download: fix plugin wasm 404 coercion + api.github.com rate-limit lockout; back off failed installs#2
Open
dywongcloud wants to merge 1 commit into
Conversation
…e-limit lockout; daemon: back off failed plugin installs Three interlocking defects that together made bert/libsql/treesitter permanently uninstallable under agentplug-runner -- which is what broke embed_text (memorize-fire "embed_text failed ... refusing silent-NULL- embedding insert", recall "invalid query embedding" LIKE-fallback; no bert sibling for host_vec_embed, and plugkit-slim.wasm deliberately carries no baked-in bert weights so the wasm-side fallback that saved the retired JS host cannot engage) and codesearch (permanent libsql_ok=false fallback_kv mode with zero chunks ever indexed; no libsql sibling): 1. ensure_plugin_installed's effective_basename decision -- written for gm's plugkit-slim-vs-plugkit fallback -- ran for EVERY plugin and coerced every non-slim download to "plugkit": bert's sha sidecar fetch succeeded under its own name, then the wasm fetch requested agentplug-bert-bin/.../plugkit.wasm, a guaranteed 404 (live-verified: bert.wasm.sha256 = 200, same-release plugkit.wasm = 404). Non-gm plugins now keep their own basename; only gm's fat-fallback case coerces (extracted to effective_wasm_basename() with unit tests). 2. "Latest" version resolution used unauthenticated api.github.com releases/latest calls (60 requests/hour/IP). Combined with defect 3 this exhausted the budget within minutes of boot and then permanently locked out every plugin install on the machine -- live-witnessed as "API rate limit exceeded" on all three plugin-bin repos while the releases themselves were fine. Both fetch_latest_plugin_version and fetch_latest_runner_version now resolve the tag from the plain github.com releases/latest/download/<asset> redirect's first-hop Location header (resolve_latest_tag_via_redirect) -- not API rate-limited, no auth needed, and the redirect fires before any asset-existence check so it needs no slim/fat fallback of its own. 3. The daemon's compile-ahead pass retried a failing plugin install on every ~200ms main-loop tick -- each attempt a fresh network round trip (the API-quota burner above) plus a five-times-a-second error log line. PluginModules now records failed_at per plugin and backs off PLUGIN_INSTALL_RETRY_BACKOFF (60s) before re-attempting, skipping silently inside the window so the tick loop stops spamming both the remote and the log. Verified live on the previously-broken machine: bert/libsql/treesitter all download+sha-verify+install via the fixed path, and against the rebuilt runner a real memorize-fire dispatch returns ok:true embedded:true, recall returns genuine cosine-ranked vector hits, and codesearch returns mode:fusion with real symbol hits (fetchWithTimeout / ensureSessionMinted from ui/lib/api.ts) instead of fallback_kv-empty. cargo test -p agentplug-runner: 4/4 passing.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Three interlocking defects made bert/libsql/treesitter permanently uninstallable under agentplug-runner — the root cause of two live-hit failures: embed_text always failing (
memorize-firerefusing NULL-embedding inserts,recallfalling back to LIKE search withinvalid query embedding) and codesearch stuck infallback_kvwithlibsql_ok=falseand zero chunks ever indexed.Every non-gm plugin's wasm URL was coerced to
plugkit.wasm.ensure_plugin_installed'seffective_basenamedecision (written for gm's plugkit-slim-vs-plugkit fallback) ran for every plugin: bert's sha sidecar fetch succeeded under its own name, then the wasm fetch requestedagentplug-bert-bin/.../plugkit.wasm— a guaranteed 404 (live-verified:bert.wasm.sha256= 200, same-releaseplugkit.wasm= 404). Non-gm plugins now keep their own basename; extracted to a pureeffective_wasm_basename()with unit tests."Latest" version resolution burned the unauthenticated api.github.com rate limit. Combined with (3), the 60-req/hour/IP budget was exhausted within minutes of boot, permanently locking out all plugin installs (live-witnessed:
API rate limit exceededon all three plugin-bin repos while the releases were fine). Version tags now come from the plaingithub.com/<repo>/releases/latest/download/<asset>redirect's first-hopLocationheader — not API-rate-limited, no auth needed.Failed installs were retried every ~200ms daemon tick — each a fresh network round trip plus a 5×/sec error log line.
PluginModulesnow backs off 60s per failed plugin, skipping silently inside the window.Why the wasm-side bert fallback (which saved the retired JS host) doesn't apply here: under agentplug the gm plugin ships as
plugkit-slim.wasm, which deliberately carries no baked-in bert weights — correctness requireshost_vec_embed's bert sibling to actually load, i.e.bert.wasmmust install.Test plan
cargo build --release— clean, zero warningscargo test -p agentplug-runner— 4/4 passing (basename coercion regression table, CDN-querystring slim detection, fat-fallback case, redirect-Location tag parsing)memorize-firedispatch returnsok:true, embedded:true,recallreturns cosine-ranked vector hits, andcodesearchreturnsmode:fusionwith real symbol hits (fetchWithTimeout/ensureSessionMintedfromui/lib/api.ts) instead of fallback_kv-empty🤖 Generated with Claude Code