Skip to content

Release Aimee 1.1.0: promote testing to main - #3095

Merged
JBailes merged 799 commits into
mainfrom
testing
Nov 16, 2026
Merged

JBailes merged 799 commits into
mainfrom
testing

Conversation

@JBailes

@JBailes JBailes commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Promote testing to main for Aimee 1.1.0, including release preparation #3093, runtime blocker repairs #3094 and promotion CI repairs #3096 and #3097.

The release ships Native memory with scoped personal/shared memory, typed facts, encrypted Vault custody for PII and opaque memory references. Memory, source records and history remain retained. See the release notes and upgrade guide. Existing-vector KB upgrades must keep their original embedder digest; re-embedding is a separate operation.

Promotion checks now use reviewed application-source and public-surface snapshots. Synthesis promotion selects the exact E2E-validated b10219 images by source commit, content tag and immutable digest; CI verifies registry tag/digest agreement before promotion. Pending synthesis wrapper edits wait for the next upstream runtime release.

The required windows-cmake context is restored as a gate over the actual Windows CMake thin-client build and Winsock smoke tests. Both check names are preserved, with no duplicate compilation. The gate rejects failures, cancellations and unexplained skips.

The owner explicitly approved the exact historical PR app identities #3030–3037, #3042–3046 and #3051, plus four exact historical author/committer identities from #3032–3035. The checker pins their repository, app, head, merge/state or commit SHA and identity values. It validates legacy failed-job evidence against those records and rechecks currently clean metadata. Unknown or changed identities, live title/body attribution, message trailers and incomplete evidence remain blocked. The four editable historical body footers were cleaned through Aimee, with originals retained in the local audit. Automatic reviews remain enabled.

Validation carried forward from #3094: all 12 portable plain-storage release gates passed, including real E4B/E2B on both roles, T1/T2/T3, PII/Vault custody, private Docker execution, Vault egress, PostgreSQL, KB upgrade/recreation, browser pairing/owner isolation, readiness/restarts, backup/key recovery and published 1.0.2 migration/rollback. One KB passed eleven complete asynchronous sequences (1,045 assertions). Sanitized evidence passed canary/decryption verification and secret scanning; the disposable LXC and its disk were removed.

Promotion CI repairs were validated locally with the attribution package and API-failure tests, replay of both actual legacy failed jobs, synthesis promotion/release tests, model-input tests, application-lock and surface-baseline tests, current lock checks, cleanup accounting and release-policy checks. GitHub CI evidence is recorded in #3096, #3097 and the local repair report; the current promotion checks must complete on its final head.

All current promotion checks must pass before merging. Main merge approval and release tagging/publication retain their separate human approval gates.

aimee compiled the ring itself: src/Makefile listed the ring's C source files,
compiled them with the ring's private c/src on the include path, the test rules
linked individual ring objects, the module bundle compiled ring client sources
by name, and the lock and go.mod both had to name one exact commit. aimee was
bound to the ring's file layout, not to what the ring promises.

Now aimee names the release range it accepts in dependencies/ouroboros-ring.json
(compatible 0.1, semver) and links an installed ring through the ring's own
package (pkg-config / CMake ouroboros-ring). scripts/provide_ouroboros_ring.py
provides that installation: OURO_RING_PREFIX uses any installed ring whose
declared version is in range, OURO_RING_SOURCE builds a ring checkout, and by
default release v0.1.0 is fetched, checked against its content digest and built
with the ring's own CMake, with the caller's compiler and extra flags (so the
sanitizer lane links an instrumented ring). A dry run provides nothing. The Go
module requires the tagged release go/v0.1.0, and make go-unit-tests checks it
is in range.

- src/Makefile and CMakeLists.txt consume the package; test rules link the
  installed archive; the bundle builder links ouroboros-ring-client.
- check_bus_blast_radius.sh fails on any reach into ring sources or objects,
  and checks the installed client package defines no ouro_host_ entry point.
- Exports build against the release tag; publish-testing and the planner
  watch the new file; cmake is in every image and job that runs make.
- fetch_ouroboros_ring.py and ouroboros-ring.lock.json are gone; an old
  fetched tree in build/deps is removed on first use.
…usage

fix: make get-help topics and bare aimee git usage real
…nd-editor

Unify internal model communication on IR and repair browser thinclient sessions
The merge from testing made the Makefile pass aimee's -fno-ipa-bit-cp into the
ring build, which put a fact about the ring's code in aimee. Ring v0.1.1 turns
the pass off in its own build for GCC, so aimee goes back to passing only its
compiler and extra flags, and names v0.1.1 (C default release and Go module).
Merging origin/testing wrote the new agent_credentials.c and
test_delegate_credentials.c into the worktree but did not stage them, so the
merge commit kept the old versions and reverted testing's change; the test
then failed to compile against testing's headers (unit-tests shards). Take
testing's versions unchanged.
Reference the ouroboros ring by contract, not by commit
… config at start

Legacy from the ring's move:
- docs/dev/EVENT_BUS_DECISIONS.md duplicated the ring's spec/decisions.md,
  EVENT_BUS_FEATURE_TREE.md was the finished build plan, and the arena guide
  moved to the ring rewritten for the current API. aimee's own decision (which
  binaries link the ring, and the audit row's durability) moves into
  docs/EVENT_BUS.md; links point at the ring.
- check_bus_d1_gate.sh guarded a decision accepted months ago.
- bus_bench.c, check_bus_perf_gate.sh and bus_baseline.json: the dispatch
  benchmark is the ring's and now runs in its CI. The audit row's enqueue
  budget is aimee's and stays as check_audit_bus_budget.sh with
  audit_bus_budget.json; the gate's other sections repeated tests unit-tests
  already runs.
- test_bus_flow.c keeps only what is aimee's, ring loss facts reaching the
  WORM ledger; the flow-control semantics it tested are now the ring's own
  test_flow.c.
- Citations of bus_host_*, src/core/event_bus and libaimee-core-event-bus.

aimee-core-c: the repository is gone and nothing consumed it. The lock no
longer pins it (and the lock check refuses a core pin), the exporter no longer
produces it, generated C module CI no longer checks it out, and generated C
module builds no longer find_package(aimee-core), which they never linked. The
core stays bundled and core-c-package CI still proves it installs.

Ring 0.1.2: obs_bus and test_module_runtime attach in-process clients with
ouro_host_attach_local and obs_bus persists capture with
ouro_capture_flush_fd, instead of carrying their own copies. compatible
accepts a floor within a line (0.1.2) so the pin says which release the
API arrived in.

Startup race: the supervisor starts every module at once, and memory and
egress read config before serving, exiting on capability absent when config
was a moment behind. config.SnapshotWhenServed retries that one answer for up
to 30 s; any other failure still stops the module.
test_obs_bus_module_concurrency compiles obs_bus.c in and reused its private
serve thread to check that an owned private pair refuses a client claiming an
external identity. ouro_host_attach_local cannot send a claimed identity, so
the test keeps a three-line serve thread of its own.
Remove ring-move legacy, drop aimee-core-c, consume ring 0.1.2, fix the memory/config startup race
The postgres module registers its stages before its database is reachable,
by design, and answers SQL in that window with "the SQL stage has no
database". The store's startup wait retried only an unreachable module, so
when postgres was attached but not yet connected the aimee module exited and
the supervisor restarted it (seen twice in the 2026-10-28 :testing QUICKSTART
on .252, the same race memory had with config).

The refusal now carries SQLSTATE 08001, as the session stage already does for
an unavailable pool, and the store client reports any class-08 refusal as
ErrStoreUnavailable with the store's reason kept. The store's bounded 30 s
wait therefore covers it; refusals on the merits are unchanged.
The parallel runner already records each failing test with its last
40 lines. On GitHub Actions, emit those records as ::error annotations
too, so the failing test and its output show on the check run itself
without opening the raw job log.
…database

The store waits for postgres to reach its database, not only to attach
The ring is Linux-only (SOCK_SEQPACKET endpoints opened with SOCK_CLOEXEC
and accept4), but since the ring moved to a contract reference the
Makefile built it on every platform, before any target. The macOS LSP
probe never links the ring, yet its build stopped in the ring's own
compiler, which has failed the macOS lsp-real-providers job on testing
since that change.

Off Linux, provide_ouroboros_ring.py --headers now supplies just the
ring's public headers, from the same release and content digest (or the
installed/source ring, range-checked as before), without building it.
The probe still needs them: the module-call API names ring result types.
A product that links the ring stops with a plain message instead.
Unit-test shards have been failing intermittently (pg 2/3 twice, pg
3/3 and sqlite 3/4 once each) and none of them could be diagnosed
without the raw job log. The per-test annotation only covers a test
that fails inside the runner loop; testing's pg 3/3 failure happened
before any test recorded itself.

These steps now keep their whole output, and if they fail a following
step posts its last 60 lines as an error annotation. The LSP probe
build gets the same treatment, so the macOS leg is readable too.
The blast-radius check attributes every OURO_RING_LIB reference to the
target that owns it; a rule whose target was $(OURO_RING_LIB) read as a
new product linking the ring. Name the stop target by its own variable.
…output

Show the tail of a failed unit-test or LSP-probe step on the check run
Every one of the 26 module processes in process-contracts.json runs as
the Go multicall runtime, and the runtime bundle's c-build.json has been
empty: no image has built a C module process since the last module
moved. What remained was the machinery for one and the C handlers kept
as "wire-parity fixtures", which tested C against C rather than
against the Go handlers that serve.

Removed:
- the 14 modules/*/module_adapter.c handlers, learning_signal_policy.c
  and the C trigger matcher in skill_trigger_policy.c, linked by no
  shipped binary (a no-LTO link of every shipped C target, with
  --print-gc-sections, shows none of them reached);
- the exporter's C-process path (generated C mains, CMake, c_build,
  c_init, c-build.json), build_c_module_runtime_bundle.py and
  generate_c_embedded_header.py with their tests, and the image step
  that ran the builder over an empty list;
- db1_stages.h, declaring 18 stage handlers of the deleted C store
  module that nothing defines or includes.

The C caller-side codec tests stay, as test_module_wire_contracts.c.
Coverage the C copies held that the Go tests lacked moves to Go: dedup
key isolation by principal, provider, model, context and behaviour
flags, and the bounded long-input key. The aimee-module-aimee fixture
rule now installs the Go multicall binary; it had been calling the C
builder, which produced nothing, so learning-loop-evidence and
self-learning-efficacy depended on a file no rule made.

Contracts and descriptors now refuse a C process runtime and the
retired c_build/c_init fields.
Sixteen C core tests linked the control-web, kb-synthesis, runtime-web,
workspace and delegates handlers as in-process stand-ins for the Go
module processes, so the code that calls those modules could be tested
without a ring. Removing the handlers broke every unit shard.

They live on as tests/support/module_handlers/*.c, each headed as a test
double for its Go module, and are built by one pattern rule. No shipped
binary links them, and no module descriptor owns them.
The tail of a failed unit-test step showed only compiler command lines
on #3076: a parallel make kept compiling after the object that failed,
so the 'No rule to make target' lines had scrolled out of the last 60.
The annotation now starts with every line that names a failure
(compiler, linker and make errors, failed tests and assertions),
deduplicated and capped, then the last 40 lines, each cut to 300
characters. Checked against that run's log, it names the missing rules.
…4aab76a68859be7

Prepare 1.1.0: ship aims-db memory baked into the image
…ockers

Fix release promotion blockers in account activation, Vault startup, and upgrades
@JBailes
JBailes deployed to main-merge-approval November 16, 2026 15:06 — with GitHub Actions Active
@JBailes
JBailes deployed to main-merge-approval November 16, 2026 19:45 — with GitHub Actions Active
@JBailes
JBailes merged commit 825fe79 into main Nov 16, 2026
151 checks passed

This branch was successfully deployed

1 active deployment
main-merge-approval — 56d9b0fa Deployed Nov 16, 2026 by JBailes via approval #78
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant