Release v1.0.0

.gemini/GEMINI.md

@@ -0,0 +1 @@
+../AGENTS.md

.github/workflows/dogfood.yml

@@ -19,11 +19,15 @@ on:
   pull_request_review_comment:
     types: [created]
 
+# NOTE: `workflows: write` is not a valid permissions scope — GitHub began
+# hard-rejecting it ~2026-05-01 ("Unexpected value 'workflows'"), which made
+# every push spawn a phantom failed dogfood run and broke /dogfood comments.
+# Workflow-file write access is granted by the App's Workflows permission
+# (see CONTRIBUTING.md), not by a job permissions key.
 permissions:
   contents: write
   issues: write
   pull-requests: write
-  workflows: write
 
 jobs:
   resolve:

.github/workflows/test.yml

@@ -2,7 +2,7 @@ name: Tests
 
 on:
   pull_request:
-    branches: [main]
+    branches: [main, dev]
 
 jobs:
   test:

CHANGELOG.md

@@ -1,5 +1,40 @@
 # Changelog
 
+## v1.0.0 (Aug 2026)
+
+- **Model aliases: three capability tiers.** Aliases are now
+  `{provider}-{small,medium,large}`: small = best value-for-money
+  agentic-coding workhorse, medium = the provider's flagship, large =
+  frontier-above-flagship where one exists. Tiers are anchored on
+  capability at agentic coding (price is an expectation, not the
+  admission test) and the ladder is monotone — going up a tier never
+  gets a weaker model. The model landscape doesn't always map cleanly
+  onto three tiers, so aliases sometimes point at the same model rather
+  than leaving a tier undefined or pointing it at something that can't
+  do the job. Repointing rules and reasoning are documented in
+  `remote-dev-bot.yaml` and the README.
+- **Zero-config default model** (`default_model: auto`): fresh installs
+  run without configuring a model at all — the default resolves from
+  whichever provider API key secret is present. Explicit configuration
+  always wins.
+- **Lots of small fixes and cleanup in preparation for 1.0**: security
+  hardening, reliability and prompt-cache correctness, cost-accounting
+  accuracy, doc repair, and test-infrastructure work, driven by a
+  comprehensive pre-1.0 review (`comprehensive-review/2026-06-11/`).
+
+### Known issues
+
+- The gemini tiers point at `gemini-2.5-flash` rather than the newer
+  3.x models: litellm's Gemini 3.x thought-signature handling
+  intermittently breaks multi-turn tool calling (#654).
+- The gpt tiers point at `gpt-5.3-codex` rather than GPT-5.6 Sol: the
+  5.6 frontier family requires OpenAI's Responses API for tool use,
+  which rdb's litellm path doesn't use yet (#653).
+
+**Breaking changes:** model aliases renamed as part of the tier system
+(pre-1.0, no compatibility guarantee was in effect). Post-1.0, alias
+names are stable and the alias set is append-only.
+
 ## v0.9.0 — Delegate mode, reconcile mode, council reviews (May 2026)
 
 - **Delegate mode** (`/agent-delegate`): Full design-to-implementation pipeline

CLAUDE.md

@@ -0,0 +1 @@
+AGENTS.md

CONTRIBUTING.md

@@ -24,11 +24,13 @@ Three separate GitHub identities are used so each role is cleanly separated:
 | Account                 | Purpose                                                                                                                                                        | Credentials stored as                                                                                 |
 | ----------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------- |
 | `gnovak`                | Repo owner. Used for normal development and testing.                                                                                                           | `RDB_PAT_TOKEN` (on both repos), or GitHub App (`RDB_APP_ID` variable + `RDB_APP_PRIVATE_KEY` secret) |
-| `remote-dev-bot`        | Dedicated bot account. Collaborator on `remote-dev-bot-test`. Posts authorized test comments that trigger agent runs without attributing activity to `gnovak`. | `RDB_TESTER_PAT_TOKEN` (on remote-dev-bot)                                                            |
+| `remote-dev-bot`        | Dedicated bot account. Collaborator on `remote-dev-bot-test`. Posts authorized test comments that trigger agent runs without attributing activity to `gnovak`. | `RDB_TESTER_PAT_TOKEN` (on remote-dev-bot) — but see the temporary-state note below                   |
 | `remote-dev-bot-tester` | Simulates an unauthorized external user. NOT a collaborator on any repo.                                                                                       | `RDB_TESTER_UNAUTHORIZED_PAT_TOKEN` (on remote-dev-bot)                                               |
 
 ### remote-dev-bot (bot account) details
 
+The intended setup:
+
 - Classic PAT with `public_repo` + `workflow` scopes, no expiration
 - **`public_repo`** — create issues, post comments, open PRs in public repos
   (standard e2e test flow)
@@ -40,6 +42,17 @@ Three separate GitHub identities are used so each role is cleanly separated:
   its trigger comments
 - Keeps test activity out of `gnovak`'s GitHub contribution stats
 
+> **Temporary state (as of 2026-08):** the `remote-dev-bot` account is
+> flagged by GitHub (its profile 404s on the public API), so tokens cannot
+> be minted on it until that's resolved with GitHub support. Until then,
+> `RDB_TESTER_PAT_TOKEN` holds a PAT minted on `gnovak` instead (verifiable:
+> e2e test issues on rdb-test are authored by `gnovak`). If minting the
+> stopgap on `gnovak`, prefer a fine-grained PAT scoped to
+> `gnovak/remote-dev-bot-test` only (Contents/Issues/Pull requests
+> read-write + Actions read-write) with a long expiry — the 2026-08-05
+> full-suite failure was a 90-day stopgap token quietly expiring. Once the
+> bot account is restored, move the token back per the intended setup above.
+
 ### remote-dev-bot-tester details
 
 - Classic PAT with `public_repo` scope, no expiration
@@ -81,7 +94,7 @@ Secrets stored on `gnovak/remote-dev-bot`:
 | `OPENAI_API_KEY` | OpenAI API key (for GPT models) |
 | `GEMINI_API_KEY` | Google AI API key (for Gemini models) |
 | `RDB_APP_PRIVATE_KEY` | GitHub App private key, for bot identity on comments/PRs. |
-| `RDB_TESTER_PAT_TOKEN` | PAT for `remote-dev-bot` account (collaborator on rdb-test). Used by e2e tests to post authorized trigger comments. |
+| `RDB_TESTER_PAT_TOKEN` | PAT for `remote-dev-bot` account (collaborator on rdb-test; temporarily a `gnovak` PAT — see account details). Used by e2e tests to post authorized trigger comments. |
 | `RDB_TESTER_UNAUTHORIZED_PAT_TOKEN` | PAT for `remote-dev-bot-tester` (not a collaborator). Used by security e2e tests to verify unauthorized users are blocked. |
 
 Variables stored on `gnovak/remote-dev-bot`:

README.md

@@ -21,7 +21,7 @@ to design agents that can autonomously handle real development tasks.
 ## How It Works
 
 1. Create a GitHub issue describing a feature or bug
-2. Comment `/agent-resolve` (or `/agent-resolve-claude-large`, etc.) to trigger
+2. Comment `/agent-resolve` (or `/agent-resolve-claude-medium`, etc.) to trigger
    implementation
 3. A GitHub Action spins up an AI agent that runs a custom LiteLLM agent loop
    that:
@@ -65,11 +65,11 @@ comment (no code changes).
 | Command                         | What it does                                         |
 | ------------------------------- | ---------------------------------------------------- |
 | `/agent-resolve`                | Resolve the issue and open a PR (default model)      |
-| `/agent-resolve-claude-large`   | Resolve with a specific model                        |
+| `/agent-resolve-claude-medium`  | Resolve with a specific model                        |
 | `/agent-design`                 | Explore codebase and post a design analysis comment  |
-| `/agent-design-claude-large`    | Design analysis with a specific model                |
+| `/agent-design-claude-medium`   | Design analysis with a specific model                |
 | `/agent-review`                 | Post a code review comment on a PR (no code changes) |
-| `/agent-review-claude-large`    | Code review with a specific model                    |
+| `/agent-review-claude-medium`   | Code review with a specific model                    |
 | `/agent-workshop[-<model>]`     | Design analysis + multi-model council critique       |
 | `/agent-build[-<model>]`        | Implement issue + multi-model council code review    |
 | `/agent-delegate[-<model>]`     | Full design-to-implementation cycle with no human pauses |
@@ -78,8 +78,8 @@ comment (no code changes).
 Modes and model aliases are configured in `remote-dev-bot.yaml`.
 
 **Mobile-friendly syntax:** Commands are case-insensitive and you can use spaces
-instead of dashes: `/agent resolve claude large` works the same as
-`/agent-resolve-claude-large`.
+instead of dashes: `/agent resolve claude medium` works the same as
+`/agent-resolve-claude-medium`.
 
 ### Per-Invocation Arguments
 
@@ -113,13 +113,30 @@ Model aliases (like `claude-small`) map to **LiteLLM model identifiers** in
 `remote-dev-bot.yaml`. Remote Dev Bot uses LiteLLM to talk to different LLM
 providers through a unified interface.
 
+Aliases come in three capability tiers per provider, so you never have to
+track model releases:
+
+| Tier       | Meaning                                              | Typical use |
+| ---------- | ---------------------------------------------------- | ----------- |
+| `*-small`  | Best value-for-money coding workhorse                | The default; most tasks |
+| `*-medium` | The provider's flagship                              | Complex multi-file features |
+| `*-large`  | Frontier model above the flagship, where one exists  | Hardest long-horizon work |
+
+Tiers are anchored on capability at agentic coding, not price, and the ladder
+is monotone — going up a tier never gets you a weaker model. Where a provider
+has no distinct model for a tier, the alias points at the tier below it, so
+every alias always runs. The model landscape doesn't always divide cleanly
+into three tiers; these mappings are judgment calls, and you can always
+define your own aliases (see "Adding Models" below) if you'd draw the lines
+differently. An abstraction doesn't have to be perfect to be useful.
+
 **Model ID format:** `provider/model-name`
 
 | Provider           | Prefix       | Example                       |
 | ------------------ | ------------ | ----------------------------- |
-| Anthropic (Claude) | `anthropic/` | `anthropic/claude-sonnet-4-5` |
-| OpenAI (GPT)       | `openai/`    | `openai/gpt-5.1-codex-mini`   |
-| Google (Gemini)    | `gemini/`    | `gemini/gemini-2.5-flash`     |
+| Anthropic (Claude) | `anthropic/` | `anthropic/claude-sonnet-5`   |
+| OpenAI (GPT)       | `openai/`    | `openai/gpt-5.3-codex`        |
+| Google (Gemini)    | `gemini/`    | `gemini/gemini-3.6-flash`     |
 
 ### Supported Providers
 
@@ -161,12 +178,15 @@ Prefix the model string with the provider name in `remote-dev-bot.yaml` (e.g.,
 **For most tasks:** Use the default (`/agent-resolve`). Claude Sonnet
 (`claude-small`) offers a good balance of capability and cost.
 
-**For complex multi-file features:** Use `/agent-resolve-claude-large` (Opus) or
-`/agent-resolve-gpt-large` (GPT Codex). These models handle larger contexts and
-more intricate reasoning.
+**For complex multi-file features:** Use `/agent-resolve-claude-medium` (Opus)
+or `/agent-resolve-gpt-medium`. Flagship models handle larger contexts and more
+intricate reasoning.
+
+**For the hardest long-horizon work:** `/agent-resolve-claude-large` (Fable)
+is the frontier tier — expect roughly 2-4x the cost of a medium run.
 
 **For coding-heavy tasks:** Models with "codex" in the name (e.g.,
-`openai/gpt-5.1-codex-mini`) are specifically tuned for code generation and may
+`openai/gpt-5.3-codex`) are specifically tuned for code generation and may
 perform better on implementation tasks.
 
 ### Commit Trailers
@@ -435,7 +455,7 @@ The agent ran, posted a comment with its evaluation, but didn't open a PR. This
 means the agent judged that it couldn't fully resolve the issue — it hit the
 iteration limit, got confused, or determined its changes were incomplete.
 
-Try a more capable model (`/agent-resolve-claude-large`) or add more detail to
+Try a more capable model (`/agent-resolve-claude-medium`) or add more detail to
 the issue description. The agent's evaluation comment will say what it attempted
 and why it stopped.