Repository navigation
chore(release): cut 1.1.5 - #2262
Conversation
Move every site that tracks the packaged version to 1.1.5 and write the 1.1.5 section of docs/release-notes.md from the PRs merged between the 1.1.4 branch point (60b5770) and the 1.1.5 branch point (9da8788), plus #2185. The Unreleased entry moves into that section, and #2186, merged to main after the cut, is listed as a known issue. Co-authored-by: Cursor <cursoragent@cursor.com>
Code metrics gate: PASSEDJudged: the 9 files this change touches. Files it does not touch never fail it; anything found there is listed under Outside this change for information.
116 findings in files this change does not touchOutside this change (informational): 116
...and 91 more; the job log lists every row.
Thresholds and sources: |
xiaofei-zheng
left a comment
xiaofei-zheng
left a comment
There was a problem hiding this comment.
PR #2262 -- chore(release): cut 1.1.5
What it does: Cuts release 1.1.5 on release/v1.1.5. It moves the eight sites that track the packaged version (pyproject.toml, docs/conf.py, the README badge, the compatibility matrix row, the three pip install hyperloom-inference-optimizer== pins, and install_baremetal.sh's HYPERLOOM_WHEEL_TAG) from 1.1.4 to 1.1.5, and turns the empty Unreleased entry into a full 1.1.5 section in docs/release-notes.md written from the ~115 PRs merged between the 1.1.4 and 1.1.5 branch points, plus #2185.
Blocking issues: 1
- [X7]
docs/compatibility.rst's matrix row now claims 1.1.5 is validated with a Magpie release this PR's own notes declare unsupported [verified]
Problem: the version cell moves 1.1.4 -> 1.1.5 atdocs/compatibility.rst:79, while the row's Magpie cell atdocs/compatibility.rst:87still readsMagpie 0.2.0(linkingdocs-0.2.0atdocs/compatibility.rst:94). The table's lead-in calls it the list of "validated Hyperloom version and component combinations". At v1.1.4 that cell was exact:install.shpinnede6833b81, identical to Magpie'sv0.2.0tag. For 1.1.5 the pin is6a1ee26f(install.sh:183), whose ownpyproject.tomlreadsversion = "0.3.0"and which contains Magpie'sv0.3.0tag (comparev0.3.0...6a1ee26f: ahead 9), andinstall.sh:178plusdocs/release-notes.md(Installation prerequisites move) state the minimum supported revision rose frome6833b81(v0.2.0) to1eb3df69-- itself 4 commits pastv0.3.0. So the matrix and the release notes shipped in the same PR contradict each other about Magpie.
Impact: an operator reading the matrix for 1.1.5 takes Magpie 0.2.0 as the validated component and follows thedocs-0.2.0documentation; withMAGPIE_PATHorMAGPIE_REFpointing at such a tree, the notes say a DeepSeek-V4 round fails withclient_tokenizer_unpatchable.
Action: update the Magpie cell and its link target to the release the 1.1.5 pin carries (0.3.0) in this PR, as earlier cuts' component moves did (6bcc43161moved the cell 0.1.0 -> 0.2.0).
Checked: the eight version sites (tree-wide grep: no 1.1.4 pin remains outside docs/release-notes.md; docs/reference/upgrade.md prose and sphinx-book-theme==1.1.4 are unrelated); the whole 1.1.5 notes section against the release tree -- coordinator_db.SCHEMA_VERSION = 7, DEFAULT_CLAUDE_MODEL = claude-opus-5-5, DEFAULT_CODEX_MODEL = gpt-6-astra, MAGPIE_REF/minimum, TRACELENS_REF 614c597b, TraceLens_generate_perf_report_from_config, validate_server_args, agentx_baseline_prefix_cache_off, speculative_context_copy_on_synthetic, magpie_nonzero_invalid_measurement, client_tokenizer_unpatchable, e2e_norm_intvty_p50, hyperloom-kb, schema_ref, and the absence of every removed flag and variable the notes list (--no-eval, --pr-feed-window-days, --local-kb-root, --return-after-read-kb, --only-aiperf, multi-node kernel-bench, HYPERLOOM_LOCAL_KB_ROOT, TRACELENS_SOURCE_ROOT, the two enablement allowlists, INSTALL_AIPERF, extend_explore_budget); the RUN_EVAL claim against _workload_envs.py:751 and :1340; all 121 first-parent PR subjects in 60b5770..9da8788 against the notes' coverage and the body's explicit left-out list; the Known-issues entry (#2186 merged to main 2026-10-26, not contained in release/v1.1.5). TraceLens's 1.0.0 cell was already approximate at v1.1.4 (pin 57 commits off the tag) and is pre-existing, not reported.
Ran: git grep over the head tree for the version and identifier sweeps above; git show v1.1.4: for the base-state comparison; gh api compares on AMD-AGI/Magpie (v0.2.0...e6833b81 identical, v0.3.0...6a1ee26f ahead 9, v0.3.0...1eb3df69 ahead 4) and AMD-AGI/TraceLens. No local test run: no production code changed and CI is green at head.
Skill: 2d0399bbe97b | Model: claude-opus-5 | Base: f62f5c4 | Head: b5ff580
SKIPPED: V4 -- "# PARTIAL: #1867 changes 199 files, only its first 100 were compared", "# PARTIAL: #1869 changes 908 files, only its first 100 were compared", "# PARTIAL: #2253 changes 317 files, only its first 100 were compared": overlap there unchecked.
|
Two statements in this cut do not match the release tree (
Checked: |
…he forge-loop refusal The component matrix names the Magpie release the 1.1.5 pin carries (0.3.0), the support matrix lists the Radeon AI PRO R9700 that 1.1.5 adds, and the upgrade note quotes the message forge-loop --resume prints for a 1.1.4 checkpoint. Co-authored-by: Cursor <cursoragent@cursor.com>
|
Thanks, all three are fixed in 0ae2e09.
|
Fixed at 0ae2e09: the compatibility matrix now names Magpie 0.3.0, matching the pinned revision.
PR #2262 -- chore(release): cut 1.1.5 (re-review at 0ae2e09)What it does: Cuts release 1.1.5 on Blocking issues: none Prior: X7 (compatibility matrix named Magpie 0.2.0 for 1.1.5) -- fixed. Checked (second commit): the three added changes. Checked (first commit, unchanged and re-confirmed): the eight version sites (tree-wide grep: no LGTM |
xiaofei-zheng
left a comment
xiaofei-zheng
left a comment
There was a problem hiding this comment.
Re-reviewed at 0ae2e09. The compatibility matrix now names Magpie 0.3.0, matching the pinned revision, and the two other changes in that commit (the R9700 row and the quoted forge-loop refusal) check out against the tree. No blocking issues; details in the review comment above.
* fix(orchestrator): refuse a contradicted skip_to_close and show the initial-roofline hold (#2175) (#2185) * fix(orchestrator): refuse a skip_to_close the session's own facts contradict A skip_to_close ends the run on the agent's word and is booked as global_converged whenever budget is left. Session d11ccead closed at cycle 0 with a measured baseline and 9.3 h of budget left, and release qualification read that close as PASSED. The intent router now denies the hint (rule skip_to_close_contradicted) when a baseline is measured and the remaining budget still funds another cycle, so a live skip_to_close can only earn baseline_failed or time_exhausted. GPU work that keeps failing after the baseline still ends the run, through the sweep. compute_next_phase is unchanged: recorded decisions replay as they were made, and a hint already pending still closes the run. Part of #2167. * fix(orchestrator): tighten the skip_to_close refusal Drop the unread now_unix parameter, build the denial before recording it, and keep the PHASE CONTRACT rewording to the lines it replaces. * fix(orchestrator): show the agent the initial-roofline hold in the Phase block While PRELUDE or ENABLEMENT waits for the initial roofline it enqueued, the `=== Phase ===` block now carries a `hold` line: the roofline it waits for, how long the wait can still last, that no hint advances the phase, that `remaining_sec` is the phase's share rather than the session's, and that a `send_message` observation is the expected answer each tick. The PHASE CONTRACT no longer says `baseline_tput > 0` alone exits PRELUDE, and orchestration.md stops reading that share as urgency under a hold. One helper reads the hold's two inputs for both phase predicates and for the block, so the line the agent is shown and the exit decision cannot disagree. Part of #2167. --------- (cherry picked from commit e7b8bc7) Co-authored-by: Cursor <cursoragent@cursor.com> * chore(release): cut 1.1.5 (#2262) * chore(release): cut 1.1.5 Move every site that tracks the packaged version to 1.1.5 and write the 1.1.5 section of docs/release-notes.md from the PRs merged between the 1.1.4 branch point (60b5770) and the 1.1.5 branch point (9da8788), plus #2185. The Unreleased entry moves into that section, and #2186, merged to main after the cut, is listed as a known issue. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(release): correct the Magpie version, list the R9700 and quote the forge-loop refusal The component matrix names the Magpie release the 1.1.5 pin carries (0.3.0), the support matrix lists the Radeon AI PRO R9700 that 1.1.5 adds, and the upgrade note quotes the message forge-loop --resume prints for a 1.1.4 checkpoint. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: fengshaoyi-amd <shafeng@amd.com> Co-authored-by: Cursor <cursoragent@cursor.com>
Cuts release 1.1.5 on
release/v1.1.5(current basef62f5c4). It moves every site that tracks the packaged version to 1.1.5 and turns the release-notesUnreleasedentry into a complete 1.1.5 section.Version bump. Every site that tracks the packaged version moves together:
pyproject.toml[project].versiondocs/conf.pyversion_numberpyproject.tomlREADME.mdversion badgedocs/compatibility.rstcomponent matrix rowdocs/install/install.md,examples/README.md,hyperloom-setup/SKILL.mdpip install hyperloom-inference-optimizer==pinsinstall_baremetal.shHYPERLOOM_WHEEL_TAGSupport matrix.
docs/compatibility.rstlists the Radeon AI PRO R9700 (gfx1201, vLLM only), which 1.1.5 adds.Release notes.
docs/release-notes.mdis the only document to cut. A## Hyperloom 1.1.5 releaseheading is added above 1.1.4, an empty## Unreleasedstays for the next cycle, and the page header names 1.1.5 as the packaged version.Unreleasedcarried a single entry (the accuracy eval can no longer be turned off), so the section is written from the titles and descriptions of the PRs merged between the 1.1.4 branch point (60b5770) and the 1.1.5 branch point (9da8788), plus #2185:state.json, manifest, report and log-line changes that session and report consumers have to follow.--apply-gpu-power-settingsand the power-throttle share, the operable Experience KB withhyperloom-kb, the PrimaTune predictor, Radeon AI PRO R9700 support, the Claude–Codex forge-loop relay, the Claude subscription credential, the breakdown variants funnel and the Experience KB identity.Unreleased, reworded (the baseline already did not retry withRUN_EVAL=falsein 1.1.4), the resume cut-off and the deprecated options,kbmine, native Magpie AgentX, the phase-budget and phase-transition changes, the operator-defined specialist, explore dedup, and the broad-exception-handler series.mainafter the cut. On a recipe that pins golden acceptance, a removal KEEP rewrites the launch arguments, so later benches fail asmagpie_nonzero_invalid_measurement. The entry says the fix ships in the next release.main'sUnreleasedentries merged after the cut (the SGLang 0.5.22 bump,optimize --session-dir,hyperloom hosted run, GEAK processes after KERNEL_AGENT) are not in 1.1.5 and are not described here.None.
No production code change.
pytest src/hyperloom/inference_optimizer/tests/test_baremetal_doc_version_consistency.py src/kernelforge/tests/test_rename_completeness.py: 6 passed locally on Windows.test_baremetal_defaults_match_compat_doccannot import on Windows (fcntl), so CI on Linux runs it. The rename-completeness allowlist needs no change, because the 1.1.5 notes use no retired spelling.python scripts/check_cli_references.py: clean.No
1.1.4package pin remains outsidedocs/release-notes.md.docs/reference/upgrade.mdstill names 1.1.4 when it describes sessions that release started, andsphinx-book-theme==1.1.4is an unrelated dependency.Size/complexity triggers crossed: split, or why the shape is right
None crossed in code. The diff is long only in
docs/release-notes.md, because 1.1.5 carries about 115 PRs. The version bumps are one line each, the same shape as #1905.N/A.
Installs pin
1.1.5, the bare-metal installer downloads its hotfix asset from thev1.1.5release, and the release notes list everything an operator upgrading from 1.1.4 has to change.No for this PR. The breaking changes in 1.1.5 come from the PRs listed above and are documented under "Before upgrading from 1.1.4".
Yes, the 1.1.5 release cut.
N/A.
After merging
v1.1.5onrelease/v1.1.5and attachrocm-profiler-hotfix-libs.tar.gz(the same file as v1.1.4) to the GitHub release, becauseHYPERLOOM_WHEEL_TAGnow points the bare-metal installer at that tag. Attach the wheel as v1.1.4 did.mainwith a merge commit.main'sUnreleasedkeeps its post-cut entries and drops its accuracy-eval entry, which the 1.1.5 section now carries.Test plan
scripts/check_cli_references.pyis clean.f62f5c4, with both top-model AgentX recipes: DeepSeek-V4-Pro (SGLang) session20261026T123138Z-08d91b18, +11.86% (1077.90 → 1101.98); DeepSeek-V4.1-Flash session20261027T023953Z-ea463689, +8.28% (1263.77 → 1290.58). Both stopped withsweep_done.Release-note entries were drafted with AI assistance from the PR titles and descriptions, and their identifiers were spot-checked against the release tree.