Commit Graph

134 Commits

Author SHA1 Message Date
Agent Workbench Dev 5acb657157 fix: retain installation receipts without a size ceiling 2026-11-06 08:35:07 +08:00
Agent Workbench Dev 0c4895b7c8 fix: honor complete REA contract and synchronize UI verification 2026-11-06 07:26:36 +08:00
Agent Workbench Dev b8a5f95018 fix: advance offline demo from original conversation 2026-11-06 06:55:29 +08:00
Agent Workbench Dev c9c77c7ee6 fix: complete unrestricted waits catalog and network flows 2026-11-06 06:21:59 +08:00
Agent Workbench Dev c93c434808 fix(runtime): apply open policy when continuing legacy tasks 2026-11-06 05:08:02 +08:00
Agent Workbench Dev 2fa220735d feat(runtime): default to passthrough and persist execution permissions 2026-11-06 04:58:57 +08:00
Agent Workbench Dev 67c034f4b6 fix(context): pair route capacities and preserve source evidence within bounded requests 2026-11-06 02:37:52 +08:00
Agent Workbench Dev d5de90930e chore(release): freeze 1.25.5 capacity and cache delivery checks 2026-11-06 01:27:39 +08:00
Agent Workbench Dev 72115647d9 fix(runtime): preserve model capacity history and cache evidence with resumable file proposals 2026-11-06 01:18:37 +08:00
Agent Workbench Dev 2b17821138 fix(models): accept streamed thinking signatures and publish valid plan tools 2026-11-06 00:34:04 +08:00
Agent Workbench Dev 74b0030888 chore(release): bump to 1.25.4 for output-truncation recovery
Pin package.json, package-lock.json and the literal APP_VERSION together,
add 1.25.4 to the controlled-release version allowlist, and record the
output-truncation recovery in the changelog.
2026-11-05 08:55:48 +08:00
Agent Workbench Dev 0fefa65eeb feat(runtime): recover a bounded number of output-ceiling truncations
When a turn stops at the output-token ceiling (output_truncated), the run no
longer dies. The truncated turn is still dropped whole - no tool is executed
and the invalid assistant message is never appended - but the model is given a
standalone instruction to split its proposals and the run continues. Recoveries
are capped per run (2); past that the original diagnostic still fails the run.
The connection is not cooled down for a truncation, since it responded fine and
only the model overran its budget, so the recovery turn can reach it. Both the
failed attempt and the recovering turn are accounted; neither is unknown.
Fixture regression covers recover-then-continue and the exhausted cap.
2026-11-05 05:56:10 +08:00
Agent Workbench Dev dfb6029f33 fix(models): give an actionable diagnostic when output hits the token ceiling
A response stopped at finish_reason=length (or Messages max_tokens /
Responses max_output_tokens) is still rejected whole - a complete leading
tool call is never applied because a later call may be cut mid-argument.
Previously the model only saw a generic incompleteness message and would
repeat the same oversized multi-tool turn. Now these paths raise a shared
output_truncated diagnostic telling the model to split proposals into one
file per response. Regression covers streaming/non-streaming Chat,
Responses and Messages, including the AB sample shape.
2026-11-05 04:47:42 +08:00
Agent Workbench Dev d038191832 feat(models): discover upstream catalogs and select shared-source model pools 2026-11-05 00:44:35 +08:00
Agent Workbench Dev 109e51247a chore(release): freeze 1.25.3 reasoning acceptance and retained upgrade 2026-11-05 00:16:46 +08:00
Agent Workbench Dev f1242ee800 feat(models): pair reasoning controls with provider profiles and tool history 2026-11-05 00:13:53 +08:00
Agent Workbench Dev 3ab3a6e46a fix(models): reject empty Messages responses with retained evidence 2026-11-04 23:53:49 +08:00
Agent Workbench Dev 4a49f02daf fix(context): size sealed browser-result reads without mutation reserve 2026-11-04 20:29:02 +08:00
Agent Workbench Dev e3f2d8e79a chore(release): prepare 1.25.2 validation and fresh bounded sample X 2026-11-04 19:57:24 +08:00
Agent Workbench Dev 3647296a8d perf(reading): share bounded file and log budgets across independent reads 2026-11-04 19:53:13 +08:00
Agent Workbench Dev d2703ad340 fix(reading): preserve mixed-turn progress without disabling no-op guards 2026-11-04 19:43:48 +08:00
Agent Workbench Dev b4d7d2c965 fix(tasks): honor explicit prohibitions on running engineering checks 2026-11-04 19:43:47 +08:00
Agent Workbench Dev 1037031e82 fix(install): recheck file ownership after the manifest scan 2026-11-04 19:40:33 +08:00
Agent Workbench Dev dcea76b732 fix(budget): verify reservation history in ledger order and preserve output bounds 2026-11-04 19:34:36 +08:00
Agent Workbench Dev 0a82636d16 fix(release): advance served APP_VERSION to 1.25.1 and pin it to package.json
The 1.25.0 literal in diagnostics.mjs lagged the package bump, so a hot upgrade served 1.25.0 and the installer correctly rolled back on HOST_VERSION_MISMATCH. Advance the constant, add a release-efficiency guard that fails when the constant or compiled bundles drift from package.json, and add the 1.25.1 formal-release driver that refuses a candidate whose compiled server version is stale.

No product behaviour change beyond the version string; the reservation and install-scan fixes are unchanged.
2026-11-04 18:38:54 +08:00
Agent Workbench Dev 19f919ce52 perf(install): verify each manifest file with one native realpath instead of re-walking ancestors
The root chain is checked before and after each scan; each file uses one lstat+realpath equality, nlink and stream-name checks, and size is rechecked after read. Directory inventory keeps full checks for directories. Same hashes and manifest scope; junctioned subdirectory, hard link, stream path and escape stay rejected. Warm per-file cost on a real 1.16 tree 2.08ms -> 0.17ms with identical hashes.
2026-11-04 16:59:47 +08:00
Agent Workbench Dev 5801f9cabb fix(budget): set aside a single isolated input-ratio outlier in dispatch reservation
observed-route-max-v2 excludes at most one receipt whose reported/local input ratio is >=1.5x every other receipt on the route, with >=3 receipts and a later normal one. Excluded receipts stay in the ledger and are listed in excludedSources with their position; repeated highs, the latest receipt and short histories are never excluded. Sealed v1 reservations still verify byte-for-byte. Regression replays the original R evidence: v1 29623 > remaining 17429, v2 15285.
2026-11-04 16:59:47 +08:00
Agent Workbench Dev cb54b75abe chore(release): prepare immutable task agreement 1.25 delivery 2026-11-04 14:00:26 +08:00
Agent Workbench Dev 929d12ba66 feat(tasks): bind diagnostics to original failures and explain delivery gaps 2026-11-04 13:53:54 +08:00
Agent Workbench Dev d389b23c99 feat(ui): review original-text drafts and explicit task scope 2026-11-04 13:38:08 +08:00
Agent Workbench Dev b963e73301 feat(tasks): preserve reviewable original-text drafts and scope agreements 2026-11-04 13:28:00 +08:00
Agent Workbench Dev 1561560cbb build: prepare independent adoption A/B release and retained-host gates 2026-11-04 12:18:12 +08:00
Agent Workbench Dev 6fe04042c5 feat(runtime): observe turn evidence without changing execution controls 2026-11-04 11:26:47 +08:00
Agent Workbench Dev 785dd0c54d feat(tasks): review sourced requirements against current original evidence 2026-11-04 10:36:50 +08:00
Agent Workbench Dev 7e77fd65c3 fix(context): retain observed file fragments and current engineering facts 2026-11-04 07:52:59 +08:00
Agent Workbench Dev 5dcd6d7c3f fix: retain source continuation and explain request reservation pauses 2026-11-04 00:45:53 +08:00
Agent Workbench Dev 54dbe3d6fc feat: add bounded connection network diagnostics 2026-11-03 23:11:33 +08:00
Agent Workbench Dev 13eb37efa7 fix: deduplicate mandatory task context and clarify check-plan evidence 2026-11-03 21:39:34 +08:00
Agent Workbench Dev 65ddee16b0 feat(models): add native Anthropic Messages and observed endpoint diagnostics 2026-11-03 18:58:59 +08:00
Agent Workbench Dev 6bd605ad15 fix(models): unify model selection entries and preserve routes on budget saves 2026-11-03 14:40:45 +08:00
Agent Workbench Dev 1153e19a4c release: prepare 1.24.2 identity and delivery integrity candidate 2026-11-03 12:20:50 +08:00
Agent Workbench Dev 983af1416e fix(engines): preserve verified review pauses and always close native sessions 2026-11-03 12:20:20 +08:00
Agent Workbench Dev ecdc939c29 fix(context): preserve progress counters and reserve fresh read evidence 2026-11-03 12:03:42 +08:00
Agent Workbench Dev 10dbd8f500 fix(runtime): provide current delivery evidence after reviewed application 2026-11-03 11:57:42 +08:00
Agent Workbench Dev 3659f5e858 fix(models): retain upstream identity and pause strict mismatches 2026-11-03 11:53:07 +08:00
Agent Workbench Dev c1521e6cd2 fix: hand off complete changes and retain verified source context 2026-11-03 07:58:13 +08:00
Agent Workbench Dev 9106655d46 feat: integrate pinned original Purge with host execution gates 2026-11-02 22:32:31 +08:00
Agent Workbench Dev bb45dabac4 fix: bind observed models and calibrate every model dispatch budget 2026-11-02 20:50:42 +08:00
Agent Workbench Dev e73bca94a0 fix: isolate model transport and stop replaying unknown requests 2026-11-02 20:11:21 +08:00
Agent Workbench Dev 0a58c0d09f fix(engine): bound native history and preserve verified continuation context 2026-11-02 19:42:29 +08:00