cawdex 1.35.74 → 1.35.76

package/README.md

@@ -81,7 +81,7 @@ If you have multiple keys for the same provider (e.g. several free OpenRouter ac
 
 Cool-down policy: 60s for rate-limit (`429`, `rate.?limit`), 1h for quota / auth / 403. 404 model-not-found and 5xx server errors are NOT treated as key problems — they're surfaced upward without burning a key.
 
-For a free-tier-only OpenRouter account, use `/openrouter-free` to reset the provider and primary model to `openrouter/free`. The interactive `/model` picker floats currently free models to the top. Automatic fallback retry is explicit: choose it in `/config` or set it later with `/fallback <model-id>`.
+For a free-tier-only OpenRouter account, use `/openrouter-free` to reset the provider and primary model to `openrouter/free`. The interactive `/model` picker floats currently free models to the top. `/model alias <name> <model>` creates shortcuts, and `/model once <model-or-alias> [--effort high]` applies a single-turn override without changing `config.json`. Automatic fallback retry is explicit: choose it in `/config` or set it later with `/fallback <model-id>`.
 
 ---
 
@@ -115,7 +115,7 @@ Agents are pulled from the bundled ECC harness. Each runs against an empty tool
 | `/walkthrough` | Agent-led tour. Aliases: `/tour`, `/guide`. |
 | `/help` | Full command list. |
 | `/mode <name>` | Switch mode. |
-| `/model [name]` | Show or switch model. |
+| `/model [name]` | Show or switch model. Also supports aliases and one-shot overrides. |
 | `/openrouter-free` | Switch OpenRouter back to the free model router. |
 | `/openai-login status\|smoke` | Check Codex OAuth status or run an OAuth stream smoke test. |
 | `/theme`, `/palette`, `/palette <id>`, `/palettes` | Pick display density and color palette, or switch directly by palette id. `/pallete` is accepted as a typo alias. |
@@ -125,7 +125,7 @@ Agents are pulled from the bundled ECC harness. Each runs against an empty tool
 | `/keys add\|remove\|status\|clear` | Manage the key-rotation pool. |
 | `/swarm <task>` | Infer roles, run analysis-only agents, and synthesize a handoff. Expert: `/swarm <agents> <task>`. |
 | `/sources <topic> [--json]` | Direct source scan against arXiv, GitHub, Hugging Face, and Kaggle. Defaults to recent 90-day leaderboard-oriented coverage and does not call the model. Add `--json` for machine-readable source packets. |
-| `/benchmark-repos [query]` | Offline public Terminal-Bench repo catalog for source mining, including known no-public-source gaps; follow positive hits with `/repo-digest <owner/repo>`. |
+| `/benchmark-repos [query]` | Offline public Terminal-Bench repo catalog for source mining, including known no-public-source gaps; use `--top-open-source --from-top 20 --docs-only` for verified public repo targets from the top of the leaderboard, then follow positive hits with `/repo-digest <owner/repo> --docs-only`. |
 | `/repo-digest <github-url\|owner/repo>` | Direct GitHub repo digest: metadata, file tree signals, manifests, likely commands, key excerpts, and AHE-style component surfaces without cloning or a model call. |
 | `/source-research <topic>` | Research arXiv, GitHub repos/issues/PRs/code, Hugging Face papers/models/datasets, and Kaggle datasets/competitions before synthesis. |
 | `/benchmark <task>` | Run a benchmark-grade issue/terminal/general-agent workflow. Profiles: `swe-bench`, `terminal-bench`, `terminalworld`, `swe-context`, `swe-chain`, `swe-cycle`, `swe-ci`, `swe-prbench`, `tml-bench`, `pi-bench`, `ci-repair`, `wildclaw`, `arc-agi`, `specbench`, `reward-hacking`, `roadmapbench`, `saasbench`, `swe-bench-mobile`, `appworld`, `browsecomp`, `tau2`, `generic`. |
@@ -151,7 +151,7 @@ See **[COMMANDS.md](COMMANDS.md)** for the full reference.
 
 Typing `/` at an empty prompt opens a bounded inline command selector, not a full-screen overlay. It stays under the prompt, keeps the widget under roughly half the terminal height, supports typing to narrow, arrows/PageUp/PageDown to scroll, Home/End to jump, and Enter to fill the prompt with the highlighted command so you can edit it or press Enter again to run it. If a slash prefix is already typed, Tab reopens the same bounded selector with that filter instead of dumping every command into the terminal. The selector only erases its own rows, so the surrounding transcript remains visible.
 
-While a model call is running, typed text stays visible and editable in the fixed bottom footer and is restored into the next prompt if you cancel. The first interactive prompt opens with `/model` prefilled by default; change that with `/footer opening <prompt>`. Footer templates support `{version}` for the installed Cawdex version. On every interactive launch, Cawdex checks npm for a newer package and starts `npm install -g cawdex@latest --force` in the background when one is available. Known-flaky OpenRouter preview/free models such as `openrouter/owl-alpha` also get a first-token watchdog: if no stream event arrives, Cawdex cancels the stuck call and retries once only when a fallback model is configured; otherwise it records the timeout and returns the prompt instead of leaving the terminal looking frozen.
+While a model call is running, typed text stays visible and editable in the fixed bottom footer and is restored into the next prompt if you cancel. The first interactive prompt opens with `/model` prefilled by default; change that with `/footer opening <prompt>`. Footer templates support `{version}` for the installed Cawdex version. The active-turn indicator uses a subtle Edo/neon-style animation and keeps Esc/F5 visible as the interrupt path. On every interactive launch, Cawdex checks npm for a newer package and starts `npm install -g cawdex@latest --force` in the background when one is available. Known-flaky OpenRouter preview/free models such as `openrouter/owl-alpha` also get a first-token watchdog: if no stream event arrives, Cawdex cancels the stuck call and retries once only when a fallback model is configured; otherwise it records the timeout and returns the prompt instead of leaving the terminal looking frozen.
 
 The 12 built-in color palettes are the first-page Coolors trending schemes: `olive-garden-feast`, `fiery-ocean`, `refreshing-summer-fun`, `ocean-blue-serenity`, `pastel-dreamland-adventure`, `sunny-beach-day`, `dark-sunset`, `fiery-red-sunset`, `fiery-palette`, `rustic-earthy-tones`, `golden-summer-fields`, and `vibrant-tones`.
 
@@ -209,7 +209,7 @@ Targeted-fix manifest tracking is live alongside decision observability: repair
 
 Trajectory cleanup tracking is also live: redacted traces now flag base64/data-URI blobs, high-entropy encoded or minified output, repeated duplicate output, and excessive truncation as `trajectory_cleanup_*` signals. Risky cleanup events create an execution-control process defect, completion reminder, and `experienceCard.trajectoryCleanup` entry so AHE-style runs can deduplicate noisy observations before reusing prior evidence.
 
-Interactive turns are acknowledged before provider I/O starts, and active turns can be cancelled with Esc or F5/Shift+F5, including raw Windows/xterm F5 escape sequences that readline may not classify correctly. Any type-ahead captured before a cancellation or permission interruption is preserved and restored at the prompt instead of being silently submitted. Known-stuck OpenRouter preview models such as `openrouter/owl-alpha` are auto-healed to the configured fallback model on startup and before the first API call unless `CAWDEX_ALLOW_FLAKY_MODELS=1` is set.
+Interactive turns are acknowledged before provider I/O starts, and active turns can be cancelled with Esc or F5/Shift+F5, including raw Windows/xterm F5 escape sequences that readline may not classify correctly. Any type-ahead captured before a cancellation or permission interruption is preserved and restored at the prompt instead of being silently submitted. Known-stuck OpenRouter preview models such as `openrouter/owl-alpha` keep a shorter first-token watchdog, but Cawdex never replaces a user-configured model on startup or before the first API call. Automatic retry only happens after a real failure when you explicitly configure `/fallback`.
 
 Bundled ECC skills are exposed with progressive disclosure: the system prompt shows only Level-0 skill names and descriptions, and `skill_view` loads the full prompt body only after a fit check. In benchmark mode the prompt asks the agent to inspect `benchmark_context` and local repo evidence before loading a skill, prefer one strongly domain/version-matched skill, and treat local files plus verifier output as authoritative.
 

package/bin/anycode.js

@@ -1,2 +1,2 @@
-#!/usr/bin/env node
-import('../dist/index.js');
+#!/usr/bin/env node
+import('../dist/index.js');