@te-river/opencode-team-mode 1.7.6 → 1.7.8

package/docs/v2-port-notes.md

@@ -497,83 +497,83 @@ survivors were written. Each is recorded here so the next reader does not re-der
   providerExecuted,cache,metadata,providerMetadata`; `media` → `type,media,filename,metadata`.
   The binary's union also names `compaction` and `effort`, which were never observed in a turn.
   This is why the prune layer's first version was inert: it accepted only `type:"text"`, and the
-  bulk of a real request is `tool-result`/`tool-call`. `[L]`
-
-### 2026-10-07 — the 1.7.1 cut: the 75% trigger is gone, and why the readers stayed `[L][R][D]`
-
-**Decision (user, 2026-10-07).** The early-compaction trigger shipped in 1.7.0 was removed
-outright: the host owns when a conversation is summarized, and a plugin that summarizes a live
-session on its own takes a context-losing action the user did not ask for. This is a
-`[D]`+product decision, not a measurement — recorded here because the *other* two facts below
-are measurements, and a future reader will otherwise look for a falsified claim that does not
-exist.
-
-**What was removed.** `src/host/v2-compaction.ts` (the whole layer: `applyV2EarlyCompaction`,
-`resolveCompactConfig`, `CompactReport`, the `session.hook("context")` registration and the
-`observeUsage` tap), the `onUsage` option on `applyV2EventFeed`, the `compact` input and the
-`ctx.session.compact` capability row in `v2-capabilities.ts`, the `compact_*` counters on both
-the `v2-surface` and `v2-shutdown` rows, the `v2-compact` trajectory line, the three
-file-overridable registry keys (`compactTrigger` / `compactAtPercent` / `compactMinMs`), and
-the `compact` seam on the test fake. An operator's `team-mode.jsonc` that still names those
-keys now collects them as UNKNOWN keys (a boot warning), which is the honest outcome: a knob
-whose feature is gone must not keep reading as if it did something. `[R]`
-
-**What was kept, and it is not sentiment.** The four pure usage readers (`readUsedTokens` /
-`readModelKey` / `lastUsageOf` / `percentOf`) moved to `src/host/v2-usage.ts` because
-`src/host/v2-prune.ts` still computes ITS OWN threshold from the same formula and the same
-`ctx.model.list()` → `limit.context` denominator. Copying them into the prune layer instead
-would have left two definitions of "how full is this window" free to drift, and the evidence
-below is the reason each branch of those readers exists. `[R]`
-
-**The evidence that shaped them (unchanged, re-stated so it is not lost with the module).**
-- Context usage lives ONLY in the `session.usage.updated` event. `session.hook("context")`
-  hands over the assembled messages and those messages carry no token counts: a second turn
-  that DID contain an assistant message still logged `source=no_usage messages=3` (run
-  `r-20261005-235343-114738`). The denominator is `ctx.model.list()`'s `limit.context`; the
-  numerator is the event. `[L]`
-- The readers therefore accept every location the host has actually used (`tokens` /
-  `metadata.tokens` / `info.tokens` / `usage`) and return `null` — never `0` — when they
-  find none, because "we did not find them" and "usage is zero" must not collapse into one
-  number. A model the catalog does not describe gets `percentOf` = 0, never a guessed window. `[R]`
-
-**What 1.7.1 does NOT change.** The host's own compaction is still supported, not intercepted:
-`TM_COMPACTION_CONTEXT` (the survival list on `session.hook("compaction")`) and
-`TM_COMPACTION_AUTOCONTINUE` are untouched, `compaction_lines` still rides `tm_stats`, and the
-`ctx.session.compact` seam stays declared in `src/host/v2-types.ts` with a "DECLARED, NOT
-CALLED" note — a future layer would otherwise have to re-probe a seam the 2.0.23 probe already
-settled. Context Pruning (`TM_PRUNE`, default 70%) is a different mechanism: it rewrites settled
-message BODIES into pointers inside an outgoing request and never asks the host to summarize
-anything. `[R]`
-
-### 2026-10-07 — the agent ID is the FILE NAME, and a case-only difference is not cosmetic `[L][R]`
-
-**Measured on 2.0.24 / 2026-10-07 (this machine).** The global config directory held
-`agents/team.md` — lowercase, left over from the pre-1.7 naming — while the generator had been
-"writing" `agents/Team.md` on every install since. The host was asked directly:
-
-- `opencode-cli run --standalone --agent Team --model nope/nope "x"` → `Agent not found: "Team"`
-- `opencode-cli run --standalone --agent team --model nope/nope "x"` → resolves (fails only on
-  the fake model: `Model unavailable: nope/nope`)
-- `opencode-cli api get /api/agent?directory=…` → `{"id":"team","name":"team",…}` for the lead
-
-So the on-disk spelling IS the registered id, and the installer's own
-`default_agent: "Team"` named an agent that did not exist. Per the agents doc the fallback for
-an unresolvable default is `build` — silent, which is exactly the failure the installer's
-"write `default_agent` LAST, and read it back" rule exists to prevent; the read-back cannot see
-it, because the key IS on disk and the ID is what is wrong. `[L]`
-
-**Why every install failed to fix it.** `writeFileSync(<dir>/Team.md)` on a case-insensitive,
-case-PRESERVING filesystem opens the existing `team.md` and leaves that name alone — and the
-reclaim pass deliberately refused to delete the case-only entry (deleting it would delete the
-file just written: the #47/#48 guard). Nothing in the pipeline could ever change the spelling.
-The fix is an explicit two-step rename (`team.md` → temp → `Team.md`; a case-only `rename` on
-win32 is itself a no-op), performed ONLY on an entry the identity verdict has already PROVEN is
-the same file as a target, with a restore path if the second step throws. `[R]`
-
-**The lesson, stated in the general form.** "Is the tool loaded" and "is the file on disk" were
-both yes here — the two checks the installer prints. What no check covered was *which id the
-host derives from it*, and that is the difference between Team being the default and a silent
-`build`. A verifiable claim has to name the artefact the host actually keys on (the FILE NAME),
+  bulk of a real request is `tool-result`/`tool-call`. `[L]`
+
+### 2026-10-07 — the 1.7.1 cut: the 75% trigger is gone, and why the readers stayed `[L][R][D]`
+
+**Decision (user, 2026-10-07).** The early-compaction trigger shipped in 1.7.0 was removed
+outright: the host owns when a conversation is summarized, and a plugin that summarizes a live
+session on its own takes a context-losing action the user did not ask for. This is a
+`[D]`+product decision, not a measurement — recorded here because the *other* two facts below
+are measurements, and a future reader will otherwise look for a falsified claim that does not
+exist.
+
+**What was removed.** `src/host/v2-compaction.ts` (the whole layer: `applyV2EarlyCompaction`,
+`resolveCompactConfig`, `CompactReport`, the `session.hook("context")` registration and the
+`observeUsage` tap), the `onUsage` option on `applyV2EventFeed`, the `compact` input and the
+`ctx.session.compact` capability row in `v2-capabilities.ts`, the `compact_*` counters on both
+the `v2-surface` and `v2-shutdown` rows, the `v2-compact` trajectory line, the three
+file-overridable registry keys (`compactTrigger` / `compactAtPercent` / `compactMinMs`), and
+the `compact` seam on the test fake. An operator's `team-mode.jsonc` that still names those
+keys now collects them as UNKNOWN keys (a boot warning), which is the honest outcome: a knob
+whose feature is gone must not keep reading as if it did something. `[R]`
+
+**What was kept, and it is not sentiment.** The four pure usage readers (`readUsedTokens` /
+`readModelKey` / `lastUsageOf` / `percentOf`) moved to `src/host/v2-usage.ts` because
+`src/host/v2-prune.ts` still computes ITS OWN threshold from the same formula and the same
+`ctx.model.list()` → `limit.context` denominator. Copying them into the prune layer instead
+would have left two definitions of "how full is this window" free to drift, and the evidence
+below is the reason each branch of those readers exists. `[R]`
+
+**The evidence that shaped them (unchanged, re-stated so it is not lost with the module).**
+- Context usage lives ONLY in the `session.usage.updated` event. `session.hook("context")`
+  hands over the assembled messages and those messages carry no token counts: a second turn
+  that DID contain an assistant message still logged `source=no_usage messages=3` (run
+  `r-20261005-235343-114738`). The denominator is `ctx.model.list()`'s `limit.context`; the
+  numerator is the event. `[L]`
+- The readers therefore accept every location the host has actually used (`tokens` /
+  `metadata.tokens` / `info.tokens` / `usage`) and return `null` — never `0` — when they
+  find none, because "we did not find them" and "usage is zero" must not collapse into one
+  number. A model the catalog does not describe gets `percentOf` = 0, never a guessed window. `[R]`
+
+**What 1.7.1 does NOT change.** The host's own compaction is still supported, not intercepted:
+`TM_COMPACTION_CONTEXT` (the survival list on `session.hook("compaction")`) and
+`TM_COMPACTION_AUTOCONTINUE` are untouched, `compaction_lines` still rides `tm_stats`, and the
+`ctx.session.compact` seam stays declared in `src/host/v2-types.ts` with a "DECLARED, NOT
+CALLED" note — a future layer would otherwise have to re-probe a seam the 2.0.23 probe already
+settled. Context Pruning (`TM_PRUNE`, default 70%) is a different mechanism: it rewrites settled
+message BODIES into pointers inside an outgoing request and never asks the host to summarize
+anything. `[R]`
+
+### 2026-10-07 — the agent ID is the FILE NAME, and a case-only difference is not cosmetic `[L][R]`
+
+**Measured on 2.0.24 / 2026-10-07 (this machine).** The global config directory held
+`agents/team.md` — lowercase, left over from the pre-1.7 naming — while the generator had been
+"writing" `agents/Team.md` on every install since. The host was asked directly:
+
+- `opencode-cli run --standalone --agent Team --model nope/nope "x"` → `Agent not found: "Team"`
+- `opencode-cli run --standalone --agent team --model nope/nope "x"` → resolves (fails only on
+  the fake model: `Model unavailable: nope/nope`)
+- `opencode-cli api get /api/agent?directory=…` → `{"id":"team","name":"team",…}` for the lead
+
+So the on-disk spelling IS the registered id, and the installer's own
+`default_agent: "Team"` named an agent that did not exist. Per the agents doc the fallback for
+an unresolvable default is `build` — silent, which is exactly the failure the installer's
+"write `default_agent` LAST, and read it back" rule exists to prevent; the read-back cannot see
+it, because the key IS on disk and the ID is what is wrong. `[L]`
+
+**Why every install failed to fix it.** `writeFileSync(<dir>/Team.md)` on a case-insensitive,
+case-PRESERVING filesystem opens the existing `team.md` and leaves that name alone — and the
+reclaim pass deliberately refused to delete the case-only entry (deleting it would delete the
+file just written: the #47/#48 guard). Nothing in the pipeline could ever change the spelling.
+The fix is an explicit two-step rename (`team.md` → temp → `Team.md`; a case-only `rename` on
+win32 is itself a no-op), performed ONLY on an entry the identity verdict has already PROVEN is
+the same file as a target, with a restore path if the second step throws. `[R]`
+
+**The lesson, stated in the general form.** "Is the tool loaded" and "is the file on disk" were
+both yes here — the two checks the installer prints. What no check covered was *which id the
+host derives from it*, and that is the difference between Team being the default and a silent
+`build`. A verifiable claim has to name the artefact the host actually keys on (the FILE NAME),
 not the one we wrote (the path we computed). `[R]`
 
 ### 2026-10-18 — the request-layer `temperature` fill is gone, because a field we added broke other models `[L][R]`
@@ -608,4 +608,47 @@ overwrite path. `[R]`
 no-op because it is "only a default". The host assembled a request that worked; our fill made
 it a request some providers refuse. A plugin that improves determinism on one model can
 therefore remove compatibility on another, and the compatibility cost is invisible in our own
-logs — it shows up as somebody else's API error. `[R]`
+logs — it shows up as somebody else's API error. `[R]`
+
+### 2026-10-21 — host 2.1.x: where its release notes actually live, and the #57260 contract renames that bind our seams `[D][L]`
+
+**Where to look.** The desktop shipped 2.1.2 (`opencode-cli.version` = `2.1.2`; boot log
+`msg="cli starting" version=2.1.2 channel=latest`). `api.github.com/repos/sst/opencode/releases/latest`
+answers **v1.19.3** — the 2.x desktop line is not published as a GitHub release of that repo,
+so `/releases/latest` (which also skips prereleases) is the wrong source and a search there
+finds nothing. The authoritative notes are `https://opencode.ai/changelog`. `[D]`
+
+**v2.1.0 is a BREAKING contract release (#57260)** — "intentional changes to public SDK, HTTP
+message/event, and plugin contracts". The renames that touch our readers: tool parts lose
+separate `providerState`/`providerResultState` — call and result data share `tool.native`
+(success/failure events shallow-merge native into the call's native, result wins on duplicate
+keys); `SessionCompactionResult.providerState` → `native`; compaction `recent` →
+`metadata.keepID` reference; event names `session.tool.called` / `session.tool.success` /
+`session.tool.failed` / `session.step.ended` / `session.step.failed` /
+`session.compaction.started|ended`; `Credential.Key.configuration` removed (writes use
+`metadata`). Our seams named in the risk list: `v2-offload.ts` (native-result shape measured as
+`{content, metadata, output}` on 2.0.16), `v2-subagent.ts` (`result.metadata.sessionID` off the
+ack), `v2-prune.ts` (parts-by-shape), `v2-events.ts` (whitelisted event NAMES — a renamed event
+arrives as unrecognised and is counted, never forwarded on a guess), `v2-session.ts`
+(compaction hook `output.context`). The design law caps the blast radius: an unrecognized shape
+is left alone and COUNTED (`unknownParts`, `unmatched`, `scope_unknown`), so the expected failure
+mode on a drifted field is silent degradation visible in `tm_stats`, not corruption.
+Post-install checks owed on 2.1.2: the `v2-boot` row, a native `shell`/`read` result actually
+offloading (`native:shell offloaded:true`), a background child settling via injection, and the
+`v2-surface` counters' `unmatched`/`unknownParts` deltas. `[D]`
+
+**v2.1.2 (#57146, #57589) adds `retryAfterMs` to structured session errors** and fixes retry
+timing being dropped from quota/rate-limit/provider-internal errors. Our retry governor
+(`v2-retry.ts`) parses `Retry-After` out of the error TEXT only — the structured field is a
+follow-up read point (the 1.7.8 ladder is fixed 5s/10s/20s per the user's ask regardless, so
+this is an enrichment, not a break). Also relevant: #54276 fixed Code Mode async jobs
+interleaving (our `execute` programs), and #57416 keeps locations loaded while long tools run
+(the new HTTP ladder's worst case is ~115 s single-URL, ~165 s across redirect hops).
+v2.1.0's `experimental.todo` (a TUI `todowrite`) is config-gated and off by default —
+`tm_ledger` remains our lead's ledger. `[D]`
+
+**Live evidence, same day `[L]`:** this session ran on 2.1.2 with the installed 1.7.7 plugin —
+three forced-background `subagent` dispatches each settled and their full bodies were collected
+(the completion envelope injection path works on 2.1.2), and the boot probe against the real
+config shows `loading plugin … @te-river/opencode-team-mode@latest` from the host cache with the
+`[team-mode] v2 人格已启动` note and no load failure.

package/index.js

@@ -1,7 +1,7 @@
-// The host resolves a plugin DIRECTORY as `<dir>/index.js` (measured on 2.0.16:
-// `msg="loading plugin" … entrypoint=file:///…/team-mode/index.js`, while a directory that
-// only declares package.json#exports is skipped with no error at all). This shim is
-// therefore part of the contract, not a convenience: without it neither
-// `plugins: ["…/opencode-team-mode"]` in a config nor a node_modules install under that
-// key can load the plugin. The dual-personality shape itself lives in dist/index.js.
-export { default } from "./dist/index.js"
+// The host resolves a plugin DIRECTORY as `<dir>/index.js` (measured on 2.0.16:
+// `msg="loading plugin" … entrypoint=file:///…/team-mode/index.js`, while a directory that
+// only declares package.json#exports is skipped with no error at all). This shim is
+// therefore part of the contract, not a convenience: without it neither
+// `plugins: ["…/opencode-team-mode"]` in a config nor a node_modules install under that
+// key can load the plugin. The dual-personality shape itself lives in dist/index.js.
+export { default } from "./dist/index.js"

package/package.json

@@ -1,64 +1,64 @@
-{
-  "name": "@te-river/opencode-team-mode",
-  "version": "1.7.6",
-  "description": "Team collaboration mode plugin for OpenCode Desktop — injects specialized agents (architect, reviewer, implementer, tester, researcher) and team workflow commands into your OpenCode workspace.",
-  "type": "module",
-  "main": "dist/index.js",
-  "types": "dist/index.d.ts",
-  "exports": {
-    ".": {
-      "import": "./dist/index.js",
-      "types": "./dist/index.d.ts"
-    }
-  },
-  "files": [
-    "dist",
-    "index.js",
-    "scripts/lib",
-    "scripts/gen-v2-config.mjs",
-    "README.md",
-    "README.zh-CN.md",
-    "LICENSE",
-    "docs"
-  ],
-  "scripts": {
-    "build": "tsc",
-    "test": "node scripts/run-tests.mjs",
-    "test:serial": "node scripts/run-tests.mjs --serial",
-    "test:one": "node scripts/run-tests.mjs --no-build",
-    "prepublishOnly": "tsc",
-    "dev": "tsc --watch"
-  },
-  "keywords": [
-    "opencode",
-    "opencode-plugin",
-    "plugin",
-    "team-mode",
-    "collaboration",
-    "multi-agent",
-    "desktop",
-    "ai-coding"
-  ],
-  "author": "",
-  "license": "Apache-2.0",
-  "publishConfig": {
-    "access": "public"
-  },
-  "optionalDependencies": {
-    "playwright-core": "1.63.0"
-  },
-  "peerDependencies": {
-    "@opencode-ai/plugin": ">=0.1.0"
-  },
-  "devDependencies": {
-    "@types/node": "^26.4.0",
-    "typescript": "^5.5.0"
-  },
-  "repository": {
-    "type": "git",
-    "url": "https://github.com/Te-River/Opencode-TeamMode.git"
-  },
-  "engines": {
-    "node": ">=18.0.0"
-  }
-}
+{
+  "name": "@te-river/opencode-team-mode",
+  "version": "1.7.8",
+  "description": "Team collaboration mode plugin for OpenCode Desktop — injects specialized agents (architect, reviewer, implementer, tester, researcher) and team workflow commands into your OpenCode workspace.",
+  "type": "module",
+  "main": "dist/index.js",
+  "types": "dist/index.d.ts",
+  "exports": {
+    ".": {
+      "import": "./dist/index.js",
+      "types": "./dist/index.d.ts"
+    }
+  },
+  "files": [
+    "dist",
+    "index.js",
+    "scripts/lib",
+    "scripts/gen-v2-config.mjs",
+    "README.md",
+    "README.zh-CN.md",
+    "LICENSE",
+    "docs"
+  ],
+  "scripts": {
+    "build": "tsc",
+    "test": "node scripts/run-tests.mjs",
+    "test:serial": "node scripts/run-tests.mjs --serial",
+    "test:one": "node scripts/run-tests.mjs --no-build",
+    "prepublishOnly": "tsc",
+    "dev": "tsc --watch"
+  },
+  "keywords": [
+    "opencode",
+    "opencode-plugin",
+    "plugin",
+    "team-mode",
+    "collaboration",
+    "multi-agent",
+    "desktop",
+    "ai-coding"
+  ],
+  "author": "",
+  "license": "Apache-2.0",
+  "publishConfig": {
+    "access": "public"
+  },
+  "optionalDependencies": {
+    "playwright-core": "1.63.0"
+  },
+  "peerDependencies": {
+    "@opencode-ai/plugin": ">=0.1.0"
+  },
+  "devDependencies": {
+    "@types/node": "^26.4.0",
+    "typescript": "^5.5.0"
+  },
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/Te-River/Opencode-TeamMode.git"
+  },
+  "engines": {
+    "node": ">=18.0.0"
+  }
+}