@codetreker/daynight 0.21.1 → 0.21.3

package/README.md

@@ -0,0 +1,1153 @@
+# Daynight CLI
+
+The `@codetreker/daynight` package provides the `daynight` command and its `dn`
+alias. Use it to manage server-backed projects, tasks and attachments, install
+the `using-daynight` skill for coding agents, and run a local GitHub PR reviewer
+through Copilot CLI. Both command names share configuration and behavior.
+
+See the [project overview](../../README.md) for the browser workspace and service,
+and the [development guide](../../docs/development.md) for source builds,
+validation, package versions and npm publication. Each command's `--help` owns
+its exact usage. Review execution additionally requires Git, an authenticated
+GitHub CLI and Copilot CLI on `PATH`.
+
+- [Install and update](#install-and-update)
+- [Client commands](#client-commands)
+- [Agent skill installation](#agent-skill-installation)
+- [Shell completion](#shell-completion)
+- [Configuration setup](#configuration-setup)
+- [Project management](#project-management)
+- [Task management](#task-management)
+- [Task deletion](#task-deletion)
+- [Task attachments](#task-attachments)
+- [Review configuration and startup](#review-configuration-and-startup)
+- [Review status listing](#review-status-listing)
+- [Durable review state](#durable-review-state)
+- [GitHub observations](#github-observations)
+- [Repository clones and worktrees](#repository-clones-and-worktrees)
+- [Copilot execution](#copilot-execution)
+- [Coordination and shutdown](#coordination-and-shutdown)
+
+## Install and update
+
+Use Node.js 24 LTS. Install the CLI with npm:
+
+```bash
+npm install --global @codetreker/daynight
+dn --version
+```
+
+The installation provides both `daynight` and `dn`; every command works with
+either name, for example `dn task list` or `dn skill install`. Help and diagnostics
+use the canonical `daynight` spelling.
+
+Update that npm global installation with:
+
+```bash
+dn update
+```
+
+The command checks `@codetreker/daynight`'s `latest` version in the configured npm
+registry and installs that exact version only when it is newer. It updates the
+installation that is running, even if npm's current default global prefix has
+changed. An equal or newer installed version succeeds without installation or
+downgrade. User and global npm registry/authentication settings apply; the current
+working directory's project `.npmrc` does not select the update source.
+
+`daynight update --help` explains the command. Self-update requires an npm-owned
+global package and launcher. Source checkouts, local dependencies, `npx`, pnpm and
+`npm link` installations receive guidance to use their original package manager.
+If an update is needed, run the command from outside its installation directory.
+The installer streams its output, enforces the target's Node.js engine requirement
+and reports permission, network and installation errors. It does not request
+administrator privileges automatically. npm owns replacement and recovery; a
+failed installation is not a promise that every original file remains unchanged.
+
+The command verifies the installed manifest, launcher and a fresh invocation's
+version, then refreshes existing `using-daynight` skills from that installed
+package. It checks User destinations and the current Git worktree's Project
+destinations, including calls from subdirectories and linked worktrees. Outside
+Git it checks User destinations only. It does not search other projects or need
+agent launchers on PATH. Missing skills remain uninstalled. The same refresh
+runs when the CLI is already current or ahead, using its installed bundle.
+
+Refresh replaces complete skill directories, including local edits and obsolete
+files. Shared physical destinations are written once. Each result identifies
+its agents, scope, path and package version. A skill failure exits nonzero and
+reports the package outcome separately; a successfully installed npm package is
+retained, and rerunning `update` retries the skill refresh. All phases must
+succeed before the command reports success. It requires no review configuration and does
+not open SQLite, modify sessions or restart a running reviewer. The next CLI
+invocation uses the installed version. Restart any running reviewer to use it;
+on Windows, a reviewer holding a native addon open can prevent npm from replacing
+that file and cause the update to fail. The
+[self-update decision](../../.agents/notes/implemented/feature/2026-10-20-cli-self-update.md)
+records installation ownership and verification requirements.
+
+## Client commands
+
+```text
+daynight                            # Show root help
+daynight --help                      # Also -h
+daynight --version                   # Also -v
+daynight update
+daynight login [--no-browser]         # Authorize CLI access through the browser
+daynight logout                      # Revoke and remove the current CLI credential
+daynight review                     # Show review help
+daynight review run
+daynight review list                 # Also review ls
+daynight review add-repo <remote> [--path <path>]
+daynight completion                 # Show completion help
+daynight completion bash
+daynight completion zsh
+daynight completion setup [--shell bash|zsh] [--rc-file <path>]
+daynight skill                      # Show skill help
+daynight skill install              # Select Agents, then User or Project scope
+daynight project
+daynight project create <name> [--repo <repo>]
+daynight project list [--json]
+daynight project show <project-id> [--json]
+daynight task
+daynight task create --title <title> [options]
+daynight task list [--all | --status <status>] [--ready] [--json] [--project <project-id>]
+daynight task ready [--limit <N>] [--json] [--project <project-id>]
+daynight task show <task-id> [--json] [--project <project-id>]
+daynight task update <task-id> [options]
+daynight task set-status <task-id> <status> [--if-version <version>] [--project <project-id>]
+daynight task delete <task-id> [--remove-depends] [--if-version <version>] [--project <project-id>]
+daynight task graph [--json] [--project <project-id>]
+daynight task attachment
+daynight task attachment put <task-id> <file> [--path <relative-path>] [--project <project-id>]
+daynight task attachment list <task-id> [--tree | --json] [--project <project-id>]
+daynight task attachment read <task-id> <relative-path> [--project <project-id>]
+daynight task attachment remove <task-id> <relative-path> [--project <project-id>]
+```
+
+Groups and operations show help when `--help` or `-h` is their sole argument.
+Help, version and completion generation do not resolve the user home, load review
+configuration, open SQLite or require review tools. `review add-repo` accepts one
+`--path <path>` before or after its remote; `--path=value` and `--` are invalid.
+Unknown commands and invalid arguments exit with status 2; operation failures
+exit with status 1 and retain their cause. The old `--daemon-review` and root
+`add-repo` forms are invalid. The
+[client design](../../docs/design/daynight.md) owns command composition and lazy loading;
+the [composition decision](../../.agents/notes/implemented/architecture/2026-10-20-client-command-composition.md)
+records module ownership and the preserved data layout.
+
+## Agent skill installation
+
+Install the public `using-daynight` skill through an interactive terminal:
+
+```bash
+daynight skill install
+```
+
+The command detects available codex, copilot and claude launchers without
+starting them or checking login. Select one or more detected Agents with Space,
+then Enter; the initial selection is empty. Next select User or Project,
+with User initially selected. Installation starts after the scope is submitted.
+Repeating the command refreshes the full bundle from the running package.
+`daynight update` automatically refreshes existing User and current-worktree
+Project copies with the same replacement semantics; it also repairs existing
+skill directories whose entrypoint was removed and replaces dangling leaf links.
+A regular file at the expected skill-directory path is an error. See
+[CLI updates](#install-and-update) for package ownership and failure handling.
+
+| Agent | User destination | Project destination |
+|---|---|---|
+| Codex | `~/.agents/skills/using-daynight/` | `<worktree>/.agents/skills/using-daynight/` |
+| Copilot CLI | `~/.agents/skills/using-daynight/` | `<worktree>/.agents/skills/using-daynight/` |
+| Claude Code | `~/.claude/skills/using-daynight/` | `<worktree>/.claude/skills/using-daynight/` |
+
+User scope works outside Git. Project scope uses the current Git worktree root,
+including calls from subdirectories or linked worktrees, without requiring an
+origin, server configuration or project association. Installing the skill and
+associating a project remain separate operations. Terminal stdin and stderr are
+required; prompts use stderr and installation results use stdout. Help and
+installation do not depend on review configuration or credentials.
+
+Each installation replaces the complete using-daynight leaf, including local
+edits and obsolete files. It then removes the legacy daynight leaf in the same
+selected physical skill root. The old name is retained until publication
+succeeds. Other skills, opposite-scope copies and parent-directory links are
+preserved; leaf links are replaced or removed without following their targets.
+Several Agent paths resolving to one physical root share one write. Destructive
+nesting involving an installation or legacy-cleanup path is rejected before
+writing. Discovery paths may be shared with unselected Agents, and host precedence
+continues to govern copies at other scopes.
+
+The command reports per-Agent destinations and results. Failure stops later
+targets, retains earlier completed installations and exits nonzero. If legacy
+cleanup fails after publication, the new bundle stays installed and the command
+reports the incomplete migration. Replacement failure after removing an existing
+using-daynight leaf may leave that destination absent; diagnostics preserve the
+actual phase and reason. Rerunning installs the complete packaged bundle.
+Cancellation before publication cleans its stage and newly created empty parents.
+
+Codex normally discovers changed skills automatically; restart its session if
+needed. For Copilot CLI run `/skills reload` and inspect
+`/skills info using-daynight`. Claude Code watches existing roots; use
+`/reload-skills` when a top-level skills directory was created during the session.
+The [skill design](../../docs/design/daynight-agent-skill.md) owns mechanisms, routing
+and the simulated-launcher native installation matrix.
+
+## Shell completion
+
+Bash 4.2 or newer and zsh have native completion scripts. Install startup
+configuration with:
+
+```bash
+daynight completion setup
+```
+
+Setup detects the calling shell from its direct parent process. If it cannot
+identify Bash or zsh reliably, select one explicitly:
+
+```bash
+daynight completion setup --shell bash
+daynight completion setup --shell zsh
+```
+
+The login-shell preference in `SHELL` does not establish which shell is running.
+For a custom startup file, use `--rc-file`; only that file is managed:
+
+```bash
+daynight completion setup --shell bash --rc-file /path/to/bashrc
+```
+
+Bash setup manages `.bashrc` and the first readable existing login file in the
+order `.bash_profile`, `.bash_login`, `.profile`. It creates `.bash_profile` only
+when none of those files exists; if existing profiles are all unreadable, it
+reports an error. The login block runs only in interactive Bash and can safely
+appear in a `.profile` shared with other POSIX shells. It does not
+make the login shell load the rest of `.bashrc`.
+
+zsh setup uses `$ZDOTDIR/.zshrc`, or `$HOME/.zshrc` when `ZDOTDIR` is unset.
+An exported empty `ZDOTDIR` selects `/.zshrc`; setup resolves a relative value
+from its current working directory. A later zsh started in another directory
+resolves that relative value from its own startup directory. A value set only
+inside zsh is not visible to the installer; use `--rc-file` in that case.
+The managed block reuses an initialized completion system or runs `compinit`
+with its normal security checks before loading daynight.
+
+Setup recognizes its managed block and the documented standalone manual loading
+statement. Repeating it leaves an installed file unchanged. A managed block
+checks that both `daynight` and `dn` use the expected callable completion
+function before skipping loading. A missing or replaced registration for either
+command reloads the shared script and restores both commands. Complete
+registrations are reused, including when startup files load each other. Preserved
+manual loading statements keep their unconditional behavior: if one runs after a managed block, it generates and loads the script
+again. Existing content, line endings, permissions and startup-file symbolic
+links are retained. Added blocks use LF line endings because a carriage return
+is shell token data. Malformed managed blocks, invalid targets and file access failures produce
+explicit errors. Setup does not migrate blocks from earlier templates. If an
+earlier block is reported as unrecognized, remove its marked section and rerun
+setup, or use the source command below to activate the current script directly.
+
+Setup reports the startup files it handled. It installs persistent configuration;
+an ordinary CLI child process cannot inspect or change its parent shell's loaded
+completion functions. Use the existing loading command to enable completion in
+the current session:
+
+```bash
+source <(daynight completion bash)
+```
+
+For zsh, initialize its completion system first if the session has not already
+done so:
+
+```zsh
+autoload -Uz compinit
+compinit
+source <(daynight completion zsh)
+```
+
+The `bash` and `zsh` generation operations write only scripts to stdout and do
+not change startup files. Loading the zsh script without initialized `compinit`
+fails with a diagnostic. The outputs are designed for `source`; using the same
+output directly as an `fpath` autoload file is outside the loading contract.
+Setup checks file contents without executing them. Custom early returns or later
+commands that remove registrations can prevent a startup file from enabling
+completion; verify such configuration by starting the shell.
+
+The generated scripts register completion for both `daynight` and `dn`.
+Completion suggests commands, nested operations, aliases, options, shell names
+and local directories for `review add-repo --path`.
+zsh also displays descriptions.
+The scripts run in the shell on each Tab without starting Node or reading review
+configuration, SQLite or the network. Directory suggestions preserve quoted and
+special-character paths; an uncreated target directory remains valid input.
+
+New shells generate definitions from the currently installed package. After an
+update, run the applicable `source` command again to refresh an existing shell.
+A saved generated script must be regenerated after updating. The
+[completion design](../../docs/design/daynight-completion.md) explains the shared grammar
+and shell integration. The
+[completion decision](../../.agents/notes/implemented/feature/2026-10-20-shell-completion.md)
+records compatibility and execution trade-offs; the
+[setup decision](../../.agents/notes/implemented/feature/2026-10-21-completion-setup.md)
+records startup-file ownership and installation limits.
+
+## Configuration setup
+
+Project and task commands use `https://daynight.cc` when the configuration file or
+`serverUrl` is absent. A production connection needs no server configuration.
+Run `daynight login` to display an authorization code and open the verification
+page. Check that the page's code matches the terminal, confirm the account and
+approve access. The CLI saves the resulting credential without displaying its
+secret. `daynight login --no-browser` displays the URL and code for authorization
+in another browser. The device authorization expires after ten minutes; rejection, expiry and
+cancellation leave an existing saved credential unchanged.
+
+The CLI credential is named `Daynight CLI (<hostname>)`, using the machine that
+runs `daynight login`, including when another device approves access. Hostnames
+longer than 85 Unicode characters are shortened to their first 85 characters to
+keep the credential name within 100 characters. The credential never expires by default.
+The Web account's token manager can set a future expiration, restore “Never expires”, or
+revoke it. Changing expiration keeps the same credential, so the saved CLI login
+continues without replacing its secret. Expired or revoked credentials cannot be
+revived. The [token manager mockup](../../packages/daynight-web/tests/mockups/tokens.html)
+shares the production components and styles for creating tokens and editing
+expiration, including conflict and uncertain-result states. A credential
+represents the account's current project permissions. `daynight logout` revokes the current credential and removes
+its local value after a confirmed response; an already invalid credential is
+removed locally. Network or unconfirmed failures remain visible. Configuration
+changes made during login or logout cannot overwrite another service origin or
+another saved credential. A lost successful login response requires a fresh
+login; the account token list can revoke a credential whose response was lost.
+
+Without a local credential, a service command fails before dispatch and directs
+the user to `daynight login`. A server `401` gives the same guidance. Business
+commands do not start login, follow redirects with credentials, or automatically
+repeat a write. Reconcile an unconfirmed business mutation before another one.
+
+The first `review run`, `review list` or `review ls` asks for a repository when
+the file or `repositories` field is absent. To configure one in advance, run:
+
+```bash
+daynight review add-repo https://github.com/OWNER/REPO.git
+daynight review add-repo git@github.com:OWNER/OTHER.git --path ./review-other
+```
+
+Interactive guidance requires terminal stdin and stderr. Prompts and save
+messages use stderr so stdout retains its normal text or JSON format.
+Noninteractive missing review configuration fails with a `review add-repo`
+command. Shared YAML/schema errors, invalid selected values, explicit empty
+review lists and read failures retain their errors. Unrelated capability values
+remain outside the current command's validation. Help, version, completion,
+update and invalid arguments do not prompt.
+
+Test services and noninteractive automation use the existing
+`~/.daynight/config.yml` fields. For a test service, provision its origin before
+running `daynight login`:
+
+```yaml
+serverUrl: http://127.0.0.1:8787
+```
+
+For unattended access, a secret-provisioning process can write `serverToken`
+together with its intended `serverUrl`; omit `serverUrl` only when the credential
+belongs to `https://daynight.cc`. Keep that file private, with mode `0600` on
+POSIX, and never put credentials in command arguments, logs or repository files.
+Provisioners must preserve unrelated settings and pair the origin with its own
+credential. Replacing `serverUrl` manually does not clear an old token; remove
+or replace it before sending another request. External editors and provisioners
+do not participate in the CLI's configuration lock.
+
+Tokens are opaque visible ASCII and require HTTPS except for HTTP origins whose
+host is literal `localhost`, a numeric IPv4 address in `127.0.0.0/8`, or `::1`.
+Invalid explicit origins and credentials remain errors. Normal CLI use requires
+no manual PAT creation. There is no top-level `daynight setup` command; shell
+startup installation remains available through `daynight completion setup`.
+
+Review guidance collects values before one atomic save under the configuration
+write lock, preserving comments, unrelated settings and completed concurrent
+changes to unrelated fields. A selected review-field conflict requires a fresh
+decision. Login and logout check the observed origin and credential together
+under the same lock. Credential-bearing files use POSIX mode `0600`, including
+review writes; files without a token retain existing permissions. Symlink
+rejection and atomic replacement apply to CLI configuration writers.
+Ctrl+C, termination signals or EOF cancel prompts without partial saves or
+business execution. Human input consumes no business timeout. Review guidance
+saves configuration without network, login or authentication checks; cloning and
+review state initialization belong to the business operation.
+
+The [configuration decision](../../.agents/notes/implemented/feature/2026-10-21-guided-configuration-setup.md)
+owns review guidance and persistence; the
+[device-login decision](../../.agents/notes/implemented/architecture/2026-10-25-cli-device-login.md)
+owns the default service, explicit login/logout and credential lifecycle. The
+[top-level setup removal decision](../../.agents/notes/implemented/simplification/2026-10-25-remove-top-level-setup.md)
+records the command removal and configuration-file provisioning boundary.
+
+## Project management
+
+After `daynight login`, create, list or inspect projects on `https://daynight.cc`.
+To use a test server, set `serverUrl` in `~/.daynight/config.yml`:
+
+```yaml
+serverUrl: http://127.0.0.1:8787
+```
+
+The origin includes its protocol, host and optional port; the client adds
+`/api/v1`. When changing the origin externally, remove the old `serverToken`
+before logging in to the new service; manual edits do not run CLI credential
+cleanup. Then use the same commands:
+
+```bash
+daynight project create "Example"
+daynight project create "Example" --repo https://github.com/OWNER/REPO.git
+daynight project list
+daynight project list --json
+daynight project show <project-id> --json
+```
+
+Creation requires one positional name; quote names containing spaces. Omit
+`--repo` to use the current Git worktree's effective `origin`, including Git
+`insteadOf` expansion. Supply `--repo` to use that address without inspecting Git;
+this form, listing and details work outside Git repositories. For a name beginning
+with `-`, use `daynight project create -- -name` or put an explicit repository
+option before the terminator: `daynight project create --repo <repo> -- -name`.
+The name is sent unchanged.
+
+Creation first scans every project page and compares canonical repository
+identities. With no match it registers the original name and repository address.
+With one match it reports `Project already exists.` and the existing project,
+leaving that project's name and address unchanged. Multiple matches fail with
+their IDs; lookup failures and exhausted command budgets stop before creation.
+This client preflight does not provide atomic uniqueness against concurrent
+creators, and the service continues to allow duplicate repository registrations.
+Project commands do not write repository bindings or ignore rules. Lists traverse
+all pages. Human output uses complete labeled project blocks with visible
+terminal-control escapes. JSON preserves resource fields and nanosecond timestamp
+strings; a later-page failure produces an error with empty stdout. The
+[project/task requirements](../../docs/spec/daynight-project-task.md) own the complete
+behavior and remaining task-command scope.
+
+## Task management
+
+Task commands use the same `serverUrl` configuration as project commands:
+
+```bash
+daynight task create --title "Implement sign-in" --description '## Sign-in
+
+Implement OAuth sign-in for the web workspace. Deliver the callback flow and
+visible failure recovery; preserve the existing session and provider rules.
+Complete when a valid login opens the workspace and a rejected login can retry.'
+daynight task create --title "Build the sign-in page" --depends-on <design-task-id> --external-depends-on <authentication-task-id>
+daynight task ready
+daynight task ready --limit 5 --json
+daynight task ready --allow-agents true --compatible-platform linux --json
+daynight task list --allow-agents false --ready
+daynight task list --ready
+daynight task list --all --json
+daynight task list --status in_progress --json
+daynight task show <task-id> --json
+daynight task update <task-id> --description=
+daynight task update <task-id> --allow-agents true --if-version <version>
+daynight task create --title "Check Windows build" --required-platform windows
+daynight task update <task-id> --required-platform macos --if-version <version>
+daynight task update <task-id> --clear-required-platform --if-version <version>
+daynight task set-status <task-id> <status> [--if-version <version>] [--project <project-id>]
+daynight task update <task-id> --depends-on <task-a> --depends-on <task-b>
+daynight task update <task-id> --clear-dependencies
+daynight task graph --json
+```
+
+Every operation accepts `--project <project-id>`, which selects a project for one
+invocation without reading Git or local bindings. Otherwise, the CLI reads the
+current worktree's effective `origin`, normalizes equivalent HTTPS/SSH/SCP
+addresses, and matches the configured server's projects. A missing match fails
+with `Project does not exist`; multiple matches require explicit selection.
+A unique discovery is saved in the worktree's `.daynight/project.yml` with the
+server origin and repository identity, after ensuring `/.daynight/` is effective
+in the root `.gitignore`. A cached binding is verified against the server on each
+use and remains stable if another project later shares that repository.
+
+Creation starts in `todo` with `allowAgents=false` and unrestricted
+`requiredPlatform=null` unless explicitly set. `task set-status <id> <status>`
+sets `todo`, `in_progress` or `done`, including reopening and direct completion.
+Every project member may use ordinary status setting. Active worker completion
+requires the current owner and matching claim ID through the typed completion
+operation; release and continuation of released work use their protected
+operations. Status does not grant execution or delivery authority. Generic updates cannot write status; old lifecycle commands and routes
+are removed without aliases. `--if-version` uses the exact reviewed int64 value;
+without it, the CLI reads once immediately before submitting. No write retries
+occur. Version is checked before same-state handling: current same-state keeps
+version, update time and timing unchanged; actual changes increment once.
+
+Entering in_progress or done requires all direct prerequisites, ordinary and
+external, to be done. Reopening done requires no direct in-progress or done
+dependents. Refusals return sorted blocking task IDs without cascading changes.
+The project transaction serializes status and dependency writes and preserves
+materials, properties/PRs, permission, both dependency sets and attachments.
+
+Migration `014_task_status_assignment.sql` converts historical proposed rows
+to todo, increments only those versions and updates their update time while
+preserving every other field, including requiredPlatform and historical execution timestamps. Default,
+data, CHECK and ledger change atomically. Other rows and applied migrations remain
+unchanged; restart does not repeat the conversion. Runtime readers reject the old state.
+
+Task resources expose required nullable `startedAt` and `completedAt`. Actual entry
+into in_progress starts a new interval and clears the end; return to todo clears
+both. In-progress to done preserves a nullable start and records a real end;
+todo to done records only an end, leaving duration unknown. Metadata and same-state
+writes preserve time. Complete-graph and task-sync responses supply a database
+`observedAt` sample; Web synchronization updates its separate clock without
+rebuilding unchanged graph/index data.
+Read-only timing fields remain in JSON; human CLI output keeps its field choices.
+The [status requirements](../../docs/spec/daynight-task-status.md) and
+[architecture](../../docs/design/daynight-task-status.md) own the complete rules.
+
+Task and graph resources also require `externalDependsOn`, a sorted,
+duplicate-free subset of all `dependsOn` IDs. CLI JSON preserves it and the
+shared human dependency block shows `External dependencies (N): ...`. Current
+API dependency fields are independently optional: a dependsOn-only PATCH
+retains classification on surviving edges, while an externalDependsOn-only PATCH
+replaces the exact external subset. CLI --depends-on and --external-depends-on
+compile both complete sets whenever either option is supplied; ordinary-only
+clears the external subset and external-only replaces all relationships. Omit
+both CLI options to preserve them. Metadata and lifecycle actions retain
+unmodified classification. Every edge still participates in blocking and cycles.
+
+Every Project resource includes `settings` with an exact positive int64 `version`
+and `executionWarningMinutes` from 1 through 2147483647, initially 60. Project
+create/read/list, the project in graph responses, and a successful settings PATCH
+return this same shape. The member-readable settings page lets administrators
+save or cancel a draft; ordinary members see the saved value without editing.
+The server checks current administrator authority after the project lock, then
+the observed settings version. An equal-value save preserves version and
+`updatedAt`; a real change increments version and updates `updatedAt`. Migration
+preserves historical project timestamps. CLI JSON preserves settings and nested
+versions without rounding; human output keeps its established fields.
+
+An unconfirmed settings PATCH response requires a Project and membership reread
+before another write. Failed reconciliation keeps Save blocked and offers read
+retry; network, protocol, 5xx or navigation cancellation never prove rollback.
+Confirmed refusals and unsaved/canceled drafts do not apply local settings.
+Project-only changes update graph metadata while retaining the task index and
+layout when tasks and summary are unchanged. Running known-start tasks subscribe
+to one visible-page ticker; completed duration uses stored endpoints and unknown
+starts show no inferred duration. Warnings use an unrounded strict comparison
+and disappear on completion without changing task status. The server-sample
+clock avoids client wall-clock skew but does not remove network sampling delay.
+
+
+The Web detail header places a three-item status menu beside the actual task
+title on both tabs. The current item remains selectable. Confirmation names the
+task, current/target states and timing effects; Cancel and Escape send no write.
+Status, **Automatic execution / 自动执行** and platform confirmation share one per-task mutation owner.
+Pending, stale or unknown results block conflicting writes. Current-version same-state responses
+confirm no stored change; actual changes refresh the graph. Unknown outcomes use
+GET-only reconciliation, preserving errors and never attributing the observed
+state to the uncertain request. Closing details does not roll back a submission.
+
+Task details use one platform select in the left status group to display the
+confirmed requirement and offer Any platform, Windows, Linux and macOS. Choosing
+a different value opens confirmation showing the current and target platforms;
+only confirmation submits the reviewed exact task version. Choosing the saved
+value, Cancel, close, backdrop dismissal and Escape send no request. Cancellation
+returns focus to the select. The confirmed value remains displayed until success,
+and Any platform clears requiredPlatform. Overview has no separate platform
+editor or Save action. The control stays available on either detail tab, is
+locked during in_progress, and retains shared pending/unknown reconciliation
+across navigation. Rows and graph nodes remain read-only and show each task's
+own requirement; root values do not define chain-wide platform requirements.
+The [confirmed platform control decision](../../.agents/notes/implemented/simplification/2026-10-24-task-platform-confirmation-control.md)
+records its allocation and focus behavior.
+
+The permission switch has confirmation in both directions, remains false by
+default and is locked against actual changes while in_progress. It governs
+automatic discovery; explicit task execution may use either value. Rows and graph
+nodes display each task's own permission. Permission does not control worker
+ownership; ordinary status setting cannot release active execution or resume
+released resources. Claim-bound done validates ownership. None of these
+operations stops a remote process or replaces PR review and delivery authority.
+The [status decision](../../.agents/notes/implemented/architecture/2026-10-24-task-status-assignment.md)
+records migration and shared-session trade-offs.
+
+`task ready` selects up to 10 actionable tasks by
+default; `--limit <N>` accepts decimal integers from 1 through 100. It makes one
+task-list request with `ready=true`, the chosen limit and optional
+`--allow-agents true|false` and `--compatible-platform windows|linux|macos`
+filters, without fetching later pages. Compatibility includes unrestricted tasks
+and tasks requiring that platform. The service applies all filters before
+pagination and binds them to the cursor. The service orders results by creation time and then task ID, both
+ascending. A ready task is `todo` with all direct dependencies `done`. This
+bounded selection supports choosing the next work without retrieving the whole
+project; `task graph` provides a dependency overview when needed. `task list`
+reads every matching page, including with `--ready`, and excludes `done` by
+default. `--all` includes every status; `--status <status>` selects one status,
+including `done`. These two options conflict. `--ready` intersects the selected
+statuses and can be combined with `--all`. List results place `in_progress`
+first, then ready tasks, other `todo` tasks, and `done`; each group retains
+service order across all pages. This order applies to human and JSON results.
+`task ready` accepts neither `--status` nor `--cursor`. `task list` supports the
+same permission and compatibility filters across every matching page. Both are
+automatic-selection conditions independent of readiness; explicitly designated
+execution retains the manual exception.
+
+Updates preserve omitted fields; an empty description clears it. Repeated
+`--depends-on` and `--external-depends-on` declare the complete ordinary and
+external sets. Either option replaces all relationships and classifications;
+ordinary-only makes every supplied edge ordinary, while external-only leaves
+only those external edges. IDs are canonicalized and sorted; malformed and
+case-equivalent duplicate/cross-declared UUIDs fail before configuration or
+network access. `--clear-dependencies` sends two empty sets and conflicts with
+either dependency option. An update needs at least one mutable
+field. `--allow-agents true|false` explicitly changes automatic-execution
+permission; omitted preserves it. Creation rejects the option. Existing tasks
+receive true through migration without changing versions or business timestamps.
+A matching same-value permission PATCH retains version and time, while a stale
+version still conflicts and an actual in-progress toggle fails atomically.
+Create and update accept `--required-platform windows|linux|macos`; update
+accepts `--clear-required-platform` to send null, mutually exclusive with setting
+a platform. Omission preserves an existing requirement. New and migrated tasks
+are unrestricted unless explicitly set. Exact-version equal platform updates
+preserve version and time; stale updates conflict and actual in-progress changes
+are refused atomically. Explicit manual status setting remains available for every
+requirement. The query platform selects candidates; it does not configure or
+prove the environment of a future worker.
+By default the CLI reads the task version before PATCH; `--if-version <n>`
+uses that exact positive int64 version. Conflicts and failed writes are reported
+without automatic retries.
+
+Text options accept both separated values and `--option=value`. Inline assignment
+splits at the first equals sign and supports leading dashes, such as
+`--title=-example`; `--description=` is the explicit empty form. Separated values
+starting with a dash are rejected. Directory options, switches and help reject
+inline assignment. Attachment `put` declares its `--path` as text, accepting
+`--path=value`; attachment `put/read/remove` also accept `--` before positional
+paths starting with a dash. Other task operations retain their declared syntax.
+
+Human task lists use labeled blocks containing ID, project, title, PR references,
+status, readiness, automatic-execution permission, required platform, dependencies and blocking.
+Description, version and timestamps
+remain available in show, ready, create and update output. Graph nodes include
+state, permission, required platform, dependencies, blocking and title-preview truncation flags.
+Task and graph decoders require a boolean `allowAgents`; missing, null or
+nonboolean values are protocol errors. They also require `requiredPlatform`
+as null or one of the three platform enums; missing keys or other values fail
+without a runtime default. Human output uses `Required platform: any` for null
+and the enum otherwise. Queries support
+`--json`: task `list` returns every matching task in one array, omitting only
+`description` while retaining every other API field, including raw PR strings,
+version and timestamps. This applies to the default list and every filter,
+including `--all`. Project lists retain all API fields. Task `ready` returns its
+selected full tasks as an array, details return one full resource, and graphs
+return the complete API object. An empty human ready result is `No ready tasks.`
+and empty JSON is `[]`. Versions remain JSON numbers with full int64 precision.
+Local preparation fails before any task mutation; a lost write response is
+reported as unconfirmed. Installed CLI integration uses the real Go/PostgreSQL
+service; deadline behavior uses short test-only budgets and byte probes retain
+the full production limits. [PR #55](https://github.com/codetreker/daynight/pull/55)
+records the original integration and real-default-budget validation, with its
+historical source identity.
+
+## Task deletion
+
+Read the task and use its observed version when deleting it:
+
+```bash
+daynight task show <task-id> --json
+daynight task delete <task-id> --if-version <version>
+```
+
+Deletion removes the task from normal queries and prevents later reads and
+changes. Stored task data, its own prerequisite references and attachment files
+remain. An active worker claim on the target prevents deletion. Without
+`--remove-depends`, an undeleted task depending on the target also prevents it.
+
+To explicitly remove every other task's ordinary and external dependencies on
+the target in the same transaction, use:
+
+```bash
+daynight task delete <task-id> --remove-depends --if-version <version>
+```
+
+This includes references added before that transaction executes and may make
+affected tasks ready. Each changed dependency owner receives a new version;
+its status, timing and other data stay unchanged. Deletion requires explicit
+authorization for the target; removing other tasks' dependencies requires that
+additional authorization.
+
+Without `--if-version`, the CLI reads the current version immediately before the
+write. Conflicts and failed writes are not retried. After an interrupted or lost
+response, read the task to reconcile before considering another request. An
+authorized read returning `410 task_deleted` confirms deletion; a live resource
+requires a new decision, while a permission or transport failure leaves the
+outcome unknown. The [deletion requirements](../../docs/spec/daynight-task-deletion.md)
+and [design](../../docs/design/daynight-task-deletion.md) own retained history,
+reference removal and cache invalidation.
+
+## Task attachments
+
+Attachment commands use the existing project selection and server configuration:
+
+```bash
+daynight task attachment put <task-id> ./design.md --path docs/design.md
+daynight task attachment put <task-id> ./mockup.html --path mockup/index.html
+daynight task attachment list <task-id> --tree
+daynight task attachment list <task-id> --json
+daynight task attachment read <task-id> docs/design.md
+daynight task attachment read <task-id> images/mockup.png > ./mockup.png
+daynight task attachment remove <task-id> docs/design.md
+```
+
+Each upload accepts one local regular file. Without `--path`, its basename is the
+target path; the same complete task-relative path is overwritten. Names preserve
+case and Unicode spelling, with `/` separators and no absolute or dot-segment
+paths. The content type comes from the target path's extension, even when it
+differs from the local filename; unknown extensions use
+`application/octet-stream`. Uploads and deletions affect one file at a time.
+
+`list` reads every page before emitting its complete metadata result. `--tree`
+and `--json` select mutually exclusive formats. `read` writes the original bytes
+without labels, decoding or an added newline. Diagnostics remain on stderr;
+a failed transfer can leave incomplete stdout and exits nonzero. File transfers
+stream with backpressure and retain the shared deadlines and cancellation;
+metadata JSON keeps its existing byte budgets. Use `--` for a dash-prefixed
+positional path, for example:
+
+```bash
+daynight task attachment put --project <project-id> <task-id> -- -design.md
+daynight task attachment read --project <project-id> <task-id> -- -design.md
+```
+
+Failed writes are not retried. File storage and PostgreSQL metadata are not
+atomic; an unsuccessful operation may already have changed storage. The
+[attachment requirements](../../docs/spec/daynight-task-attachments.md) define this
+known issue and the trusted same-origin preview model. The
+[authentication architecture](../../docs/design/daynight-auth.md) defines the required
+authentication and project-membership checks for attachment requests. Trusted
+executable HTML retains the application origin. The
+[CLI delivery decision](../../.agents/notes/implemented/feature/2026-10-21-task-attachment-cli.md)
+records binary transport and output ownership.
+
+## Review configuration and startup
+
+Register a repository explicitly, or let the first review operation guide registration:
+
+```text
+daynight review add-repo https://github.com/OWNER/REPO.git
+daynight review add-repo git@github.com:OWNER/OTHER.git --path ./review-other
+daynight review run
+```
+
+For a local build, run `pnpm daynight review run` from the repository root.
+These commands select `~/.daynight/config.yml` using the current OS user's home
+directory. `review add-repo`
+creates the configuration when absent. It reads, validates and saves configuration;
+cloning, authentication checks, GitHub calls, SQLite initialization and PR boundary
+selection belong to review startup. An already running reviewer keeps its loaded
+configuration until its next startup.
+
+The same YAML also accepts the server client's `serverUrl` and optional
+`serverToken` fields. Review validates its own fields independently of server
+business values. The server projection requires a valid HTTP(S) origin and
+validates any configured token and its transport. The shared parser still rejects
+invalid YAML, duplicate keys and unknown root fields. Registration can add the
+first review repository to a server-only file while preserving its server
+settings and comments.
+
+A configuration with an explicit addition root is:
+
+```yaml
+repositoryRoot: ~/daynights
+copilotModel: auto
+copilotReasoningEffort: high
+repositories:
+  - remote: https://github.com/OWNER/REPO.git
+    path: ../daynights/github.com/owner/repo
+```
+
+`copilotModel` and `copilotReasoningEffort` are independent optional fields for
+all review repositories in the running process. Each configured value must be a
+string containing non-whitespace text and no NUL; null and other value types are
+rejected. Accepted strings are preserved exactly. Copilot owns model availability,
+effort levels and model/effort compatibility; Daynight maintains no static lists.
+`auto` delegates model selection to Copilot. Omitting either field sends no
+corresponding override and preserves Copilot's choice, so neither, either or both
+may be configured. Changes take effect on the next `review run` startup. Guided configuration
+and repository registration preserve both fields and YAML comments. The server
+configuration projection recognizes these shared keys without validating their
+review-owned values.
+
+`repositoryRoot` defaults to `~/daynights`. It is used only when `add-repo` has no
+`--path`: the command derives filesystem-safe components from the normalized
+host/owner/name and saves the resulting absolute repository path. A configured
+root may be absolute, config-directory-relative, or begin with `~` for the current
+OS user's home. Other users' tilde forms are rejected. Changing the root affects
+future additions only. An explicit `--path` resolves against the command's working
+directory and is also saved as an absolute path. Manually written repository paths
+remain absolute or config-directory-relative and do not expand `~`; every repository
+entry requires its own `path`.
+
+Adding an existing normalized repository without `--path` reuses its saved path
+and succeeds without rewriting configuration, even after the root changes.
+An explicit path to the same actual location is also idempotent; an explicit
+different path for that repository is an error.
+Existing YAML settings and comments survive additions. The command validates the
+complete resulting configuration before replacing the file atomically. Validation
+and pre-replacement write failures preserve the original configuration; cleanup
+errors after replacement remain visible. A separate
+`~/.daynight/config.yml.lock` excludes another configuration writer and fails
+clearly when occupied; it is independent of the review instance lock. After an
+abnormal exit, confirm no configuration writer remains before removing this lock.
+Configuration mutation rejects symlinks and nonregular files. A resulting
+configuration containing `serverToken` uses POSIX mode `0600`; without a token,
+existing file-mode behavior is preserved.
+
+Numeric options use the documented [configuration defaults](../../docs/design/daynight-review.md#使用方式与配置)
+when omitted. Explicit invalid values, unsupported fields, duplicate YAML keys and
+malformed documents produce nonzero exits with useful errors. Missing configuration
+triggers the [configuration setup](#configuration-setup) rules before review
+startup. Duplicate repository identities,
+duplicate local paths, paths overlapping the retained state directory, and paths
+with an existing nondirectory component are rejected. Missing clone directories
+remain valid for later repository creation. Existing paths are compared by
+filesystem identity. Missing destinations with the same existing ancestor and
+component count are compared by temporarily creating one candidate's directory
+names inside a unique probe directory there and looking up the other spelling.
+This observes actual case and Unicode rules. Validation removes only the
+directories it created, using nonrecursive removal. Probe permission, inspection,
+or cleanup failures stop validation with an error.
+
+The selected retained-state location is `~/.daynight/` for SQLite, sessions,
+locks, logs, cache, and temporary files. PR worktrees belong to each monitored
+repository root's `.worktrees/` directory. `maxConcurrentWorker` limits concurrent
+review preparation and execution. Startup acquires an exclusive instance lock,
+opens SQLite and prepares the configured clones. Each repository is polled
+independently; access failures retain its last complete observation and remain
+visible in diagnostics. The
+[repository registration decision](../../.agents/notes/implemented/feature/2026-10-18-add-repo-configuration.md)
+records the fixed-path and configuration-only choices.
+
+## Review status listing
+
+Print the configured repositories' locally tracked open PRs, including Draft PRs:
+
+```text
+daynight review list
+daynight review ls
+```
+
+Both forms use the same persisted-snapshot operation and print repositories in
+configuration order, with PRs ordered numerically. The columns are `PR`,
+`PR STATE`, `REVIEW` and `ATTEMPTS`. Completed or failed reviews remain listed while their PR is open;
+closed PRs and history belonging to removed configuration entries are omitted.
+
+The output identifies its local snapshot source and each repository's
+`Last observed` successful complete poll and saved poll error when present.
+PRs at or below the permanent tracking boundary are not stored individually and
+cannot appear in the listing. `No tracked ongoing PRs.` describes an initialized
+repository with no saved open PRs above the boundary; `Not initialized.` describes
+a missing database, an unregistered repository or an incomplete initialization.
+The listing makes no claim about GitHub changes since the saved observation.
+
+Review states are `Preparing`, `Running`, `Pending`, `Retry waiting`, `Completed`,
+`Failed`, `Interrupted` and `Paused` for Draft PRs. `Running` is a saved state,
+not proof that a process remains alive. Retry deadlines and saved errors appear
+under the corresponding row. A newer desired revision remains `Pending` while an
+older revision is active; the older revision's recorded state, attempt budget,
+retry deadline and error are printed separately. The current row's attempt count
+always belongs to its own desired revision, head and target branch. A base-tip
+change alone does not invalidate a completed review.
+
+Pending worktree cleanup is a separate saved resource barrier, including after a
+PR reopens. Such rows show `Worktree cleanup: Pending` and any saved cleanup error.
+With no current review, an open non-Draft row shows `Pending (cleanup blocked)`;
+Draft remains `Paused`. Older active reviews retain their separate details. The
+listing reports the barrier without attempting cleanup or recovery.
+
+The command loads the current configuration and opens an existing SQLite database
+read-only for one consistent transaction. It needs no Git, GitHub CLI or Copilot
+execution, instance lock, clone or review startup. It does not create an absent
+database, migrate a schema, recover interrupted work or update application state.
+SQLite may create or update its WAL coordination sidecar files while reading an
+existing database. Missing or invalid configuration, unreadable or corrupt state,
+and unsupported schema versions fail visibly. The
+[listing decision](../../.agents/notes/implemented/feature/2026-10-18-review-status-listing.md)
+records status attribution and snapshot limits.
+
+## Durable review state
+
+The command and source layout preserve existing YAML fields, retained-state
+locations and resource identities. Runtime opening explicitly migrates earlier
+SQLite state to retain verified review publications and conversation follow-ups;
+existing cutoffs, revision budgets, sessions, process evidence and worktrees remain
+owned by their original records. Read-only listing does not migrate state. The
+review module owns its domain types independently of its SQLite implementation.
+
+The SQLite module keeps repository identity, the permanent startup PR boundary,
+observed PRs, captured review targets, completion revisions, sessions, worktree
+registration, retry state, and verified GitHub review/thread identities. Repository
+identity includes the GitHub host and stable repository ID. Completing a captured
+revision leaves later observations pending; retries retain their original review
+identity and consumed budget. Each successful comment result arms one durable
+follow-up that is consumed once when conversations are resolved, even at the same
+commit. A verified approval settles its current target; description edits retain
+that approval. Reusing a stored approval to settle a pending target requires a
+fresh snapshot confirming that exact remote review is still `APPROVED`, with the
+captured actor, commit and stable identity. A dismissed historical review cannot
+suppress that pending work.
+
+Runtime database opening applies explicit schema migrations and enables WAL with
+FULL synchronization and a bounded busy timeout. The listing reader checks the
+existing schema, accepting versions 1 and 2 without applying migrations. The
+instance lock checks whether its recorded CLI owner still exists; that check does not establish whether its agents
+or tools have stopped. Interrupted reviews and pending cleanup remain blocked
+after automatic lock recovery until process exclusion is explicitly confirmed.
+Database state does not replace physical Git/worktree or session-history
+verification.
+
+See the [review architecture](../../docs/design/daynight-review.md) for transitions and the
+[state decision](../../.agents/notes/implemented/architecture/2026-10-17-durable-review-state.md)
+for persistence choices. The coordinator opens this state before scheduling work
+and advances only the revision captured by each completed attempt.
+
+## GitHub observations
+
+The GitHub module reads one repository pass through authenticated `gh api` GET
+requests. Initialization scans every page and PR state before committing the
+permanent cutoff. Later passes track open PRs, including Draft, and confirm
+missing tracked PRs individually before recording closure.
+
+Failed pages, malformed responses, and failed closure confirmations preserve
+the previous repository observation. A completed pass commits its observations
+and success time atomically. Rate-limit failures expose the server's retry
+deadline and block later same-host calls in that client; the coordinator schedules
+the next pass after the applicable deadline.
+The [observation decision](../../.agents/notes/implemented/architecture/2026-10-17-github-observation.md)
+records pagination and failure handling.
+
+CI admission uses a read-only GraphQL query for the current PR head and
+available test-merge commit. It combines GitHub's check/status rollups with
+registered workflow suites, including approval-blocked workflows without jobs.
+All reported CI must finish successfully; GitHub's `success`, `skipped` and
+`neutral` outcomes pass. No CI allows immediate review. API errors and partial
+results keep the PR waiting. Latest workflow runs replace historical attempts;
+waiting or failing workflows do not consume review capacity.
+
+Fresh review also requires a complete, paginated review snapshot with no unresolved
+native `PullRequestReviewThread` from any author. Outdated threads still block
+until resolved. Threads proven to belong to unpublished `PENDING` reviews are
+excluded from this author-facing gate; unknown ownership cannot imply resolution.
+Review summaries and ordinary PR comments have no resolution state. A pending round and no active or uncertain execution are also required, so
+repeated passing-CI/zero-thread observations do not repeatedly review a settled
+target. CI and conversation waiting consume no review Worker or model attempt.
+API failure or incomplete data cannot prove that no conversations remain.
+
+## Repository clones and worktrees
+
+The Git module creates missing clones and validates existing repository identity.
+Prepared GitHub CLI authentication supplies command-scoped HTTPS credentials,
+including repositories configured with SSH addresses or API ports. Existing
+checkout files, origin configuration, branches, remote-tracking refs, and
+FETCH_HEAD remain untouched by review preparation.
+
+Configured clones and their PR worktrees are dedicated to this system. Each PR
+uses its fixed repository-local `.worktrees/pr-<number>/` directory. Physical
+root/common-directory checks, validated linked administration and backpointers,
+and a session-labelled Git lock establish the target before destructive reuse
+or removal. Initial preparation and new revisions provide a complete checkout
+even when the base checkout uses sparse settings. Confirmation retries retain
+local context. Files, nested clones and administration inside the owned PR
+worktree are disposable; cleanup does not inventory their external consumers.
+Other PR worktrees and retained sessions remain separate. Cleanup rechecks remote
+closure before removal. The [exclusive-environment decision](../../.agents/notes/implemented/architecture/2026-10-17-exclusive-review-environment.md)
+defines this ownership boundary.
+
+See the [worktree ownership decision](../../.agents/notes/implemented/architecture/2026-10-17-owned-review-worktrees.md)
+for target validation and recovery boundaries.
+
+## Copilot execution
+
+The review module ships `daynight-review` under its own `skills/` directory.
+After taking the instance lock, the daemon captures the installed bundle once
+and prepares an immutable content-addressed snapshot under
+`~/.daynight/cache/review-skills/`. Package updates cannot replace the content
+used by that running daemon. Snapshots remain available for retained sessions
+and uncertain executions; automatic collection is not provided.
+
+Each attempt first runs native Copilot skill discovery in the prepared PR
+worktree. A single `--plugin-dir=<snapshot root>` argument supplies a generated
+local plugin wrapper containing `plugin.json` and `skills/daynight-review/`.
+This preserves paths containing commas, spaces and Unicode; inherited
+`COPILOT_SKILLS_DIRS` variables are removed from the child environment.
+Exactly one enabled `daynight-review` must resolve to that snapshot's directory;
+a disabled skill, project override or malformed response fails the attempt.
+The review prompt explicitly invokes the skill and identifies its current
+absolute entrypoint and digest on both new sessions and resume. The bundled
+skill owns review methodology; the prompt owns the immutable target, publication
+eligibility, retry reconciliation and completion protocol. The prompt also
+explicitly requests the skill's native subagent delegation and synthesis when
+fresh review is required; confirming a valid matching submitted result retains
+its reconciliation exception.
+
+The Copilot adapter then launches the review child in that same worktree and
+environment. It selects the persisted per-PR UUID on first launch and explicitly resumes that
+same session thereafter. An interrupted first launch also resumes existing
+history; missing established history is an error. Only opaque session-directory
+presence is inspected. Copilot history and agent JSONL are not parsed.
+
+Every review child receives configured `copilotModel` as separate `--model` and
+value arguments, and configured `copilotReasoningEffort` as separate
+`--reasoning-effort` and value arguments. New sessions, resumes, retries and later
+PR revisions use the same construction. Skill discovery receives neither override.
+Values pass through the existing `shell: false` execution path; unsupported or
+unavailable models, effort levels or combinations retain Copilot's normal failure
+result and the existing attempt/error handling. The
+[model options decision](../../.agents/notes/implemented/feature/2026-10-22-review-copilot-model.md)
+records the validation and default-selection choices.
+
+Prompts travel through stdin in noninteractive mode. `cross-spawn` handles native
+commands and Windows command shims; prompt text is not embedded in shell code or
+command arguments. Runtime paths contain Copilot state, cache, temporary files
+and per-attempt diagnostics. Existing GitHub CLI authentication remains available
+through Copilot's supported fallback. Discovery and review are sequential phases
+of one persisted attempt, sharing one retry budget and timeout deadline. The
+discovery child must be excluded before its exact persisted PID and identity are
+cleared and a review child can start. Cancellation, timeout or uncertain process
+exclusion prevents that next launch. Discovery stdout and stderr stream to
+separate `skill-list` logs and never enter completion-marker parsing.
+
+The agent reads the current PR title, description and linked scope requirements,
+then checks the complete captured diff against one concrete problem and outcome.
+A clear description and changes within the original scope are independent
+approval gates. Independent objectives, an unreviewable combined PR or unrelated
+changes block approval even without a Critical/Major code defect. Necessary
+implementation, structural changes, tests and documentation for the same outcome
+remain in scope; fixed line or file counts do not decide reviewability. Expanding
+the description alone cannot justify unrelated work.
+
+For every fresh review, the main agent first inspects that complete fixed diff,
+then chooses at least two complementary perspectives based on the PR's actual
+changes, contracts and risks. It invokes independent subagents, preferably in
+parallel, with exact head/base/merge-base context, scope, applicable repository
+instructions, evidence standards and shared worktree, temporary-file and process
+rules. Source and Git state remain read-only; subagents return findings and
+coverage limits to the main agent without publishing or emitting completion
+markers. The main agent waits for all required perspectives, verifies candidates,
+resolves disagreements and deduplicates root causes. It records the perspective
+rationale, completed coverage, limits, fix/defer/reject dispositions and confirmed
+blockers in attempt output, while GitHub comments contain concise findings only.
+An unavailable tool, failed child or incomplete required perspective is an
+operational failure; retries or reassignment stay within the attempt's existing
+budget. An incomplete review publishes neither a result nor a completion marker.
+
+These subagents belong to one attempt within the daemon-managed main review
+session; they do not create more review Workers, PR sessions or worktrees.
+Reconciliation of a valid matching submitted result confirms that result without
+repeating delegation. The selected questions must fit each PR, including
+documentation-only changes.
+
+Before new publication, the agent rereads the current PR information and
+reevaluates those gates. A successful review either approves or publishes
+`COMMENT` with at least one actual resolvable review thread. Suitable code findings
+use diff lines. Description and other PR-level blockers use a clearly labeled
+file-level conversation on the principal changed file; a review summary alone
+cannot satisfy this requirement. `REQUEST_CHANGES` remains unused. The reviewer
+must verify corrections after authors resolve feedback; resolution is a request
+for rereview, not evidence that the issue was fixed. Operational failures retain
+bounded retries and cannot be represented as quality findings.
+
+The daemon verifies remote publication using its stable review identity, captured
+commit and publishing actor, without interpreting the quality judgment or parsing
+findings. Marker acceptance is persisted before this read-only verification; API
+errors retry verification across restart without another model attempt. A complete
+lookup with no valid matching publication is a protocol failure and uses the
+remaining bounded model retry budget. A verified `COMMENT` records its real
+review/thread IDs and arms exactly one follow-up. Once
+all submitted native conversations are resolved and the current CI passes, that follow-up
+runs even without new commits or metadata edits. New targets and feedback wakeups
+coalesce. Same-round confirmation can proceed through open conversations only
+when exact-round remote evidence proves an approval, commented review with owned
+threads, or a recoverable matching pending draft under the captured publisher. A
+failed attempt alone supplies no exemption. If the authenticated reader changes,
+it can confirm the captured actor's submitted result; an unpublished result or
+pending draft waits for that actor without consuming model attempts. Verified
+approval settles the current target, and
+description edits do not revoke it or start another round.
+
+The adapter streams separate stdout/stderr logs and accepts only normal zero
+exit, complete capture and the exact final nonempty stdout line
+`PR N Review Completed`. Incomplete output, timeout, cancellation and logging
+errors cannot assert execution completion. A marker alone is insufficient: the
+coordinator verifies the corresponding remote approval or resolvable feedback
+before completing the round. It owns scheduling and final state transitions;
+runtime process supervision supplies evidence under the foreground execution
+contract.
+
+The [execution decision](../../.agents/notes/implemented/architecture/2026-10-17-copilot-session-execution.md)
+records process and publication boundaries; the
+[PR-quality decision](../../.agents/notes/implemented/process/2026-10-21-review-pr-quality-gates.md)
+records the approval policy, and the
+[multi-perspective decision](../../.agents/notes/implemented/process/2026-10-22-pr-specific-review-perspectives.md)
+records delegation and synthesis. The
+[conversation-cycle decision](../../.agents/notes/implemented/architecture/2026-10-21-review-conversation-cycle.md)
+records durable follow-ups and publication verification. The
+[skill design](../../docs/design/daynight-review-skill.md) owns packaging, snapshots and
+loading, with rationale in the
+[bundled-skill decision](../../.agents/notes/implemented/architecture/2026-10-20-bundled-review-skill.md).
+Native probes with Copilot CLI 1.0.100 verify actual bundled-skill activation and
+changed-bundle resume with the same UUID, including comma, space and Unicode
+paths. The bundled-skill decision records the probe's conditions and limits.
+Live Copilot/GitHub publication acceptance
+remains separate from those probes and controlled subprocess tests.
+
+## Coordination and shutdown
+
+A FIFO pending queue coalesces observations for each PR while preventing a
+frequently updated PR from starving other repositories or PRs. One reservation
+spans preparation, execution and cleanup for each PR. Fresh-round CI and
+conversation admission precedes that reservation and is refreshed at
+`pollIntervalMs`; waiting candidates retain their queue position without blocking
+ready PRs. Admission evidence is kept in memory for the exact revision, head and
+base branch and is refreshed after restart. New review launches recheck PR state,
+CI and conversations after worktree preparation. Same-round reconciliation can
+confirm proven exact-round publication while its own feedback remains open;
+matching pending drafts require the captured publishing actor. An unsuccessful
+attempt without publication still waits for unrelated open conversations. A regression
+defers fresh preparation without consuming an attempt or losing the retained
+session. CI changes alone do not retrigger completed revisions.
+Readiness is a launch-time observation; later CI registration or reruns do not
+automatically cancel an active agent.
+
+Cleanup does not consume a review worker slot, but cannot overlap another
+operation on the same worktree. Draft/closure observations cancel active work; head changes wait for the current
+attempt and remain pending. Failed revisions retain their exhausted budget.
+
+Ctrl+C stops new work and polling, terminates current managed execution, waits
+for operations and closes SQLite before releasing the instance lock. Agents must
+run tools in the foreground and stop their services before returning. Normal
+unforced exit and drained streams are evidence under that contract. POSIX uses
+current process groups; Windows cancels live trees with `taskkill /T /F`.
+Uncertain termination retains the lock and interrupted state.
+
+Startup checks the PID in an existing `~/.daynight/instance.lock` with a
+signal-zero probe. A live owner, permission denial, invalid lock metadata or an
+unclassified probe error preserves the lock and stops startup. Only an explicit
+no-such-process result permits automatic replacement of a dead owner's lock.
+Competing starts and releases serialize through `instance.lock.sqlite`; this
+guard database remains in place and must not be deleted while a CLI can use it.
+Lock age never establishes that an owner has stopped. PID reuse conservatively
+blocks startup until the recorded owner can be inspected.
+
+A crash with no unresolved review or cleanup can therefore restart directly.
+When interrupted reviews exist, automatic lock recovery blocks their PRs.
+Pending cleanup remains recorded with its error even if the PR reopens; only
+conclusively stopped cleanup operations and excluded agents allow that state to
+clear. Automatic recovery blocks only each affected PR's cleanup and reuse;
+polling and other PRs in the same repository continue normally, as do other
+repositories. Uncertain
+resources retain the lock on shutdown. After stopping the daemon, confirm that
+the previous CLI and all its agents/tools have stopped, then remove only
+`instance.lock` and restart. Keep `instance.lock.sqlite`, SQLite state, sessions
+and logs. That explicitly confirmed restart reconciles interrupted work with
+the same captured target and session; it never declares a saved running attempt
+complete or kills a saved process ID.
+
+The [coordination decision](../../.agents/notes/implemented/architecture/2026-10-17-foreground-review-coordination.md)
+explains scheduling and the foreground execution boundary. The
+[stale-lock recovery decision](../../.agents/notes/implemented/bug-fix/2026-10-18-stale-instance-lock-recovery.md)
+records owner-liveness checks and the separate process-exclusion requirement.