axstack 0.49.0 → 0.50.0

package/README.md

@@ -124,18 +124,16 @@ host validation. See [PR-manager setup and safety](docs/host-operations.md#optio
 
 ## Automatic merge boundaries
 
-Eligible own PRs use guarded automatic merge; see [merge boundaries, exclusions, account-selection carve-out and accepted risks](docs/workflows.md#automatic-merge-boundaries).
+Eligible own PRs use guarded automatic merge; see [merge boundaries, exclusions and accepted risks](docs/workflows.md#automatic-merge-boundaries).
 
 ## Some notes
 
 - T3 Code is the only supported active runtime. Axstack adds no daemon or runtime
-  database. The exceptions are exactly two opt-in systemd user timers:
-  `axstack-default-sync` rewrites only T3 `defaultModelSelection`, and
+  database. The exception is exactly one opt-in systemd user timer:
   `axstack-single-server` removes duplicate `t3 serve` process trees for one
   base dir while keeping the ownership-lock holder, otherwise the recorded
-  server. Both run stateless scripts, keep
-  no workflow state and launch no agent; see
-  [New-chat default sync](docs/host-operations.md#new-chat-default-sync) and
+  server. It runs a stateless script, keeps
+  no workflow state and launches no agent; see
   [Single T3 server](docs/host-operations.md#single-t3-server).
 - The recorded owning watch thread merges eligible own PRs; excluded PRs use a merge card.
 - This is an early project; expect the workflows to evolve.
@@ -145,7 +143,7 @@ Eligible own PRs use guarded automatic merge; see [merge boundaries, exclusions,
 | Page | Use it for |
 | --- | --- |
 | [Getting started](docs/getting-started.md) | Install, understand `check`, and run a first bounded task. |
-| [Concepts](docs/concepts.md) | Driver threads, phases, roles, presets, holds, evidence and account selection. |
+| [Concepts](docs/concepts.md) | Driver threads, phases, roles, presets, holds, evidence and provider bindings. |
 | [Guides](docs/guides.md) | Step-by-step journeys for features, fixes, reviews, watches, debugging and releases. |
 | [Workflows](docs/workflows.md) | Routing, review policy, merge boundaries, exclusions and accepted risks. |
 | [Installation](docs/installation.md) | CLI commands, harness paths, configuration, upgrades and troubleshooting. |

package/docs/concepts.md

@@ -57,58 +57,18 @@ do not belong in it.
 
 ## T3 runtime boundary
 
-Delegation uses the `t3-code` MCP. Read-only roles use `delegate_task`; authors
-use `t3_thread_launch` with their own pinned worktrees. Configuration read-back,
+Delegation uses the `t3-code` MCP. Use async `delegate_task` for every role, including authors with
+`isolation:"worktree"` in their own SHA-pinned worktrees.
+Top-level drivers use `t3_thread_launch`.
+Use `t3_thread_launch` as a rare author fallback only when
+worktree isolation is unavailable and record each use in the run record. Configuration read-back,
 run receipts and revision-bound completion are separate from capability
 discovery. Read the [runtime contract](../skills/axstack/references/t3-runtime.md)
 immediately before dispatch, receipt consumption or recovery.
 
-## Account selection
-
-Before each Claude or Codex dispatch or launch, the driver runs
-`bun skills/axstack/scripts/pick-instance.js --provider claude|codex`.
-The picker selects the enabled account of that provider's T3 driver with the
-highest tier-weighted headroom. Save its `--json` output in private evidence
-and record the chosen instance ID. `--settings <path>` overrides
-`~/.t3/userdata/settings.json`.
-
-For known usage, the score equals `(100 − max known window utilization) × tier weight`.
-For unknown utilization, the score equals the tier weight alone.
-The maximum uses the known utilization windows; an unavailable window is not
-assumed to be empty.
-
-| Provider / tier | Weight |
-| --- | --- |
-| Claude Max 20x: `default_claude_max_20x` | 4 |
-| Claude Max 5x: `default_claude_max_5x` | 1 |
-| Codex: `pro` | 4 |
-| Codex: `prolite` | 1 |
-
-For an unknown tier, the weight equals 1.
-An account with any window at ≥95% or a reported limit reached is excluded.
-Accounts with missing credentials are skipped.
-Usage is cached for five minutes. Failed requests use stale usage when present,
-otherwise a tier-only unknown score. Codex's plan is unknown until a successful
-usage response, so an uncached failed Codex request has weight 1.
-
-Exit 0 prints the selected instance or JSON.
-For dispatched roles, only error exit 1 permits fallback to the canonical instance after validating availability.
-The canonical instances are `codex` and `claudeAgent`.
-Exit 2 (no eligible provider instances) must hold the work without fallback.
-If the driver's own current instance is eligible, it must stay despite headroom differences.
-On error exit 1, the driver must keep its current instance.
-At a turn boundary, an excluded driver with an eligible same-provider sibling
-switches its own calling thread and records the change. Provider, model, class
-and effort remain fixed. Dispatched roles never fail over mid-thread.
-See [Provider bindings](../skills/axstack/references/t3-runtime.md#preflight-and-binding)
-for schedule rebinding and the [policy carve-out](workflows.md#automatic-merge-boundaries).
-
-The opt-in `sync-default.js` rotates the New-chat default by elapsed-minus-used pace at a ten-point lead or current exclusion. `AXSTACK_SYNC_POOL` enables cross-driver weekly comparison and model/effort selection; without it, only the same-driver account changes.
-
-## Account selection environment
-
-`XDG_CACHE_HOME` selects the cache directory; the default is `~/.cache`.
-The usage cache lives at `<cache-dir>/axstack/usage.json`.
-`AXSTACK_CLAUDE_USAGE_URL` is a test-only endpoint override for Claude usage.
-`AXSTACK_CODEX_USAGE_URL` is a test-only endpoint override for Codex usage.
-Local tests use loopback endpoints; these variables are not provider setup steps.
+## Provider bindings
+
+Claude binds to T3 instance `claudeAgent`; Codex binds to `codex`.
+Grok binds to `grok`.
+A usage-limited instance holds the work without failover.
+See [Provider bindings](../skills/axstack/references/t3-runtime.md#preflight-and-binding).

package/docs/host-operations.md

@@ -7,10 +7,10 @@ instructions and local tests do not prove live provider readiness or delivery.
 
 ## T3 setup
 
-Set `worktreeCleanup` to `off` for every Axstack project before dispatch;
-Axstack preserves author worktrees and salvages evidence before retirement.
-The driver reads back this setting via `t3_project_read` where exposed, or
-records the setup limitation.
+Hosts can delete author worktrees at merge through `storageCleanup.worktreeOnMerge`.
+`t3_project_read` cannot read `storageCleanup.worktreeOnMerge`.
+Evidence never lives in a worktree and belongs in private `<run>/evidence/<key>/`.
+A missing merged or closed author worktree is expected, not a hold.
 
 For remote and Android access on the existing tailnet, run:
 
@@ -23,13 +23,6 @@ Run `t3 serve --tailscale-serve` as a VPS user service and pair the Android
 app with `t3 pair`. T3 Connect is outside this setup. Service installation,
 network access, and pairing require their own authorized host checks.
 
-Install Antigravity through T3 provider settings using its managed runtime,
-then complete the user's browser sign-in before its canary. T3 uses Google's
-Antigravity ACP agent (`agy_acp_server`); the IDE/`agy` CLI skill paths in the
-installation reference are separate installer targets and do not configure this managed runtime.
-Antigravity roles receive self-contained briefs; its runtime does not read
-`~/.agents/skills`. A missing runtime, sign-in, or canary holds those roles.
-
 Grok CLI must be >=1.0.13 on desktop and VPS. T3 advertising Grok does not
 prove the CLI runs.
 
@@ -43,10 +36,9 @@ the affected operation. Provider/model routing, Linear documents through the
 executor MCP, and live schedule behavior need separate preflights.
 
 Installation creates no production schedule and never enables a timer.
-Axstack allows exactly two opt-in systemd user timers as scheduler exceptions:
-`axstack-default-sync` rewrites only T3 `defaultModelSelection`, and
+Axstack allows exactly one opt-in systemd user timer as a scheduler exception:
 `axstack-single-server` removes duplicate server processes for one base dir.
-Both run stateless scripts, keep no workflow state and launch no agent.
+It runs a stateless script, keeps no workflow state and launches no agent.
 Every verified own-PR publication arms or joins the driver's chat-run watch.
 Its bound T3 schedule uses the cadence in [Chat-run watch activation](#chat-run-watch-activation).
 See [Chat-run PR watch](workflows.md#chat-run-pr-watch) for authority, schedule identity and stop conditions.
@@ -58,11 +50,23 @@ live behavior. See [Review manager](../skills/axstack/references/automations.md)
 
 ## Chat-run watch activation
 
-A bound T3 schedule wakes the driver thread on a normal 30-minute cadence.
+For parked holds, follow [Parked user-reply holds](../skills/axstack/references/contracts.md#parked-user-reply-holds).
+A parking unwatch never triggers native re-arm or the 5-minute recovery cadence.
+
+A bound T3 schedule wakes the driver thread on a 60-minute backstop cadence.
 Unavailable or unarmed native watches for any open member, failed re-arming,
-or a resumed question run restore five minutes. Human-only waits use 60 minutes
+or an unsettled writer or follow-up without a native completion path restore
+five minutes. A user Stop never activates this recovery. Other waits use 60 minutes
 under [Chat-run watch runtime](../skills/axstack-watch/references/watch-runtime.md#chat-run-watch).
-A T3 "stopped watching" notice triggers a PR re-check and native re-arm in that same turn.
+For a non-user stop while the PR stays open, follow
+[Native PR links and watches](../skills/axstack/references/t3-runtime.md#native-pr-links-and-watches)
+in that same turn: unsettle the driver thread before native re-arm.
+User Stop stays stopped; the linked procedure owns stop-reason recording,
+diagnostic rate-limit checks and recovery on the next wake after reset.
+Never wait for a GraphQL reset inside a turn. After one PR merge auto-settles a
+multi-PR driver, re-arm the remaining watches through that procedure.
+The digest runs only on schedule wakes; native PR wakes reconcile their event
+and a current-state read.
 The run record holds the schedule ID and driver thread.
 Each wake reconciles all unsettled dispatch attempts
 and runs the own-PR maintenance loop: feedback, base movement, required CI,
@@ -103,7 +107,7 @@ binding. Continuity lives outside worktrees at
 `~/.local/share/axstack/runs/review-manager/progress.md`. Per-PR detached
 review checkouts come from existing host clones; a missing clone holds that job.
 At pass start, follow [Provider bindings](../skills/axstack/references/t3-runtime.md#preflight-and-binding)
-for account selection and schedule recreation.
+for provider binding.
 
 Every pass reconciles saved, GitHub, and native T3 state across the lane before
 admission and reads all discovery pages. Incomplete inventory or unknown
@@ -130,9 +134,8 @@ retirement. No broad scratch deletion or forced worktree removal applies.
 Manual review and user-driven `axstack-watch` remain outside this schedule.
 Requested peer reviews cover any accessible repository. T3 owns schedules,
 threads, runs, and delegated tasks; Axstack adds no queue engine, cursor files,
-or historical runtime fallback. The only scheduler exceptions are the two opt-in
-systemd user timers described in [New-chat default sync](#new-chat-default-sync)
-and [Single T3 server](#single-t3-server).
+or historical runtime fallback. The only scheduler exception is the opt-in
+systemd user timer described in [Single T3 server](#single-t3-server).
 
 ## Review automation
 
@@ -155,84 +158,6 @@ Scenario fixtures are behavioral-evaluation inputs, not model-evaluation
 results, and neither form is live proof; activation still requires the native
 canary described by the operational contract.
 
-## New-chat default sync
-
-This opt-in script changes only root `defaultModelSelection` in
-`~/.t3/userdata/settings.json`. Without a pool it changes only the instance
-among enabled accounts of the current Claude/Codex driver.
-`AXSTACK_SYNC_POOL` compares drivers by weekly clocks (>=604800s), without
-plan weights. Weekly remaining quota is `R = 100 - used` for the selected
-weekly window. Prefer the highest band: `R >= 15`, `10 <= R < 15`, then `R < 10`.
-A higher band switches without a pace margin; within the same band, rotation
-needs a ten-point elapsed-minus-used pace lead and the current account wins ties.
-This also applies when every eligible account is below 10% remaining: known
-usable quota stays usable, ranked by pace. Without a pool, only pace applies.
-This is a best-effort weekly reserve preference, not a guaranteed budget for
-future work. The current hard guard still excludes any window at 95% used
-or a reached limit. Excluded/unauthorized/skipped current accounts can escape
-to known candidates before the unknown hold; otherwise any unknown holds,
-even below 10% remaining. No eligible candidate leaves the default untouched.
-Non-current absent credentials
-exclude; malformed/unreadable credentials and unknown/stale usage hold.
-Cache expires after five minutes. Missing defaults stay untouched; OAuth tokens
-are never refreshed. Each host's pool uses only enabled, credentialed drivers:
-
-```sh
-AXSTACK_SYNC_POOL=claudeAgent=claude-opus-5-5,codex=gpt-6-astra:medium \
-  bun ~/.agents/skills/axstack/scripts/sync-default.js --dry-run
-```
-
-Exit 0 prints one token-free JSON decision with proposed selection; dry-run
-never writes. `--settings <path>` selects another file. Pool entries are
-`driver=model[:effort]`: Claude `effort` or Codex `reasoningEffort` options;
-omitted effort omits options. Rotation writes the full pool selection,
-overwriting manual model/effort picks only then. Pin with `AXSTACK_SYNC_DEFAULT=0`
-or a default driver outside the pool. Exit 1 means invalid pool/settings or write
-failure. Save the full selection for rollback. Other bytes stay intact.
-A fsynced sibling preserves mode; original bytes are checked before rename and
-read back afterward. The accepted sub-ms T3 write race can lose an edit;
-read-back mismatch is an error.
-
-Validate pool models and efforts against each host's catalog before activation.
-Preserve that host's current Opus effort by including it explicitly when set;
-the illustrative Claude entry above omits effort. If the current Codex default
-still uses Sol, the driver performs a one-time same-account selection migration
-to `gpt-6-astra:medium` during authorized host activation, saving rollback bytes
-and verifying the write/readback. The helper does not enforce models when the
-account stays selected. Only new chats use the changed default; running threads,
-role routing, schedules and `roles.json` stay outside this policy.
-
-Under host-mutation authority, install and enable:
-
-```sh
-mkdir -p ~/.config/systemd/user
-cp ~/.agents/skills/axstack/systemd/axstack-default-sync.service ~/.config/systemd/user/axstack-default-sync.service
-cp ~/.agents/skills/axstack/systemd/axstack-default-sync.timer ~/.config/systemd/user/axstack-default-sync.timer
-systemctl --user daemon-reload
-systemctl --user enable --now axstack-default-sync.timer
-systemctl --user list-timers axstack-default-sync.timer
-journalctl --user -u axstack-default-sync.service
-```
-
-The timer starts about one minute after user-manager startup, then every ten
-minutes. Its monotonic `Persistent=true` timer never replays calendar ticks.
-Override Bun/skills paths via `ExecStart`/`Environment` in
-`systemctl --user edit axstack-default-sync.service`. Add the per-host pool:
-
-```ini
-[Service]
-Environment=AXSTACK_SYNC_POOL=claudeAgent=claude-opus-5-5,codex=gpt-6-astra:medium
-```
-
-Reload after edits. Logout needs authorized linger; installation never activates units.
-
-Disable with `systemctl --user disable --now axstack-default-sync.timer` and
-stop any active one-shot with `systemctl --user stop axstack-default-sync.service`.
-Rollback: remove the pool drop-in, reload, and restore the full selection in T3 Settings.
-The sibling `<settings>.axstack-default-sync.lock` directory prevents overlap
-(unchanged/locked). Empty locks older than sixty seconds are removed and retried
-once; fresh contention reports locked.
-
 ## Single T3 server
 
 This opt-in timer checks once per minute for this user's `t3 serve` processes

package/docs/installation.md

@@ -81,23 +81,17 @@ independently reviewed through the selected preset's explanation roles.
 Repeat installs leave unchanged files untouched. Legacy tool copies are
 released by ownership on upgrade and uninstall; retained copies report why.
 
-See [Account selection environment](concepts.md#account-selection-environment)
-for the picker cache directory and test-only usage endpoint variables.
-
 ## Packaged scripts
 
 Run these with Bun from the installed `<skills-dir>/axstack/scripts/` directory:
-`resolve-models.js` (model IDs), `pick-instance.js` (dispatch account),
-`pr-digest.js` (PR discovery), `archive-evidence.js` (evidence archive), and
-`sync-default.js [--settings <path>] [--dry-run]` (New-chat default; `AXSTACK_SYNC_POOL` opts into cross-driver rotation).
+`resolve-models.js` (model IDs), `pr-digest.js` (PR discovery), and
+`archive-evidence.js` (evidence archive).
 `single-server.js [--base-dir <dir>] [--proc-root <dir>] [--grace-seconds N] [--dry-run]` removes
 duplicate `t3 serve` process trees while keeping the ownership-lock holder,
 otherwise the recorded runtime pid.
-`account-usage.js` shares usage retrieval and the five-minute cache between
-picker and sync. The opt-in systemd user unit files ship in `axstack/systemd/`;
-installation copies them as assets and never enables either timer.
-See [New-chat default sync](host-operations.md#new-chat-default-sync) and
-[Single T3 server](host-operations.md#single-t3-server) for activation.
+The opt-in systemd user unit files ship in `axstack/systemd/`;
+installation copies them as assets and never enables the timer.
+See [Single T3 server](host-operations.md#single-t3-server) for activation.
 
 ## Role presets
 
@@ -219,9 +213,13 @@ The deterministic `<!-- axstack:begin v1 -->` / `<!-- axstack:end -->` block
 contains the target-derived Axstack entry path and model-free routing prose. It
 also routes every subagent, delegated worker, reviewer, and cross-harness dispatch
 through the `t3-code` MCP. It forbids harness-native subagent tools. The current
-T3 thread is the driver; read-only roles use async `delegate_task`, while authors
-use `t3_thread_launch` in their own SHA-pinned worktrees. The standing block
-authorizes writer launch; it grants no push, merge, release, or host mutation.
+T3 thread is the driver. Use async `delegate_task` for every role, including authors with
+`isolation:"worktree"` in their own SHA-pinned worktrees.
+Top-level drivers use `t3_thread_launch`.
+Use `t3_thread_launch` as a rare author fallback only when
+worktree isolation is unavailable and record each use in the run record. The standing block
+authorizes top-level driver threads, delegated writers and worktrees;
+it grants no push, merge, release, or host mutation.
 The [runtime reference](../skills/axstack/references/t3-runtime.md) owns the dispatch
 and receipt protocol.
 
@@ -249,12 +247,11 @@ The complete bundle is validated before writes:
 - each preset is a real JSON file with version 1, a non-empty `roles` array,
   the filename's selected identity supplied by the caller, and the same role-ID
   set as its peers;
-- every role has valid preserved fields, while the mixed checker,
-  `axstack-research-web-google`, `axstack-research-x`, and both arena candidate launch-by-agent-id
-  routes explicitly permit `model: null`;
+- every role has valid preserved fields; `axstack-research-x` and the Grok
+  arena candidate explicitly permit `model: null`;
   in each single-provider preset, the unavailable adviser and round-2 seat
-  explicitly permit `model: null`, as do both cross-provider research routes
-  and both arena candidate seats;
+  explicitly permit `model: null`, as do the X research route
+  and the Grok arena candidate;
 - obsolete runtime configuration flags fail before mutation with migration
   guidance.
 
@@ -290,30 +287,23 @@ the user explicitly changes it and accepts the resulting evidence invalidation.
 The `mixed` and `claude-only` presets assign `axstack-auditor`,
 `axstack-research-requirements`, `axstack-research-code`, `axstack-research-web`,
 `axstack-explore-execution` to the Claude Sonnet class at high effort.
-`axstack-explore-codebase` and `axstack-monitor` adopt Haiku 5.5 high with
+`axstack-explore-codebase` adopts Haiku 5.5 high with
 `modelClass: haiku` and the exact `model: claude-haiku-5-5` pin.
 Activation requires that exact ID and high effort in saved T3 capabilities;
-missing either holds these roles without substitution, including Haiku 4.5.
+missing either holds this role without substitution, including Haiku 4.5.
 The `codex-only` assignments for these roles are unchanged.
 Scout adds `axstack-explore-codebase-luna` at Codex Luna xhigh in `mixed` and
-`codex-only`. The optional `axstack-explore-codebase-flash` seat uses
-Antigravity `gemini-3.8-flash-medium` at medium effort in `mixed`.
-The missing Scout seats have explicit null models: Luna and Flash in
-`claude-only`, and Flash in `codex-only`.
+`codex-only`. Luna has an explicit null model in `claude-only`.
 The three `-sol` pair seats for auditor, research-code, and explore-execution
 use Sol high in `mixed` and `codex-only`; `claude-only` records intentional
 absences. The paired seats run independently on one brief and the driver
 reconciles their findings.
 
-The mixed checker and `axstack-research-web-google` have provider
-`antigravity`; mixed `axstack-research-x` has provider `grok`. All three use
-`model: null` with notes authorizing their agent-ID routes; T3 resolves Grok's
-exact model from the first entry for that provider in saved capabilities.
-For Antigravity `model:null` roles, select the first listed model whose ID ends
-with `-<effort>` from saved capabilities and record its exact ID.
-Missing effort-suffix matches hold resolution for Antigravity, including empty
-model catalogs. The single-provider presets configure the checker and keep
-both cross-provider research routes as intentional absences. Their
+The mixed checker uses the Claude Sonnet class at low effort, resolved from
+saved capabilities. Mixed `axstack-research-x` uses `grok` with
+`model: null`; T3 resolves its exact model from the first entry for that provider
+in saved capabilities. The single-provider presets configure the checker and
+keep X research as an intentional absence. Their
 unavailable adviser and round-2 seat remain explicit same-provider
 `model: null` roles, which do not make installation unready;
 Align and Spec still hold until both Astra and Opus can return independent
@@ -326,7 +316,7 @@ the affected work. Codex and Claude class resolution reads the saved T3 capabili
 the command below; missing or malformed catalogs hold.
 
 ```text
-bun skills/axstack/scripts/resolve-models.js --provider <codex|claude|grok|antigravity> --capabilities <saved-json> <--class <class>|--model <id|null>> --effort <level> [--exclude <id>]
+bun skills/axstack/scripts/resolve-models.js --provider <codex|claude|grok> --capabilities <saved-json> <--class <class>|--model <id|null>> --effort <level> [--exclude <id>]
 ```
 
 `--class` resolves a class; `--model` selects an exact ID or an explicitly
@@ -336,11 +326,10 @@ and reports a resolution hold on exit 1.
 
 A preset model is used as given; class rows resolve to the newest
 matching catalog ID. Resume retains the recorded snapshot without re-resolution.
-Rejection, timeout, quota, and auth failures hold; outside bounded same-provider,
-same-model account selection among one driver's instances via `pick-instance.js`,
+Rejection, timeout, quota, and auth failures hold;
 no subscription inference, quota routing, or alternative-model retry applies.
 Follow [Provider bindings](../skills/axstack/references/t3-runtime.md#preflight-and-binding)
-for driver account re-selection at turn boundaries and schedule rebinding.
+for fixed provider instances and usage-limit holds.
 
 `modeId` and similar permission fields remain conservative declared intent.
 They do not prove effective T3 `runtimeMode`, sandboxing, or permission parity.
@@ -356,20 +345,6 @@ installed shared-root path. The [T3 runtime reference](../skills/axstack/referen
 owns model/account authority; [Cleanup](../skills/axstack-cleanup/SKILL.md)
 owns evidence retirement authority.
 
-### Account picker
-
-```text
-bun skills/axstack/scripts/pick-instance.js --provider <claude|codex> [--settings <t3-settings-json>] [--json]
-```
-
-`--settings` overrides the default `~/.t3/userdata/settings.json`; `--json`
-prints the selected instance and scored inventory rather than only its ID.
-The picker uses exit 0 for an eligible selection, exit 1 for invalid input/settings
-or an unexpected failure, and exit 2 when no eligible provider instance remains.
-Follow [Provider bindings](../skills/axstack/references/t3-runtime.md#preflight-and-binding)
-for dispatch fallback, driver stay rules, and holds; selection alone grants no
-provider or model substitution authority.
-
 ### PR digest
 
 ```text

package/docs/workflows.md

@@ -15,10 +15,9 @@ The directly invoked phase loads the applicable shared references for routing,
 lifecycle, T3 runtime boundaries, role/model/risk contracts, the run record,
 and PR shape.
 
-For the picker score, account eligibility, dispatch exits and driver stay rule,
-see [Account selection](concepts.md#account-selection). Follow
+Follow
 [Provider bindings](../skills/axstack/references/t3-runtime.md#preflight-and-binding)
-immediately before dispatch or account re-selection.
+immediately before dispatch.
 
 Direct routes need no spec ceremony:
 
@@ -132,15 +131,13 @@ role.
 
 In `mixed` and `claude-only`, `axstack-research-requirements`,
 `axstack-research-code`, `axstack-research-web`, and `axstack-explore-execution`
-use the Claude Sonnet class at high effort. `axstack-explore-codebase` and
-`axstack-monitor` adopt Haiku 5.5 high, with class `haiku` and exact pin
+use the Claude Sonnet class at high effort. `axstack-explore-codebase` adopts
+Haiku 5.5 high, with class `haiku` and exact pin
 `claude-haiku-5-5`. Missing that ID or high effort in saved T3 capabilities
-holds those roles without substitution. `codex-only` keeps its Codex
+holds that role without substitution. `codex-only` keeps its Codex
 assignments for those roles. Scout adds a Luna xhigh codebase seat in `mixed`
-and `codex-only`, and an optional Antigravity Flash medium seat in `mixed`.
-In `claude-only`, Luna and Flash are intentionally absent; Flash is also
-intentionally absent in `codex-only`. Mixed web-google and X retain their
-source-specific Antigravity and Grok routes.
+and `codex-only`. In `claude-only`, Luna is intentionally absent.
+The mixed X route uses Grok.
 The mixed `axstack-auditor` uses Haiku 5.5 xhigh, pinned to `claude-haiku-5-5`;
 missing that ID or xhigh effort holds without substitution. The mixed
 `axstack-explainer-review` uses Sol medium; `axstack-explainer` pins
@@ -171,15 +168,14 @@ in codebase findings mode.
 The unavailable adviser in each single-provider preset stays explicitly
 `model: null` within that provider's bounds. Installer readiness accepts that
 intentional absence, but Align and Spec hold because both independent receipts
-are required. The mixed checker and Google web-research route use provider
-`antigravity`; the X route uses `grok`. Those agent-ID routes retain
-`model: null` notes and resolve the exact model from the first provider entry
-in saved T3 capabilities. Empty Antigravity catalogs hold. Missing or
+are required. The mixed checker uses the Claude Sonnet class at low effort,
+resolved from saved capabilities. The X route
+uses `grok` with `model: null` and resolves the exact model from the first
+provider entry in saved T3 capabilities. Missing or
 unavailable roles hold only affected work. Requested model, effort, and
 permission values need actual T3 configuration read-back; stored `modeId` is
 neither permission parity nor a sandbox. Rejection, timeout, quota, and auth
-failures hold; outside bounded same-provider, same-model account selection among
-one driver's instances via `pick-instance.js`, no subscription inference, quota
+failures hold; no subscription inference, quota
 routing, or alternative retry applies.
 
 ## T3 runtime boundary
@@ -193,9 +189,13 @@ Capability discovery alone is not execution proof.
 
 All subagent, delegated-worker, reviewer, and cross-harness work uses T3
 orchestration through the `t3-code` MCP. Do not use harness-native subagent
-tools. Read-only roles use async `delegate_task`; reviewers and investigators
-receive driver-made disposable detached checkouts pinned to candidate and base.
-Authors use `t3_thread_launch` in their own SHA-pinned worktrees. Repairs return
+tools. Use async `delegate_task` for every role, including authors with
+`isolation:"worktree"` in their own SHA-pinned worktrees.
+Top-level drivers use `t3_thread_launch`.
+Use `t3_thread_launch` as a rare author fallback only when
+worktree isolation is unavailable and record each use in the run record.
+Reviewers and investigators receive driver-made disposable detached checkouts pinned to candidate
+and base. Repairs return
 to the same author and worktree. The current T3 driver owns coordination and
 forge mutations and never writes an author's tracked files.
 
@@ -203,10 +203,23 @@ Preserve native `taskId/childThreadId/childRunId` or
 `threadId/runId/worktree/branch/base SHA`, dispatch key, requested/effective
 configuration, and private evidence. Persist `task_status` before
 `t3_thread_read`; delegated completion needs terminal success, available result,
-settled child runs, and the current `AXSTACK-DONE` marker. Launched writers
-send their marker to the driver, which also verifies terminal `t3_thread_wait`,
+settled child runs, and the current `AXSTACK-DONE` marker. Delegated authors also
+require a clean tree, non-empty diff and red/green logs.
+Launched fallback writers send their marker to the driver, which also verifies
+terminal `t3_thread_wait`,
 a clean tree, non-empty diff, and red/green logs. An older attempt never
-completes a newer one. Questions remain incomplete until the resumed run settles.
+completes a newer one. Questions remain incomplete until the answered task or resumed run settles.
+
+For a delegated child, detect questions with `t3_pending_request_list(childThreadId)`.
+Never rely on `hasPendingQuestion` alone.
+Codex `request_user_input_async` does not set it.
+Use `t3_pending_request_read`, then confirm the request is still pending with
+`t3_pending_request_list(childThreadId)`, then answer once within existing authority
+using `t3_pending_request_respond`; a second ask holds.
+Independent review rounds require fresh `delegate_task` calls.
+Only author continuation and an answer to a child's own question use the same child.
+Providers without a question tool (grok) and launched writers keep
+the `AXSTACK-QUESTION` marker route.
 
 Trust, permission, authentication, and provider safety prompts are holds;
 never answer trust or permission prompts for a worker. Unknown liveness,
@@ -223,8 +236,8 @@ for explicit ownership transfer.
   Rung 1 or 2 design question it loads `axstack-brainstorm` inline and reuses
   its receipt instead of consulting twice; Align owns the interview.
 - `axstack-brainstorm` validates an approach standalone or inline in the driver,
-  report-only. Every invocation compares independent Astra, Opus, Grok and
-  Antigravity candidates, including premise and smallest-change/do-nothing
+  report-only. Every invocation compares independent Astra, Opus and Grok
+  candidates, including premise and smallest-change/do-nothing
   checks. The driver scores, picks and grafts; Rung 1 uses no judges. At Rung 2,
   `axstack-arena-judge-opus` judges round 1; disagreement on the base or caller
   re-invocation with the user's rejection triggers fresh Fable/Astra round 2.
@@ -310,18 +323,19 @@ keeps the current owner and a resumable record.
 
 ## Decisions and holds
 
+For parked holds, follow [Parked user-reply holds](../skills/axstack/references/contracts.md#parked-user-reply-holds).
+
 Raise user-decision holds and serious-risk holds in the T3 driver thread,
-and the driver ends its turn parked on the decision.
+and the driver ends its turn waiting on the decision.
 These include `verification not proven`, merge cards needing user input,
 spec approval and npm stage approval.
 Raise serious risks immediately; attempt bounded safe recovery for blocked operations.
 After bounded safe recovery, raise a genuine blocked operation that still needs
-the user in the T3 driver thread, and the driver ends its turn parked on it.
+the user in the T3 driver thread, and the driver ends its turn waiting on it.
 Routine questions stay in the T3 driver thread. Progress, CI pending, and completion always stay
 in the T3 driver thread.
 Post the report-only unresolved-rating notice there, deduplicated per comment and
 head. It grants no reply authority and never clears the hold.
-The standalone monitor never sends; the chat-run schedule resumes the driver.
 Healthy watch observations remain quiet.
 
 ## Chat-run PR watch
@@ -353,7 +367,8 @@ An incomplete scan leaves readiness `UNKNOWN`.
 The chat-run watch never expires or waits for re-authorization while PRs remain open.
 The chat-run watch ends only when all watched PRs merge or close, launched work
 is settled, and release is settled or not applicable, or the user cancels.
-A required PR closed without merging keeps its decision hold and wake.
+A required PR closed without merging keeps its decision hold.
+For a required PR closed without merging, apply parking eligibility.
 Follow [Chat-run watch runtime](../skills/axstack-watch/references/watch-runtime.md#chat-run-watch)
 for native lifetime re-arming and quiet cadence changes on the recorded schedule ID.
 Implementation candidates are published and read back before independent
@@ -417,7 +432,7 @@ review, or return the work to the author. Only a gap the driver cannot close
 becomes a `verification not proven` card bound to head and base.
 The card names what is unproven, why, and the action that would settle it.
 
-Raise `verification not proven` in the T3 driver thread and end the turn parked
+Raise `verification not proven` in the T3 driver thread and end the turn waiting
 on the decision, bound to the exact PR head and base.
 A changed PR head or base must supersede its open verification decision card,
 reissuing only if a decision is still needed.
@@ -515,7 +530,7 @@ those SHAs, counted collaborator approvals and bot votes with each vote's SHA
 and stale flag, and the causes holding this merge.
 A card that needs the user's decision (work promotion or work deploying/unknown base,
 `Auto-merge: off`, provenance, or an unclosable verification gap) is a decision
-hold raised in the T3 driver thread; the driver ends its turn parked on the decision.
+hold raised in the T3 driver thread; the driver ends its turn waiting on the decision.
 Revalidate the current head and base; the reply clears only its bound cause.
 Person-wait cards follow the clock below.
 A changed head or base requires a refreshed card.
@@ -559,10 +574,8 @@ Npm publication still requires human stage approval; agents never run
 This policy grants no release, npm publish, or host install
 authority. See [Preview authority and operations](host-operations.md#private-pr-previews).
 
-Excluded: CLI proxy, account pooling behind a proxy or shared session, and IP
-routing; local CI contention handling is deferred. Quota-driven scheduling or
-model routing (provider/model substitution) is excluded. Per-dispatch selection
-among the user's own same-provider, same-model accounts is permitted. Automatic
+Excluded: changes to proxy or hub infrastructure and IP routing; local CI contention handling is deferred. Quota-driven scheduling or
+model routing (provider/model substitution) is excluded. Automatic
 merge of work promotion, deploying-base and unknown-base PRs is excluded.
 Automatic merge of peer PRs is excluded. Previews
 outside the VPS, public previews, and production data are excluded. Nightly triage
@@ -571,7 +584,7 @@ never sends user messages.
 Accepted risks: two agents can miss the same defect while CI is green; spec
 approval is the user's main checkpoint. A head guard does not atomically guard
 base freshness; the concurrent-merge race is held by the post-merge push-failure
-rule. A watch with a 30-minute fallback, 5-minute recovery, and 60-minute human wait
+rule. A watch with a 60-minute backstop and 5-minute recovery
 until PRs land has an accepted token cost. Preview code runs under the same VPS user as agents and is not isolated;
 tests already do, so the added risk is small.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "axstack",
-  "version": "0.49.0",
+  "version": "0.50.0",
   "description": "Axstack installer and setup CLI: installs owned chat skills and role data, configures supported harness settings, and checks T3 Code capabilities.",
   "keywords": [
     "claude-code",