Merge branch 'main' into chat-hide-notice-timeout

This commit is contained in:
2026-10-18 23:29:59 +01:00
69 changed files with 2662 additions and 1098 deletions
@@ -0,0 +1,25 @@
# Android target pool allocation — 2026-10-18, phase one not activated
This item adds a reusable store contract only. Pool declarations are not
activated in live jobs, runner handoff, server configuration, Android maps,
shared documentation, or deployment.
`Store.ConfigureResourcePool(ctx, protocol.ResourcePool)` replaces a bounded
pool declaration. Each member has a unique 1–128 byte resource key, an enabled
flag, quarantine state, and at most 16 bounded ASCII-keyed metadata values
(each value is at most 128 bytes). A pool has 1–32 explicit members. Invalid,
empty, duplicate, unknown, or malformed declarations are rejected. Updating a
pool never transfers a live owner; queued requests use the new declaration on
the next eligible release.
`Store.RequestResourcePool(ctx, pool, visit)` atomically returns a lease for one
free eligible concrete member, or a FIFO wait channel. The lease's `Target`
contains only the concrete member and bounded non-secret metadata. Capability
tokens remain private to the store and are never protocol/model metadata.
Direct concrete-resource leases and pool leases share the owner table, so they
cannot overlap. Disabled or quarantined members are skipped; if none are
eligible, the request queues without inference. Release, cancellation, stale
credentials, recovery, and restart retain the existing no-expiry semantics.
This is scheduler coordination only: no device probing, emulator creation,
credentials, activation, or live job changes are included.
+40
View File
@@ -0,0 +1,40 @@
# Background task output API
Changed: 2026-10-18. The chat websocket now has `chat_shells` for current,
operator-session-owned running shells and `chat_shell_output` for a selected,
bounded (64 KiB) absolute-byte output page. Ownership requires both the
operator-session registry and provider shell metadata `sessionID`; a shared cwd
is not evidence. Output reads first verify provider detail and explicitly mark
removed/provider-unavailable shells unavailable. This is not completed-task
history or restart retention.
Stale shared documentation: the packaged OpenCode operational references under
`internal/author/skill/references/` need a later docs-fold update describing the
two additive websocket messages, 64 KiB maximum page, and current-only history.
Fixture evidence (2026-10-18): `go build -o <scratch>/pipelines-large-chat-fixture
./cmd/pipelines-large-chat-fixture`; then `<scratch>/pipelines-large-chat-fixture
-scratch-dir <scratch>/run -shell-exited`, stopped with SIGTERM and awaited. The
foreground run exercised authenticated `chat_sessions`, three reconnecting history
pages, `chat_shells`, bounded `chat_shell_output`, and four executable negative
readbacks (`foreign-list`, `unknown-list`, `foreign-output`, `unknown-output`) against
scratch DB/workspace only. Its JSON contains `clean_shutdown: true` and
`shutdown_exit_status: 0`, plus the selected exited shell's `exit: 0` readback.
The evidence revision was `ded79277783b6ccdc25eb4ea25fe55d8632b89c7`; session
`ses_large_chat_fixture`, shell `sh_large_chat_fixture`, and fixed marker
`fixture shell output marker\n`. The current list remains running-shell discovery.
Integration coverage added: `TestChatShellOutputRejectsBoundsBeforeRegistryOrProvider`,
`TestChatShellOutputRealHandlerWebsocketOwnedAndForeign`, and
`TestChatShellOutputRealHandlerExitedReadback`,
`TestChatShellOutputRealHandlerUnavailableAndReconnect`, and
`TestChatShellOutputRejectsMalformedAndOversizedProviderPages`, and
`TestChatShellOutputRealHandlerCancellationStopsProvider` in
`internal/server/chat_shell_output_test.go`. Run with
`go test ./internal/server -run 'TestChatShellOutput' -count=1 -v` and
`go test -race ./internal/server -run 'TestChatShellOutput(RealHandler|RejectsBounds)' -count=1 -v`.
These tests use an authenticated in-process HTTP provider, a scratch SQLite operator
registry, and real `/connect` websocket traffic. The fixture executes foreign
session/shell refusal requests and records positive/negative JSON readbacks under
`shell_readbacks`; it remains scratch-only evidence, not evidence of a real provider,
model, production workspace, or device.
@@ -0,0 +1,12 @@
# Binary snapshots no longer consume the source patch budget
Tracked screenshots exhausted the 32 MiB retained snapshot budget before source
files were visited. Captures could therefore show empty file diffs for real code edits.
Binary files now retain only hashes and metadata; text remains bounded to 32 MiB,
with existing per-file, file-count, time and patch limits unchanged.
The operations skill reference's capture-limit explanation is stale: distinguish
retained text bytes from transient bounded binary reads. Existing historical empty
captures are not backfilled and remain incomplete. Commit-only stages may have no
stage-local content change; run summaries should retain earlier patches rather
than treating a commit or absent capture as proof of no code changes.
+3
View File
@@ -0,0 +1,3 @@
# Remove run Evidence tab
The Android run-detail UI no longer exposes its redundant Evidence tab or automatically fetches session evidence when a run opens. No shared README, topic document, or packaged OpenCode skill reference currently documents that tab; this note records the UI behavior change for the manual `docs-fold` pass.
@@ -0,0 +1,11 @@
# Resource-pool runner handoff — 2026-10-18
Stage `resources` declarations now accept exactly one of `exclusive` or `pool`.
Pool stages wait on the named pool, then receive the selected concrete member and
bounded metadata through explicit runtime context. The model input remains
unchanged and capability tokens remain private to the store. Retained control
evidence records the concrete target; queued run state names the pool while an
active owner names the selected member.
This is parser/runner handoff only. It does not add Android definitions, pool
configuration rollout, device setup, probing, or deployment.
+11
View File
@@ -0,0 +1,11 @@
# Retry from failed stage
As of 2026-10-18, retry requests default to `resume`: the server records a
durable checkpoint for the terminal failed visit and starts the new attempt at
that visit with its completed-prefix handoff context. `--mode=restart` retains
the older full-restart behavior. Older clients that omit the mode therefore
resume rather than restart.
The CLI and protocol now expose the mode and retry provenance. The shared CLI
documentation and packaged skill references should be folded from this note by
the manual `docs-fold` pipeline.
+12
View File
@@ -0,0 +1,12 @@
# Retry memory input override ordering
On 2026-10-18, selected-stage retry input overrides are persisted separately
from the immutable source-stage input. Retry reconstruction now resolves the
selected stage's memory bindings from the original prefix first, then applies
the validated override once to that invocation. This keeps the persisted
`run_steps.input` and downstream handoff aligned with the requested correction
without changing the source history or unrelated later stages.
The shared retry documentation and packaged OpenCode skill reference should be
folded by the manual `docs-fold` pipeline; this note records the behavior change
without editing those shared documents.
@@ -0,0 +1,18 @@
# Retry selected stage and input overrides
As of 2026-10-18, `retry` can select a stage recorded in a terminal failed run
with `--stage` and, when that stage was visited repeatedly, `--visit`. The
default remains the last failed visit. `--version` may select an explicitly
published version of the same pipeline; the server rejects a changed or missing
completed prefix.
`--input-file FILE` (or `-` for stdin) accepts a bounded JSON object. Its keys
are shallow-merged into the reconstructed selected-stage input, replacing values
with the same key while preserving all other handoff fields. The override is
stored in the retry checkpoint and downstream stages receive the selected
stage's output normally. Runtime, ownership, pipeline, workspace and scope
fields are reserved. The selected stage may repeat its partial side effects.
The shared documentation and packaged OpenCode skill reference for retry options
are now stale and should be folded from this note by the manual `docs-fold`
pipeline.
+7
View File
@@ -0,0 +1,7 @@
# Runs default home
The Android app now starts at Runs. Top-level navigation and pipeline search pop up
to the Runs home route, while explicit run deep links continue to select their run.
The shared Android navigation documentation may still describe Pipelines as the
initial destination; the `docs-fold` documentation pass should update that reference.
@@ -0,0 +1,11 @@
# Shell output default wire fields
The Android shell-output query now always serializes its bounded `cursor` and
`limit` fields, including when callers use the request helper defaults. The
server requires both bounds explicitly; unrelated protocol DTO defaults remain
unchanged.
The packaged OpenCode skill references and any shared protocol documentation
that describe Android shell-output request serialization are stale. The
manual-only `docs-fold` pipeline should update those references; this item does
not edit shared docs or packaged skill files.