[Handsome-wzw]

Summary

Two more reasoning-effort rungs, xhigh and max, for a conversation whose
model is served by the Anthropic Messages adapter -- the follow-up #1113 left
out on purpose.

Base: cut from feat/session_reasoning_effort (#1113), which is not merged
yet, so this PR shows #1113's commits beside its own until #1113 lands. The
change of this PR is the one feat(agent) commit; the merges after it bring
in origin/main and #1113's later head (its merge with main and its fix that
reads the opened conversation's effort), so the diff against main is #1113
plus this.

Design, settled from the code:

• The words a session may pick come from the provider of the session's
  model, not from one global tuple. New LLMProvider.reasoning_efforts(model)
  (raven/providers/base.py) answers the four every wire takes
  (REASONING_EFFORTS); AnthropicMessagesProvider answers the keys of its
  _ADAPTIVE_EFFORTS table, which is six (xhigh, max over high), one
  answer per wire rather than per model (a model that refuses a rung says so in
  a 400 and rewrite_on_400 already turns the label into a budget there); the
  OpenAI-shaped adapters and litellm keep the four. LazyProvider,
  PerModelProvider, ResolvingProvider and EndpointRotorProvider forward it
  the way they forward reasoning_wire_keys, pinned by name in a test.
• session.set_effort answers options from the provider of the session's
  own binding (a draft gets the default model's), and refuses a word the
  provider does not take with a sentence naming the model and the words it
  takes. With no loop or a provider that cannot answer, the four stand.
• The console checker for agents.defaults.reasoningEffort stays at the four,
  on purpose: that key reaches every model call on every vendor and every
  sub-agent (RAVEN_PARENT_REASONING_EFFORT), so a rung only one wire takes
  would be refused everywhere else with no per-session way out. Written on
  REASONING_EFFORTS and at the checker. The settings page therefore keeps
  its four rungs.
• The record may carry any rung this raven knows: REASONING_EFFORT_RUNGS
  (config/schema.py) is the retry descent's REASONING_EFFORT_LADDER read
  upward, so a pick and a retry never disagree about which rung is higher, and
  session_effort.of reads xhigh/max back after a restart.
• The model can move after the pick (/model, the picker), so the turn asks
  its own provider again: session_effort.fit runs a pick the new wire does
  not take at the highest rung it does take (max on a Claude, then a switch
  to a four-word model, runs at high), logged, rather than at a request the
  wire would refuse or at the default the person moved away from. The pick
  only: a mode's reasoningEffort is the catalogue's word and goes out as
  named, as before. A provider that cannot answer leaves the pick as named,
  and nothing on that path fails a turn.
• lower_reasoning_effort already descended from max/xhigh; the tests now
  say so explicitly, on the label-only wire and on the Anthropic one.

Surfaces: the picker row names the two rungs (gui.effort.xhigh / .max,
en + zh) and, after a model switch, re-reads the row so the words are the new
model's; a pick the words no longer carry is drawn after them, ticked, so it
can be moved off rather than a row with no rung pressed. The TUI's /effort
help and usage list the six words and say where the top two apply. The offline
page library's conversation runs on a Claude, so its fixture offers six. The
using-raven guide (en + zh), CHANGELOG (Unreleased, its own bullet) and
CONTEXT.md's Session Reasoning Effort entry say the above; openrpc.json
describes effort and options as the model's words, and both generated
contracts were regenerated.

Type

• Fix
• Feature
• Docs
• CI / tooling
• Refactor
• Other

Verification

• pytest tests/test_session_effort.py tests/test_rpc_session.py tests/test_agent_loop_usage_sink.py tests/test_agent_loop_empty_recovery.py tests/test_provider_protocol.py tests/test_rpc_settings.py tests/test_rpc_console.py tests/test_import_cycle_budget.py tests/test_rpc_schema_match.py tests/test_rpc_registration.py tests/test_rpc_session_init_bundle.py -- pass

• full pytest tests -- see the first comment for the count and the baseline failures

• scripts/coverage_gate.py diff --base-ref origin/main -- see the first comment

• ruff check / ruff format --check / ty check on the touched modules, lint-imports -- clean

• ui-web: npm run gen (contract regenerated), vitest run (all gates), type-check, lint -- clean

• ui-tui: npm run gen:rpc, npm run gen:i18n, vitest run, type-check, lint -- clean

• mkdocs build --strict -- clean

• The picker row was checked on the built page's offline library (?stub=1) in light, dark and zh; the screenshots were shown to the owner before this PR was opened

• Relevant tests pass locally

• Relevant lint / type checks pass locally

• User-facing docs or screenshots are updated when needed

Risk

• Security impact considered -- the method still writes one word (or null) to a session's own metadata; the word is now checked against the provider's answer rather than a tuple, and a provider that cannot answer falls back to the four
• Backward compatibility considered -- the four words every wire takes are unchanged; a page or TUI that never saw the new rungs shows a word it does not know as the server spelt it; the global key accepts what it accepted; a record carrying xhigh/max is read as no pick by a build without this change
• Rollback path is clear for risky changes -- revert the PR; the record's key is unchanged, and the previous reader ignores the two words

Related Issues

Follows #1113.

[Handsome-wzw]

Verification on the pushed head (merged with origin/main and with #1113's current head 516c7e7d4):

• full pytest tests (merged tree, -n auto): 35982 passed, 3 failed, 255 skipped, 2 xfailed, 7m32s. The 3 failures are this Mac's known baseline, unrelated to the change: test_rpc_files::test_a_host_without_libreoffice_says_so, test_simulation_scenario::test_a_fresh_trial_refreshes_the_subagent_homes_before_it_runs, test_simulation_suite::test_a_stopped_run_is_marked_in_its_record_and_finished_with_its_evidence.
• scripts/coverage_gate.py diff --base-ref origin/main: 99.26% (134/135 executable changed lines). The one uncovered line is feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113's _set_effort dispatcher wrapper in raven/rpc/methods/session.py, which the diff against main still counts until feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113 merges.
• related set (test_session_effort, test_rpc_session, test_agent_loop_usage_sink, test_agent_loop_session_effort, test_agent_loop_empty_recovery, test_provider_protocol, test_rpc_settings, test_rpc_console, test_import_cycle_budget, test_rpc_schema_match, test_rpc_registration, test_rpc_session_init_bundle): pass. ruff check / ruff format --check / ty check on the touched modules, lint-imports: clean; the import-cycle budget is unchanged (providers/base now imports config/schema, which reaches nothing back).
• ui-web: npm run gen (contract regenerated), vitest run 258 files / 4419 tests, type-check clean, lint 0 errors (the 4 warnings main has), npm run build + build.py (boot snapshots match), check-page / check-css / check-class-namespace OK.
• ui-tui: npm run gen:rpc, npm run gen:i18n, vitest run 146 files / 2286 tests, type-check, lint: clean. (One timing case in modelPicker.test.tsx failed once under a loaded first full run, passed alone and did not recur on the merged tree.)
• mkdocs build --strict: clean.
• Range gates run locally as CI runs them (check_source_language, check_large_files, check_commit_messages on origin/main..HEAD, check_pr_title, check_pr_body): pass.
• The picker row was checked on the built page's offline library (?stub=1): en light / dark (seven rungs wrap as whole words, 5 + 2) and zh (seven rungs on one line); the screenshots were shown to the owner before this PR was opened.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed github/main...HEAD at 35489d530c28, including the repository rules and canonical Session Reasoning Effort term, the full diff, provider and loop callers, RPC/schema/generated-client boundaries, WebUI/TUI state flows, relevant history, backward compatibility with older providers/gateways, and the test changes. I specifically tried to refute the model-switch fitting behavior, provider-wrapper routing, draft staging, stale async reads, persisted session precedence, and the claim that tests were not weakened; I found no concrete failure.

Verification:

• uv run pytest tests/test_session_effort.py tests/test_provider_protocol.py tests/test_rpc_session.py tests/test_agent_loop_empty_recovery.py tests/test_agent_loop_usage_sink.py tests/test_rpc_session_init_bundle.py tests/test_rpc_settings.py -q: 595 passed.
• uv run pytest tests/test_rpc_schema_match.py tests/test_l4_entrances.py tests/test_source_language_check.py -q: 559 passed.
• WebUI focused Vitest run: 221 passed; TUI focused Vitest run: 12 passed.
• WebUI generated RPC check and type check passed. TUI generated RPC/i18n checks and type check passed.
• Source-language, large-file, diff-whitespace, and generated-file cleanliness checks passed.

[LivXue]

Windows compatibility review — PR #1123 @ 35489d530c28bb4ebf6073cda0cb6d6d6610ff1f

Scope: full base 4718421324 → this head (55 files). The PR is stacked on #1113; its effort core is identical to the 516c7e7d head reviewed there. Its own delta on top is 33 files (+688/−60): reasoning_efforts() capability reporting with provider forwarding (base, anthropic_messages, endpoint_rotor, lazy, per_model_provider, resolving_provider), the xhigh/max rungs in the session/RPC validation and pickers, docs, and tests.

Result: no Windows-specific issue.

• The provider changes are capability metadata only — each adapter/wrapper answers which effort labels its wire takes and forwards to the inner provider that would serve the model. No new HTTP client, TLS/verify, proxy, path, subprocess, signal, encoding, or platform primitive is introduced; the single scan hit was a source-comment sentence.
• The rest is the same browser/TUI effort-row logic already reviewed in feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113 and fix(ui-web): read the opened conversation's tier, not the one being left #1122.

Verification (native Windows 11, CPython 3.12.13, isolated worktree D:\tmp\rv-1123 at this head; scratch HOME/USERPROFILE, PYTHONUTF8=1; node_modules borrowed read-only from the main checkout):

vitest run src/state/effort.test.ts src/features/model/source.test.ts
→ 2 files, 57 passed

pytest tests/test_session_effort.py tests/test_rpc_session.py \
       tests/test_agent_loop_usage_sink.py tests/test_agent_loop_empty_recovery.py \
       tests/test_provider_protocol.py -q
→ 409 passed in 19.21s

Limitations: TUI tests not run (harness unavailable in this checkout); no live model provider was called, so the xhigh/max wire behaviour is exercised through the provider-protocol tests only.

[0xKT]

Merge conflict: @Handsome-wzw this PR no longer merges cleanly into main.

Measured just now with git merge-tree --write-tree of main 26352da0 and this head 35489d53. Main moved with #1110 (Agent Plugins loader), #1064 (conditional branches) and #1066 (DAG file references), each of which changed CHANGELOG.md. Conflicting files: CHANGELOG.md.

After a rebase onto current main I re-check what moved before the grade comes back.

[gloryfromca]

Blocking: resolve the CHANGELOG.md conflict against current GitHub main before merge.

Confirmed independently at head 35489d530c28 against GitHub main 26352da09a33: git merge-tree --write-tree HEAD github/main exits 1 with a content conflict in CHANGELOG.md, and GitHub reports mergeable: false / mergeable_state: dirty. The branch code is otherwise unchanged; uv run pytest tests/test_session_effort.py -q passes (9 passed).

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the delta from the previously reviewed head 35489d530c28 to d2e73da75f01. The new commit merges current GitHub main; the only hand-resolved conflict was CHANGELOG.md, and the resolution preserves both effort entries and the upstream entries. The PR diff against its new merge base remains the expected 55-file effort change, git merge-tree --write-tree HEAD github/main succeeds, and no effort implementation or test was weakened by the refresh. I rechecked the repository/domain rules, merge history and resolution, overlapping runtime/RPC/provider/frontend callers, generated contracts, backward-compatible surfaces, architecture checks, and test changes.

Verification on this head:

• Focused Python runtime/RPC/provider/schema/layer/source-language suite: 1,169 passed.
• Focused WebUI suite: 221 passed; generated RPC check and type check passed.
• Focused TUI suite: 12 passed; generated RPC/i18n checks and type check passed.
• PR diff whitespace and merge-tree checks passed.

[LivXue]

Windows compatibility re-review — PR #1123 @ d2e73da75f01e9ed92cc08b68f1b2c37e2c190eb

Scope: the head moved from the previously reviewed 35489d53. Against the new main tip 26352da0, the PR's own effort/provider code is byte-identical to the reviewed head — raven/agent/session_effort.py, raven/config/schema.py and the six provider files total the same +188 lines; the rpc/methods/session.py / rpc/models.py changes in the raw delta are main's #1119 merge, not this PR.

Result: no Windows-specific issue; the previous conclusion holds.

Verification (native Windows 11, isolated worktree D:\tmp\rv-1123-n2 at this head; scratch HOME/USERPROFILE, PYTHONUTF8=1; node_modules borrowed read-only from the main checkout):

vitest run src/state/effort.test.ts src/features/model/source.test.ts
→ 2 files, 57 passed

pytest tests/test_session_effort.py tests/test_rpc_session.py \
       tests/test_agent_loop_usage_sink.py tests/test_agent_loop_empty_recovery.py \
       tests/test_provider_protocol.py -q
→ 424 passed in 14.37s

The provider additions remain capability metadata only (reasoning_efforts forwarding), with no HTTP/TLS/proxy, path, subprocess, encoding, or platform code.

Limitations: TUI tests not run; no live model provider was called.

[0xKT]

Blocking: carried from #1113. This PR targets main and contains #1113's earlier head 516c7e7d (merged in through 35489d5), so merging it as it stands merges #1113's open defect too: the Provider default rung clears agents.defaults.reasoningEffort, but a running gateway keeps sending the effort it booted with while session.info / session.set_effort report the clear (thread on #1113, anchored to raven/agent/loop/turn_path.py; the judge reproduced it there).

This PR's own commit a23029c0 reads fine on its own:

• LLMProvider.reasoning_efforts defaults to the four words; anthropic_messages answers its _ADAPTIVE_EFFORTS keys (low to max).
• The lazy, per-model, resolving and rotor wrappers forward it the way they forward reasoning_wire_keys.
• _effort_words falls back to the four when the provider cannot answer.
• session_effort.fit runs a pick the new wire does not take at the highest rung it does take, and logs it. effective_model is bound before the call at turn_path.py:1003.
• The console key stays at the four on purpose.

This clears once #1113's fix lands and this branch is rebased onto it, so its diff against main carries the fixed code. Read only; local tests not run.

[gloryfromca]

Blocking: merge the #1113 clear fix and resolve the current CHANGELOG.md conflict before merge.

Confirmed on this head: gen_overrides still omits reasoning_effort when both the session pick/mode and live configured default are null, so the provider can reuse its frozen boot-time word while the RPC surfaces report the provider default. The current #1113 branch fixes this in 5057b49f by resolving the effort once and always sending None explicitly; that fix is absent here. The affected existing suite passes (uv run pytest tests/test_agent_loop_usage_sink.py tests/test_rpc_session.py tests/test_rpc_settings.py -q: 407 passed), consistent with the transition lacking coverage on this head. I am confirming the facts, not re-grading or resolving the blocker opened here.

[Handsome-wzw]

Merged rather than rebased (branches here merge, never rebase): 754ff3160 brings in #1113's head 01f245518, which holds the clear fix 5057b49f0, so this PR's diff against main now carries the fixed turn -- session_reasoning_effort ranks pick > mode's word > configured default and the call sends it every time, None included. The pick fitting from a23029c0 sits in front of it: turn_effort = self.session_reasoning_effort(session_key, _fit_reasoning_effort(...)) in raven/agent/loop/turn_path.py. Head c325201e2; the verification comment on the PR has the local counts.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Confirmed on c325201e2d6a: #1113 fix 5057b49f0 is now in the ancestry, the conflict resolution fits the session pick before session_reasoning_effort ranks it, and gen_overrides always carries the resulting value, including explicit None. The new clear-transition coverage passes. Current GitHub main is also in the ancestry and git merge-tree --write-tree HEAD github/main succeeds. My prior blocker is withdrawn; I am leaving resolution of this thread to its opener.

[0xKT]

Merge conflict: @Handsome-wzw this PR no longer merges cleanly into main.

Measured just now with git merge-tree --write-tree of main 984d47d4 and this head d2e73da7. Main moved with #1124 (add buttons on the skills and plugins pages that lead into the chat), which changed CHANGELOG.md. Conflicting files: CHANGELOG.md.

After a rebase onto current main I re-check what moved before the grade comes back.

[gloryfromca]

Blocking: merge the #1113 clear fix and resolve the current CHANGELOG.md conflict before merge.

The second condition also reproduces independently: after fetching GitHub main at 984d47d4b74c, git merge-tree --write-tree HEAD github/main exits 1 with a content conflict in CHANGELOG.md, and GitHub reports mergeable: false / mergeable_state: dirty. The branch remains at d2e73da75f01.

[Handsome-wzw]

Merged on top of the previous head d2e73da75, no rebase and no force-push:

• feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113's current head 01f245518 (merge 754ff3160): its clear fix 5057b49f0 -- the turn sends the effort in force on every call, None included, through session_effort.in_force / session_reasoning_effort -- plus origin/main as feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113 carries it (feat(ui-web): add buttons on the skills and plugins pages that lead into the chat #1124). Three conflicts, all between that fix and this PR's pick fitting:
  • raven/agent/loop/turn_path.py: the pick is fitted to the turn's wire first (_fit_reasoning_effort) and then handed to session_reasoning_effort to rank -- the pick only; a mode's word goes out as named, as before. gen_overrides sends reasoning_effort on every call as feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113 now has it.
  • raven/rpc/methods/session.py: feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113's _mode_effort / _configured_effort / _default_effort / _effort_in_force kept as they are, this PR's _effort_words beside them.
  • CONTEXT.md: the Session Reasoning Effort entry carries feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113's in_force sentences and this PR's provider-words paragraph, and says the fit runs before in_force ranks the pick.
• origin/main 06509f8e4 (merge 70b83a601, clean): fix(*): report a delegated turn's real outcome and cost #1121 and the 0.2.6 bump.

git diff origin/feat/session_reasoning_effort -- CHANGELOG.md is this PR's bullet alone; the Added order stays this PR's bullet, then #1113's, then main's.

Local checks on c325201e2:

• full pytest tests (-n auto): 36319 passed, 3 failed, 258 skipped, 2 xfailed, 7m32s -- run on the tree one blank line short of this head (ruff format wanted one more between two functions in rpc/methods/session.py; folded into 47f5d64e7, with the related group and ruff rerun on the final tree). The 3 are this Mac's known baseline, failing the same way on main: test_rpc_files::test_a_host_without_libreoffice_says_so, test_simulation_scenario::test_a_fresh_trial_refreshes_the_subagent_homes_before_it_runs, test_simulation_suite::test_a_stopped_run_is_marked_in_its_record_and_finished_with_its_evidence
• related group (test_session_effort, test_agent_loop_session_effort, test_agent_loop_iteration_hooks, test_watch_work, test_agent_loop_empty_recovery, test_agent_loop_usage_sink, test_provider_protocol, test_rpc_session, test_import_cycle_budget): 533 passed (325 of them, the five session/effort files, rerun on the final head); ruff check / ruff format --check clean; lint-imports 10 kept, 0 broken
• ui-web: gen:check in sync, tsc --noEmit clean, eslint 0 errors / 4 warnings (as on main), vitest run 261 files / 4452 tests passed
• ui-tui: lint:rpc / lint:i18n in sync, tsc --noEmit clean, vitest run 146 files / 2286 tests passed
• check_commit_messages / check_source_language / check_large_files / commitlint over origin/main..HEAD: all pass

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the delta from d2e73da75f01 to c325201e2d6a, including the repository and domain rules, merge history and three hand-resolved overlaps, affected loop/RPC/provider callers, backward-compatible None semantics, architecture boundaries, and whether tests were weakened. The prior clear defect is fixed: provider-aware fitting feeds the shared in-force ranking, and every model call receives the resolved effort, including explicit None. Current GitHub main is in the ancestry and the merge-tree check is clean. I found no new concrete failure.

Verification on this head:

• Focused session-effort, loop, provider, RPC, and import-boundary suite: 533 passed.
• WebUI generated RPC check and type check passed.
• TUI generated RPC/i18n checks and type check passed.
• Source-language, large-file, diff-whitespace, and merge-tree checks passed.

[LivXue]

Windows compatibility re-review — PR #1123 @ c325201e2d6a9d8bb0fac5ea8541dc4345ca4f7d

Scope: the head moved from the previously reviewed d2e73da7. The only change in the PR's own files is the base PR's session_effort.in_force() helper (+23 lines, stacked from #1113); the reasoning_efforts() capability additions in raven/providers/** and the RPC/UI effort rows are unchanged. The rest of the raw delta is main's merges.

Result: no Windows-specific issue; the previous conclusion holds.

Verification (native Windows 11, isolated worktree D:\tmp\rv-1123-n3 at this head; scratch HOME/USERPROFILE, PYTHONUTF8=1; node_modules borrowed read-only from the main checkout):

vitest run src/state/effort.test.ts src/features/model/source.test.ts
→ 2 files, 57 passed

pytest tests/test_session_effort.py tests/test_rpc_session.py \
       tests/test_agent_loop_usage_sink.py tests/test_agent_loop_empty_recovery.py \
       tests/test_provider_protocol.py -q
→ 426 passed in 19.03s

The provider additions remain capability metadata (reasoning_efforts forwarding); no HTTP/TLS/proxy, path, subprocess, encoding, or platform code.

Limitations: TUI tests not run; no live model provider was called.

[Handsome-wzw]

Merged #1113's current head 1a0e9c2e6 on top of c325201e2 (merge f050cb46d; no rebase, no force-push). That brings in its three latest commits -- 251cdb551 (/effort refused in a direct chat, /mode's reset words), bffb9803f (a refused draft pick put back; applyStagedEffort re-reads under the promotion's ticket), 1a0e9c2e6 (the side calls' effort pinned, raven_config clears the key, a fork carries the pick). origin/main has not moved since 06509f8e4.

Five conflicts, none in Python; each was the two branches adding a sentence or a fake at the same place, kept together:

• CONTEXT.md: the Session Reasoning Effort entry keeps this PR's provider-words paragraph and takes feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113's two additions into it (the raven_config door stays at the four with the console's checker; /effort is refused inside a direct chat).
• docs-site/docs/using-raven.md / .zh.md: both sentences on the /effort bullet (the row offers the model's words and a refusal names them; a direct chat refuses the command).
• ui-tui/src/app/slash/commands/session.ts: one comment with both facts; feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113's RESET_WORDS and direct-chat refusal and this PR's usage line merged without conflict.
• ui-web/src/features/model/source.test.ts: the harness's src/state/effort fake now records every load(sid, gen) in its own list, which feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113's effortReads() returns and this PR's effortLoads() counts -- off calls, whose exact shape other cases pin. feat(agent): per-conversation reasoning effort (picker row, /effort, session.set_effort) #1113's five new cases and this PR's two run unchanged.
• One comment in raven/config/self_surface.py (auto-merged) said the session's pick reads REASONING_EFFORTS too; on this branch the pick's words are the model's provider's, so the comment now says that.

git diff origin/feat/session_reasoning_effort -- CHANGELOG.md is still this PR's bullet alone. Generated files (ui-web gen:check, ui-tui lint:rpc / lint:i18n) are in sync.

Local checks on f050cb46d:

• full pytest tests (-n auto): 36334 passed, 3 failed, 258 skipped, 2 xfailed, 6m22s; the 3 are this Mac's known baseline, failing the same way on main: test_rpc_files::test_a_host_without_libreoffice_says_so, test_simulation_scenario::test_a_fresh_trial_refreshes_the_subagent_homes_before_it_runs, test_simulation_suite::test_a_stopped_run_is_marked_in_its_record_and_finished_with_its_evidence
• related group (session effort, loop session effort, rpc session, config self_surface, raven_config tool, session manager, usage sink, empty recovery, provider protocol, import cycle budget, iteration hooks, watch work): 957 passed; ruff check / ruff format --check clean; lint-imports 10 kept, 0 broken; mkdocs build --strict clean
• ui-web: tsc --noEmit clean, eslint 0 errors / 4 warnings (as on main), vitest run 261 files / 4457 tests passed (source.test.ts + effort.test.ts alone: 62)
• ui-tui: tsc --noEmit clean, eslint clean, vitest run 146 files / 2289 tests: 2288 passed and one load-timing case, subagentsHub.test.tsx > a rename carries the prior probe status forward under the new name, which failed once under the concurrent Python run and passes 3/3 alone (no file it touches changed on this branch)
• check_commit_messages / check_source_language / check_large_files / commitlint over origin/main..HEAD: all pass

LivXue's thread on #1113 about the words being the model's is left as it is; nothing here answers it.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the delta from c325201, including all five conflict resolutions and the three newly merged #1113 fixes. I checked the repository rules and domain terms; the resulting diff and affected callers; the merge history and current main; backward compatibility; test changes for weakening; and the provider/session/UI architecture boundaries. The resolutions preserve the provider-specific xhigh/max behavior while correctly integrating draft recovery, promotion-ticket guarding, side-call effort pinning, config clearing, fork inheritance, and the TUI direct-chat guard.

Verification on this head: 953 affected Python tests passed; 133 focused Web tests passed; 12 focused TUI tests passed; Web and TUI type checks passed; generated Web RPC plus TUI RPC/i18n checks passed; git diff --check passed; current GitHub main is in the branch ancestry and git merge-tree --write-tree HEAD github/main succeeds. I found no new defect, compatibility break, weakened test, or unresolved conflict introduced by this merge.

The only open review thread was opened by another reviewer; its defect is fixed and my prior reply already withdrew my supporting objection, so I did not resolve someone else's thread.

[LivXue]

Windows compatibility re-review — PR #1123 @ f050cb46d1d585e2c96b90c62f1dc303cea4874b

Scope: the head moved from the previously reviewed c325201e. The delta is the stacked #1113 changes (effort source/runtime keyed to the opened conversation, session fork carrying reasoning_effort, self-surface nullable effort) plus main's merges; #1123's own additions — the reasoning_efforts() capability forwarding in raven/providers/** and the effort rows — are unchanged.

Result: no Windows-specific issue; the previous conclusion holds. The delta is browser/TUI state and config-catalogue logic, with no path, subprocess, encoding, or platform primitive.

Verification (native Windows 11, isolated worktree D:\tmp\rv-1123-n4 at this head; scratch HOME/USERPROFILE, PYTHONUTF8=1; node_modules borrowed read-only from the main checkout):

vitest run src/state/effort.test.ts src/features/model/source.test.ts
→ 2 files, 62 passed

pytest tests/test_session_effort.py tests/test_rpc_session.py \
       tests/test_agent_loop_usage_sink.py tests/test_agent_loop_empty_recovery.py \
       tests/test_provider_protocol.py tests/test_agent_loop_session_effort.py -q
→ 441 passed in 25.12s

Limitations: TUI tests not run; no live model provider was called.