v0.9.0 Open PR harvest: merge, supersede, or close long-lived branches

[Hmbown]

Problem

The repo has many open PRs that may already contain useful v0.9.0 pieces or stale/conflicting work. Agents should not reimplement features without checking these branches first.

Scope

Create a PR-harvest lane that reviews open PRs and assigns each one a disposition:

• merge after tests
• cherry-pick useful parts
• superseded by main
• blocked by conflict/design issue
• close as stale
• defer past v0.9.0

High-priority PRs to inspect first:

• feat: add WhaleFlow — declarative multi-agent workflow orchestration #2482 WhaleFlow declarative multi-agent orchestration
• Feat/whaleflow cost tracking #2486 WhaleFlow cost tracking
• feat(config): collapse ProviderKind/ApiProvider dual enums behind Provider trait #2479 collapse ProviderKind/ApiProvider dual enums behind Provider trait
• feat(engine): project mode prompts per request #2687 mode-agnostic system prompt with append-only mode/approval messages
• fix: Windows sub-agent completion halves TUI render width #2708 Windows sub-agent completion width fix
• fix(tui): add scroll support to plan prompt modal #2623 plan prompt modal scroll support
• perf(project-context): cache load_project_context_with_parents by mtime signature #2636 project-context cache
• perf(history): cache output_rows and selected_output_indices per cell #2635 output rows cache
• perf(prefix-cache): cache tool-catalog JSON serialization across checks #2632 tool-catalog JSON cache
• perf(engine): memoize estimated_input_tokens via content-keyed cache #2631 token estimate memoization
• feat(client): add hard compaction option preserving system segment #2522 hard compaction preserving system segment
• feat(client): add cross-session prompt base section disk cache #2520 cross-session prompt base disk cache
• refactor(client): relocate turn_meta to tail of user message for prefix stability #2517 relocate turn_meta to tail for prefix stability
• fix(tui): count unreconciled subagents against the cap #2505 subagent cap accounting
• feat: enforce allowed tools for custom slash commands #2326/docs: clarify shell tool mode availability #2526 shell tool availability docs if still relevant

Acceptance criteria

• A table lists every open PR reviewed with owner, branch, v0.9.0 relevance, test status, and disposition.
• Any PR merged into v0.9.0 has fresh local verification against current main.
• Any stale PR closed has a short explanation and links to replacement issues/PRs.
• No agent starts WhaleFlow/provider/prompt-prefix work without checking the relevant PRs above.
• The harvest outcome is linked from the v0.9.0 execution map.

[Hmbown]

Status update: #2687 by @LeoAlex0 has been harvested via #2801 and is included in green #2762 at fbe8d9ee5d0e83f1e95196c66659a43b91d2ac75. The landed slice projects mode, approval, and tool-taxonomy prompt metadata per request while keeping stable system prompts and stored history byte-stable.

Leaving this tracker open for the remaining open-PR disposition table and any follow-up harvest decisions.

[Hmbown]

2026-06-06 v0.9 stewardship harvest map update from the current open-PR sweep:

PR	Contributor	Current disposition	v0.9 action
#2808	@gaord	Partially harvested; mutating runtime API endpoints remain unsafe	Keep open. Read-only snapshot/listing and VS Code resilience slices landed via #2822/#2848; save/undo/retry mutation semantics still need a separate design/verification pass.
#2786	@h3c-hexin	Safe slice harvested	Keep open. Static prompt-composer intent was harvested in #2819; do not merge the remaining override surface into v0.9 without embedder semantics nailed down.
#2781	@punkcanyang	Defer	Keep open. Ghost-text suggestions need opt-in/provider/privacy/cost semantics before v0.9.
#2773	@idling11	Safe foundation harvested	Keep open. Config/RFC/foundation landed; automatic provider fallback remains too risky for v0.9 defaults.
#2753	@ljm3790865	Broad/blocked	Keep open for later design; no narrow v0.9 harvest selected yet.
#2579	@encyc	Broad/dirty	Keep open. AppendLog replacement is too invasive for the current v0.9 cut unless a very small read-only/test slice is identified.
#2522	@HUQIANTAO	Safe planner harvested, runtime path deferred	Keep open. #2838 landed the default-disabled planner and tests; do not merge #2522's separate hard-compaction executor/system-summary path before v0.9.
#2486	@AdityaVG13	WhaleFlow cost/memo intent harvested in pieces	Keep open as draft. Mock telemetry/cost/budget/replay slices have landed; live billing/provider behavior remains deferred.
#2482	@AdityaVG13	WhaleFlow IR/Starlark intent harvested in pieces	Keep open as draft. Typed IR, examples, mock executor, replay, teacher/student gates, and cutlines have landed; live workflow_run remains deferred.
#2479	@sximelon	Safe provider-registry slice harvested	Keep open. #2820 covers the v0.9-safe registry/drift layer; full trait collapse remains larger than this release cut.
#2239	@gordonlu	Broad/dirty	Keep open; i18n is valuable but too wide for current v0.9 stabilization without a scoped sub-issue.
#1893	@wavezhang	Safe TLS config harvested	Keep open. #2834 landed provider-scoped insecure_skip_tls_verify; remaining PR state is stale/conflicting relative to current config surfaces.
#1865	@dumbjack	Broad/dirty/defer	Keep open for later mode architecture; not a v0.9 release blocker.

Additional notes:

• New issue Abnormal stop working while coding or analysis #2847 by @qamranmushtaq-collab received a narrow triage/fix PR in fix(tui): classify stream decode failures as network errors #2849 (35cd09a9f) for stream/body decode error classification, but the issue remains open pending reporter repro details.
• feat(client): add hard compaction option preserving system segment #2522 was re-audited read-only after feat(compaction): add dormant hard compaction planner #2838; verdict is still "planner harvested, runtime hard compaction deferred" because executing hard compaction can split tool-call/result pairs or mutate system/developer prompt semantics.
• No tag, publish, release, or main-merge action was taken.

[Hmbown]

Harvest update for #2808: audited as partially harvested/superseded; do not merge directly for v0.9.

Already harvested/satisfied:

• POST /v1/sessions landed earlier from @gaord's GUI runtime direction in a safer completed-thread-only shape with active-work guards.
• GET /v1/snapshots was harvested into feat(runtime-api): add read-only snapshot listing #2822 and consumed by the VS Code Agent View restore-points slice in feat(vscode): show runtime restore points #2828.

Deferred / not harvested:

• restore/retry/undo/patch-undo mutation endpoints from feat(runtime-api): add session save, undo/retry, and snapshot endpoints for GUI #2808 remain unmerged pending atomicity, workspace/thread scoping, provider-valid history reconstruction, auth/read-only boundary, and error-path tests.
• feat(runtime-api): add session save, undo/retry, and snapshot endpoints for GUI #2808 itself is stale, blocked by checks, and broad enough that any future mutation work should be redesigned as a fresh narrow maintainer PR rather than ported wholesale.

Credit: keep @gaord credited for the GUI runtime endpoint direction and session/save API trail; keep @AiurArtanis and @ygzhang-cn credited on the VS Code/GUI demand trail. Bots/review tools supplied useful risk evidence but are not contributor credits.

[Hmbown]

Harvest update for #2851 / #2791: audited as too broad for a safe v0.9 harvest as-is.

Why not harvest directly:

• Refactor TUI command groups into focused implementations #2851 rewrites central TUI command dispatch across ~247 files and is 100 commits ahead / 181 behind the current stewardship branch.
• The branch does not include the current v0.9 /hf command group, so a broad merge/harvest risks regressing Hugging Face command work that is already in the cutline.
• It removes older every-registered-command / every-alias dispatch smoke coverage and replaces it with narrower registry/help/unknown-command checks.
• It also moves config/model/share, shell/sandbox, engine, sidebar, and runtime behavior beyond the scope of Refactor command dispatch from monolithic match to modular strategy pattern #2791.

Safe follow-up: keep #2791 open and use Paulo's proposal as inspiration for a test-first maintainer refactor: strengthen command parity coverage on the current branch, then move one low-risk command group at a time with before/after dispatch, alias, help, palette, and migration-hint tests.

Credit: @aboimpinto / Paulo Aboim Pinto should be credited as #2791 author and #2851 PR author/source of the command-refactor direction. No code from #2851 is harvested in v0.9 at this point.

[aboimpinto]

I haven't done... but we should PR this to v0.9.0?

[Hmbown]

Open-PR harvest status after the #2862 pass:

• feat(runtime-api): add session save, undo/retry, and snapshot endpoints for GUI #2808 (GUI runtime API): keep open. Safe read-only slices are now harvested as feat(runtime-api): add read-only snapshot listing #2822 (GET /v1/snapshots) and feat(runtime-api): expose git status metadata for Agent View #2862 (head / dirty metadata in thread summary and workspace status). Mutation endpoints remain deferred pending atomicity and provider-message reconstruction tests. Thanks @gaord.
• feat: add WhaleFlow — declarative multi-agent workflow orchestration #2482 / Feat/whaleflow cost tracking #2486 (WhaleFlow): keep open as source/design branches. Current v0.9 branch has harvested typed IR, mock executor, replay, model policy, telemetry, teacher/student artifacts, and dogfood trace evidence; live workflow_run, provider calls, worktree application, TraceStore writes, and pod monitor stay deferred. Thanks @AdityaVG13.
• feat(prompts): add static prompt composer override for embedders #2786, feat(client): add hard compaction option preserving system segment #2522, feat(config): collapse ProviderKind/ApiProvider dual enums behind Provider trait #2479, feat(provider): complete provider fallback chain (#2574) #2773: safest v0.9 slices already harvested; leave originals open for their broader architecture/runtime work. Thanks @h3c-hexin, @HUQIANTAO, @sximelon, @idling11, and Feature Request: Provider fallback chain — auto-switch on API failure #2574 reporter @hsdbeebou.
• feat(tui): ghost-text follow-up prompt suggestion #2781 (ghost-text prompt suggestions): leave open for direct community iteration. It is thoughtful and CI-green, but the actual feature still needs an explicit opt-in/provider/privacy/cost contract before v0.9 should ship automatic post-turn network calls. A dead default-off config knob without behavior does not feel valuable enough to harvest tonight. Thanks @punkcanyang.
• feat(tui): multi-tab system with cross-tab collaboration #2753 and other broad architecture PRs remain too large for direct v0.9 merge without splitting and focused verification.

Current disposition: no additional direct merges from the open PR set; continue with narrow, tested maintainer slices only when they ship real behavior or release evidence without expanding unsafe runtime mutation surfaces.