Skip to content

Releases

v0.4.1 — Routing and benchmark hardening

Choose a tag to compare

@santapong santapong released this 09 Sep 13:05
f60c7fa

Accretion v0.4.1 hardens routing evidence, utility alignment, and benchmark reporting following v0.4.0.

Changes

  • Training snapshots: resolve run-projected records through experience_id, allowing evidence from routed runs to enter a training snapshot instead of being rejected as an empty window (#161).
  • Utility alignment: align the selector’s default weight vector with the registered benchmark weights, 1.0 / 0.3 / 0.15. Persisted objective contracts retain their original weights (#162).
  • Explicit benchmark pooling: add a registered PoolingRule and report both conjunction and per-trial rates while preserving the original v0.4.0 results and thresholds (#163).
  • Amendment 1: preserve the confirmed, frozen amendment and its one authorized locked/drift re-read in the research record. Results include the expectations that failed; this is not evidence of an overall positive outcome or live-provider benefit (#164).
  • Release CI: make the release-bridge exception for computed-style base comparisons explicit (#160).

Compatibility and evidence

No database migrations were added relative to v0.4.0. Node routing remains disabled by default, with ACCRETION_ENABLE_NODE_ROUTING=false and ACCRETION_NODE_ROUTING_MODE=BASELINE_ONLY.

The recorded acceptance line is unchanged: 167 criteria in scope, 159 proven by claiming tests, 5 frontend witnesses, 3 recorded manual/live-provider witnesses, and 0 unmet MUST criteria; all 5 inherited release-gate conditions passed. These are historical release results; publication did not rerun the test suites or research study.

The routing benchmark remains synthetic/replay-only, and no priced real-provider routing study was run. Original results remain alongside Amendment 1. This tag inherits v0.4.0’s disclosed limitations, including restart-sensitive in-memory exploration accounting and no exactly-once guarantee for external execution. See the tagged audit and research results.

Existing annotated tag: v0.4.1 → f60c7faae68a4ebae7ad4e15696746957690ce7b (release bridge #166). This page documents that original released tree; later develop changes are not included.

Full release notes and v0.4.1 addendum · Changelog · Changes since v0.4.0

v0.4.0 — Evidence-aware node configuration routing

Choose a tag to compare

@santapong santapong released this 09 Sep 13:04
dd1d9f3

Accretion v0.4.0 delivers evidence-aware node configuration routing for its provider-neutral control plane for local AI coding agents.

Highlights

  • Deterministic compatibility checks and receipt-first selection across agent, tool and verifier configurations, with immutable contracts and recorded decision explanations.
  • Independent verification, bounded recovery, and an experience/feedback pipeline for routing evidence.
  • Offline ranking and calibration, project adapters, shadow evaluation, guarded exploration, and recorded model promotion/rollback.
  • Experiment Studio routing views and operator UI styling, plus a registered replay benchmark with preserved results and ablations.

Defaults and limitations

Node routing remains opt-in: ACCRETION_ENABLE_NODE_ROUTING=false and ACCRETION_NODE_ROUTING_MODE=BASELINE_ONLY by default. Requested modes or exact tool bindings that the runtime cannot honor fail closed.

Research evidence is synthetic/replay-only; no priced real-provider routing study was run. Both safety gates fail on the original locked and drift corpora, so these results do not establish overall routing superiority or live-provider benefit. Exploration accounting in this tag is in-memory and loses its history on restart. External execution is not guaranteed exactly once. See the tagged audit and limitations and original research results.

Upgrade and validation

Run uv run alembic upgrade head. The migration chain is 0017 → 0018 → 0020 → 0019, with 0019 as the head. Downgrade checks require a clean database; populated experience history is protected against destructive downgrade.

Recorded release acceptance: 167 criteria in scope; 159 proven by claiming tests, 5 by the frontend suite, and 3 by recorded manual/live-provider witnesses; 0 unmet MUST criteria. All 5 inherited release-gate conditions passed. These are the original release records, not a new test run for this publication.

Existing annotated tag: v0.4.0 → dd1d9f300ba05dc5875b964322c34b47a5288c39 (release bridge #158). This page documents that original released tree.

Full release notes · Changelog · Changes since v0.3.0

v0.3.0 — Plugin, MCP & Identity Integration Platform

Choose a tag to compare

@santapong santapong released this 01 Sep 05:36
bf5b774

Plugin, MCP & Identity Integration Platform.

v0.3.0 turns the v0.2 control plane into an integration platform: capabilities resolve through connectors, connections and bindings rather than environment variables; principals sign in through OIDC; credentials live in an encrypted broker instead of the process environment; and remote MCP servers, plugins and a research capability set are governed by the same policy, audit and isolation rules as everything before them.

Every v0.3 feature is disabled by default, and none of them can expand permissions, bypass approvals or verifiers, expose credentials, or erase durable execution history.

Highlights

  • M0 connection-aware capabilities — connector, connection and binding contracts with a capability resolver; every v0.1/v0.2 capability resolves unchanged.
  • M1 identity and SSO — principals keyed by issuer and subject, workspaces and memberships, an OIDC Authorization Code + PKCE client, session middleware with a LOCAL_PRINCIPAL default for local development.
  • M2 token broker — an encrypted secret store, single-use OAuth transactions, broker-backed capability invocation, and a GitHub connector with the full connect / callback / reauthorize / revoke / health lifecycle.
  • M3 remote MCP — authenticated MCP SDK v2 HTTP discovery and invocation, durable per-connection discovery snapshots, server lifecycle and audit records, health and circuit-breaker state.
  • M4 plugin manager — a plugin version registry, workspace-scoped installations, manifest trust and dependency resolution, and a resolver gate.
  • M5 research intelligence — an MCP-backed research capability set with evidence normalization, connector-attributed sources, and trust that the connector cannot assert for itself.
  • M6 operator administration — Plugins, Connections, MCP Servers, Capability Inspector and Identity pages, plus run-detail inspectors that project real audit provenance onto the graph.
  • M7 enterprise-managed authorization — optional EMA, off by default, in which a centrally issued token still passes through capability policy, connection isolation and the audit trail unchanged.
  • M8 release hardening — the release gate became executable and every inherited acceptance criterion is now claimed.

Acceptance

Count
Criteria across the three SDDs 117
Proven by a passing claiming test 111
Proven by the frontend suite 3
Proven by a recorded live-provider run 3
Uncovered 0
Unmet MUST 0

make acceptance exits PASS and CI gates that full harness rather than the eight stage-scoped subsets it replaces. make release-gate evaluates SDD §24.8 as five separately failable conditions, all of which pass.

axe-core 4.10.2 reports zero violations across all seventeen routes, no route scrolls horizontally at 390 px, and none of 1,421 measured text nodes falls below WCAG AA — discharging the v0.2 accessibility exception (#52).

Honest limitations

This release does not claim more than it proved.

  • The P5/P6/P7 benchmarks are replays of frozen traces, not live experiments. Every published figure is pinned literally and shown to be derived from the corpus it summarizes, but a replay is a reproducibility guarantee, not a new measurement.
  • The P7 stale-rejection figure is served from the declared corpus. V02-P7-003 is proven directly against the real ExperienceService.assess() across all 19 reason codes, but the benchmark's two API routes still report stale_rejection_source: DECLARED. Closing the document half is a v0.4 item.
  • Three criteria are manual. V01-P0-002, V01-P0-004 and V01-P4-008 describe real vendor CLIs and cannot run in CI. They are backed by a dated evidence document from a real signed-in run and expire on 2027-02-28.
  • The §24.8 counters are derived, not telemetered. secret_exposure_incidents and capability_policy_bypass are computed from test outcomes and audit rows.

Upgrading

Migrations 0010–0016 apply in order and are reversible against a clean database:

uv run alembic upgrade head

The LOCAL_PRINCIPAL session mode is the default, so an existing local deployment keeps working without an identity provider.

Release integrity

Item Value
Release commit bf5b774eb964252d448b44ec3ea9d6b7b7511213
Release tree 902c5c75aa899ecb2306cf26696bbccb867fc797
Authorized develop 0bf1d747eb9428efe13d8e71b13e3866e9cebb92 (#102)
Protected bridge b849ca0846b12977042d72cbc5a5841504ca7fbd (#103)

The v0.1.0 and v0.2.0 tags are unchanged and remain the static control and previous release.

Docs: release notes · audit · frozen baseline · accessibility evidence · release-hardening runbook

Full changelog: v0.2.0...v0.3.0

Accretion v0.2.0

Choose a tag to compare

@santapong santapong released this 24 Aug 07:40
de146cd

Accretion v0.2.0 release notes

Version: v0.2.0. Release date: 2026-08-24.

Accretion v0.2.0 extends the immutable v0.1 static control plane with opt-in,
deterministically governed dynamic orchestration. Dynamic features remain
disabled by default and cannot expand permissions, bypass approvals or
verifiers, expose credentials, or erase durable execution history.

v0.2 release evidence flowing from the immutable v0.1 control through P5 dynamic, P6 search, P7 transfer, and the full clean-checkout gate before promotion to main

Highlights

  • P5 validated dynamic workflows: typed proposals, deterministic validation,
    one bounded repair, immutable graph revisions, safe replanning, observable
    runtime selection, and a validated v0.1 static fallback.
  • P6 bounded candidate search: best-of-N, hypothesis, cross-provider, and
    generator-reviewer modes with isolated worktrees/sessions, shared budgets,
    independent ranking, explicit stop reasons, and crash-safe promotion.
  • P7 verified experience: immutable terminal evidence, deterministic local
    embeddings, compatibility and transfer-risk scoring, negative procedural
    knowledge, controlled replay seeds, fresh controls, and repeated revalidation.
  • Complete operator frontend: eleven routes spanning task planning, live
    execution, governance, ACR-ARCH, and dedicated P5, P6, and P7 research views.
  • Frozen research evidence: reproducible ACR-ARCH, P5 static/dynamic, P6
    quality-versus-compute, and P7 experience-transfer suites with exact hashes
    and preserved negative/null results.

Research results

Gate Frozen result
P5 dynamic workflow +0.224631 utility on heterogeneous/uncertain tasks; predictable cohort +0.009195; static fallback PASS
P6 bounded search verified acceptance rises from 8/12 at N=1 to 12/12 at N=4; mean quality 0.472500 → 0.768333
P7 verified experience replay quality +0.070500, 20% fewer tools, 95% stale rejection, 3.33% negative transfer

These are deterministic replay-fixture results. Signed-in Codex and Claude
calibration is recorded separately in the release audit and does not rewrite
the frozen metrics.

Compatibility and migration

  • Python 3.12 or newer and Node.js 22 or newer remain required.
  • PostgreSQL 16 with pgvector is the supported database used by the release gate.
  • Apply Alembic migrations through 0009_p7_experience_contracts.
  • Existing v0.1 endpoints remain available; v0.2 behavior is additive under
    /api/v2 and independent project/deployment feature flags.
  • Validated live CLI ranges are Codex CLI >=0.148,<0.149 and Claude Code
    >=2.1.231,<2.2.
  • Claude runs use safe mode to exclude ambient hooks/plugins and honor an
    explicit typed model selection; the release calibration records sonnet.

Install

git clone --branch v0.2.0 https://github.com/santapong/Accretion.git
cd Accretion
cp .env.example .env
uv sync --all-groups
npm ci
docker compose up -d postgres
uv run alembic upgrade head

Run make api and make ui, then open http://localhost:5173. Start with the
deterministic fake runtime. See the frontend guide,
developer guide, and showcase.

Evidence and limitations

The release audit identifies the exact candidate, checks, tool
versions, provider calibration, protected promotion, and release decision. The
maintainer authorized this tag with one documented exception: the supported
browser-control surface exposed no connected browser, so no rendered route,
responsive, keyboard, focus, or automated accessibility PASS is claimed. That
post-release validation is tracked in
#52. The v0.1 tag remains
immutable and reproducible through its baseline record.

v0.2 does not include learned routing, self-modifying policy, automatic
experience promotion, unrestricted production deployment, or the v0.3
identity/plugin/connection platform. The current frontend bundle-size advisory
is non-blocking and remains documented technical debt.

Accretion v0.1.0

Choose a tag to compare

@santapong santapong released this 22 Aug 14:49
6324c8f

Accretion v0.1.0 release notes

Accretion v0.1.0 is the first stable observable static meta-harness release. It
supervises signed-in Codex and Claude Code runtimes through a provider-neutral,
local-first control plane while keeping architecture selection, workflow topology,
policy, and independent acceptance authoritative in code.

Shipped scope

  • Structured Codex App Server and Claude stream-JSON adapters with normalized,
    append-only events, health classification, interrupt/resume, and isolated Git
    worktrees.
  • Versioned prompt/context contracts, typed deterministic profiling, explainable
    DIRECT/LOOP/GRAPH/HYBRID selection, audited overrides, and a safe unknown fallback.
  • Bounded verifier-gated feedback loops and five validated static templates,
    including checkpoint/replay and approval-gated GRAPH/HYBRID execution.
  • Immutable capability, skill, plugin, and policy registries; task-scoped MCP tool
    exposure; deny-by-default policy; approval-bound idempotent side effects; and
    credential injection only at the executor boundary.
  • Complete operator routes for dashboard, task profiling, live runs, runtimes,
    trace history, approvals, capabilities, and ACR-ARCH. SSE reconnect begins from
    an authoritative snapshot and a monotonic cursor.
  • A frozen ACR-ARCH v1 corpus with 30 versioned tasks, 68 replay scenarios balanced
    34/34 across Claude and Codex, independently versioned environments/configuration,
    raw cost/latency/risk/human-burden dimensions, utility, and selector regret.

Reproducibility

The v0.1 release gate was reproduced from a clean checkout of candidate
934ea75cf06be820ac6fdd0946e3203982c779c4 on PostgreSQL 16.

Artifact SHA-256
ACR-ARCH task corpus 9251bb918912e73a2dade20189f93cc26cd7bc217a0dea03713ef252843b9dd7
ACR-ARCH replay traces 2f62f87eaf079914d41f47bea57a4dd04ce469d0e54f1d6a38faa6de0dd6f051

The replay produced 30 tasks and 68 scenario metrics. The optional signed-in
calibration selected two tasks from every category, split them evenly across both
providers, and deterministically verified 10/10 provider-written artifacts. Live
calibration does not mutate the frozen replay dataset.

Runtime compatibility

  • Python 3.12 or newer
  • Node.js 22 or newer
  • PostgreSQL 16
  • Codex CLI >=0.148,<0.149 for live Codex execution
  • Claude Code >=2.1.231,<2.2 for live Claude execution

Live providers remain disabled by default. Accretion checks CLI health and sign-in
state without reading or copying raw authentication tokens.

Upgrade and verification

uv sync --all-groups
npm ci
uv run alembic upgrade head
make check
make test
npm run build

For signed-in provider acceptance and the separate ACR-ARCH calibration:

ACCRETION_LIVE_PROVIDERS=1 PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 \
  uv run --no-sync pytest -p pytest_asyncio.plugin -m live

ACCRETION_LIVE_PROVIDERS=1 \
  uv run --no-sync python scripts/run_acr_arch_live_sample.py

Known constraints

  • v0.1 executes only validated static templates; dynamic workflow synthesis and
    learned routing remain locked to later SDDs.
  • The visual browser pass could not run in the release environment because no
    controllable browser instance was available. The formal P4 page/functionality
    criterion is covered by route/component tests, API tests, type checking, and the
    production build; no visual-browser result is claimed.