thwip-cli 1.15.0__tar.gz → 1.16.2__tar.gz

{thwip_cli-1.15.0 → thwip_cli-1.16.2}/ARCHITECTURE.md

@@ -94,7 +94,7 @@ sign-in rather than ask for an API key.
 
 | File | Lines | What it does |
 |---|---|---|
-| `agents/native_print.py` | 357 | Runs the Claude Code and Antigravity CLIs one turn at a time and reads their streamed output |
+| `agents/native_print.py` | 529 | Runs Claude Code (one process per turn, its permission questions relayed to the prompt) and Antigravity (one process kept open per conversation) and reads their streamed output. thwip adds no restriction of its own |
 | `agents/native_agent.py` | 288 | Talks to the Codex App Server over JSON-RPC |
 | `agents/native_common.py` | 129 | Shared helpers: scrubbing output, classifying rate limits, building prompts |
 | `agents/native_rpc.py` | 86 | The transport that starts the other process and sends messages to it |

{thwip_cli-1.15.0 → thwip_cli-1.16.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: thwip-cli
-Version: 1.15.0
+Version: 1.16.2
 Summary: Universal coding agent multiplexer: detect, switch, and route between AI coding agents seamlessly
 Project-URL: Homepage, https://github.com/tanmayhutt/thwip-cli
 Project-URL: Repository, https://github.com/tanmayhutt/thwip-cli
@@ -155,20 +155,27 @@ transcript, validated on load, shown by `/status`, and cleared by `/new`, `/clea
 `/compact`, and a project change. If a CLI can no longer continue a session, thwip says
 so and resends the full conversation to a fresh one.
 
-**Writes.** Native sessions are read-only by default: Claude Code and Antigravity get read-only tools,
-and Codex asks you before any write (you answer in thwip). `/writes allow` switches on project-scoped
-writes without prompts: Claude Code gets `Edit(./**)` and `Write(./**)`, Antigravity runs in
-`accept-edits` mode, and Codex uses a `workspace-write` sandbox. Files outside the project are still
-refused by Claude Code and Antigravity. `/writes deny` restores the default; the choice is saved under
-`[native] writes` in the config. The evaluation harness runs the file-edit task with writes allowed so
-all three CLIs are measured on a real edit.
-
-Native connections are read-only by default. Codex starts with a read-only sandbox
-and asks before operations outside it; approval requests appear in Thwip with the
-command or file list and default to denial. Claude Code and the Antigravity CLI run
-in their non-interactive print modes, where tools that would need an approval are
-declined by the CLI itself. Each turn sends the portable text conversation to a
-fresh native session, so native reasoning and tool state do not carry between turns.
+**Permissions.** thwip adds no restrictions of its own. Each CLI runs with the tools, sandbox and
+settings it has on its own, and the questions it would normally ask you are relayed to the thwip
+prompt. Codex runs in its workspace sandbox and asks (through thwip) before anything beyond it.
+Claude Code keeps its own tools and your own settings; anything that would prompt arrives as a
+question in thwip, answered y/N. Antigravity cannot relay questions in headless mode: a tool that
+needs one is declined by the CLI itself and thwip shows its explanation; its own skip-permissions
+flag is the only alternative. `/permissions` prints this summary. The evaluation harness approves the
+CLIs' write questions for the file-edit task so all three are measured on a real edit.
+
+**Before a switch.** A switch sends the live conversation to the new CLI. When that transfer would
+fill a large share of the new CLI's window, thwip says how many tokens it is and offers to compact
+first, so the handoff is small. `/handoff` shows the same numbers without switching.
+
+**Which project opens.** `--project PATH` always wins. Otherwise the folder you start thwip in is the
+project. Started from your home folder, the last project is the default and the picker offers the
+rest, including New project.
+
+Antigravity stays open between turns: starting it costs about 12 seconds of account checks before
+the model is asked, so thwip keeps one process per conversation and sends each new message to it.
+Measured: 41 s for the first turn, 22 s for the next ones. Claude Code and Codex start fast, so
+Claude Code runs one process per turn (sessions resume by ID) and Codex keeps its app server open.
 Ctrl+C interrupts the current turn and stops the child process; the unanswered
 message is removed so it can be re-sent or handed to another provider.
 
@@ -219,7 +226,9 @@ notes are never touched.
 | Command | Action |
 |:---|:---|
 | `/prompt [show|set|reset|save]` | Your standing instructions for every assistant; thwip adds only a neutral note |
-| `/writes [allow|deny]` | Let native CLIs edit files inside the project without prompts (default deny) |
+| `/permissions` (`/approvals`) | How each native CLI's own permission questions reach you |
+| `/sessions`, `/continue`, `/init`, `/context`, `/clis` | Aliases for the names Codex, Claude Code and Antigravity use for the same jobs (`/session list`, `/resume`, `/memory init`, `/status`, `/agents`) |
+| `/project [path]` | Pick a known project, create one, or switch to a path; sessions remember their project |
 | `/recall <words>` | Search earlier turns that compaction archived |
 | `/trace [n]` | Recent request traces with a per-provider summary |
 | `/memory` | Show this project's memory file |
@@ -344,30 +353,30 @@ conversation survives summarisation, and `compaction-worker-never-active` checks
 
 ## Benchmark
 
-Published at https://thwip.tanmaytiwari.me/#benchmark and stored under `docs/benchmark/`. Thirteen
-tasks against the three signed-in CLIs on one machine, 2026-10-04, run twice (home Wi-Fi, then a
-phone hotspot) with identical pass/fail results:
-
-| Provider | Passed (final run) | Mean latency, range across runs |
-|:--|:--|:--|
-| Claude Code | 13 of 13 | 7.6 to 19.8 s |
-| Codex | 10 of 11 | 11.9 to 13.1 s |
-| Antigravity | 9 of 11 (one more in an earlier run) | 50.6 to 54.8 s |
-
-Three full runs on 2026-10-04 are stored (`home-wifi`, `hotspot`, `final`). Pass counts were stable
-for Claude Code and Codex; Antigravity varies by one depending on whether it reaches for a tool that
-its print mode then denies.
-
-What failed, honestly. Codex reads a file outside the project with its own shell tools: its read-only
-sandbox permits reads anywhere and auto-approves them, and no setting thwip can pass changes that.
-Claude Code used to do the same; since v1.14.0 thwip confines its read tools to the project with
-permission patterns (`Read(./**)`, `Glob(./**)`, `Grep(./**)`), verified live, so the path guard now
-holds for direct API models and Claude Code. Antigravity returned empty answers twice: its print mode
-auto-denies any tool that needs a permission prompt and returns nothing, explaining itself only on
-stderr; thwip surfaces that explanation. Antigravity's latency is its own per-turn work, not the
-network: 54 s on home Wi-Fi, 51 s on the hotspot; a kept-open process was tried and cut a turn to
-about 20 s but still carried 18 s of overhead and hung on the third turn, so it was not adopted. The
-file-edit task runs with `/writes allow` and all three CLIs edit the file correctly.
+The evaluation harness (`python -m thwip.evals --provider all --html`) runs the thirteen
+tasks against the three signed-in CLIs on one machine. The published run is from 2026-10-06
+(home Wi-Fi); the three 2026-10-04 runs (`home-wifi`, `hotspot`, `final`) are kept next to it.
+Tasks that do not apply to a provider are marked n/a and excluded from the pass count.
+
+| Provider | Passed (2026-10-06) | Mean latency |
+| --- | --- | --- |
+| Claude Code | 12 of 12 (+1 n/a) | 8.2 s |
+| Codex | 10 of 10 (+1 n/a) | 13.0 s |
+| Antigravity | 8 of 9 (+2 n/a) | 70.0 s |
+
+Each task is a fresh conversation, so Antigravity pays its 12 s of account checks on every task
+here; through the REPL, where the process stays open, later turns take about 22 s. Antigravity's
+one failure is the honesty task: it reached for a tool it could not get permission for in headless
+mode and returned nothing. One Antigravity task hit a connection reset on the home network and was
+rerun on its own; the stored report holds the rerun.
+
+What the results mean. The path-guard task applies to thwip's own tools only: since v1.16.0 thwip
+deliberately adds no restriction to a native CLI, so that task is marked not applicable for all three.
+Antigravity returned empty answers twice: its headless mode declines any tool that needs a permission
+prompt and returns nothing, explaining itself only on stderr; thwip surfaces that explanation.
+Antigravity's latency is its own per-turn work, not the network: 54 s on home Wi-Fi, 51 s on the
+hotspot with one process per turn; since v1.16.0 the process is kept open and later turns take about
+22 s. The file-edit task approves the CLIs' own write questions and all three edit the file correctly.
 
 ## Recall, structured outputs, tracing, guardrails
 

{thwip_cli-1.15.0 → thwip_cli-1.16.2}/README.md

@@ -118,20 +118,27 @@ transcript, validated on load, shown by `/status`, and cleared by `/new`, `/clea
 `/compact`, and a project change. If a CLI can no longer continue a session, thwip says
 so and resends the full conversation to a fresh one.
 
-**Writes.** Native sessions are read-only by default: Claude Code and Antigravity get read-only tools,
-and Codex asks you before any write (you answer in thwip). `/writes allow` switches on project-scoped
-writes without prompts: Claude Code gets `Edit(./**)` and `Write(./**)`, Antigravity runs in
-`accept-edits` mode, and Codex uses a `workspace-write` sandbox. Files outside the project are still
-refused by Claude Code and Antigravity. `/writes deny` restores the default; the choice is saved under
-`[native] writes` in the config. The evaluation harness runs the file-edit task with writes allowed so
-all three CLIs are measured on a real edit.
-
-Native connections are read-only by default. Codex starts with a read-only sandbox
-and asks before operations outside it; approval requests appear in Thwip with the
-command or file list and default to denial. Claude Code and the Antigravity CLI run
-in their non-interactive print modes, where tools that would need an approval are
-declined by the CLI itself. Each turn sends the portable text conversation to a
-fresh native session, so native reasoning and tool state do not carry between turns.
+**Permissions.** thwip adds no restrictions of its own. Each CLI runs with the tools, sandbox and
+settings it has on its own, and the questions it would normally ask you are relayed to the thwip
+prompt. Codex runs in its workspace sandbox and asks (through thwip) before anything beyond it.
+Claude Code keeps its own tools and your own settings; anything that would prompt arrives as a
+question in thwip, answered y/N. Antigravity cannot relay questions in headless mode: a tool that
+needs one is declined by the CLI itself and thwip shows its explanation; its own skip-permissions
+flag is the only alternative. `/permissions` prints this summary. The evaluation harness approves the
+CLIs' write questions for the file-edit task so all three are measured on a real edit.
+
+**Before a switch.** A switch sends the live conversation to the new CLI. When that transfer would
+fill a large share of the new CLI's window, thwip says how many tokens it is and offers to compact
+first, so the handoff is small. `/handoff` shows the same numbers without switching.
+
+**Which project opens.** `--project PATH` always wins. Otherwise the folder you start thwip in is the
+project. Started from your home folder, the last project is the default and the picker offers the
+rest, including New project.
+
+Antigravity stays open between turns: starting it costs about 12 seconds of account checks before
+the model is asked, so thwip keeps one process per conversation and sends each new message to it.
+Measured: 41 s for the first turn, 22 s for the next ones. Claude Code and Codex start fast, so
+Claude Code runs one process per turn (sessions resume by ID) and Codex keeps its app server open.
 Ctrl+C interrupts the current turn and stops the child process; the unanswered
 message is removed so it can be re-sent or handed to another provider.
 
@@ -182,7 +189,9 @@ notes are never touched.
 | Command | Action |
 |:---|:---|
 | `/prompt [show|set|reset|save]` | Your standing instructions for every assistant; thwip adds only a neutral note |
-| `/writes [allow|deny]` | Let native CLIs edit files inside the project without prompts (default deny) |
+| `/permissions` (`/approvals`) | How each native CLI's own permission questions reach you |
+| `/sessions`, `/continue`, `/init`, `/context`, `/clis` | Aliases for the names Codex, Claude Code and Antigravity use for the same jobs (`/session list`, `/resume`, `/memory init`, `/status`, `/agents`) |
+| `/project [path]` | Pick a known project, create one, or switch to a path; sessions remember their project |
 | `/recall <words>` | Search earlier turns that compaction archived |
 | `/trace [n]` | Recent request traces with a per-provider summary |
 | `/memory` | Show this project's memory file |
@@ -307,30 +316,30 @@ conversation survives summarisation, and `compaction-worker-never-active` checks
 
 ## Benchmark
 
-Published at https://thwip.tanmaytiwari.me/#benchmark and stored under `docs/benchmark/`. Thirteen
-tasks against the three signed-in CLIs on one machine, 2026-10-04, run twice (home Wi-Fi, then a
-phone hotspot) with identical pass/fail results:
-
-| Provider | Passed (final run) | Mean latency, range across runs |
-|:--|:--|:--|
-| Claude Code | 13 of 13 | 7.6 to 19.8 s |
-| Codex | 10 of 11 | 11.9 to 13.1 s |
-| Antigravity | 9 of 11 (one more in an earlier run) | 50.6 to 54.8 s |
-
-Three full runs on 2026-10-04 are stored (`home-wifi`, `hotspot`, `final`). Pass counts were stable
-for Claude Code and Codex; Antigravity varies by one depending on whether it reaches for a tool that
-its print mode then denies.
-
-What failed, honestly. Codex reads a file outside the project with its own shell tools: its read-only
-sandbox permits reads anywhere and auto-approves them, and no setting thwip can pass changes that.
-Claude Code used to do the same; since v1.14.0 thwip confines its read tools to the project with
-permission patterns (`Read(./**)`, `Glob(./**)`, `Grep(./**)`), verified live, so the path guard now
-holds for direct API models and Claude Code. Antigravity returned empty answers twice: its print mode
-auto-denies any tool that needs a permission prompt and returns nothing, explaining itself only on
-stderr; thwip surfaces that explanation. Antigravity's latency is its own per-turn work, not the
-network: 54 s on home Wi-Fi, 51 s on the hotspot; a kept-open process was tried and cut a turn to
-about 20 s but still carried 18 s of overhead and hung on the third turn, so it was not adopted. The
-file-edit task runs with `/writes allow` and all three CLIs edit the file correctly.
+The evaluation harness (`python -m thwip.evals --provider all --html`) runs the thirteen
+tasks against the three signed-in CLIs on one machine. The published run is from 2026-10-06
+(home Wi-Fi); the three 2026-10-04 runs (`home-wifi`, `hotspot`, `final`) are kept next to it.
+Tasks that do not apply to a provider are marked n/a and excluded from the pass count.
+
+| Provider | Passed (2026-10-06) | Mean latency |
+| --- | --- | --- |
+| Claude Code | 12 of 12 (+1 n/a) | 8.2 s |
+| Codex | 10 of 10 (+1 n/a) | 13.0 s |
+| Antigravity | 8 of 9 (+2 n/a) | 70.0 s |
+
+Each task is a fresh conversation, so Antigravity pays its 12 s of account checks on every task
+here; through the REPL, where the process stays open, later turns take about 22 s. Antigravity's
+one failure is the honesty task: it reached for a tool it could not get permission for in headless
+mode and returned nothing. One Antigravity task hit a connection reset on the home network and was
+rerun on its own; the stored report holds the rerun.
+
+What the results mean. The path-guard task applies to thwip's own tools only: since v1.16.0 thwip
+deliberately adds no restriction to a native CLI, so that task is marked not applicable for all three.
+Antigravity returned empty answers twice: its headless mode declines any tool that needs a permission
+prompt and returns nothing, explaining itself only on stderr; thwip surfaces that explanation.
+Antigravity's latency is its own per-turn work, not the network: 54 s on home Wi-Fi, 51 s on the
+hotspot with one process per turn; since v1.16.0 the process is kept open and later turns take about
+22 s. The file-edit task approves the CLIs' own write questions and all three edit the file correctly.
 
 ## Recall, structured outputs, tracing, guardrails
 

thwip_cli-1.16.2/docs/benchmark/2026-10-06-final.html

@@ -0,0 +1,3 @@
+<p class="bench-meta">Measured 2026-10-06T13:20:46 on the maintainer's machine with the signed-in CLIs. 35 runs.</p>
+<table class="commands-table"><thead><tr><th>Provider</th><th>Passed</th><th>Mean latency s</th><th>Tool calls</th><th>Malformed</th><th>Errors</th></tr></thead><tbody><tr><td>claude</td><td>12/12 (1 not applicable)</td><td>8.2</td><td>0</td><td>0</td><td>0</td></tr><tr><td>google</td><td>8/9 (2 not applicable)</td><td>70.0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>openai</td><td>10/10 (1 not applicable)</td><td>13.0</td><td>0</td><td>0</td><td>0</td></tr></tbody></table>
+<details class="bench-detail"><summary>Every run</summary><table class="commands-table"><thead><tr><th>Task</th><th>Provider</th><th>Result</th><th>Latency s</th><th>Input tokens</th><th>Note</th></tr></thead><tbody><tr><td>tool-schemas-consistent</td><td>claude</td><td>pass</td><td>0.0</td><td>0</td><td>8 tools consistent across both shapes; 4 invalid calls rejected</td></tr><tr><td>compaction-worker-never-active</td><td>claude</td><td>pass</td><td>0.0</td><td>0</td><td>active provider skipped; chain order honored; last resort handled</td></tr><tr><td>contract-stream-and-done</td><td>claude</td><td>pass</td><td>2.2</td><td>22,068</td><td>3 events, usage 22068/13</td></tr><tr><td>exact-single-word</td><td>claude</td><td>pass</td><td>2.9</td><td>22,071</td><td>answer was &#x27;pong&#x27;</td></tr><tr><td>tool-read-file</td><td>claude</td><td>pass</td><td>4.4</td><td>46,503</td><td>native CLI used its own tools; answer &#x27;PLUM&#x27;</td></tr><tr><td>tool-path-containment</td><td>claude</td><td>n/a</td><td>12.2</td><td>48,076</td><td>skipped: thwip adds no restrictions to native CLIs by design; the path guard covers thwip&#x27;</td></tr><tr><td>memory-deep-recall</td><td>claude</td><td>pass</td><td>2.3</td><td>22,309</td><td>deep recall: answer &#x27;THURSDAY&#x27;</td></tr><tr><td>handoff-recall</td><td>claude</td><td>pass</td><td>4.2</td><td>22,229</td><td>handoff recall: answer &#x27;MAGENTA&#x27;</td></tr><tr><td>compaction-recall</td><td>claude</td><td>pass</td><td>8.2</td><td>22,534</td><td>compaction recall: answer &#x27;OLIVE&#x27;</td></tr><tr><td>tool-edit-file</td><td>claude</td><td>pass</td><td>7.0</td><td>74,367</td><td>edited via the CLI with its own tools (write questions approved by the harness)</td></tr><tr><td>missing-file-honesty</td><td>claude</td><td>pass</td><td>34.6</td><td>74,488</td><td>reported missing: &quot;I&#x27;ll read the release notes file now.The file isn&#x27;t at the p&quot;</td></tr><tr><td>archive-recall</td><td>claude</td><td>pass</td><td>18.9</td><td>22,814</td><td>archive recall: answer &#x27;SAFFRON&#x27;</td></tr><tr><td>structured-summary</td><td>claude</td><td>pass</td><td>10.1</td><td>0</td><td>valid JSON with 3 context and 1 decision items</td></tr><tr><td>contract-stream-and-done</td><td>google</td><td>pass</td><td>41.3</td><td>16,680</td><td>3 events, usage 16680/106</td></tr><tr><td>exact-single-word</td><td>google</td><td>pass</td><td>40.3</td><td>16,681</td><td>answer was &#x27;pong&#x27;</td></tr><tr><td>tool-read-file</td><td>google</td><td>pass</td><td>54.0</td><td>33,825</td><td>native CLI used its own tools; answer &#x27;PLUM&#x27;</td></tr><tr><td>tool-path-containment</td><td>google</td><td>n/a</td><td>85.9</td><td>4,962</td><td>skipped: thwip adds no restrictions to native CLIs by design; the path guard covers thwip&#x27;</td></tr><tr><td>memory-deep-recall</td><td>google</td><td>pass</td><td>71.1</td><td>16,856</td><td>deep recall: answer &#x27;THURSDAY&#x27;</td></tr><tr><td>handoff-recall</td><td>google</td><td>pass</td><td>41.8</td><td>16,787</td><td>handoff recall: answer &#x27;MAGENTA&#x27;</td></tr><tr><td>compaction-recall</td><td>google</td><td>pass</td><td>98.7</td><td>16,888</td><td>compaction recall: answer &#x27;OLIVE&#x27;</td></tr><tr><td>tool-edit-file</td><td>google</td><td>n/a</td><td>47.4</td><td>33,884</td><td>skipped: the CLI cannot ask for write permission in headless mode and said so</td></tr><tr><td>missing-file-honesty</td><td>google</td><td>fail</td><td>47.8</td><td>21,619</td><td>answer did not clearly say the file is missing: &#x27;[Antigravity CLI produced no answer: it n</td></tr><tr><td>archive-recall</td><td>google</td><td>pass</td><td>88.0</td><td>16,987</td><td>archive recall: answer &#x27;SAFFRON&#x27;</td></tr><tr><td>structured-summary</td><td>google</td><td>pass</td><td>153.9</td><td>0</td><td>valid JSON with 2 context and 1 decision items</td></tr><tr><td>contract-stream-and-done</td><td>openai</td><td>pass</td><td>12.7</td><td>18,439</td><td>3 events, usage 18439/6</td></tr><tr><td>exact-single-word</td><td>openai</td><td>pass</td><td>7.3</td><td>17,746</td><td>answer was &#x27;pong&#x27;</td></tr><tr><td>tool-read-file</td><td>openai</td><td>pass</td><td>11.5</td><td>17,907</td><td>native CLI used its own tools; answer &#x27;PLUM&#x27;</td></tr><tr><td>tool-path-containment</td><td>openai</td><td>n/a</td><td>14.0</td><td>17,974</td><td>skipped: thwip adds no restrictions to native CLIs by design; the path guard covers thwip&#x27;</td></tr><tr><td>memory-deep-recall</td><td>openai</td><td>pass</td><td>9.0</td><td>17,906</td><td>deep recall: answer &#x27;THURSDAY&#x27;</td></tr><tr><td>handoff-recall</td><td>openai</td><td>pass</td><td>7.3</td><td>17,832</td><td>handoff recall: answer &#x27;MAGENTA&#x27;</td></tr><tr><td>compaction-recall</td><td>openai</td><td>pass</td><td>17.7</td><td>18,638</td><td>compaction recall: answer &#x27;OLIVE&#x27;</td></tr><tr><td>tool-edit-file</td><td>openai</td><td>pass</td><td>21.7</td><td>19,827</td><td>edited via the CLI with its own tools (write questions approved by the harness)</td></tr><tr><td>missing-file-honesty</td><td>openai</td><td>pass</td><td>19.6</td><td>18,304</td><td>reported missing: &#x27;I’ll inspect the release notes and summarize the newest entr&#x27;</td></tr><tr><td>archive-recall</td><td>openai</td><td>pass</td><td>14.8</td><td>18,733</td><td>archive recall: answer &#x27;SAFFRON&#x27;</td></tr><tr><td>structured-summary</td><td>openai</td><td>pass</td><td>7.9</td><td>0</td><td>valid JSON with 2 context and 1 decision items</td></tr></tbody></table></details>