock. Unsupported custom/no-op coordinators fail admission, and rejected installation restores presenter/writer bindings. Node policy mappings are frozen.

Component transactions reject unrelated-slot insertion and changes to a live idle effect ledger. Finalization verifies the staged navigation reservation, canonical identities and retained receipts; interrupted prepare retries preserve existing render intent. A cached resume reservation cannot be reused to restore a different screen. That eligibility-change case fails closed and needs explicit operator recovery rather than automatically retargeting an uncertain operation.

Separate idle effect recorders in one coordination scope now share stage and checkpoint locks, preventing repeated replay-safe effects in delegated tasks. Observational hook failures remain separate from committed business actions.

Verification: 213 affected integration/component/core tests passed against a private Redis 7.4 server before the final effect-locking regression; the final 134-case focused suite (including all new review cases, Redis recovery, memory contracts and the new recorder regression) passed after those final changes. Ruff and both strict Pyright configurations passed. Wheel/sdist builds and Twine checks passed, and the wheel's Redis extra imported memory/Redis backends in a clean environment with aiogram-nav absent. The earlier full-matrix figures above remain historical evidence; this follow-up did not rerun that entire matrix.