continuous-refactoring 0.2.0__tar.gz → 1.0.0__tar.gz

{continuous_refactoring-0.2.0 → continuous_refactoring-1.0.0}/.gitignore

@@ -1,4 +1,6 @@
 .scratchpad/
+tmpdir/
+.pytest_cache/
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

{continuous_refactoring-0.2.0 → continuous_refactoring-1.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: continuous-refactoring
-Version: 0.2.0
+Version: 1.0.0
 Summary: Continuous refactoring loop for AI coding agents
 Project-URL: Repository, https://github.com/bigH/continuous-refactoring
 Project-URL: Issues, https://github.com/bigH/continuous-refactoring/issues
@@ -32,6 +32,8 @@ Small, test-gated cleanup commits by an AI coding agent.
 
 Think of it as a supervised janitor loop: the agent proposes a cleanup, your tests decide if it stays.
 
+Here's [an article](https://artisincode.com/essays/how-i-use-unspent-tokens/) I wrote about it.
+
 ## Install
 
 Try it without installing:
@@ -67,6 +69,10 @@ uv pip install -e .
 
 That gives you the `continuous-refactoring` command.
 
+The CLI itself can be installed without `uv`, but the default validation command
+is `uv run pytest`. Pass `--validation-command pytest` or a project-specific
+script when the target repo does not use `uv`.
+
 Maintainers: see the [release checklist](https://github.com/bigH/continuous-refactoring/blob/main/docs/release.md).
 
 ## Fastest way to get one refactor
@@ -92,15 +98,18 @@ continuous-refactoring run \
   --max-attempts 2
 ```
 
-That keeps sweeping targets until it runs out, hits your caps, or starts failing.
+That runs up to 10 refactor actions, then stops sooner if the finite target file
+runs out or the loop starts failing. Use `run --focus-on-live-migrations` when
+you want the loop to work only on eligible live migrations; it bypasses target
+selection and `--max-refactors`.
 
 ## What it does
 
-- Resolves a target from `--targets`, `--globs`, `--extensions`, or `--paths`, with optional natural-language scoping via `--scope-instruction`.
+- Resolves each source action from `--targets`, `--globs`, `--extensions`, or `--paths`, with optional natural-language scoping via `--scope-instruction`.
 - Runs the agent with a refactoring prompt + your "taste" guidelines.
 - Runs your validation command (default: `uv run pytest`).
 - If green and there's a diff, it commits locally and leaves the branch for you to inspect.
-- Repeats until it runs out of targets, hits the retry budget, or stacks too many failures.
+- Repeats until it spends the action budget, exhausts a finite target file, hits the retry budget, or stacks too many failures.
 
 ## Requirements
 
@@ -119,8 +128,8 @@ continuous-refactoring init --in-repo-taste
 
 # 2. (Optional) Write your refactoring taste — either edit the file, have an agent interview you,
 #    or refine an existing draft collaboratively
-continuous-refactoring taste --interview --with codex --model gpt-5 --effort high
-continuous-refactoring taste --refine --with codex --model gpt-5 --effort high
+continuous-refactoring taste --interview --with codex --model gpt-5
+continuous-refactoring taste --refine --with codex --model gpt-5
 
 # 3. Do one pass
 continuous-refactoring run-once \
@@ -136,17 +145,23 @@ continuous-refactoring run \
   --sleep 5
 ```
 
+Active taste agent modes require `--with` and `--model`; taste agent sessions
+always run at fixed `medium` effort.
+
 ## Subcommands
 
 | Command | What it does |
 |---|---|
 | `init` | Registers this directory as a project, creates a default `taste.md`, and can store `--live-migrations-dir` or `--in-repo-taste`. |
-| `taste` | Prints the active taste file path. Add `--interview` to have an agent author it, `--refine` to iteratively improve an existing taste doc, `--upgrade` to refresh stale taste dimensions, `--global` for the shared file, and `--force` to let `--interview` overwrite custom content after writing a `.bak`. |
+| `taste` | Prints the active taste file path. Add `--interview` to have an agent author it, `--refine` to iteratively improve an existing taste doc, `--upgrade` to refresh stale taste dimensions, `--global` for the shared file, and `--force` to let `--interview` overwrite custom content after writing a `.bak`. Active agent modes require `--with` and `--model`, then run at fixed `medium` effort. |
 | `run-once` | Single pass on one resolved target. No retry. If there is a diff and validation passes, it commits locally and prints the diffstat. |
-| `run` | The loop. Iterates targets, retries on failure, and commits successful targets locally. |
+| `run` | The loop. Iterates refactor actions, retries on failure, and commits successful changes locally. Add `--focus-on-live-migrations` to bypass targeting and work only on eligible live migrations. |
 | `upgrade` | Checks that the global config manifest is current, rewrites it idempotently, and warns if the global taste file is stale. |
-| `review list` | Lists migrations flagged for human review (`awaiting_human_review`). |
-| `review perform <migration>` | Starts an interactive agent session to resolve a flagged migration's review. Requires `--with`, `--model`, and `--effort`. |
+| `migration list` | Lists visible migrations as TSV with headers. Add `--status <status>` or `--awaiting-review` to filter, or `--no-headers` for parsing. |
+| `migration doctor <slug-or-path>` | Validates one visible migration's consistency. |
+| `migration doctor --all` | Validates every visible migration plus internal transaction state. |
+| `migration review <slug-or-path>` | Starts staged review for a migration awaiting human review. Requires `--with` and `--model`; review runs at fixed internal `high` effort. |
+| `migration refine <slug-or-path>` | Records feedback for a planning or unexecuted ready migration and runs one staged planning revision. Requires `--message <text>` or `--file <path>`, plus `--with` and `--model`; refine runs at fixed internal `high` effort. Add `--show-agent-logs` to mirror the planning agent. |
 
 ## Targeting / Useful flags
 
@@ -157,29 +172,56 @@ Target resolution is first-match-wins:
 
 These flags are not mutually exclusive, but only the highest-priority populated source is used.
 
-- `--targets path/to/targets.jsonl` — explicit list; one JSON object per line with `description`, `files`, optional `scoping`, `model-override`, `effort-override`. Effort overrides use `low`, `medium`, `high`, or `xhigh`.
-- `--globs 'src/**/*.py:tests/**/*.py'` — colon-separated globs; each matched file becomes its own target.
-- `--extensions .py,.ts` — shorthand that expands to `**/*.py`, `**/*.ts`; each matched file becomes its own target.
-- `--paths a.py:b.py` — literal paths, all treated as one target.
-- `--scope-instruction "clean up the auth module"` — extra free-text scoping. If file-based targeting resolves nothing, this becomes the useful fallback context.
+- `--targets path/to/targets.jsonl` — explicit finite list; one JSON object per line with `description`, `files`, optional `scoping`, `model-override`, `effort-override`. Effort overrides use `low`, `medium`, `high`, or `xhigh`. If `--max-refactors` is omitted, `run` processes the file once and stops.
+- `--globs 'src/**/*.py:tests/**/*.py'` — colon-separated globs matched once against tracked files from `git ls-files`; each refactor action samples one matched file, so files can repeat.
+- `--extensions .py,.ts` — shorthand that expands to `**/*.py`, `**/*.ts` against tracked files from `git ls-files`; each refactor action samples one matched file, so files can repeat.
+- `--paths a.py:b.py` — literal user-provided paths, all treated as one grouped target; each refactor action reuses that group.
+- `--scope-instruction "clean up the auth module"` — extra free-text scoping. If selected file patterns resolve nothing, this becomes the useful fallback context.
 
-If you provide none of `--targets`, `--globs`, `--extensions`, or `--paths`, then `run` and `run-once` require `--scope-instruction`.
+If `--globs` or `--extensions` match no tracked files and there is no
+`--scope-instruction`, `run` completes successfully with zero refactor actions;
+`run-once` falls back to a no-file `general refactoring` target. `--paths` is
+literal input and is not filtered through `git ls-files`.
+
+If you provide none of `--targets`, `--globs`, `--extensions`, or `--paths`,
+then `run` and `run-once` require `--scope-instruction`; the driver still
+random-samples tracked files from `git ls-files` for each action and uses the
+scope text as context for that target.
 
 ### Migrations & taste flags
 
 - `init --live-migrations-dir PATH` — enables the larger-refactoring workflow for this project. The path is stored repo-relative in the project registry and created if missing.
 - `init --in-repo-taste [PATH]` — stores this project's taste file in the repo and remembers the repo-relative path. Defaults to `.continuous-refactoring/taste.md`; re-run `init --in-repo-taste ...` to choose a different path.
-- `taste --refine` — opens a collaborative editing session for the taste file. The agent keeps refining until you tell it to write, then the session ends automatically after the settled write.
-- `taste --upgrade` — re-interviews for taste dimensions added since your last version. No-op when already current; use `taste --refine` if you want to rework the doc anyway.
+- `migration list` — shows visible migrations as TSV with headers; `--awaiting-review` narrows to human-review handoffs and `--no-headers` keeps parser-friendly rows only.
+- `migration doctor <slug-or-path>` / `migration doctor --all` — read-only consistency checks. Doctor reports problems; it does not repair them.
+- `migration review <slug-or-path> --with ... --model ...` — resolves an `awaiting_human_review` migration through a staged workspace at fixed internal `high` effort.
+- `migration refine <slug-or-path> (--message <text>|--file <path>) --with ... --model ... [--show-agent-logs]` — adds user feedback to a planning or unexecuted ready migration and resumes planning through the `revise` step when reopening ready work at fixed internal `high` effort.
+- `taste --refine --with ... --model ...` — opens a collaborative editing session for the taste file. The agent keeps refining until you tell it to write, then the session ends automatically after the settled write.
+- `taste --upgrade --with ... --model ...` — re-interviews for taste dimensions added since your last version. No-op when already current; use `taste --refine` if you want to rework the doc anyway.
+- Taste agent sessions always use fixed `medium` effort.
 - `taste --force` — only applies to `--interview`; it allows a customized taste file to be overwritten after backing it up to `taste.md.bak`.
 
+Canonical migration commands:
+
+```bash
+continuous-refactoring migration list
+continuous-refactoring migration list --status planning
+continuous-refactoring migration list --awaiting-review
+continuous-refactoring migration list --no-headers
+continuous-refactoring migration doctor <slug-or-path>
+continuous-refactoring migration doctor --all
+continuous-refactoring migration review <slug-or-path> --with codex --model gpt-5
+continuous-refactoring migration refine <slug-or-path> --message "split the risky phase" --with codex --model gpt-5
+continuous-refactoring migration refine <slug-or-path> --file feedback.md --with codex --model gpt-5
+```
+
 ### Shared `run` / `run-once` flags
 
 - `--with`, `--model` — required agent backend/model selection.
 - `--default-effort` — default effort for run calls. Defaults to `low`. Valid labels are `low`, `medium`, `high`, `xhigh`.
 - `--max-allowed-effort` — cap for target overrides and migration escalation. Defaults to `xhigh`.
 - `--repo-root PATH` — repository root; defaults to the current directory.
-- `--validation-command` — defaults to `uv run pytest`. Swap it for whatever keeps your repo honest.
+- `--validation-command` — defaults to `uv run pytest`. This is parsed with `shlex.split` and run without a shell, so simple commands like `pytest -q` work, but shell syntax such as `cmd && cmd`, pipes, redirects, or leading `VAR=value` assignments is not interpreted. Put compound validation in a script.
 - `--timeout` — per-agent-call timeout in seconds.
 - `--show-agent-logs` / `--show-command-logs` — mirror output to your terminal instead of just logging.
 - `--refactoring-prompt` — override the default refactoring prompt.
@@ -187,10 +229,11 @@ If you provide none of `--targets`, `--globs`, `--extensions`, or `--paths`, the
 
 ### `run`-only flags
 
-- `--max-attempts N` — per-target retry budget. `1` = no retry, `0` = unlimited (which means permanently broken targets will never give up).
-- `--max-refactors N` — cap the number of targets per run. Required unless you use `--targets`.
-- `--max-consecutive-failures N` — bail after N targets fail in a row. Default 3.
-- `--sleep SECONDS` — pause between completed targets. Useful when you want a long batch without hammering the repo or your agent budget.
+- `--max-attempts N` — per-action retry budget. `1` = no retry, `0` = unlimited (which means permanently broken actions will never give up).
+- `--max-refactors N` — cap the number of refactor actions per run. Required unless you use `--targets` or `--focus-on-live-migrations`.
+- `--focus-on-live-migrations` — bypass target selection and `--max-refactors`; iterate eligible live migrations until they are done, deferred, blocked, or the failure budget trips.
+- `--max-consecutive-failures N` — bail after N actions fail in a row. Default 3.
+- `--sleep SECONDS` — pause between completed actions. Useful when you want a long batch without hammering the repo or your agent budget.
 - `--commit-message-prefix TEXT` — subject prefix for successful refactor or migration-plan commits. Default `continuous refactor`.
 
 ## Safety behaviors
@@ -207,13 +250,27 @@ If you provide none of `--targets`, `--globs`, `--extensions`, or `--paths`, the
 Each run writes to `$TMPDIR/continuous-refactoring/<run-id>/`:
 
 - `summary.json` — rolling status, counts, per-attempt stats
-- `events.jsonl` — structured event log
+- `events.jsonl` — structured event log with call roles such as
+  `scope-expansion`, `classify`, `planning.<step>`, `planning.publish`,
+  `phase.ready-check`, `phase.execute`, and `phase.validation`
 - `run.log` — human-readable log
 - `attempt-NNN/[retry-NN/]refactor/` — per-attempt agent + test stdout/stderr
+- `baseline/initial/` — baseline validation stdout/stderr before work starts
+- `classify/` — classifier agent stdout/stderr
+- `scope-expansion/` — scope candidates, selection, and bypass reason
+- `attempt-NNN/[retry-NN/]planning/<step>/` — planning agent stdout/stderr for
+  migration planning steps
+- `phase-ready-check/` — phase precondition agent stdout/stderr
+- `attempt-NNN/[retry-NN/]phase-execute/` — phase agent and validation logs
+- `migration-probes/action-NNN/` — migration probe logs during normal `run`
+  actions, including planning, phase ready-checks, and phase execution
 
 Mixed-effort runs are auditable: summaries and call events record the default effort, max allowed effort, requested effort, effective effort, source, and whether the request was capped.
 
-The path prints at startup. Grep it when something goes sideways.
+The path prints at startup. Grep it when something goes sideways. Failed
+non-commit decisions also write durable XDG snapshots under the project failure
+directory, usually
+`~/.local/share/continuous-refactoring/projects/<uuid>/failures/`.
 
 ## Taste files
 
@@ -222,7 +279,7 @@ The taste file is a short bullet list of your refactoring preferences. It gets i
 - Project taste: `~/.local/share/continuous-refactoring/projects/<uuid>/taste.md`, or the repo-local path chosen with `init --in-repo-taste [PATH]`
 - Global taste: `~/.local/share/continuous-refactoring/global/taste.md`
 
-Project taste wins over global. Use `taste` to print the active path, `taste --interview` to bootstrap one, `taste --refine` to rework it with an agent, or edit the file directly any time.
+Project taste wins over global. Use `taste` to print the active path, `taste --interview --with ... --model ...` to bootstrap one, `taste --refine --with ... --model ...` to rework it with an agent, or edit the file directly any time.
 
 ## Larger refactorings
 
@@ -241,7 +298,7 @@ This tells the CLI where to store migration artifacts. The path is repo-relative
 Each `run` / `run-once` tick now checks for eligible migration work before falling back to single-commit cleanups:
 
 1. **Classify** — a classifier agent reads the target and decides: `cohesive-cleanup` (one-shot path) or `needs-plan` (migration path).
-2. **Plan** — for `needs-plan` targets, a six-stage planning workflow runs: generate approaches → pick best → expand into phases → review → revise → final review. Artifacts land under `<live-migrations-dir>/<migration-name>/`.
+2. **Plan** — for `needs-plan` targets, each automation action runs exactly one planning step: approaches, pick-best, expand, review, optional revise/review-2, then final-review. Accepted steps update `.planning/state.json`, store stdout under `.planning/stages/`, and publish through a staged transaction. Failed current-step output stays in run artifacts and is not resume input.
 3. **Execute** — each phase is a self-contained unit of work. The tick picks the oldest eligible migration, checks whether its current phase precondition is satisfied, and executes it on the current branch. Phase completion is judged against the phase file's `## Definition of Done`; commit message identifies the migration as `migration/<name>/<phase-file>.md`.
 
 ### Migration directory layout
@@ -250,14 +307,20 @@ Each `run` / `run-once` tick now checks for eligible migration work before falli
 <live-migrations-dir>/
   <migration-name>/
     manifest.json          # status, phases, wake-up schedule
+    .planning/
+      state.json           # durable planning cursor and accepted step refs
+      stages/              # accepted planning stdout, suffixed on repeats
     plan.md                # the expanded plan
     approaches/            # candidate approaches considered during planning
     phase-1-<name>.md      # per-phase specification
     phase-2-<name>.md
     ...
+  __transactions__/        # internal staged publish state
   __intentional_skips__/   # migrations rejected at final review
 ```
 
+Do not edit `.planning/` or `__transactions__/` by hand. Use `migration doctor` when the shape looks wrong.
+
 ### Wake-up rules
 
 Migrations don't run on every tick. The scheduler now separates **activity** from
@@ -291,7 +354,7 @@ Before executing a phase, a ready-check agent verifies that the current phase pr
 
 - **ready: yes** — phase executes; on green tests, the phase is marked done, any prior deferral markers are cleared, and the migration advances immediately to the next phase.
 - **ready: no** — manifest activity is bumped, a retry cooldown is started, and a future `wake_up_on` is recorded when needed; the tick moves on.
-- **ready: unverifiable** — the migration is flagged `awaiting_human_review` and put on cooldown. Automated migration ticks skip flagged migrations until review clears the flag. Use `review list` to find it and `review perform <migration> --with ... --model ... --effort ...` to resolve it interactively.
+- **ready: unverifiable** — the migration is marked `awaiting_human_review` and put on cooldown. Automated migration ticks skip migrations awaiting human review until review clears the flag. Use `migration list --awaiting-review` to find it and `migration review <slug-or-path> --with ... --model ...` to resolve it interactively.
 
 Human-facing migration references use the relative phase spec path, for example `phase-2-failure-report.md`. The manifest cursor stores the phase `name`, not a numeric index.
 

{continuous_refactoring-0.2.0 → continuous_refactoring-1.0.0}/README.md

@@ -9,6 +9,8 @@ Small, test-gated cleanup commits by an AI coding agent.
 
 Think of it as a supervised janitor loop: the agent proposes a cleanup, your tests decide if it stays.
 
+Here's [an article](https://artisincode.com/essays/how-i-use-unspent-tokens/) I wrote about it.
+
 ## Install
 
 Try it without installing:
@@ -44,6 +46,10 @@ uv pip install -e .
 
 That gives you the `continuous-refactoring` command.
 
+The CLI itself can be installed without `uv`, but the default validation command
+is `uv run pytest`. Pass `--validation-command pytest` or a project-specific
+script when the target repo does not use `uv`.
+
 Maintainers: see the [release checklist](https://github.com/bigH/continuous-refactoring/blob/main/docs/release.md).
 
 ## Fastest way to get one refactor
@@ -69,15 +75,18 @@ continuous-refactoring run \
   --max-attempts 2
 ```
 
-That keeps sweeping targets until it runs out, hits your caps, or starts failing.
+That runs up to 10 refactor actions, then stops sooner if the finite target file
+runs out or the loop starts failing. Use `run --focus-on-live-migrations` when
+you want the loop to work only on eligible live migrations; it bypasses target
+selection and `--max-refactors`.
 
 ## What it does
 
-- Resolves a target from `--targets`, `--globs`, `--extensions`, or `--paths`, with optional natural-language scoping via `--scope-instruction`.
+- Resolves each source action from `--targets`, `--globs`, `--extensions`, or `--paths`, with optional natural-language scoping via `--scope-instruction`.
 - Runs the agent with a refactoring prompt + your "taste" guidelines.
 - Runs your validation command (default: `uv run pytest`).
 - If green and there's a diff, it commits locally and leaves the branch for you to inspect.
-- Repeats until it runs out of targets, hits the retry budget, or stacks too many failures.
+- Repeats until it spends the action budget, exhausts a finite target file, hits the retry budget, or stacks too many failures.
 
 ## Requirements
 
@@ -96,8 +105,8 @@ continuous-refactoring init --in-repo-taste
 
 # 2. (Optional) Write your refactoring taste — either edit the file, have an agent interview you,
 #    or refine an existing draft collaboratively
-continuous-refactoring taste --interview --with codex --model gpt-5 --effort high
-continuous-refactoring taste --refine --with codex --model gpt-5 --effort high
+continuous-refactoring taste --interview --with codex --model gpt-5
+continuous-refactoring taste --refine --with codex --model gpt-5
 
 # 3. Do one pass
 continuous-refactoring run-once \
@@ -113,17 +122,23 @@ continuous-refactoring run \
   --sleep 5
 ```
 
+Active taste agent modes require `--with` and `--model`; taste agent sessions
+always run at fixed `medium` effort.
+
 ## Subcommands
 
 | Command | What it does |
 |---|---|
 | `init` | Registers this directory as a project, creates a default `taste.md`, and can store `--live-migrations-dir` or `--in-repo-taste`. |
-| `taste` | Prints the active taste file path. Add `--interview` to have an agent author it, `--refine` to iteratively improve an existing taste doc, `--upgrade` to refresh stale taste dimensions, `--global` for the shared file, and `--force` to let `--interview` overwrite custom content after writing a `.bak`. |
+| `taste` | Prints the active taste file path. Add `--interview` to have an agent author it, `--refine` to iteratively improve an existing taste doc, `--upgrade` to refresh stale taste dimensions, `--global` for the shared file, and `--force` to let `--interview` overwrite custom content after writing a `.bak`. Active agent modes require `--with` and `--model`, then run at fixed `medium` effort. |
 | `run-once` | Single pass on one resolved target. No retry. If there is a diff and validation passes, it commits locally and prints the diffstat. |
-| `run` | The loop. Iterates targets, retries on failure, and commits successful targets locally. |
+| `run` | The loop. Iterates refactor actions, retries on failure, and commits successful changes locally. Add `--focus-on-live-migrations` to bypass targeting and work only on eligible live migrations. |
 | `upgrade` | Checks that the global config manifest is current, rewrites it idempotently, and warns if the global taste file is stale. |
-| `review list` | Lists migrations flagged for human review (`awaiting_human_review`). |
-| `review perform <migration>` | Starts an interactive agent session to resolve a flagged migration's review. Requires `--with`, `--model`, and `--effort`. |
+| `migration list` | Lists visible migrations as TSV with headers. Add `--status <status>` or `--awaiting-review` to filter, or `--no-headers` for parsing. |
+| `migration doctor <slug-or-path>` | Validates one visible migration's consistency. |
+| `migration doctor --all` | Validates every visible migration plus internal transaction state. |
+| `migration review <slug-or-path>` | Starts staged review for a migration awaiting human review. Requires `--with` and `--model`; review runs at fixed internal `high` effort. |
+| `migration refine <slug-or-path>` | Records feedback for a planning or unexecuted ready migration and runs one staged planning revision. Requires `--message <text>` or `--file <path>`, plus `--with` and `--model`; refine runs at fixed internal `high` effort. Add `--show-agent-logs` to mirror the planning agent. |
 
 ## Targeting / Useful flags
 
@@ -134,29 +149,56 @@ Target resolution is first-match-wins:
 
 These flags are not mutually exclusive, but only the highest-priority populated source is used.
 
-- `--targets path/to/targets.jsonl` — explicit list; one JSON object per line with `description`, `files`, optional `scoping`, `model-override`, `effort-override`. Effort overrides use `low`, `medium`, `high`, or `xhigh`.
-- `--globs 'src/**/*.py:tests/**/*.py'` — colon-separated globs; each matched file becomes its own target.
-- `--extensions .py,.ts` — shorthand that expands to `**/*.py`, `**/*.ts`; each matched file becomes its own target.
-- `--paths a.py:b.py` — literal paths, all treated as one target.
-- `--scope-instruction "clean up the auth module"` — extra free-text scoping. If file-based targeting resolves nothing, this becomes the useful fallback context.
+- `--targets path/to/targets.jsonl` — explicit finite list; one JSON object per line with `description`, `files`, optional `scoping`, `model-override`, `effort-override`. Effort overrides use `low`, `medium`, `high`, or `xhigh`. If `--max-refactors` is omitted, `run` processes the file once and stops.
+- `--globs 'src/**/*.py:tests/**/*.py'` — colon-separated globs matched once against tracked files from `git ls-files`; each refactor action samples one matched file, so files can repeat.
+- `--extensions .py,.ts` — shorthand that expands to `**/*.py`, `**/*.ts` against tracked files from `git ls-files`; each refactor action samples one matched file, so files can repeat.
+- `--paths a.py:b.py` — literal user-provided paths, all treated as one grouped target; each refactor action reuses that group.
+- `--scope-instruction "clean up the auth module"` — extra free-text scoping. If selected file patterns resolve nothing, this becomes the useful fallback context.
 
-If you provide none of `--targets`, `--globs`, `--extensions`, or `--paths`, then `run` and `run-once` require `--scope-instruction`.
+If `--globs` or `--extensions` match no tracked files and there is no
+`--scope-instruction`, `run` completes successfully with zero refactor actions;
+`run-once` falls back to a no-file `general refactoring` target. `--paths` is
+literal input and is not filtered through `git ls-files`.
+
+If you provide none of `--targets`, `--globs`, `--extensions`, or `--paths`,
+then `run` and `run-once` require `--scope-instruction`; the driver still
+random-samples tracked files from `git ls-files` for each action and uses the
+scope text as context for that target.
 
 ### Migrations & taste flags
 
 - `init --live-migrations-dir PATH` — enables the larger-refactoring workflow for this project. The path is stored repo-relative in the project registry and created if missing.
 - `init --in-repo-taste [PATH]` — stores this project's taste file in the repo and remembers the repo-relative path. Defaults to `.continuous-refactoring/taste.md`; re-run `init --in-repo-taste ...` to choose a different path.
-- `taste --refine` — opens a collaborative editing session for the taste file. The agent keeps refining until you tell it to write, then the session ends automatically after the settled write.
-- `taste --upgrade` — re-interviews for taste dimensions added since your last version. No-op when already current; use `taste --refine` if you want to rework the doc anyway.
+- `migration list` — shows visible migrations as TSV with headers; `--awaiting-review` narrows to human-review handoffs and `--no-headers` keeps parser-friendly rows only.
+- `migration doctor <slug-or-path>` / `migration doctor --all` — read-only consistency checks. Doctor reports problems; it does not repair them.
+- `migration review <slug-or-path> --with ... --model ...` — resolves an `awaiting_human_review` migration through a staged workspace at fixed internal `high` effort.
+- `migration refine <slug-or-path> (--message <text>|--file <path>) --with ... --model ... [--show-agent-logs]` — adds user feedback to a planning or unexecuted ready migration and resumes planning through the `revise` step when reopening ready work at fixed internal `high` effort.
+- `taste --refine --with ... --model ...` — opens a collaborative editing session for the taste file. The agent keeps refining until you tell it to write, then the session ends automatically after the settled write.
+- `taste --upgrade --with ... --model ...` — re-interviews for taste dimensions added since your last version. No-op when already current; use `taste --refine` if you want to rework the doc anyway.
+- Taste agent sessions always use fixed `medium` effort.
 - `taste --force` — only applies to `--interview`; it allows a customized taste file to be overwritten after backing it up to `taste.md.bak`.
 
+Canonical migration commands:
+
+```bash
+continuous-refactoring migration list
+continuous-refactoring migration list --status planning
+continuous-refactoring migration list --awaiting-review
+continuous-refactoring migration list --no-headers
+continuous-refactoring migration doctor <slug-or-path>
+continuous-refactoring migration doctor --all
+continuous-refactoring migration review <slug-or-path> --with codex --model gpt-5
+continuous-refactoring migration refine <slug-or-path> --message "split the risky phase" --with codex --model gpt-5
+continuous-refactoring migration refine <slug-or-path> --file feedback.md --with codex --model gpt-5
+```
+
 ### Shared `run` / `run-once` flags
 
 - `--with`, `--model` — required agent backend/model selection.
 - `--default-effort` — default effort for run calls. Defaults to `low`. Valid labels are `low`, `medium`, `high`, `xhigh`.
 - `--max-allowed-effort` — cap for target overrides and migration escalation. Defaults to `xhigh`.
 - `--repo-root PATH` — repository root; defaults to the current directory.
-- `--validation-command` — defaults to `uv run pytest`. Swap it for whatever keeps your repo honest.
+- `--validation-command` — defaults to `uv run pytest`. This is parsed with `shlex.split` and run without a shell, so simple commands like `pytest -q` work, but shell syntax such as `cmd && cmd`, pipes, redirects, or leading `VAR=value` assignments is not interpreted. Put compound validation in a script.
 - `--timeout` — per-agent-call timeout in seconds.
 - `--show-agent-logs` / `--show-command-logs` — mirror output to your terminal instead of just logging.
 - `--refactoring-prompt` — override the default refactoring prompt.
@@ -164,10 +206,11 @@ If you provide none of `--targets`, `--globs`, `--extensions`, or `--paths`, the
 
 ### `run`-only flags
 
-- `--max-attempts N` — per-target retry budget. `1` = no retry, `0` = unlimited (which means permanently broken targets will never give up).
-- `--max-refactors N` — cap the number of targets per run. Required unless you use `--targets`.
-- `--max-consecutive-failures N` — bail after N targets fail in a row. Default 3.
-- `--sleep SECONDS` — pause between completed targets. Useful when you want a long batch without hammering the repo or your agent budget.
+- `--max-attempts N` — per-action retry budget. `1` = no retry, `0` = unlimited (which means permanently broken actions will never give up).
+- `--max-refactors N` — cap the number of refactor actions per run. Required unless you use `--targets` or `--focus-on-live-migrations`.
+- `--focus-on-live-migrations` — bypass target selection and `--max-refactors`; iterate eligible live migrations until they are done, deferred, blocked, or the failure budget trips.
+- `--max-consecutive-failures N` — bail after N actions fail in a row. Default 3.
+- `--sleep SECONDS` — pause between completed actions. Useful when you want a long batch without hammering the repo or your agent budget.
 - `--commit-message-prefix TEXT` — subject prefix for successful refactor or migration-plan commits. Default `continuous refactor`.
 
 ## Safety behaviors
@@ -184,13 +227,27 @@ If you provide none of `--targets`, `--globs`, `--extensions`, or `--paths`, the
 Each run writes to `$TMPDIR/continuous-refactoring/<run-id>/`:
 
 - `summary.json` — rolling status, counts, per-attempt stats
-- `events.jsonl` — structured event log
+- `events.jsonl` — structured event log with call roles such as
+  `scope-expansion`, `classify`, `planning.<step>`, `planning.publish`,
+  `phase.ready-check`, `phase.execute`, and `phase.validation`
 - `run.log` — human-readable log
 - `attempt-NNN/[retry-NN/]refactor/` — per-attempt agent + test stdout/stderr
+- `baseline/initial/` — baseline validation stdout/stderr before work starts
+- `classify/` — classifier agent stdout/stderr
+- `scope-expansion/` — scope candidates, selection, and bypass reason
+- `attempt-NNN/[retry-NN/]planning/<step>/` — planning agent stdout/stderr for
+  migration planning steps
+- `phase-ready-check/` — phase precondition agent stdout/stderr
+- `attempt-NNN/[retry-NN/]phase-execute/` — phase agent and validation logs
+- `migration-probes/action-NNN/` — migration probe logs during normal `run`
+  actions, including planning, phase ready-checks, and phase execution
 
 Mixed-effort runs are auditable: summaries and call events record the default effort, max allowed effort, requested effort, effective effort, source, and whether the request was capped.
 
-The path prints at startup. Grep it when something goes sideways.
+The path prints at startup. Grep it when something goes sideways. Failed
+non-commit decisions also write durable XDG snapshots under the project failure
+directory, usually
+`~/.local/share/continuous-refactoring/projects/<uuid>/failures/`.
 
 ## Taste files
 
@@ -199,7 +256,7 @@ The taste file is a short bullet list of your refactoring preferences. It gets i
 - Project taste: `~/.local/share/continuous-refactoring/projects/<uuid>/taste.md`, or the repo-local path chosen with `init --in-repo-taste [PATH]`
 - Global taste: `~/.local/share/continuous-refactoring/global/taste.md`
 
-Project taste wins over global. Use `taste` to print the active path, `taste --interview` to bootstrap one, `taste --refine` to rework it with an agent, or edit the file directly any time.
+Project taste wins over global. Use `taste` to print the active path, `taste --interview --with ... --model ...` to bootstrap one, `taste --refine --with ... --model ...` to rework it with an agent, or edit the file directly any time.
 
 ## Larger refactorings
 
@@ -218,7 +275,7 @@ This tells the CLI where to store migration artifacts. The path is repo-relative
 Each `run` / `run-once` tick now checks for eligible migration work before falling back to single-commit cleanups:
 
 1. **Classify** — a classifier agent reads the target and decides: `cohesive-cleanup` (one-shot path) or `needs-plan` (migration path).
-2. **Plan** — for `needs-plan` targets, a six-stage planning workflow runs: generate approaches → pick best → expand into phases → review → revise → final review. Artifacts land under `<live-migrations-dir>/<migration-name>/`.
+2. **Plan** — for `needs-plan` targets, each automation action runs exactly one planning step: approaches, pick-best, expand, review, optional revise/review-2, then final-review. Accepted steps update `.planning/state.json`, store stdout under `.planning/stages/`, and publish through a staged transaction. Failed current-step output stays in run artifacts and is not resume input.
 3. **Execute** — each phase is a self-contained unit of work. The tick picks the oldest eligible migration, checks whether its current phase precondition is satisfied, and executes it on the current branch. Phase completion is judged against the phase file's `## Definition of Done`; commit message identifies the migration as `migration/<name>/<phase-file>.md`.
 
 ### Migration directory layout
@@ -227,14 +284,20 @@ Each `run` / `run-once` tick now checks for eligible migration work before falli
 <live-migrations-dir>/
   <migration-name>/
     manifest.json          # status, phases, wake-up schedule
+    .planning/
+      state.json           # durable planning cursor and accepted step refs
+      stages/              # accepted planning stdout, suffixed on repeats
     plan.md                # the expanded plan
     approaches/            # candidate approaches considered during planning
     phase-1-<name>.md      # per-phase specification
     phase-2-<name>.md
     ...
+  __transactions__/        # internal staged publish state
   __intentional_skips__/   # migrations rejected at final review
 ```
 
+Do not edit `.planning/` or `__transactions__/` by hand. Use `migration doctor` when the shape looks wrong.
+
 ### Wake-up rules
 
 Migrations don't run on every tick. The scheduler now separates **activity** from
@@ -268,7 +331,7 @@ Before executing a phase, a ready-check agent verifies that the current phase pr
 
 - **ready: yes** — phase executes; on green tests, the phase is marked done, any prior deferral markers are cleared, and the migration advances immediately to the next phase.
 - **ready: no** — manifest activity is bumped, a retry cooldown is started, and a future `wake_up_on` is recorded when needed; the tick moves on.
-- **ready: unverifiable** — the migration is flagged `awaiting_human_review` and put on cooldown. Automated migration ticks skip flagged migrations until review clears the flag. Use `review list` to find it and `review perform <migration> --with ... --model ... --effort ...` to resolve it interactively.
+- **ready: unverifiable** — the migration is marked `awaiting_human_review` and put on cooldown. Automated migration ticks skip migrations awaiting human review until review clears the flag. Use `migration list --awaiting-review` to find it and `migration review <slug-or-path> --with ... --model ...` to resolve it interactively.
 
 Human-facing migration references use the relative phase spec path, for example `phase-2-failure-report.md`. The manifest cursor stores the phase `name`, not a numeric index.
 

{continuous_refactoring-0.2.0 → continuous_refactoring-1.0.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "continuous-refactoring"
-version = "0.2.0"
+version = "1.0.0"
 description = "Continuous refactoring loop for AI coding agents"
 readme = "README.md"
 requires-python = ">=3.10"