pi-freeflow 1.17.1 → 1.19.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -1,284 +1,296 @@
1
- # Changelog
2
-
3
- ## 1.17.1
4
-
5
- ### Patch Changes
6
-
7
- - Assistant and subagent requests now stay working when the shared free-tier gateway refuses sessions: retried and resumed conversations automatically continue on a healthy fallback model instead of repeating the refusal, and normal routing resumes on its own once the gateway recovers.
8
-
9
- ## 1.17.0
10
-
11
- ### Minor Changes
12
-
13
- - Automatic upstream degradation: when the shared free-tier gateway starts refusing new sessions, new chat sessions seamlessly use a healthy fallback model on the same API, affected errors carry an actionable hint, in-progress chats are never rerouted, and normal routing resumes automatically on recovery. `/freeflow status` now shows live upstream state.
14
-
15
- ## 1.16.0
16
-
17
- ### Minor Changes
18
-
19
- - New free model: Union Alpha Free joins the picker. It answers through a different upstream endpoint than the other free models, handled automatically 262K context, image input, tool calling, and zero cost like the rest of the catalog.
20
-
21
- ## 1.15.2
22
-
23
- ### Patch Changes
24
-
25
- - Relay traffic is now spent only on the free models the pool exists for: requests for other models go straight upstream even with the relay pool on, cutting relay bandwidth. Applies in automatic and always-on modes; relay-off behavior is unchanged.
26
- - The disabled-deployment rollover now triggers only on genuine hosting verdicts, so ordinary payment and quota refusals always surface immediately without cooling a healthy relay. Relay-list imports also no longer show a negative removal count when the file holds more relays than the current pool.
27
-
28
- ## 1.15.1
29
-
30
- ### Patch Changes
31
-
32
- - Requests now roll over to the next relay when the current one answers with a disabled-deployment verdict, instead of failing on the first relay and requiring a manual switch. Other payment and quota refusals still surface immediately.
33
-
34
- ## 1.15.0
35
-
36
- ### Minor Changes
37
-
38
- - Free-tier requests now use native session identifiers and an up-to-date client version, after the upstream gateway began rejecting older formats with a free-tier error. `/freeflow test` also gains an end-to-end chat check that verifies a relay can actually run inference, not just list models.
39
-
40
- ## 1.14.0
41
-
42
- ### Minor Changes
43
-
44
- - c28b7b5: Local proxy recovers on its own: if the background proxy stops unexpectedly, the extension now notices the refused local connection and starts a fresh proxy within seconds instead of leaving every model failing until the next session; shutdowns and crashes are also recorded in the log so the cause is visible.
45
-
46
- ### Patch Changes
47
-
48
- - /freeflow update now goes through the host plugin managers first — reinstall under OMP, package update under Pi — and only then falls back to a global install, using Bun when npm is missing. Previously the first step called an update action the plugin manager does not define, so managed installs always fell through to the wrong target.
49
-
50
- ## 1.13.0
51
-
52
- ### Minor Changes
53
-
54
- - 39a06b9: Carry a relay pool between machines: `/freeflow export` saves the relays to a file and `/freeflow import` loads them back, adding to the current pool by default or swapping it whole on confirmation.
55
-
56
- ## 1.12.1
57
-
58
- ### Patch Changes
59
-
60
- - b32f208: Sessions survive a relay change without a rejected turn.
61
-
62
- Responses models sign each thinking block for the upstream backend that produced it, and only that backend can read it back. When a later turn reached a different backend, the replayed blocks were rejected as unreadable ("reasoning `encrypted_content` was not issued to this caller"), the host repeated the same failing request, and the session could not continue.
63
-
64
- pi-freeflow now keeps each conversation on the relay that issued its reasoning while that relay is healthy, and when a relay change is unavoidable (rate limit, relay removed or redeployed, direct-mode switch) it sends that turn without the signed thinking blocks, so the new backend accepts it on the first attempt. If a rejection still happens (the provider can change backends behind a relay), the request is retried once without the blocks, and only the rejected blocks are dropped on later turns, so the model keeps whatever reasoning the current backend can read. Messages, tool calls and tool results are always preserved.
65
-
66
- ## 1.12.0
67
-
68
- ### Minor Changes
69
-
70
- - Free-model picker grows from 24 to 26: adds Nex N2.5 Pro, Nex N2.5 Mini, and Ling 3.0 Flash VL, and removes the discontinued Inkling entry. A new Stealth Models Watchlist docs page tracks masked preview IDs users ask about.
71
-
72
- ## 1.11.2
73
-
74
- ### Patch Changes
75
-
76
- - 549b66e: Oversized requests no longer fail at the relay hop: when a relay answers 413 payload limit, the proxy transparently tries the next relay and then the direct route, keeping the stream alive. Long sessions that outgrow the relay payload cap now complete instead of surfacing a function payload error.
77
-
78
- ## 1.11.1
79
-
80
- ### Patch Changes
81
-
82
- - Apologies for the daemon disconnect bug introduced in v1.11.0: in multi-agent workflows or when subagents, evaluations, and background tasks completed, the extension prematurely detached from the local proxy daemon, causing the proxy to shut down while your main session remained active. The heartbeat connection now persists throughout your active terminal session.
83
- - Increased local health check and liveness probe timeouts from 800ms and 500ms to 2500ms and 1500ms, with an automatic probe retry to avoid false-alarm daemon restarts during heavy concurrent streaming.
84
- - Added unhandled error logging inside the daemon to prevent silent process exits.
85
- - Fresh installs now log full HTTP lifecycle debug output by default so diagnostic reports contain complete request context; turn it off anytime with `/freeflow debug off`.
86
-
87
- ## 1.11.0
88
-
89
- ### Minor Changes
90
-
91
- - Proxy daemon now stays up while any session uses it, even when idle. It shuts down only after the last session leaves, instead of retiring after a short quiet window.
92
-
93
- ## 1.10.0
94
-
95
- ### Minor Changes
96
-
97
- - cf48598: Hands-free proxy recovery plus catalog refresh: the proxy now watches its own health, restarts with backoff instead of leaving a dead port, and can start on login. Removed two discontinued free models from the picker.
98
-
99
- All notable changes to pi-freeflow. Public, user-visible behavior only.
100
-
101
- ## 1.9.9 - 2026-09-06
102
-
103
- ### Fixes
104
-
105
- - **Removed our own request cap — sorry, that one was on us.** The proxy used to enforce a built-in request limit and could answer 429 before upstream quota was actually exhausted. That was a bug, not your quota. From this version the proxy never rejects on quota itself: a 429 only surfaces when the upstream — and every relay in your pool — genuinely is rate-limited, and that response now points you at `/freeflow deploy` to add relay egress.
106
-
107
- ## 1.9.8 - 2026-09-06
108
-
109
- ### Fixes
110
-
111
- - **Stale model list heals itself (follow-up to #6).** If your saved model list predates a newly added model, the background refresh now repairs the entry (correct endpoint and details) instead of sending requests to the wrong address — no manual `/freeflow refresh` or cache deletion needed.
112
- - **Paid models stay out of the picker even from old saved lists.** Every read of the saved model list now drops non-free entries, so models requiring an API key cannot linger after an upgrade.
113
- - **Old saved lists without a sync marker now re-sync once.** A saved list that could never trigger a network check now performs one plain revalidation (then syncs normally), so newly added free models appear without manual intervention. Missing or corrupt lists still fall back silently with no network call.
114
- - **Upstream errors are now visible in the proxy log.** Failed upstream responses log their status code and model, and a model routed to the wrong endpoint logs the mismatch with the fix (restart Pi/OMP after upgrade).
115
-
116
- ### Validation
117
-
118
- - TypeScript typecheck passed cleanly (`tsc --noEmit`).
119
- - Full test suite passed on Windows (304 tests) and Ubuntu Linux (`acerblue`, 305/305 tests passed), including new regressions for the stale-cache shape from #6.
120
- - Live sweep of all 26 models through a fresh install on `acerblue`: 23/26 answered on first try (both Muse Spark models via the Responses endpoint); the 3 misses are upstream per-model daily quotas (429), zero server errors.
121
-
122
- ## 1.9.7 - 2026-09-06
123
-
124
- ### Fixes
125
-
126
- - **Model picker only shows verified free models (fixes #6).** Background catalog refresh now filters out non-free models from upstream endpoints so paid models requiring an API key no longer leak into your picker.
127
- - **Relay pool automatically recovers from dead deployments (fixes #5).** If a relay URL in your pool returns an infrastructure 404 (such as a missing or deleted Vercel deployment), the proxy marks it as failed and immediately rolls over to your next healthy relay or direct mode instead of getting trapped in a 404 loop.
128
-
129
- ### Changes
130
-
131
- - **Updated model catalog (26 verified free models: 7 OpenCode Zen + 19 KiloCode Gateway).**
132
- - Added `inclusionai/ling-3.0-flash-sante:free` (Ling 3.0 Flash Sante, 262K context, 32K output, reasoning supported). Clean CLI alias `ling-3.0-flash-sante` supported.
133
- - Removed `laguna-s-2.1-free` (OpenCode Zen) after upstream dropped free-tier access. `poolside/laguna-s-2.1:free` on Kilo remains active.
134
- - **Public by default for new relay deployments.** Relays deployed via `/freeflow deploy` are now public by default with no mandatory authentication tokens. This enables seamless copy-paste migration across proxy tools (such as 9router) while maintaining safety guards for allowed AI upstreams.
135
-
136
- ### Validation
137
-
138
- - TypeScript typecheck passed cleanly (`tsc --noEmit`).
139
- - Full test suite passed across Windows (299 tests) and Ubuntu Linux (`acerblue-local`, 300/300 tests passed).
140
- - End-to-end concurrency and failover stress testing verified (100 parallel requests, 20 subagent leases, burst failover).
141
-
142
- ## 1.9.6 - 2026-09-03
143
-
144
- ### Fixes
145
-
146
- - **Hiding the status widget now sticks.** `/freeflow hide` survives restarts, new sessions, and background update notices previously the widget could reappear on the next session.
147
- - **Calmer startup with many sessions.** Sessions starting at the same time now converge on a single background proxy instead of each spawning its own; a spawn that never becomes ready falls back to in-process mode instead of running untracked.
148
- - **Survives daemon restarts mid-session.** If the background proxy is replaced while sessions run, connected sessions re-register automatically instead of silently losing their lease (which could retire the daemon mid-use).
149
- - **Stale-version cleanup targets the right process.** Windows port matching is now exact, and Linux without `lsof`/`fuser` resolves the holder via `/proc` instead of giving up.
150
- - **`/freeflow logs --follow` no longer runs forever.** The live tail stops when the session ends and caps burst output; the update/install subprocess now has a 2-minute timeout.
151
-
152
- ### Validation
153
-
154
- - `npx tsc --noEmit` clean, `npm test` 299 (298 pass + 1 Linux-only skip) + smoke green on Windows; stress harness 7/7 on Windows and Linux `acerblue-local`. **`macOS not tested`** this cycle.
155
-
156
- ## 1.9.5 - 2026-09-03
157
-
158
- ### Changes
159
-
160
- - **Catalog refreshed against live upstream model lists (26 models: 8 OpenCode Zen + 18 KiloCode Gateway).** Re-checked both upstreams against their live endpoints and live inference probes:
161
- - **Added** `muse-spark-1.3-contributor-free` (OpenCode Zen, Responses API, 1M context, 131K output, vision) verified live: completes with reasoning, accepts effort levels, and answers vision queries.
162
- - **Removed** `hy3-free` (OpenCode Zen) upstream no longer serves it (`Model hy3-free is not supported`).
163
- - **Removed** `tencent/hy3:free` and `meituan/longcat-2.0-free` (KiloCode Gateway) — upstream free tier dropped them (model unavailable / sign-in required).
164
- - Pruned IDs added to the dead-model filter so a stale disk cache cannot resurrect them.
165
-
166
- ### Validation
167
-
168
- - `npx tsc --noEmit` clean, `npm test` full suite green on Windows.
169
-
170
- ## 1.9.4 - 2026-09-02
171
-
172
- ### Dependencies
173
-
174
- - **Zero runtime dependencies.** Removed `undici@8.10.0` — `pi` (`0.84.4`) and `omp` (`18.1.3`) already bundle `undici` 6.x/7.x and expose `global fetch` with keep-alive pooling. `relayFetch` and proxy now use `global fetch` directly (`src/relay.ts` `Agent` + `canUseCustomDispatcher` + `dispatcher: agent` removed; `src/proxy.ts` dispatcher removed). Keeps thin `11.3k` + `298 tests` + `0 deps` compatible directly with `reference/pi` + `reference/oh-my-pi`.
175
-
176
- ### Validation
177
-
178
- - `npx tsc --noEmit` clean, `npm test 298/298` on **Windows** (`omp/18.1.3`, `pi 0.84.4`) and **Linux `acerblue-local`** (`Ubuntu 6.8.0-138`, `node v22.23.2`, `pi 0.84.4`) via `/tmp/pi-freeflow-validation`. **`macOS not tested`** this cycle.
179
-
180
- ## 1.9.3 - 2026-09-02
181
-
182
- ### Fixes
183
-
184
- - **Windows console flood fixed.** Two Windows-only helpers flashed a visible `conhost`/`cmd` window on every daemon probe: `netstat -ano | findstr :28180` / `taskkill` in the stale-daemon replace path and `spawn(omp|npm, shell:true)` for `/freeflow update`. Both now use `windowsHide: true` (no-op on Linux/macOS) and `spawnWithProgress` was refactored to `Promise.withResolvers` to satisfy `ts-promise-with-resolvers`. Idle `pi` no longer spawns many visible consoles even after closing the terminal (detached daemon at `127.0.0.1:28180` survives by design; `beatOnce` 10s heartbeat now throttled 2s via `lastSpawnAt`).
185
- - Daemon spawn now throttled per-process (2s) as a storm guard when `28180` is contended or blocked; the `ensuring` guard + `waitForReady 5s` already prevented tight loops.
186
-
187
- ### Validation
188
-
189
- - `npx tsc --noEmit` clean, `npm test 279/279` on **Windows** (`omp/18.1.3`, `pi 0.84.4`) and **Linux `acerblue-local`** (`Ubuntu 6.8.0-138`, `node v22.23.2`, `pi 0.84.4`) via `/tmp/pi-freeflow-validation`. **`macOS not tested`** this cycle.
190
- - Reporter `LOYINuts` issue #3 (`pi idle creates many sessions force reboot`) `grep -r rtk src` confirms `rtk` is an external global skill (`~/.agents/skills/rtk` `Command::new("cmd")` without `CREATE_NO_WINDOW`), not `pi-freeflow`. After this fix, closing the terminal no longer leaves flashing zombies; kill via `netstat -ano | findstr :28180` → `taskkill /F /PID` or `/freeflow kill`.
191
-
192
- ## 1.9.2 - 2026-08-31
193
-
194
- ### Fixes
195
-
196
- - **Closing one session no longer stops the shared local proxy.** The proxy daemon
197
- is now a fully detached background process: it outlives any single OMP/Pi session
198
- (previously, closing the session that owned the daemon could shut it down even
199
- while other sessions were still using it). It retires by itself only when no
200
- session is connected, no request is in flight, and it has been idle for a grace
201
- period. The next use starts it again automatically.
202
- - New command: `/freeflow kill` (aliases `stop`, `shutdown`) stops the background
203
- daemon on demand. The next freeflow use restarts it.
204
-
205
- ### Improvements
206
-
207
- - The proxy tracks connected sessions and last request time; `/freeflow status`
208
- and the health endpoint now report active session leases, so you can see when
209
- other sessions are keeping the daemon alive.
210
- - Docs: the command reference now lists `kill`, and the FAQ explains the shared
211
- daemon lifecycle (survives session close; self-retires when unused).
212
-
213
- ## 1.9.1 - 2026-08-30
214
-
215
- ### Fixes
216
-
217
- - Stale-daemon guard completed: a pre-1.9.0 daemon cannot report its in-flight requests, and usage cannot be verified — so it is now left running instead of being replaced. Previously such a daemon could still be killed mid-stream while busy (a one-time window when upgrading from 1.8.x). The guarantee now holds in every case: only older, verified-idle daemons are replaced.
218
- - The stale-daemon kill ritual was consolidated into one path (it was duplicated at five call sites); behavior unchanged.
219
- - Tests: new coverage for the pre-1.9 daemon case; shared sandbox helpers extracted.
220
-
221
- ## 1.9.0 - 2026-08-30
222
-
223
- ### Development hardening
224
-
225
- - Test suite now runs fully sandboxed: every test uses a temporary data directory, so no test can ever write to your real `~/.pi/agent/` files or interfere with a running local proxy. A single env override (`PI_*_DATA_DIR`) re-roots all data files for tests/CI.
226
- - Added a complete mocked user-flow e2e suite: fresh install → onboarding → proxy health → direct chat → relay add/roll/fallback → guided deploy → update check → command surface — all deterministic, zero network.
227
- - Docs: test counts no longer hardcoded in README (they drifted with releases); release history tracked here.
228
-
229
- ### Fixes
230
-
231
- - **Stale-daemon replacement no longer interrupts running sessions.** The shared local proxy can be held by another OMP/Pi session; the old upgrade logic killed that holder on version mismatch, which could terminate the session that owned the proxy mid-stream. Replacement now only happens when the running daemon is strictly older AND idle (0 in-flight requests, reported via `/_health`); newer or busy daemons are reused with a log note instead. Optional opt-out: set the no-kill env to `1` (see README FAQ for upgrades).
232
-
233
- ## 1.8.2 - 2026-08-30
234
-
235
- ### Fixes & polish
236
-
237
- - Deploy: compare-and-swap on concurrent deployments (no duplicate relays when two sessions deploy at once).
238
- - Proxy: URL-encoding edge cases (`%` in paths), startup timeout and kill handling, port conflict fallback.
239
- - Logs: sanitized sensitive headers in debug output.
240
- - Kilo gateway models: compatibility pass for all 25 models.
241
- - Docs site build included; README per-host usage guide (Oh My Pi & Pi install + pick + manage).
242
-
243
- ## 1.8.1 - 2026-08-29
244
-
245
- - Fix: stale-daemon auto-heal is now shipped in the published package (previously only in the repo).
246
-
247
- ## 1.8.0 - 2026-08-29
248
-
249
- ### Automatic upgrades
250
-
251
- - On version upgrade, the extension detects a stale local proxy daemon and replaces it automatically — users get fixes without manual restarts or killing sessions.
252
- - Fix: client auth key stripped before reaching the opencode.ai/zen upstream (failed with 401 for some clients).
253
- - 3 new free models (28 total).
254
-
255
- ## 1.7.1 - 2026-08-29
256
-
257
- - Docs: clarified npm is the distribution channel (the package is an extension loaded by OMP/Pi, not a standalone CLI).
258
-
259
- ## 1.7.0 - 2026-08-29
260
-
261
- - 4 new free models (25 total) with live-verified specs (context/output limits, vision, thinking levels).
262
- - New alias map aligned with host model selectors.
263
-
264
- ## 1.6.1 - 2026-08-29
265
-
266
- - Catalog: thinking-level map locked per model; docs sync.
267
-
268
- ## 1.6.0 - 2026-08-29
269
-
270
- ### Onboarding & UX
271
-
272
- - First-run onboarding message; 429 guidance hint (add your own relay egress).
273
- - `/freeflow test` to probe a relay; relay latency tracking with health badges in `/freeflow list`.
274
- - Guided deploy with context picker + confirmation; post-deploy health check.
275
- - Status clarity: current mode + state file path; log text filter; per-relay usage counters; throttled roll notifications.
276
-
277
- ### Reliability
278
-
279
- - Relay state write-protection: before every save, the current state is snapshotted to `.bak`; if the main file is corrupted or missing, the backup is recovered automatically.
280
- - New-user flow never seeds a fake relay; starts in direct mode with an empty pool.
281
-
282
- ## 1.5.1 - 2026-08-28
283
-
284
- - Edge-sweep fixes: port 28180 with legacy 18080 dual-probe auto-migration, OMP/Pi compatibility.
1
+ # Changelog
2
+
3
+ ## 1.19.0
4
+
5
+ ### Minor Changes
6
+
7
+ - Every tool your agent can use now works on every free model: the proxy translates tool definitions between chat, responses, and messages formats automatically, so built-in tools, MCP server tools, and custom tools all pass through intact on all request types. Strict tool schemas are preserved, and requests sent in one format to another endpoint are converted instead of rejected.
8
+
9
+ ## 1.18.0
10
+
11
+ ### Minor Changes
12
+
13
+ - OpenCode free-tier client fingerprinting: plain chat, subagent, evaluation, and background watchdog requests now pass upstream free-tier gates without error. The proxy automatically supplies compatibility tools and streaming conventions expected by the upstream gateway, while transparently converting responses to standard non-streaming format for callers that request it.
14
+
15
+ ## 1.17.1
16
+
17
+ ### Patch Changes
18
+
19
+ - Assistant and subagent requests now stay working when the shared free-tier gateway refuses sessions: retried and resumed conversations automatically continue on a healthy fallback model instead of repeating the refusal, and normal routing resumes on its own once the gateway recovers.
20
+
21
+ ## 1.17.0
22
+
23
+ ### Minor Changes
24
+
25
+ - Automatic upstream degradation: when the shared free-tier gateway starts refusing new sessions, new chat sessions seamlessly use a healthy fallback model on the same API, affected errors carry an actionable hint, in-progress chats are never rerouted, and normal routing resumes automatically on recovery. `/freeflow status` now shows live upstream state.
26
+
27
+ ## 1.16.0
28
+
29
+ ### Minor Changes
30
+
31
+ - New free model: Union Alpha Free joins the picker. It answers through a different upstream endpoint than the other free models, handled automatically — 262K context, image input, tool calling, and zero cost like the rest of the catalog.
32
+
33
+ ## 1.15.2
34
+
35
+ ### Patch Changes
36
+
37
+ - Relay traffic is now spent only on the free models the pool exists for: requests for other models go straight upstream even with the relay pool on, cutting relay bandwidth. Applies in automatic and always-on modes; relay-off behavior is unchanged.
38
+ - The disabled-deployment rollover now triggers only on genuine hosting verdicts, so ordinary payment and quota refusals always surface immediately without cooling a healthy relay. Relay-list imports also no longer show a negative removal count when the file holds more relays than the current pool.
39
+
40
+ ## 1.15.1
41
+
42
+ ### Patch Changes
43
+
44
+ - Requests now roll over to the next relay when the current one answers with a disabled-deployment verdict, instead of failing on the first relay and requiring a manual switch. Other payment and quota refusals still surface immediately.
45
+
46
+ ## 1.15.0
47
+
48
+ ### Minor Changes
49
+
50
+ - Free-tier requests now use native session identifiers and an up-to-date client version, after the upstream gateway began rejecting older formats with a free-tier error. `/freeflow test` also gains an end-to-end chat check that verifies a relay can actually run inference, not just list models.
51
+
52
+ ## 1.14.0
53
+
54
+ ### Minor Changes
55
+
56
+ - c28b7b5: Local proxy recovers on its own: if the background proxy stops unexpectedly, the extension now notices the refused local connection and starts a fresh proxy within seconds instead of leaving every model failing until the next session; shutdowns and crashes are also recorded in the log so the cause is visible.
57
+
58
+ ### Patch Changes
59
+
60
+ - /freeflow update now goes through the host plugin managers first — reinstall under OMP, package update under Pi — and only then falls back to a global install, using Bun when npm is missing. Previously the first step called an update action the plugin manager does not define, so managed installs always fell through to the wrong target.
61
+
62
+ ## 1.13.0
63
+
64
+ ### Minor Changes
65
+
66
+ - 39a06b9: Carry a relay pool between machines: `/freeflow export` saves the relays to a file and `/freeflow import` loads them back, adding to the current pool by default or swapping it whole on confirmation.
67
+
68
+ ## 1.12.1
69
+
70
+ ### Patch Changes
71
+
72
+ - b32f208: Sessions survive a relay change without a rejected turn.
73
+
74
+ Responses models sign each thinking block for the upstream backend that produced it, and only that backend can read it back. When a later turn reached a different backend, the replayed blocks were rejected as unreadable ("reasoning `encrypted_content` was not issued to this caller"), the host repeated the same failing request, and the session could not continue.
75
+
76
+ pi-freeflow now keeps each conversation on the relay that issued its reasoning while that relay is healthy, and when a relay change is unavoidable (rate limit, relay removed or redeployed, direct-mode switch) it sends that turn without the signed thinking blocks, so the new backend accepts it on the first attempt. If a rejection still happens (the provider can change backends behind a relay), the request is retried once without the blocks, and only the rejected blocks are dropped on later turns, so the model keeps whatever reasoning the current backend can read. Messages, tool calls and tool results are always preserved.
77
+
78
+ ## 1.12.0
79
+
80
+ ### Minor Changes
81
+
82
+ - Free-model picker grows from 24 to 26: adds Nex N2.5 Pro, Nex N2.5 Mini, and Ling 3.0 Flash VL, and removes the discontinued Inkling entry. A new Stealth Models Watchlist docs page tracks masked preview IDs users ask about.
83
+
84
+ ## 1.11.2
85
+
86
+ ### Patch Changes
87
+
88
+ - 549b66e: Oversized requests no longer fail at the relay hop: when a relay answers 413 payload limit, the proxy transparently tries the next relay and then the direct route, keeping the stream alive. Long sessions that outgrow the relay payload cap now complete instead of surfacing a function payload error.
89
+
90
+ ## 1.11.1
91
+
92
+ ### Patch Changes
93
+
94
+ - Apologies for the daemon disconnect bug introduced in v1.11.0: in multi-agent workflows or when subagents, evaluations, and background tasks completed, the extension prematurely detached from the local proxy daemon, causing the proxy to shut down while your main session remained active. The heartbeat connection now persists throughout your active terminal session.
95
+ - Increased local health check and liveness probe timeouts from 800ms and 500ms to 2500ms and 1500ms, with an automatic probe retry to avoid false-alarm daemon restarts during heavy concurrent streaming.
96
+ - Added unhandled error logging inside the daemon to prevent silent process exits.
97
+ - Fresh installs now log full HTTP lifecycle debug output by default so diagnostic reports contain complete request context; turn it off anytime with `/freeflow debug off`.
98
+
99
+ ## 1.11.0
100
+
101
+ ### Minor Changes
102
+
103
+ - Proxy daemon now stays up while any session uses it, even when idle. It shuts down only after the last session leaves, instead of retiring after a short quiet window.
104
+
105
+ ## 1.10.0
106
+
107
+ ### Minor Changes
108
+
109
+ - cf48598: Hands-free proxy recovery plus catalog refresh: the proxy now watches its own health, restarts with backoff instead of leaving a dead port, and can start on login. Removed two discontinued free models from the picker.
110
+
111
+ All notable changes to pi-freeflow. Public, user-visible behavior only.
112
+
113
+ ## 1.9.9 - 2026-09-06
114
+
115
+ ### Fixes
116
+
117
+ - **Removed our own request cap — sorry, that one was on us.** The proxy used to enforce a built-in request limit and could answer 429 before upstream quota was actually exhausted. That was a bug, not your quota. From this version the proxy never rejects on quota itself: a 429 only surfaces when the upstream — and every relay in your pool — genuinely is rate-limited, and that response now points you at `/freeflow deploy` to add relay egress.
118
+
119
+ ## 1.9.8 - 2026-09-06
120
+
121
+ ### Fixes
122
+
123
+ - **Stale model list heals itself (follow-up to #6).** If your saved model list predates a newly added model, the background refresh now repairs the entry (correct endpoint and details) instead of sending requests to the wrong address — no manual `/freeflow refresh` or cache deletion needed.
124
+ - **Paid models stay out of the picker even from old saved lists.** Every read of the saved model list now drops non-free entries, so models requiring an API key cannot linger after an upgrade.
125
+ - **Old saved lists without a sync marker now re-sync once.** A saved list that could never trigger a network check now performs one plain revalidation (then syncs normally), so newly added free models appear without manual intervention. Missing or corrupt lists still fall back silently with no network call.
126
+ - **Upstream errors are now visible in the proxy log.** Failed upstream responses log their status code and model, and a model routed to the wrong endpoint logs the mismatch with the fix (restart Pi/OMP after upgrade).
127
+
128
+ ### Validation
129
+
130
+ - TypeScript typecheck passed cleanly (`tsc --noEmit`).
131
+ - Full test suite passed on Windows (304 tests) and Ubuntu Linux (`acerblue`, 305/305 tests passed), including new regressions for the stale-cache shape from #6.
132
+ - Live sweep of all 26 models through a fresh install on `acerblue`: 23/26 answered on first try (both Muse Spark models via the Responses endpoint); the 3 misses are upstream per-model daily quotas (429), zero server errors.
133
+
134
+ ## 1.9.7 - 2026-09-06
135
+
136
+ ### Fixes
137
+
138
+ - **Model picker only shows verified free models (fixes #6).** Background catalog refresh now filters out non-free models from upstream endpoints so paid models requiring an API key no longer leak into your picker.
139
+ - **Relay pool automatically recovers from dead deployments (fixes #5).** If a relay URL in your pool returns an infrastructure 404 (such as a missing or deleted Vercel deployment), the proxy marks it as failed and immediately rolls over to your next healthy relay or direct mode instead of getting trapped in a 404 loop.
140
+
141
+ ### Changes
142
+
143
+ - **Updated model catalog (26 verified free models: 7 OpenCode Zen + 19 KiloCode Gateway).**
144
+ - Added `inclusionai/ling-3.0-flash-sante:free` (Ling 3.0 Flash Sante, 262K context, 32K output, reasoning supported). Clean CLI alias `ling-3.0-flash-sante` supported.
145
+ - Removed `laguna-s-2.1-free` (OpenCode Zen) after upstream dropped free-tier access. `poolside/laguna-s-2.1:free` on Kilo remains active.
146
+ - **Public by default for new relay deployments.** Relays deployed via `/freeflow deploy` are now public by default with no mandatory authentication tokens. This enables seamless copy-paste migration across proxy tools (such as 9router) while maintaining safety guards for allowed AI upstreams.
147
+
148
+ ### Validation
149
+
150
+ - TypeScript typecheck passed cleanly (`tsc --noEmit`).
151
+ - Full test suite passed across Windows (299 tests) and Ubuntu Linux (`acerblue-local`, 300/300 tests passed).
152
+ - End-to-end concurrency and failover stress testing verified (100 parallel requests, 20 subagent leases, burst failover).
153
+
154
+ ## 1.9.6 - 2026-09-03
155
+
156
+ ### Fixes
157
+
158
+ - **Hiding the status widget now sticks.** `/freeflow hide` survives restarts, new sessions, and background update notices — previously the widget could reappear on the next session.
159
+ - **Calmer startup with many sessions.** Sessions starting at the same time now converge on a single background proxy instead of each spawning its own; a spawn that never becomes ready falls back to in-process mode instead of running untracked.
160
+ - **Survives daemon restarts mid-session.** If the background proxy is replaced while sessions run, connected sessions re-register automatically instead of silently losing their lease (which could retire the daemon mid-use).
161
+ - **Stale-version cleanup targets the right process.** Windows port matching is now exact, and Linux without `lsof`/`fuser` resolves the holder via `/proc` instead of giving up.
162
+ - **`/freeflow logs --follow` no longer runs forever.** The live tail stops when the session ends and caps burst output; the update/install subprocess now has a 2-minute timeout.
163
+
164
+ ### Validation
165
+
166
+ - `npx tsc --noEmit` clean, `npm test` 299 (298 pass + 1 Linux-only skip) + smoke green on Windows; stress harness 7/7 on Windows and Linux `acerblue-local`. **`macOS not tested`** this cycle.
167
+
168
+ ## 1.9.5 - 2026-09-03
169
+
170
+ ### Changes
171
+
172
+ - **Catalog refreshed against live upstream model lists (26 models: 8 OpenCode Zen + 18 KiloCode Gateway).** Re-checked both upstreams against their live endpoints and live inference probes:
173
+ - **Added** `muse-spark-1.3-contributor-free` (OpenCode Zen, Responses API, 1M context, 131K output, vision) — verified live: completes with reasoning, accepts effort levels, and answers vision queries.
174
+ - **Removed** `hy3-free` (OpenCode Zen) upstream no longer serves it (`Model hy3-free is not supported`).
175
+ - **Removed** `tencent/hy3:free` and `meituan/longcat-2.0-free` (KiloCode Gateway) — upstream free tier dropped them (model unavailable / sign-in required).
176
+ - Pruned IDs added to the dead-model filter so a stale disk cache cannot resurrect them.
177
+
178
+ ### Validation
179
+
180
+ - `npx tsc --noEmit` clean, `npm test` full suite green on Windows.
181
+
182
+ ## 1.9.4 - 2026-09-02
183
+
184
+ ### Dependencies
185
+
186
+ - **Zero runtime dependencies.** Removed `undici@8.10.0` — `pi` (`0.84.4`) and `omp` (`18.1.3`) already bundle `undici` 6.x/7.x and expose `global fetch` with keep-alive pooling. `relayFetch` and proxy now use `global fetch` directly (`src/relay.ts` `Agent` + `canUseCustomDispatcher` + `dispatcher: agent` removed; `src/proxy.ts` dispatcher removed). Keeps thin `11.3k` + `298 tests` + `0 deps` compatible directly with `reference/pi` + `reference/oh-my-pi`.
187
+
188
+ ### Validation
189
+
190
+ - `npx tsc --noEmit` clean, `npm test 298/298` on **Windows** (`omp/18.1.3`, `pi 0.84.4`) and **Linux `acerblue-local`** (`Ubuntu 6.8.0-138`, `node v22.23.2`, `pi 0.84.4`) via `/tmp/pi-freeflow-validation`. **`macOS not tested`** this cycle.
191
+
192
+ ## 1.9.3 - 2026-09-02
193
+
194
+ ### Fixes
195
+
196
+ - **Windows console flood fixed.** Two Windows-only helpers flashed a visible `conhost`/`cmd` window on every daemon probe: `netstat -ano | findstr :28180` / `taskkill` in the stale-daemon replace path and `spawn(omp|npm, shell:true)` for `/freeflow update`. Both now use `windowsHide: true` (no-op on Linux/macOS) and `spawnWithProgress` was refactored to `Promise.withResolvers` to satisfy `ts-promise-with-resolvers`. Idle `pi` no longer spawns many visible consoles even after closing the terminal (detached daemon at `127.0.0.1:28180` survives by design; `beatOnce` 10s heartbeat now throttled 2s via `lastSpawnAt`).
197
+ - Daemon spawn now throttled per-process (2s) as a storm guard when `28180` is contended or blocked; the `ensuring` guard + `waitForReady 5s` already prevented tight loops.
198
+
199
+ ### Validation
200
+
201
+ - `npx tsc --noEmit` clean, `npm test 279/279` on **Windows** (`omp/18.1.3`, `pi 0.84.4`) and **Linux `acerblue-local`** (`Ubuntu 6.8.0-138`, `node v22.23.2`, `pi 0.84.4`) via `/tmp/pi-freeflow-validation`. **`macOS not tested`** this cycle.
202
+ - Reporter `LOYINuts` issue #3 (`pi idle creates many sessions → force reboot`) — `grep -r rtk src` ∅ confirms `rtk` is an external global skill (`~/.agents/skills/rtk` `Command::new("cmd")` without `CREATE_NO_WINDOW`), not `pi-freeflow`. After this fix, closing the terminal no longer leaves flashing zombies; kill via `netstat -ano | findstr :28180` → `taskkill /F /PID` or `/freeflow kill`.
203
+
204
+ ## 1.9.2 - 2026-08-31
205
+
206
+ ### Fixes
207
+
208
+ - **Closing one session no longer stops the shared local proxy.** The proxy daemon
209
+ is now a fully detached background process: it outlives any single OMP/Pi session
210
+ (previously, closing the session that owned the daemon could shut it down even
211
+ while other sessions were still using it). It retires by itself only when no
212
+ session is connected, no request is in flight, and it has been idle for a grace
213
+ period. The next use starts it again automatically.
214
+ - New command: `/freeflow kill` (aliases `stop`, `shutdown`) stops the background
215
+ daemon on demand. The next freeflow use restarts it.
216
+
217
+ ### Improvements
218
+
219
+ - The proxy tracks connected sessions and last request time; `/freeflow status`
220
+ and the health endpoint now report active session leases, so you can see when
221
+ other sessions are keeping the daemon alive.
222
+ - Docs: the command reference now lists `kill`, and the FAQ explains the shared
223
+ daemon lifecycle (survives session close; self-retires when unused).
224
+
225
+ ## 1.9.1 - 2026-08-30
226
+
227
+ ### Fixes
228
+
229
+ - Stale-daemon guard completed: a pre-1.9.0 daemon cannot report its in-flight requests, and usage cannot be verified — so it is now left running instead of being replaced. Previously such a daemon could still be killed mid-stream while busy (a one-time window when upgrading from 1.8.x). The guarantee now holds in every case: only older, verified-idle daemons are replaced.
230
+ - The stale-daemon kill ritual was consolidated into one path (it was duplicated at five call sites); behavior unchanged.
231
+ - Tests: new coverage for the pre-1.9 daemon case; shared sandbox helpers extracted.
232
+
233
+ ## 1.9.0 - 2026-08-30
234
+
235
+ ### Development hardening
236
+
237
+ - Test suite now runs fully sandboxed: every test uses a temporary data directory, so no test can ever write to your real `~/.pi/agent/` files or interfere with a running local proxy. A single env override (`PI_*_DATA_DIR`) re-roots all data files for tests/CI.
238
+ - Added a complete mocked user-flow e2e suite: fresh install onboarding proxy health direct chat → relay add/roll/fallback → guided deploy → update check → command surface — all deterministic, zero network.
239
+ - Docs: test counts no longer hardcoded in README (they drifted with releases); release history tracked here.
240
+
241
+ ### Fixes
242
+
243
+ - **Stale-daemon replacement no longer interrupts running sessions.** The shared local proxy can be held by another OMP/Pi session; the old upgrade logic killed that holder on version mismatch, which could terminate the session that owned the proxy mid-stream. Replacement now only happens when the running daemon is strictly older AND idle (0 in-flight requests, reported via `/_health`); newer or busy daemons are reused with a log note instead. Optional opt-out: set the no-kill env to `1` (see README FAQ for upgrades).
244
+
245
+ ## 1.8.2 - 2026-08-30
246
+
247
+ ### Fixes & polish
248
+
249
+ - Deploy: compare-and-swap on concurrent deployments (no duplicate relays when two sessions deploy at once).
250
+ - Proxy: URL-encoding edge cases (`%` in paths), startup timeout and kill handling, port conflict fallback.
251
+ - Logs: sanitized sensitive headers in debug output.
252
+ - Kilo gateway models: compatibility pass for all 25 models.
253
+ - Docs site build included; README per-host usage guide (Oh My Pi & Pi install + pick + manage).
254
+
255
+ ## 1.8.1 - 2026-08-29
256
+
257
+ - Fix: stale-daemon auto-heal is now shipped in the published package (previously only in the repo).
258
+
259
+ ## 1.8.0 - 2026-08-29
260
+
261
+ ### Automatic upgrades
262
+
263
+ - On version upgrade, the extension detects a stale local proxy daemon and replaces it automatically — users get fixes without manual restarts or killing sessions.
264
+ - Fix: client auth key stripped before reaching the opencode.ai/zen upstream (failed with 401 for some clients).
265
+ - 3 new free models (28 total).
266
+
267
+ ## 1.7.1 - 2026-08-29
268
+
269
+ - Docs: clarified npm is the distribution channel (the package is an extension loaded by OMP/Pi, not a standalone CLI).
270
+
271
+ ## 1.7.0 - 2026-08-29
272
+
273
+ - 4 new free models (25 total) with live-verified specs (context/output limits, vision, thinking levels).
274
+ - New alias map aligned with host model selectors.
275
+
276
+ ## 1.6.1 - 2026-08-29
277
+
278
+ - Catalog: thinking-level map locked per model; docs sync.
279
+
280
+ ## 1.6.0 - 2026-08-29
281
+
282
+ ### Onboarding & UX
283
+
284
+ - First-run onboarding message; 429 guidance hint (add your own relay egress).
285
+ - `/freeflow test` to probe a relay; relay latency tracking with health badges in `/freeflow list`.
286
+ - Guided deploy with context picker + confirmation; post-deploy health check.
287
+ - Status clarity: current mode + state file path; log text filter; per-relay usage counters; throttled roll notifications.
288
+
289
+ ### Reliability
290
+
291
+ - Relay state write-protection: before every save, the current state is snapshotted to `.bak`; if the main file is corrupted or missing, the backup is recovered automatically.
292
+ - New-user flow never seeds a fake relay; starts in direct mode with an empty pool.
293
+
294
+ ## 1.5.1 - 2026-08-28
295
+
296
+ - Edge-sweep fixes: port 28180 with legacy 18080 dual-probe auto-migration, OMP/Pi compatibility.