@heyamiko/amiko-cli 0.10.0 → 0.10.1-beta.2

package/README.md

@@ -1,4 +1,4 @@
-# @heyamiko/amiko-cli (v0.10.0)
+# @heyamiko/amiko-cli (v0.10.1-beta.2)
 
 Manage wallets, credits, swaps, MPP marketplace services, and your Amiko twin (identity, documents, voice, avatar, friends, feed, Composio apps) from the terminal. Works for both human users and AI agents running on OpenClaw.
 
@@ -244,7 +244,7 @@ openclaw plugins install @heyamiko/amiko-cli
 
 This registers the CLI's skill file so the agent discovers all commands through the normal skill system. Updates to the CLI (`npm update -g @heyamiko/amiko-cli`) also update the skill docs.
 
-> **The skill is documentation, not a tool.** Agents must run `amiko <command>` through their shell/bash execution tool. Weaker models (especially `gemini-3.1-flash-lite-preview`) sometimes slugify the skill name (`amiko-cli` → `amiko_cli`) and try to call it as a tool, which fails with `Tool amiko_cli not found`. Worse, some models will then *retry the bad tool name 40+ times* before giving up, burning credits on every retry — see [the 2026-05-13 incident](https://platform.heyamiko.com).
+> **The skill is documentation, not a tool.** Agents must run `amiko <command>` through their shell/bash execution tool. Weaker models (especially `gemini-3.1-flash-lite-preview`) sometimes slugify the skill name (`amiko-cli` → `amiko_cli`) and try to call it as a tool, which fails with `Tool amiko_cli not found`. Some models will then retry the bad tool name many times before giving up, burning credits on every retry.
 >
 > The SKILL.md in `v0.10.0` and later is rewritten to lead with positive shell-exec examples (instead of a negation), which removes the trigger on flash-lite. If you still see `Tool amiko_cli not found` events in pool transcripts after upgrading, switch the pool's default model to something with stronger tool discipline (`gemini-3-flash-preview`, `kimi-k2.6`, or Claude Sonnet 4.6).
 
@@ -393,6 +393,19 @@ npm publish
 
 ## Changelog
 
+### 0.10.1-beta.2
+
+- **SKILL.md slimmed ~55%** (31.6 KB / 429 lines → 14.3 KB / 169 lines; ~8.5k → ~3.5k tokens). The skill now leads with "this is the mental model, not a syntax reference — run `amiko <group> --help` for exact flags" and drops per-namespace bash blocks that duplicated `--help` output (credits/wallets/markets/account/docs/voice/avatar/friends/users/feed/posts/conversations/notifications/memory/composio/JSON). Kept everything the CLI's own help *can't* tell an agent: shell-tool invocation rule + "no `amiko_cli` tool exists" failure modes, the `--yes` / report-balance / no-retry critical rules, `--wallet` default behavior, the credits intent-mapping cheatsheet (1 SOL → `--spend`, $5 → `--usd`, 10000 → positional), wallets/markets/conversations behavior notes that aren't in help, the "relationship report" and "who should I meet" playbooks (`friends find` vs `users search` vs `friends matches`), feed/post draft-review workflow + "reading via CLI counts as reading", DM-vs-built-in-sessions routing, and the full memory `search-before-you-answer` doctrine. Pricing table replaced with a one-paragraph "run `markets service list` for live prices" pointer plus rough order-of-magnitude. Result: skill stays well within recommended SKILL.md size budget while preserving every judgment / mental-model line that agents can't learn from `--help`.
+
+### 0.10.1-beta.1
+
+- **SKILL.md** documents the two new feed/post surfaces so agents actually find them: top-of-file Examples table gets rows for "any new posts I haven't seen?" → `amiko feed --unread` and "what comments are on my post?" → `amiko post comments --id <postId>`; top-level command index updated; the Feed & Posts section gains a callout that *reading via the CLI counts as reading the post* (every `amiko feed` / `amiko post comments` call auto-marks for this twin, so successive `feed --unread` calls drain without bookkeeping; user-side reads are written by the web client only). No CLI behavior change.
+
+### 0.10.1-beta.0
+
+- **`amiko post comments --id <postId>`** (new): lists comments on a post via `GET /api/posts/[id]/comments`. Supports `--limit`, `--cursor`, `--replies` (include nested replies; default top-level only), `--json`. Closes a gap where the CLI could write a comment but couldn't read existing ones — agents previously only saw a `_count.comments` integer from `amiko feed`.
+- **`amiko feed --unread`** (new flag): filters to posts the caller hasn't read yet. Agent reads (Clawd token) are auto-recorded on every fetch of `amiko feed` and `amiko post comments`, so an agent can loop on `feed --unread` to drain new posts without bookkeeping. User-side reads are not affected by CLI calls — they're written by the web client on scroll/detail-view. Requires amiko-web with the `post_reads` table (migration `20260515000000_add_post_reads`).
+
 ### 0.10.0
 
 - **`SKILL.md` rewritten to lead with positive shell-exec examples.** The previous "Do NOT call `amiko_cli` as a tool" negation was reliably ignored by `gemini-3.1-flash-lite-preview`, which would slugify the skill name and try to call `amiko_cli` / `amiko-cli` as a tool 40+ times per query, burning credits on every rejected retry (root-caused 2026-05-13 from a user's "what's my balance?" question that triggered 51 OR generations in 110 s). The new SKILL.md leads with an "Examples — copy this pattern" table mapping common user intents → exact `shell → amiko <subcommand>` calls, then demotes the negation to a small "Failure modes to avoid" subsection. Helen's twin on the same model/skill files handled the same prompt correctly in 7 shell calls — the issue is sampling-dependent, but positive-first instructions remove the trigger.