regent-cli 0.1.0__tar.gz → 0.2.0__tar.gz

{regent_cli-0.1.0/src/regent_cli.egg-info → regent_cli-0.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: regent-cli
-Version: 0.1.0
+Version: 0.2.0
 Summary: Autonomous conduction and mediated adversarial deliberation between AI agents, pluggable into any project.
 Author-email: Flavio Alvim <flavioalvim@gmail.com>
 License: MIT

{regent_cli-0.1.0 → regent_cli-0.2.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "regent-cli"
-version = "0.1.0"
+version = "0.2.0"
 description = "Autonomous conduction and mediated adversarial deliberation between AI agents, pluggable into any project."
 readme = "README.md"
 requires-python = ">=3.10"

{regent_cli-0.1.0 → regent_cli-0.2.0}/src/regent/__init__.py

@@ -1,3 +1,3 @@
 """regent — autonomous conduction and mediated adversarial deliberation between AI agents."""
 
-__version__ = "0.1.0"
+__version__ = "0.2.0"

{regent_cli-0.1.0 → regent_cli-0.2.0}/src/regent/initcmd.py

@@ -41,7 +41,8 @@ def _plan(project_root: Path) -> list[tuple[str, Path, str]]:
                      _template_skill(name)))
         plan.append(("symlink", project_root / ".claude" / "skills" / name,
                      f"../../.regent/skills/{name}"))
-    plan.append(("file", project_root / ".regent" / "brainstorm" / "rodadas" / ".gitkeep", ""))
+    plan.append(("file", project_root / ".regent" / "brainstorm" / "rounds" / ".gitkeep", ""))
+    plan.append(("file", project_root / ".regent" / "plans" / ".gitkeep", ""))
     return plan
 
 

regent_cli-0.2.0/src/regent/templates/skills/regent/SKILL.md

@@ -0,0 +1,119 @@
+---
+name: regent
+description: Start or resume the regent activity in this repo — brainstorm (adversarial deliberation round), plan (deliberated plan with steps and gates) or build (execute an approved plan step by step). Single state-driven entry point. Use when the owner says /regent, "retome", "próxima rodada", or asks to continue regent work.
+---
+
+# /regent — start or resume the current activity
+
+> **Capability level: v0 (file-driven).** State detection is based on files only. There is
+> no `control.json`, no turn lock, no daemon, no confined execution and no atomic abort yet.
+> Immediate interruption of a running step is the user's Esc; `/regent-stop` suspends at the
+> next boundary. Do not claim capabilities beyond this file.
+
+## 0. Locate the state roots (deterministic, never guess)
+
+- Host projects, current scheme: `ROUNDS = .regent/brainstorm/rounds/` (dirs `ROUND-NNN/`,
+  English artifact names) and `PLANS = .regent/plans/` (dirs `PLAN-NNN/`).
+- Host projects, legacy scheme (pre-0.2 seeds): `.regent/brainstorm/rodadas/` with
+  `RODADA-NNN/` and PT artifact names (PERGUNTA/RESPOSTA-CLAUDE/OPINIAO-CODEX-N/
+  REPLICA-CLAUDE/DECISAO/SUSPENSAO). A legacy host keeps operating on its PT scheme for
+  brainstorm; `plan`/`build` ALWAYS use the English `PLANS` scheme.
+- If BOTH `rounds/` and `rodadas/` exist non-empty → corrupted state: report and STOP with
+  manual migration instructions (default-deny; never auto-migrate).
+- The regent product repo itself keeps its own deliberation in `docs/brainstorm/rodadas/`
+  (dogfood location, PT names) — delimited by that fixed path only.
+
+Artifact names below use the English scheme; on a legacy host map brainstorm names to PT.
+
+## 1. Detect state (never guess)
+
+Open activities (at most ONE may exist):
+- **brainstorm**: a `ROUND-NNN/` without `DECISION.md`.
+- **planning**: a `PLAN-NNN/` without `APPROVAL.md`.
+- **building**: a plan whose `APPROVAL.md` has `status: APPROVED`, with `build/` started and
+  no `build/CONCLUSION.md`.
+
+An APPROVED plan with no `build/` is NOT open — it is a build candidate awaiting the owner's
+explicit `/regent build`. More than one open activity, or otherwise ambiguous/corrupted
+state → report and STOP. Never repair silently, never pick one arbitrarily.
+
+## 2. Decide by command argument × state
+
+- **Bare `/regent` + exactly one open activity** → resume it (sections 3–5).
+- **Bare `/regent` + nothing open** → report state (last closed round/plan, build
+  candidates) and ask what to start. Never start anything implicitly.
+- **`brainstorm "<question>"` / `plan "<goal>"` + nothing open** → create the next
+  `ROUND-NNN/` (with `QUESTION.md`) or `PLAN-NNN/` (with `REQUEST.md`), owner's words
+  verbatim, and proceed.
+- **`build`** → requires a build candidate: exactly one → start/resume it; more than one →
+  demand explicit selection `/regent build PLAN-NNN`; none → error (never build without an
+  APPROVED plan).
+- **Any mode argument + a DIFFERENT open activity** → error explaining the state. Never
+  ignore the argument silently; never create a second activity.
+
+## 3. Brainstorm mode
+
+Loop: owner asks → Claude responds → advisor opines (explicit CONCORDA/DISCORDA verdict) →
+CONCORDA: Claude executes; DISCORDA: one rebuttal cycle; persistent divergence → the owner
+arbitrates. Round content in the mediator's language (REQ-002).
+
+Resume at the first missing artifact: `QUESTION.md` → `CLAUDE-RESPONSE.md` →
+`ADVISOR-OPINION-1.md` → (if DISCORDA) `CLAUDE-REBUTTAL.md` → `ADVISOR-OPINION-2.md` →
+`DECISION.md` (what was decided, who closed it, what was executed), then execute and make an
+**operational commit** (regent-owned paths only). If resuming a suspended activity, read
+`SUSPENSION.md` first, delete it after resuming.
+
+## 4. Plan mode
+
+A plan is a deliberated artifact, driven like a round:
+1. `REQUEST.md` — the owner's goal, verbatim.
+2. `PLAN.md` — goal, scope in/out, **numbered steps, each with an acceptance criterion and
+   an executable gate command**, risks. Content in the host's native language.
+3. `ADVISOR-REVIEW-1.md` — advisor verdict; one rebuttal cycle (`CLAUDE-REBUTTAL.md` +
+   `ADVISOR-REVIEW-2.md`); persistent divergence → owner arbitrates.
+4. `APPROVAL.md` — structured: `status: APPROVED | REJECTED | CANCELLED`, actor (consensus /
+   owner), date. Without APPROVED status the plan is not executable. The owner may cancel a
+   plan at any time (record status + reason).
+
+## 5. Build mode — execute an APPROVED plan, step by step
+
+**Baseline.** On build start write `build/BASELINE.md`: the base SHA (`git rev-parse HEAD`)
+and the plan reference. The final review diff is exactly `BASE-SHA..HEAD`.
+
+**Per step (in plan order):**
+1. Worktree MUST be clean at step start; dirty → report and STOP (default-deny — never
+   absorb pre-existing user changes).
+2. Implement the step.
+3. Run the step's gate command for real; record its output. Red gate → fix and re-run, or
+   suspend; a red gate NEVER proceeds and NEVER commits.
+4. Write `build/STEP-NN.md`: base SHA of the step, files touched, gate output. NO commit
+   hash inside (it cannot know it — see trailer below).
+5. Commit deliberately: stage file-by-file ONLY the paths attributable to the step (+ the
+   step's regent artifacts), inspect the staged diff, fail without committing if anything
+   unattributable is staged. Commit message ends with trailer `Regent-Step: PLAN-NNN/STEP-NN`
+   — the artifact↔commit link is resolved via this trailer, never via a hash in the artifact.
+
+**Resume phases** (check trailer in `git log`, then STEP file, then worktree — in that order):
+- IMPLEMENTING (dirty worktree, no gate recorded) → re-run the gate before anything; never
+  trust unvalidated code.
+- GATE-RED (red output recorded) → back to implementation.
+- GATE-GREEN-UNCOMMITTED (STEP written, trailer absent) → re-verify gate, then commit.
+  Checking the trailer BEFORE committing prevents duplicate commits.
+- COMMITTED (trailer present) → step closed; advance.
+
+**Finish:** `build/ADVISOR-REVIEW.md` — advisor reviews the full `BASE-SHA..HEAD` diff
+(verdict + findings). Then `build/CONCLUSION.md`, structured: `status: ACCEPTED |
+ACCEPTED-WITH-RESERVATIONS | REJECTED`, actor, pending items. An adverse review or a
+non-arbitrated divergence can NEVER become acceptance. Any post-review fix invalidates the
+review: re-run the affected gate AND a new advisor consultation before concluding.
+
+## 6. Advisor consultations (shared contract — REQ-003 §5)
+
+Invocation (read-only, never writes, never holds the turn):
+`codex --ask-for-approval never --sandbox read-only exec --cd <repo> -o <tmpfile> "<prompt>"`
+
+Every consultation persists its full evidence tuple: the complete prompt in
+`<ARTIFACT>-PROMPT.md` next to the response artifact, and the response artifact opening with
+a structured header — `outcome: SUCCESS|TIMEOUT|FAILURE|CANCELLED`, `exit_code`,
+`timestamp`, `linkage` (ROUND-NNN / PLAN-NNN[/build]). Fail-closed: a non-SUCCESS outcome
+never advances the activity; retry means a NEW recorded consultation.

regent_cli-0.2.0/src/regent/templates/skills/regent-stop/SKILL.md

@@ -0,0 +1,32 @@
+---
+name: regent-stop
+description: Safely stop/suspend the regent activity in progress (brainstorm round, plan deliberation, or build) so it can be resumed later with /regent. Use when the owner says /regent-stop, "para", "suspende", or asks to stop regent work.
+---
+
+# /regent-stop — safe stop of the current activity
+
+> **Capability level: v0 (file-driven).** There is no daemon, lock or in-flight abort yet:
+> stopping means recording a durable suspension marker at the next message boundary. Do not
+> claim atomic cancellation — immediate interruption of a running step is the user's Esc,
+> not this command.
+
+## Steps
+
+1. Locate the state roots and the single open activity exactly as `/regent` does (its
+   sections 0–1: brainstorm round without DECISION, plan without APPROVAL, or build without
+   CONCLUSION; legacy PT scheme respected; ambiguity = report and stop, touch nothing).
+   Nothing open → report that the state is clean (nothing to stop) and stop.
+2. Write `SUSPENSION.md` (legacy hosts: `SUSPENSAO.md`) inside the open activity dir with:
+   timestamp, reason given by the owner (ask if not stated), and the resume checkpoint —
+   for brainstorm/plan: which artifact was pending; for build: the current step and its
+   phase (IMPLEMENTING / GATE-RED / GATE-GREEN-UNCOMMITTED / COMMITTED) plus any in-flight
+   consultation outcome.
+3. Persist everything already produced — never delete partial artifacts; they are evidence.
+   A build step in IMPLEMENTING phase keeps its uncommitted worktree changes untouched;
+   record in SUSPENSION.md that the worktree is intentionally dirty.
+4. Make an operational commit of ONLY regent-owned paths (`.regent/`, the `.claude/` symlink
+   integrations, and — in the regent repo only — `docs/`). Never commit host code from here:
+   deliberate build-step commits belong to `/regent` (REQ-005). A failed commit does NOT
+   prevent the suspension — report it and leave the commit pending.
+5. Confirm to the owner: what was suspended, at which checkpoint, and that `/regent`
+   resumes it.

{regent_cli-0.1.0 → regent_cli-0.2.0/src/regent_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: regent-cli
-Version: 0.1.0
+Version: 0.2.0
 Summary: Autonomous conduction and mediated adversarial deliberation between AI agents, pluggable into any project.
 Author-email: Flavio Alvim <flavioalvim@gmail.com>
 License: MIT

{regent_cli-0.1.0 → regent_cli-0.2.0}/tests/test_initcmd.py

@@ -31,7 +31,8 @@ class InitTest(unittest.TestCase):
             self.assertTrue(link.is_symlink())
             self.assertEqual(os.readlink(link), f"../../.regent/skills/{name}")
             self.assertTrue((link / "SKILL.md").is_file())  # symlink resolves
-        self.assertTrue((self.root / ".regent" / "brainstorm" / "rodadas").is_dir())
+        self.assertTrue((self.root / ".regent" / "brainstorm" / "rounds").is_dir())
+        self.assertTrue((self.root / ".regent" / "plans").is_dir())
 
     def test_rerun_is_noop(self):
         self.assertEqual(self._init()[0], EXIT_OK)

regent_cli-0.1.0/src/regent/templates/skills/regent/SKILL.md

@@ -1,60 +0,0 @@
----
-name: regent
-description: Start or resume the regent activity in this repo (brainstorm round, planning, implementation). Single state-driven entry point — detects what is open and drives the next step. Use when the owner says /regent, "retome o brainstorm", "próxima rodada", or asks to continue regent work.
----
-
-# /regent — start or resume the current activity
-
-> **Capability level: v0 (file-driven).** State detection is based on round files only.
-> There is no `control.json`, no turn lock, no daemon, no atomic abort and no structured
-> checkpoint yet — those arrive with REQ-004's full implementation. Do not claim
-> capabilities beyond this file.
-
-## 0. Locate the rounds directory (deterministic, never guess)
-
-- Host projects: `ROUNDS = .regent/brainstorm/rodadas/` (created on first use).
-- The regent product repo itself keeps its own deliberation in `docs/brainstorm/rodadas/`
-  (legacy dogfood location).
-- Resolution: if `.regent/brainstorm/` exists use it; else if `docs/brainstorm/` exists use
-  that; if BOTH exist, report the ambiguity and STOP; if neither, this is a fresh host —
-  `ROUNDS = .regent/brainstorm/rodadas/`.
-
-## 1. Detect state (never guess)
-
-- Open round = a `ROUNDS/RODADA-NNN/` directory WITHOUT `DECISAO.md`.
-- Suspended round = an open round containing `SUSPENSAO.md`.
-- If the state is ambiguous or corrupted (two open rounds, missing PERGUNTA.md), REPORT the
-  anomaly and STOP. Default-deny: never repair silently, never pick one arbitrarily.
-
-## 2. Decide by command argument × state
-
-- **No argument + exactly one open/suspended round** → resume it (step 3).
-- **No argument + nothing open** → report the state (last closed round, PRD requirement log)
-  and ask what to start. Never start anything implicitly.
-- **`brainstorm "<question>"` + nothing open** → create the next `ROUNDS/RODADA-NNN/` with
-  `PERGUNTA.md` (owner's question verbatim) and proceed to step 3.
-- **`brainstorm ...` + a DIFFERENT round already open** → error explaining the open state.
-  Never ignore the argument silently; never create a second activity.
-- **`plan` / `build`** → not available at v0; say so explicitly and point to REQ-004.
-
-## 3. Drive the round (the loop protocol is THIS section — self-contained)
-
-Loop: owner asks → Claude responds → Codex (advisor) opines with an explicit
-CONCORDA/DISCORDA verdict → if CONCORDA, Claude executes; if DISCORDA, one rebuttal cycle;
-persistent divergence → the owner arbitrates. Round language: the mediator's (REQ-002).
-
-Resume at the first missing artifact, in order:
-1. `RESPOSTA-CLAUDE.md` — Claude's reasoned position + concrete proposal.
-2. `OPINIAO-CODEX-1.md` — headless advisor consultation, saved verbatim:
-   `codex --ask-for-approval never --sandbox read-only exec --cd <repo> -o <tmpfile> "<prompt>"`
-   (prompt tells Codex which files to read and to end with a line containing only
-   CONCORDA or DISCORDA).
-3. If DISCORDA → `REPLICA-CLAUDE.md` then `OPINIAO-CODEX-2.md` (one rebuttal cycle max).
-4. Persistent divergence → the owner arbitrates (they are mediating live).
-5. `DECISAO.md` — what was decided, who closed it, what was executed. Then execute it and
-   commit ONLY regent-owned paths (`.regent/`, the `.claude/` symlink integrations, and —
-   in the regent repo only — `docs/`) — never unrelated host content. A failed commit does
-   not undo the round; report it and leave it pending.
-
-If resuming a suspended round, read `SUSPENSAO.md` first, delete it after resuming, and
-continue from the recorded pending step.

regent_cli-0.1.0/src/regent/templates/skills/regent-stop/SKILL.md

@@ -1,27 +0,0 @@
----
-name: regent-stop
-description: Safely stop/suspend the regent activity in progress (brainstorm round, planning, implementation) so it can be resumed later with /regent. Use when the owner says /regent-stop, "para o brainstorm", "suspende a rodada", or asks to stop regent work.
----
-
-# /regent-stop — safe stop of the current activity
-
-> **Capability level: v0 (file-driven).** There is no daemon, lock or in-flight abort yet:
-> stopping means recording a durable suspension marker at the next message boundary. Do not
-> claim atomic cancellation — immediate interruption of a running step is the user's Esc,
-> not this command.
-
-## Steps
-
-1. Locate the rounds directory exactly as `/regent` does (step 0 of its skill: host =
-   `.regent/brainstorm/rodadas/`; the regent repo itself = `docs/brainstorm/rodadas/`; both
-   present = ambiguity, stop). Then detect the open round: `RODADA-NNN/` without `DECISAO.md`.
-   - Nothing open → report that the state is clean (nothing to stop) and stop.
-   - Ambiguous state → report the anomaly; do not modify anything.
-2. Write `SUSPENSAO.md` in the open round with: timestamp, reason given by the owner (ask if
-   not stated), which artifact/step was pending (the resume checkpoint), and any partial
-   output worth keeping (e.g. an advisor consultation already saved).
-3. Persist everything already produced — never delete partial artifacts; they are evidence.
-4. Commit ONLY regent-owned paths (`.regent/`, the `.claude/` symlink integrations, and —
-   in the regent repo only — `docs/`). A failed commit does NOT prevent the suspension —
-   report it and leave the commit pending.
-5. Confirm to the owner: what was suspended, where it will resume from (`/regent`).