diff --git a/.changeset/brave-moons-delegate.md b/.changeset/brave-moons-delegate.md
deleted file mode 100644
index d9a8bc99..00000000
--- a/.changeset/brave-moons-delegate.md
+++ /dev/null
@@ -1,12 +0,0 @@
----
-"@openspec-ui/core": minor
-"openspec-ui-vscode": minor
-"@openspec-ui/webui": minor
----
-
-A task that needs a live check can name the agent that performs it.
-`**Delegated to <agent-id>**` sits beside `**Human-only**`: the first
-means another agent can make the check, the second that none can. The
-inbox in both hosts now carries both kinds and says who each item waits
-on, naming an agent id the registry does not carry rather than treating
-it as assigned.
diff --git a/.changeset/brown-pianos-refuse.md b/.changeset/brown-pianos-refuse.md
deleted file mode 100644
index 74de8b8b..00000000
--- a/.changeset/brown-pianos-refuse.md
+++ /dev/null
@@ -1,21 +0,0 @@
----
-"@openspec-ui/core": minor
-"@openspec-ui/server": minor
-"@openspec-ui/webui": patch
-"openspec-ui-vscode": patch
----
-
-A name arriving from a request is checked before it is used. A change
-name now passes the change-name rule before it is joined into a path,
-in core beside the path it protects, so a message naming
-`../../../../Users/me/.claude` no longer decides where a `harness.json`
-is written — the bridge answers `ok: false` and the REST routes answer
-400, both carrying the rule the name broke. A schedule entry is
-validated on the way in by the same rule the reader applies on the way
-out, so a stored row and the response that reported it can no longer
-disagree, and a body asking for an addition and a removal at once is
-refused rather than half-applied. A `customAgent` obeys the same shape
-rule as a model id, for the same reason: both reach the CLI as the value
-of a flag, and a value beginning with `-` may be read as a second one. A
-custom-agent definition whose file name that rule refuses is reported as
-found and not offered, rather than dropped in silence.
diff --git a/.changeset/olive-rows-restore.md b/.changeset/olive-rows-restore.md
deleted file mode 100644
index 2e8e8efe..00000000
--- a/.changeset/olive-rows-restore.md
+++ /dev/null
@@ -1,9 +0,0 @@
----
-"openspec-ui-vscode": patch
----
-
-A change row rebuilt to answer "what is this element's parent" now
-carries the state the tree drew, instead of `draft` written in. VS Code
-restores the tree's selection through that chain after a window reload
-and draws what it returns, so a change with every task done could read
-`draft` until the next refresh.
diff --git a/.changeset/quality-names-the-checked-agent.md b/.changeset/quality-names-the-checked-agent.md
deleted file mode 100644
index 5c5d45b6..00000000
--- a/.changeset/quality-names-the-checked-agent.md
+++ /dev/null
@@ -1,26 +0,0 @@
----
-"@openspec-ui/core": minor
-"@openspec-ui/webui": minor
----
-
-What a verifying stage's checks found is now charged to the agent whose
-work they covered. The entry gains `checkedAgent`, taken from the chain's
-resolved `apply` stage, and the quality readback groups by it — grouping
-by the entry's `agent` could only ever produce one row, named
-`verify-checks`, whatever had run the apply. An entry recorded before
-that field existed is counted and reported as such rather than charged to
-a group.
-
-A checks entry is also no longer counted as a run. It carries a terminal
-outcome and no `started` partner, so the per-change cost report listed it
-as a run refused before it started and one chain run of apply and verify
-reported two previous runs; one predicate in core now says which entries
-are runs, and both counters use it. A checks entry therefore no longer
-appears as a row in the per-change cost report — what it found is read
-back beside the run figures instead.
-
-A recommendation's gap says which nothing it is: nothing reported the
-measure, something reported it but rests on too few runs, or one
-candidate is eligible with nothing to compare against. Four runs that
-each reported a cost previously read as "no agent has reported a cost
-across 4 recorded run(s)".
diff --git a/.changeset/quiet-schedules-keep-promises.md b/.changeset/quiet-schedules-keep-promises.md
deleted file mode 100644
index 603209ff..00000000
--- a/.changeset/quiet-schedules-keep-promises.md
+++ /dev/null
@@ -1,22 +0,0 @@
----
-"@openspec-ui/core": minor
-"@openspec-ui/webui": minor
-"@openspec-ui/server": patch
-"openspec-ui-vscode": minor
----
-
-A scheduled run keeps the promise the dialog makes. Opening the
-application is now enough: the workspace is read on open, so the
-schedule is read too and a due run starts with nothing else done — it
-used to wait for a click that a real reopen never makes. The run starts
-on the path that was chosen when it was scheduled rather than reopening
-the dialog for the same choice, and the entry leaves the file only once
-the run has been opened, so a configuration that cannot be resolved
-reports itself as a run that could not be opened instead of consuming
-the schedule under the wrong message. A change archived after being
-scheduled is dropped and says it was archived, and a run due behind it
-starts on the same reading. Firing is decided once, in
-`planScheduleFiring` in core, with each host performing only the
-effects it is handed. The dialog is announced as a dialog and takes
-focus when it opens by itself, and what the schedule did is readable
-from any tab of the standalone shell.
diff --git a/.changeset/stage-override-keeps-its-custom-agent.md b/.changeset/stage-override-keeps-its-custom-agent.md
deleted file mode 100644
index 0ccaa07d..00000000
--- a/.changeset/stage-override-keeps-its-custom-agent.md
+++ /dev/null
@@ -1,33 +0,0 @@
----
-"@openspec-ui/core": patch
-"@openspec-ui/webui": patch
-"openspec-ui-vscode": patch
----
-
-A stage override keeps its custom agent, and one function decides what
-applying a named configuration writes.
-
-`mergeStepAgent` merged three named fields across a per-change override.
-`customAgent` was the fourth field a stage entry may carry, so a change
-naming the same agent plus a custom agent resolved without it and the
-chain ran with no `--agent` flag, silently. The merge now iterates
-`STEP_AGENT_KEYS` — the list the validator already reads — so the next
-field added to an entry arrives already merged, and it agrees with
-`templateConfigToWrite`, which kept the field by spread.
-
-Applying a named configuration to a change now goes through one core
-function, `changeTemplateConfigToWrite`, from all three surfaces. The
-run dialog resolved a configuration's effort against the change's
-resolved configuration and the settings view against the change's own
-override, where every stage the change does not name reads as
-"inherit" — so the two wrote different files for the same change, and
-the settings view's message said "None of the agents on screen takes an
-effort setting" when that was not the reason. That message now names the
-stages given an effort, the agents that take none, and the stages with
-no agent chosen, each only where it is true.
-
-The balanced and careful configurations describe their effort by its
-position in the agent's range ("a third of the way up", "two thirds")
-rather than as "the middle", which the thirds mapping never produced:
-for `copilot-cli` the medium level resolves to `low`, the third of
-seven. `HARNESS.md` carries the resolved value per registered agent.
diff --git a/package-lock.json b/package-lock.json
index 440202ad..917fb27f 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -11957,7 +11957,7 @@
     },
     "packages/core": {
       "name": "@openspec-ui/core",
-      "version": "0.62.0",
+      "version": "0.63.0",
       "dependencies": {
         "@agentclientprotocol/sdk": "^1.4.0",
         "cross-spawn": "^7.0.6",
@@ -11973,7 +11973,7 @@
     },
     "packages/extension": {
       "name": "openspec-ui-vscode",
-      "version": "0.46.1",
+      "version": "0.47.0",
       "license": "MIT",
       "dependencies": {
         "@openspec-ui/core": "*",
@@ -12479,7 +12479,7 @@
     },
     "packages/server": {
       "name": "@openspec-ui/server",
-      "version": "1.17.0",
+      "version": "1.18.0",
       "dependencies": {
         "@openspec-ui/core": "*",
         "ws": "^8.21.0"
@@ -12979,7 +12979,7 @@
     },
     "packages/webui": {
       "name": "@openspec-ui/webui",
-      "version": "1.33.0",
+      "version": "1.34.0",
       "dependencies": {
         "@openspec-ui/core": "*",
         "@tanstack/react-virtual": "^3.14.10",
diff --git a/packages/core/CHANGELOG.md b/packages/core/CHANGELOG.md
index b12d9d01..6edcc375 100644
--- a/packages/core/CHANGELOG.md
+++ b/packages/core/CHANGELOG.md
@@ -1,5 +1,96 @@
 # @openspec-ui/core
 
+## 0.63.0
+
+### Minor Changes
+
+- be28986: A task that needs a live check can name the agent that performs it.
+  `**Delegated to <agent-id>**` sits beside `**Human-only**`: the first
+  means another agent can make the check, the second that none can. The
+  inbox in both hosts now carries both kinds and says who each item waits
+  on, naming an agent id the registry does not carry rather than treating
+  it as assigned.
+- 9dd0767: A name arriving from a request is checked before it is used. A change
+  name now passes the change-name rule before it is joined into a path,
+  in core beside the path it protects, so a message naming
+  `../../../../Users/me/.claude` no longer decides where a `harness.json`
+  is written — the bridge answers `ok: false` and the REST routes answer
+  400, both carrying the rule the name broke. A schedule entry is
+  validated on the way in by the same rule the reader applies on the way
+  out, so a stored row and the response that reported it can no longer
+  disagree, and a body asking for an addition and a removal at once is
+  refused rather than half-applied. A `customAgent` obeys the same shape
+  rule as a model id, for the same reason: both reach the CLI as the value
+  of a flag, and a value beginning with `-` may be read as a second one. A
+  custom-agent definition whose file name that rule refuses is reported as
+  found and not offered, rather than dropped in silence.
+- 1b67bee: What a verifying stage's checks found is now charged to the agent whose
+  work they covered. The entry gains `checkedAgent`, taken from the chain's
+  resolved `apply` stage, and the quality readback groups by it — grouping
+  by the entry's `agent` could only ever produce one row, named
+  `verify-checks`, whatever had run the apply. An entry recorded before
+  that field existed is counted and reported as such rather than charged to
+  a group.
+  
+  A checks entry is also no longer counted as a run. It carries a terminal
+  outcome and no `started` partner, so the per-change cost report listed it
+  as a run refused before it started and one chain run of apply and verify
+  reported two previous runs; one predicate in core now says which entries
+  are runs, and both counters use it. A checks entry therefore no longer
+  appears as a row in the per-change cost report — what it found is read
+  back beside the run figures instead.
+  
+  A recommendation's gap says which nothing it is: nothing reported the
+  measure, something reported it but rests on too few runs, or one
+  candidate is eligible with nothing to compare against. Four runs that
+  each reported a cost previously read as "no agent has reported a cost
+  across 4 recorded run(s)".
+- c679bd4: A scheduled run keeps the promise the dialog makes. Opening the
+  application is now enough: the workspace is read on open, so the
+  schedule is read too and a due run starts with nothing else done — it
+  used to wait for a click that a real reopen never makes. The run starts
+  on the path that was chosen when it was scheduled rather than reopening
+  the dialog for the same choice, and the entry leaves the file only once
+  the run has been opened, so a configuration that cannot be resolved
+  reports itself as a run that could not be opened instead of consuming
+  the schedule under the wrong message. A change archived after being
+  scheduled is dropped and says it was archived, and a run due behind it
+  starts on the same reading. Firing is decided once, in
+  `planScheduleFiring` in core, with each host performing only the
+  effects it is handed. The dialog is announced as a dialog and takes
+  focus when it opens by itself, and what the schedule did is readable
+  from any tab of the standalone shell.
+
+### Patch Changes
+
+- ad1a8ae: A stage override keeps its custom agent, and one function decides what
+  applying a named configuration writes.
+  
+  `mergeStepAgent` merged three named fields across a per-change override.
+  `customAgent` was the fourth field a stage entry may carry, so a change
+  naming the same agent plus a custom agent resolved without it and the
+  chain ran with no `--agent` flag, silently. The merge now iterates
+  `STEP_AGENT_KEYS` — the list the validator already reads — so the next
+  field added to an entry arrives already merged, and it agrees with
+  `templateConfigToWrite`, which kept the field by spread.
+  
+  Applying a named configuration to a change now goes through one core
+  function, `changeTemplateConfigToWrite`, from all three surfaces. The
+  run dialog resolved a configuration's effort against the change's
+  resolved configuration and the settings view against the change's own
+  override, where every stage the change does not name reads as
+  "inherit" — so the two wrote different files for the same change, and
+  the settings view's message said "None of the agents on screen takes an
+  effort setting" when that was not the reason. That message now names the
+  stages given an effort, the agents that take none, and the stages with
+  no agent chosen, each only where it is true.
+  
+  The balanced and careful configurations describe their effort by its
+  position in the agent's range ("a third of the way up", "two thirds")
+  rather than as "the middle", which the thirds mapping never produced:
+  for `copilot-cli` the medium level resolves to `low`, the third of
+  seven. `HARNESS.md` carries the resolved value per registered agent.
+
 ## 0.62.0
 
 ### Minor Changes
diff --git a/packages/core/package.json b/packages/core/package.json
index 496a4836..45c56af3 100644
--- a/packages/core/package.json
+++ b/packages/core/package.json
@@ -1,7 +1,7 @@
 {
   "name": "@openspec-ui/core",
   "private": true,
-  "version": "0.62.0",
+  "version": "0.63.0",
   "type": "module",
   "main": "src/index.ts",
   "exports": {
diff --git a/packages/extension/CHANGELOG.md b/packages/extension/CHANGELOG.md
index ffe966c3..8ee4b648 100644
--- a/packages/extension/CHANGELOG.md
+++ b/packages/extension/CHANGELOG.md
@@ -1,5 +1,88 @@
 # Changelog
 
+## 0.47.0
+
+### Minor Changes
+
+- be28986: A task that needs a live check can name the agent that performs it.
+  `**Delegated to <agent-id>**` sits beside `**Human-only**`: the first
+  means another agent can make the check, the second that none can. The
+  inbox in both hosts now carries both kinds and says who each item waits
+  on, naming an agent id the registry does not carry rather than treating
+  it as assigned.
+- c679bd4: A scheduled run keeps the promise the dialog makes. Opening the
+  application is now enough: the workspace is read on open, so the
+  schedule is read too and a due run starts with nothing else done — it
+  used to wait for a click that a real reopen never makes. The run starts
+  on the path that was chosen when it was scheduled rather than reopening
+  the dialog for the same choice, and the entry leaves the file only once
+  the run has been opened, so a configuration that cannot be resolved
+  reports itself as a run that could not be opened instead of consuming
+  the schedule under the wrong message. A change archived after being
+  scheduled is dropped and says it was archived, and a run due behind it
+  starts on the same reading. Firing is decided once, in
+  `planScheduleFiring` in core, with each host performing only the
+  effects it is handed. The dialog is announced as a dialog and takes
+  focus when it opens by itself, and what the schedule did is readable
+  from any tab of the standalone shell.
+
+### Patch Changes
+
+- 9dd0767: A name arriving from a request is checked before it is used. A change
+  name now passes the change-name rule before it is joined into a path,
+  in core beside the path it protects, so a message naming
+  `../../../../Users/me/.claude` no longer decides where a `harness.json`
+  is written — the bridge answers `ok: false` and the REST routes answer
+  400, both carrying the rule the name broke. A schedule entry is
+  validated on the way in by the same rule the reader applies on the way
+  out, so a stored row and the response that reported it can no longer
+  disagree, and a body asking for an addition and a removal at once is
+  refused rather than half-applied. A `customAgent` obeys the same shape
+  rule as a model id, for the same reason: both reach the CLI as the value
+  of a flag, and a value beginning with `-` may be read as a second one. A
+  custom-agent definition whose file name that rule refuses is reported as
+  found and not offered, rather than dropped in silence.
+- 683fef4: A change row rebuilt to answer "what is this element's parent" now
+  carries the state the tree drew, instead of `draft` written in. VS Code
+  restores the tree's selection through that chain after a window reload
+  and draws what it returns, so a change with every task done could read
+  `draft` until the next refresh.
+- ad1a8ae: A stage override keeps its custom agent, and one function decides what
+  applying a named configuration writes.
+  
+  `mergeStepAgent` merged three named fields across a per-change override.
+  `customAgent` was the fourth field a stage entry may carry, so a change
+  naming the same agent plus a custom agent resolved without it and the
+  chain ran with no `--agent` flag, silently. The merge now iterates
+  `STEP_AGENT_KEYS` — the list the validator already reads — so the next
+  field added to an entry arrives already merged, and it agrees with
+  `templateConfigToWrite`, which kept the field by spread.
+  
+  Applying a named configuration to a change now goes through one core
+  function, `changeTemplateConfigToWrite`, from all three surfaces. The
+  run dialog resolved a configuration's effort against the change's
+  resolved configuration and the settings view against the change's own
+  override, where every stage the change does not name reads as
+  "inherit" — so the two wrote different files for the same change, and
+  the settings view's message said "None of the agents on screen takes an
+  effort setting" when that was not the reason. That message now names the
+  stages given an effort, the agents that take none, and the stages with
+  no agent chosen, each only where it is true.
+  
+  The balanced and careful configurations describe their effort by its
+  position in the agent's range ("a third of the way up", "two thirds")
+  rather than as "the middle", which the thirds mapping never produced:
+  for `copilot-cli` the medium level resolves to `low`, the third of
+  seven. `HARNESS.md` carries the resolved value per registered agent.
+- Updated dependencies [be28986]
+- Updated dependencies [9dd0767]
+- Updated dependencies [1b67bee]
+- Updated dependencies [c679bd4]
+- Updated dependencies [ad1a8ae]
+  - @openspec-ui/core@0.63.0
+  - @openspec-ui/webui@1.34.0
+  - @openspec-ui/server@1.18.0
+
 ## 0.46.1
 
 ### Patch Changes
diff --git a/packages/extension/package.json b/packages/extension/package.json
index 155a0eb0..e880c1da 100644
--- a/packages/extension/package.json
+++ b/packages/extension/package.json
@@ -4,7 +4,7 @@
   "displayName": "OpenSpec Workbench",
   "description": "A dashboard + VS Code extension for OpenSpec, with Claude, Copilot, Codex, and Gemini agents built in.",
   "publisher": "openspec-ui",
-  "version": "0.46.1",
+  "version": "0.47.0",
   "icon": "media/icon.png",
   "license": "MIT",
   "repository": {
diff --git a/packages/server/CHANGELOG.md b/packages/server/CHANGELOG.md
index f524b14b..b1e6b579 100644
--- a/packages/server/CHANGELOG.md
+++ b/packages/server/CHANGELOG.md
@@ -1,5 +1,48 @@
 # @openspec-ui/server
 
+## 1.18.0
+
+### Minor Changes
+
+- 9dd0767: A name arriving from a request is checked before it is used. A change
+  name now passes the change-name rule before it is joined into a path,
+  in core beside the path it protects, so a message naming
+  `../../../../Users/me/.claude` no longer decides where a `harness.json`
+  is written — the bridge answers `ok: false` and the REST routes answer
+  400, both carrying the rule the name broke. A schedule entry is
+  validated on the way in by the same rule the reader applies on the way
+  out, so a stored row and the response that reported it can no longer
+  disagree, and a body asking for an addition and a removal at once is
+  refused rather than half-applied. A `customAgent` obeys the same shape
+  rule as a model id, for the same reason: both reach the CLI as the value
+  of a flag, and a value beginning with `-` may be read as a second one. A
+  custom-agent definition whose file name that rule refuses is reported as
+  found and not offered, rather than dropped in silence.
+
+### Patch Changes
+
+- c679bd4: A scheduled run keeps the promise the dialog makes. Opening the
+  application is now enough: the workspace is read on open, so the
+  schedule is read too and a due run starts with nothing else done — it
+  used to wait for a click that a real reopen never makes. The run starts
+  on the path that was chosen when it was scheduled rather than reopening
+  the dialog for the same choice, and the entry leaves the file only once
+  the run has been opened, so a configuration that cannot be resolved
+  reports itself as a run that could not be opened instead of consuming
+  the schedule under the wrong message. A change archived after being
+  scheduled is dropped and says it was archived, and a run due behind it
+  starts on the same reading. Firing is decided once, in
+  `planScheduleFiring` in core, with each host performing only the
+  effects it is handed. The dialog is announced as a dialog and takes
+  focus when it opens by itself, and what the schedule did is readable
+  from any tab of the standalone shell.
+- Updated dependencies [be28986]
+- Updated dependencies [9dd0767]
+- Updated dependencies [1b67bee]
+- Updated dependencies [c679bd4]
+- Updated dependencies [ad1a8ae]
+  - @openspec-ui/core@0.63.0
+
 ## 1.17.0
 
 ### Minor Changes
diff --git a/packages/server/package.json b/packages/server/package.json
index 55055bd5..62434ab6 100644
--- a/packages/server/package.json
+++ b/packages/server/package.json
@@ -1,7 +1,7 @@
 {
   "name": "@openspec-ui/server",
   "private": true,
-  "version": "1.17.0",
+  "version": "1.18.0",
   "type": "module",
   "main": "src/index.ts",
   "dependencies": {
diff --git a/packages/webui/CHANGELOG.md b/packages/webui/CHANGELOG.md
index 71531507..24bbcc5b 100644
--- a/packages/webui/CHANGELOG.md
+++ b/packages/webui/CHANGELOG.md
@@ -1,5 +1,102 @@
 # @openspec-ui/webui
 
+## 1.34.0
+
+### Minor Changes
+
+- be28986: A task that needs a live check can name the agent that performs it.
+  `**Delegated to <agent-id>**` sits beside `**Human-only**`: the first
+  means another agent can make the check, the second that none can. The
+  inbox in both hosts now carries both kinds and says who each item waits
+  on, naming an agent id the registry does not carry rather than treating
+  it as assigned.
+- 1b67bee: What a verifying stage's checks found is now charged to the agent whose
+  work they covered. The entry gains `checkedAgent`, taken from the chain's
+  resolved `apply` stage, and the quality readback groups by it — grouping
+  by the entry's `agent` could only ever produce one row, named
+  `verify-checks`, whatever had run the apply. An entry recorded before
+  that field existed is counted and reported as such rather than charged to
+  a group.
+  
+  A checks entry is also no longer counted as a run. It carries a terminal
+  outcome and no `started` partner, so the per-change cost report listed it
+  as a run refused before it started and one chain run of apply and verify
+  reported two previous runs; one predicate in core now says which entries
+  are runs, and both counters use it. A checks entry therefore no longer
+  appears as a row in the per-change cost report — what it found is read
+  back beside the run figures instead.
+  
+  A recommendation's gap says which nothing it is: nothing reported the
+  measure, something reported it but rests on too few runs, or one
+  candidate is eligible with nothing to compare against. Four runs that
+  each reported a cost previously read as "no agent has reported a cost
+  across 4 recorded run(s)".
+- c679bd4: A scheduled run keeps the promise the dialog makes. Opening the
+  application is now enough: the workspace is read on open, so the
+  schedule is read too and a due run starts with nothing else done — it
+  used to wait for a click that a real reopen never makes. The run starts
+  on the path that was chosen when it was scheduled rather than reopening
+  the dialog for the same choice, and the entry leaves the file only once
+  the run has been opened, so a configuration that cannot be resolved
+  reports itself as a run that could not be opened instead of consuming
+  the schedule under the wrong message. A change archived after being
+  scheduled is dropped and says it was archived, and a run due behind it
+  starts on the same reading. Firing is decided once, in
+  `planScheduleFiring` in core, with each host performing only the
+  effects it is handed. The dialog is announced as a dialog and takes
+  focus when it opens by itself, and what the schedule did is readable
+  from any tab of the standalone shell.
+
+### Patch Changes
+
+- 9dd0767: A name arriving from a request is checked before it is used. A change
+  name now passes the change-name rule before it is joined into a path,
+  in core beside the path it protects, so a message naming
+  `../../../../Users/me/.claude` no longer decides where a `harness.json`
+  is written — the bridge answers `ok: false` and the REST routes answer
+  400, both carrying the rule the name broke. A schedule entry is
+  validated on the way in by the same rule the reader applies on the way
+  out, so a stored row and the response that reported it can no longer
+  disagree, and a body asking for an addition and a removal at once is
+  refused rather than half-applied. A `customAgent` obeys the same shape
+  rule as a model id, for the same reason: both reach the CLI as the value
+  of a flag, and a value beginning with `-` may be read as a second one. A
+  custom-agent definition whose file name that rule refuses is reported as
+  found and not offered, rather than dropped in silence.
+- ad1a8ae: A stage override keeps its custom agent, and one function decides what
+  applying a named configuration writes.
+  
+  `mergeStepAgent` merged three named fields across a per-change override.
+  `customAgent` was the fourth field a stage entry may carry, so a change
+  naming the same agent plus a custom agent resolved without it and the
+  chain ran with no `--agent` flag, silently. The merge now iterates
+  `STEP_AGENT_KEYS` — the list the validator already reads — so the next
+  field added to an entry arrives already merged, and it agrees with
+  `templateConfigToWrite`, which kept the field by spread.
+  
+  Applying a named configuration to a change now goes through one core
+  function, `changeTemplateConfigToWrite`, from all three surfaces. The
+  run dialog resolved a configuration's effort against the change's
+  resolved configuration and the settings view against the change's own
+  override, where every stage the change does not name reads as
+  "inherit" — so the two wrote different files for the same change, and
+  the settings view's message said "None of the agents on screen takes an
+  effort setting" when that was not the reason. That message now names the
+  stages given an effort, the agents that take none, and the stages with
+  no agent chosen, each only where it is true.
+  
+  The balanced and careful configurations describe their effort by its
+  position in the agent's range ("a third of the way up", "two thirds")
+  rather than as "the middle", which the thirds mapping never produced:
+  for `copilot-cli` the medium level resolves to `low`, the third of
+  seven. `HARNESS.md` carries the resolved value per registered agent.
+- Updated dependencies [be28986]
+- Updated dependencies [9dd0767]
+- Updated dependencies [1b67bee]
+- Updated dependencies [c679bd4]
+- Updated dependencies [ad1a8ae]
+  - @openspec-ui/core@0.63.0
+
 ## 1.33.0
 
 ### Minor Changes
diff --git a/packages/webui/package.json b/packages/webui/package.json
index 34ccaadd..dc701d44 100644
--- a/packages/webui/package.json
+++ b/packages/webui/package.json
@@ -1,7 +1,7 @@
 {
   "name": "@openspec-ui/webui",
   "private": true,
-  "version": "1.33.0",
+  "version": "1.34.0",
   "type": "module",
   "main": "src/index.ts",
   "dependencies": {