# Support and release status

Consumer delta2026-10-04 preserves the request failure category when a duplicate
incomplete Response exceeds the4KiB error copy. Only the copy is omitted;
the bounded original remains in RunError.partial.pending_response and never
authorizes tools. Five new regressions are included in102 affected offline
cases;58 separate runtime/persistence cases also pass. Dragonex23 focales verify
a paid inventory receipt across token retry. The failed real delivery19447/52
is preserved; no live fix or new complete framework run is claimed. The2,235
baseline below predates both October4 deltas. No publication/version bump.

Consumer delta 2026-10-04 adds one backend envelope-guidance message for
qualified tools without changing schemas, canonical history or validation.
131 affected ReqLLM tests pass offline; one Dragonex/DeepSeek4.1/Decart
dialogue succeeds with an attributed statement and zero argument retries.
The complete 2,235-case result below predates this delta. No new publication
or universal provider acceptance is implied; see the2.0.0 changelog and design record.

Updated **2026-10-04**. These docs describe **ExAgent 2.0.0**. The maintainer's
workflow uses direct commits/pushes to `main`; a published stable GitHub release
triggers package/docs publication. Release metadata and the workflow are prepared;
the maintainer configured the GitHub Hex API key on2026-10-04. Each release job
records authentication, publication and readback of the public artifact.
Availability is established by the
[Hex release](https://hex.pm/packages/exagent/2.0.0) and its versioned HexDocs,
not by the version string alone. See [publishing](development/releasing.md).

October3 release preparation verifies17 offline Git/loopback guards,18 executable
documentation cases, strict dev/test compilation and ExDoc/site/TAR checks.
A fresh installed2.0.0 minimal graph passes8 contracts; its aggregate diagnostic
remains exit1 from38 upstream warning lines. At that checkpoint all94 library
source files and the dependency lock match the accepted runtime freeze; the
October4 adapter and retention deltas have a separate combined159-case offline
qualification on Elixir1.20/OTP29, zero failures in14.3 seconds. This includes
the131 adapter cases and the retention contracts, not a new complete suite. The
[preparation receipt](https://github.com/akorda-software/exagent/blob/codex/v2-candidate-029/docs/orchestration/2026-10-01-v2-codex/RELEASE-PREPARATION.md)
distinguishes the installed staging TAR from the final documentation artifact.

The requested operational closure now integrates the Dragonex extensions, opt-in
metrics through official experimental0.6 APIs and the stock HTTP disposable-VM
recipe. A reproduced SDK-restart cache defect is corrected through a live public
provider lookup.194 focused cases pass on1.18; the complete1.20 suite passes2,235
cases with zero failures and28 exclusions in2,001.6 seconds. The original routine
exits1 at ExDoc because a callback link lacks its `c:` prefix. That prose-only
correction passes strict generation and link readback; the failed invocation is
retained, without a suite rerun. Five installed graphs pass38 functional contracts
on the frozen source. Final artifact checks and identities are recorded in the
[closure record](https://github.com/akorda-software/exagent/blob/codex/v2-candidate-029/docs/orchestration/2026-10-01-v2-codex/OPERATIONAL-CLOSURE.md).

## What is accepted

| Boundary | Evidence | Scope |
|---|---|---|
| Offline runtime | Operational FULL1.20:2,235 passes /0 failures /28 exclusions.194 integrated cases pass on1.18;53 affected tracing/metric cases pass on1.20 after the causal SDK-restart fix. | Five initial local phases pass. ExDoc initially fails on a callback link; corrected generation/link checks pass separately. No new complete1.18 run. Prior2,208/2,198/2,177 receipts retain their identities. TestModel/loopback fixtures, not real-provider compatibility. |
| Runtime targets | Elixir 1.18 / OTP 28 and Elixir 1.20 / OTP 29; strict test compile passes on both. | Tested combinations, not every patch release or dependency graph. |
| Real model | Updated dependencies: 12 scenarios pass initially; length sync/stream pass after correcting the stimulus. All 14 scenarios covered, 18 admissions / 3 effects. | Original refusal/stop failure retained; not a single new 14/14 wave. GPT-4o-mini/OpenRouter Chat tools profile only. |
| Real consumer | October 3: 27 scenarios accepted per model, reusing 26 unchanged cases plus the newly qualified required-header native case08. Consumer precommit: 20 offline passes /27 opt-in exclusions. Complex workflows retain their 4/4 receipts,46 requests /15 effects per model. | OpenRouter Chat, reasoning disabled. The old nullable extraction still fails. Smoke phases now Luna53/DeepSeek57; complex phases retain70/53. Invoice unobserved. No new complete27-case wave or universal provider qualification. |
| Durable recovery | October 3 PostgreSQL 17.4: 14 phases, including lost COMMIT ACK, competing resumers, fresh-VM resume, explicit recovery and backup/restore; cleanup confirmed. | Declared single-database/host profile, not universal HA or external exactly-once effects. |
| Langfuse and Opik | Each: prior native/API 33/33 observations, 667 attributes, 12 model usages; UI 12 cases / 248 attributes. Exporter 1.11 booleans newly verified locally. | Same finite A10 criteria. No fresh cloud/API/UI wave. Content off; synthetic TestModel tokens. |
| Package consumers | Operational staging TAR: five clean1.20 graphs,38 contracts; none/API/SDK/exporter ×8 plus six metric SDK/reader cases.94 library sources match the FULL freeze. | Staging TAR and final documentation-only artifact retain separate identities. Actual ExUnit records have zero failures/exclusions/skips. Functional acceptance; upstream strict warnings remain red. Earlier32/28/56 receipts retain their scope. |
| Integrations | Official MCP SDK 2.2.0: five profiles; LiveViewTest/Oban SQL: six cases plus crash/recovery. | Qualified recipes, not every deployment combination. |
| Load and cleanup | Finite TestModel load/soak/saturation and owned-resource cleanup. | No LLM latency SLO, cloud availability or upstream predecode RAM guarantee. |

The earlier observability lifecycle polish added an optional application-supervised
ReqLLM maintenance child and a per-instance exporter worker restart budget.
Its complete1.20 suite takes1,995.2 seconds; the nine-phase routine takes2,035.684
seconds. The final documentation/tooling correction preserves all92 library
sources from that freeze; Dragonex was outside that dated acceptance and is now
integrated by the operational closure above.
TTL must exceed live request durations. A finite budget closes admission after
exhaustion; its default remains infinity and restarting the processor resets it.
Native HTTP verifies one-profile recreation containment with budget0 while
retaining the surviving-socket limit. Stock ReqLLM metric APIs also differ from
published experimental0.6 arities; that receipt did not qualify metric export.
The new opt-in Adapter uses those public arities, with its own SDK/reader checks.
See [known limits](development/known-limits.md) and the
[lifecycle receipt](https://github.com/akorda-software/exagent/blob/codex/v2-candidate-029/docs/orchestration/2026-10-01-v2-codex/OBSERVABILITY-LIFECYCLE.md).

The critical review's one P1/four P2 findings were corrected and verified.
The single final R9 integration/distribution review is closed. No known
unresolved finding from those scoped reviews is being hidden by this summary.

The subsequent owner alignment pass corrects stale documentation rather than
changing execution: exhausted TestModel scripts fail, SQL/C7 and both tracing
backends have scoped acceptance, and API references resolve against the candidate.
A fresh 146-case critical run passes; the documentation probe passes 18 cases on
both runtimes. TestModel's executable AST and six runtime BEAM chunks are unchanged,
so the earlier complete runtime evidence retains its scope. See the
[alignment receipt](https://github.com/akorda-software/exagent/blob/codex/v2-candidate-029/docs/orchestration/2026-10-01-v2-codex/RELEASE-READINESS.md).
This pass does not reopen the independent R9 review or expand provider guarantees.

The subsequent [known-limit investigation](development/known-limits.md) isolates
native extraction to the caller's optional header contract. The final E2E schema
requires merchant/currency and case08 passes with both models, while reverting
to nullable reproduces DeepSeek's failure. ExAgent runtime and dependency lock
remain unchanged. Stock warnings have identified upstream fixes/release blockers;
the strict diagnostic remains red. On2026-10-03 the user chose to proceed with
publication while retaining these external limitations, without changing upstream
projects or suppressing their diagnostics. This is a release decision, not a
new green diagnostic or an expansion of qualified profiles.

## Read the CI result accurately

[Compatibility run03](https://github.com/akorda-software/exagent/actions/runs/37019949320)
tested commit `f69aedea1a2f7c94c546fa95b631a71664110b4d`.
Both complete suites passed with warnings-as-errors:
**2,176 passes, zero failures, 28 exclusions per runtime**. Package and finite
harness checks passed, as did all functional consumer contracts.

The **overall run is failure**, because strict dependency diagnostics report
stock TOML/WebSockex/gproc warnings. There are no ExAgent compiler warnings in
those accepted checks. G5 strict diagnostics remain open; there is no suppression,
fork or forced unsupported upgrade.

The local `bin/check` routine subsequently passed all eight phases. It took
**33m44s**, including a 1,979.7-second suite, almost entirely synchronous work.
This does not demonstrate a 20× local speedup. New documentation checks are
recorded separately; they do not relabel an old TAR or test receipt as new evidence.

## Dependency review — October 3

All 43 locked Hex packages were checked against official stable releases.
Eleven are updated; 42 are current. Gproc remains at 1.2 because stock grpcbox
requires `~> 1.2.0`, excluding 1.3. No unsupported override is introduced.
The [dependency record](development/dependencies.md) lists every version,
compatibility change and official source. `hex.audit` reports no retired packages
or advisories for this lock at the time of the query.

The new local routine passes all nine phases in 2,073.45 seconds. The complete
1.20 and 1.18 suites pass in 2,007.4 and 2,024.7 seconds respectively. Compile
checks report no ExAgent warnings. The previous failed 1.18 receipt remains below;
these new runs qualify the updated dependency graph independently.
Fresh G2 covers all fourteen scenarios with the initial failure retained and two
causal length controls; PostgreSQL and all eight clean package graphs also pass
their functional checks. Strict upstream diagnostics remain open.

The subsequent [application model comparison](development/real-consumer-e2e.md)
uses GPT-6-Luna and DeepSeek V4.1 Flash instead of Mini, with five additional
negative/lifecycle/composed-approval scenarios. Luna's initial 23/23 wave passes;
affected exact-marker controls also pass. DeepSeek's 22 accepted scenarios span
several finite waves; native JSON receipt output still has missing merchant and
currency despite schema validation. Its native extraction profile remains
unqualified in that receipt. The later required-header control above qualifies
the updated case08 without relabelling that nullable failure. No private parser,
coercion, guard relaxation or model fallback is
used to convert that failure to a pass. The library runtime and root lock are
unchanged by this comparison; the earlier G2 Mini receipt remains historical.

Four additional combined workflows now exercise six typed stages, two levels of
delegation, failure collection, lost effect ACK with explicit uncertain recovery,
and a complete application journey across fresh VMs and a PostgreSQL restart.
The last combines memory, shared state/turns, cancellation, parallel delegation,
two approvals, finalization and PubSub streaming. All four pass offline and with
both real models. An initial DeepSeek Session-marker failure is retained; a
clearer ASCII-copy instruction passes the unchanged oracle in both profiles.
Owned processes and databases close; confirmed work is not replayed. These
receipts extend the application matrix without rerunning unrelated qualifications.

## ReqLLM 1.26 qualification — October 2 receipt

The candidate requires ReqLLM 1.26.0 and its llm_db 2026.9.8 catalogue. The fresh
local routine passes all nine phases, including the complete 1.20 suite, ExDoc,
link readback, TAR and isolation. Fresh G2, application E2E, PostgreSQL and clean
consumer results are listed above. Historical 1.24 receipts retain their identity.

The 1.18 integrated run executes all 2,205 cases: 2,176 pass, one fails and 28 are
excluded. The failure is a 100 ms tool-readiness assertion before owner-death
cancellation, using TestModel. A controlled 150 ms startup reproduces it; the
existing 1,000 ms readiness barrier used by neighbouring ownership tests passes.
The cancellation assertion and runtime remain unchanged. All 12 ownership cases
plus that slow-start control pass on both runtimes; the original full-suite failure
is retained. This October 2 receipt contains no second green full run; the separate
October 3 dependency-baseline runs above qualify their later sources.

Argument, accounting, continuation and capability guards remain in place. Stock
Anthropic now preserves redacted provider blocks; ExAgent rejects their unqualified
continuation explicitly. Cache reads/writes remain separate without double-counting
inclusive input or reasoning. This update does not expand supported profiles.
Langfuse/Opik retain their previous native/API/UI qualification; their SDK,
exporter and ExAgent bridge are unchanged, with local tests exercised by this run.

## What remains before publication

- Resolve or explicitly decide the strict stock-dependency diagnostic policy
  using compatible upstream releases. Preserve the red diagnostic evidence.
- Finalize versioning, release notes, tag and the separately authorized Hex
  publication pipeline. The current draft PR is not a published v2.
- Qualify any additional provider, modality, deployment or scaling profile before
  claiming support for it. These are future qualification goals, not silently
  implemented guarantees.

The October 3 [instrumentation ownership implementation](development/backend-evaluation.md#reqllm-and-exagent-instrumentation-ownership)
keeps ExAgent as generation owner during its observed Model requests. The
host opt-in `ExAgent.Observability.ReqLLM.attach/1` uses ReqLLM's public adapter to
enrich that span and preserve standalone tracing. Incompatible/duplicate bridges
reject before provider IO in the ExAgent ReqLLM adapter; no foreign handler is
removed. Twenty-one bridge cases within 146 integrated cases pass on both
Elixir 1.18/OTP 28 and 1.20/OTP 29 with identical runtime sources,
including concurrency, sampling/named tracer, once-only pricing, raw-capture
privacy and owner death/public prune. The complete local 1.20 routine passes all
nine phases: 2,198 offline tests pass, zero fail and 28 opt-in cases are excluded.
Four clean package graphs pass 28 contracts; strict dependency diagnostics remain
red for stock TOML/WebSockex/gproc. No new cloud/UI or metrics export acceptance
is inferred; equal prior Langfuse/Opik A10 evidence remains. See the
[combination receipt](https://github.com/akorda-software/exagent/blob/codex/v2-candidate-029/docs/orchestration/2026-10-01-v2-codex/OBSERVABILITY-COMBINATION.md).

Routine verification now runs locally before commit/push. The candidate branch's
six-job CI workflow is manual; `main` adopts that change when the PR is integrated.
The [verification guide](development/verification.md) describes `bin/check`,
its optional package consumers and focused documentation checks.

## Exclusions and observations

The 28 offline exclusions are **22 real-provider tests and six PostgreSQL tests**.
Their bodies did not run; exclusions are not passes or timeouts. Real-provider and
SQL acceptance above come from separate explicit profiles.

Langfuse/Opik preview Input `null` and Output `undefined` are expected with
content capture disabled. Inspect IDs/parentage, status, request/tool counters,
usage quality/provenance and estimated cents in attributes. Neither export ACKs,
API readback nor UI visibility proves durable retention, invoice accuracy or
browser/cloud reliability.

## Source of detail

- [Roadmap](development/roadmap.md): the current R0–R9 board and evidence boundaries.
- [Release scope](development/release-scope.md) and
  [production acceptance](development/production-acceptance.md): promised profiles and oracles.
- [Consumer E2E](development/real-consumer-e2e.md) and
  [backend acceptance](development/backend-evaluation.md): exact external profiles.
- [Execution policy](development/execution-flow.md): at most one independent review per objective.
- Historical status is preserved under
  `docs/archive/2026-10-02-status-history.md` in the checkout. The
  [original dated record](https://github.com/akorda-software/exagent/blob/70c4a0c9d4b3f69805f24def7e61def5b0f5ab08/docs/status.md)
  retains its original paths and results. Historical pending labels do not reopen
  accepted work.

Receipts and private tooling remain outside the consumer package. Read task guides
to integrate the library; read these evidence records when evaluating support.