Repository navigation
feat: Codex CLI subprocess provider + web-fetch & Steam Deck-aware research - #24
Merged
Merged
Conversation
Let users on an OpenAI Codex CLI ChatGPT subscription run funfa with no API key and no metered cost. When no provider key is set (or FUNFA_PROVIDER=codex is forced) and codex is installed + logged in, funfa delegates investigate->propose to `codex exec`: codex runs the hardware probes via its own shell and researches via its own web search, returning the plan as `--output-schema` JSON. funfa keeps the consent gate, the on-disk write, and the verbatim render client-side. Spike (codex 0.135.0) showed the app-server protocol exposes no client-tool registration, so funfa's ToolLoopAgent cannot drive codex with its own tools; codex owns the loop on this path. The API-key Vercel path is unchanged. Precedence: forced codex -> API key wins -> codex fallback -> missing-keys. - codex-cli.ts: CodexCli, makeBunCodexCli (Bun.spawn, detect via Bun.which + auth.json, model from config.toml/FUNFA_CODEX_MODEL), proposeViaCodex, buildCodexInstructions/Prompt (inlines the 4 skills + a no-tools/JSON-output override), CODEX_OUTPUT_SCHEMA, CodexUnavailableError. - provider.ts: findApiKey, isCodexForced, resolveProviderIntent. - submit-plan.ts: extract shared applySubmitPlan (consent-gate parity). - registry.ts: export skillBodyByName. deps.ts: new error kinds. - index.ts: branch on provider intent. - FUNFA_FAKE_HARDWARE: env-gated dev affordance to run the loop off-Linux by injecting hardware facts (skips probes). Off by default. - compile.test guarded so the no-key path still hits missing-keys when a real codex is present on the dev/CI machine. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two linked improvements driven by a Steam Deck dogfood (AC Revelations, controller broken): the loop must (a) know it's on a Deck and (b) read the latest community fix, not the stale popular one. - New fetch_page tool (Vercel/BYOK path): fetches a cited URL and returns stripped readable text so the agent reads past search snippets to the accepted answer AND the most recent comments. makeBunFetchPage is the production impl (http(s) only, size/time capped, light HTML->text). Threaded via experimental_context like the other tools; codex path uses its own browsing instead. - Lift the old "do not fetch" rule in game-research (it blocked finding the real fix) and add "Read the source, and read the latest": paginate threads (?ctp=N), weight recent comments, a newer "no longer works / new fix" comment wins over the popular one. Keep the untrusted-data stance: ignore embedded instructions, never recommend a forum-only download. - Steam Deck awareness across skills: hardware detects the Deck (ID=steamos / VARIANT_ID=steamdeck / Van Gogh GPU); game-research reads ProtonDB's Deck reports and treats controller/input as a first-class fix (Steam Input, community keyboard-mapped layouts); funfa PHASE 1 pursues the input fix; plan allows a Steam Input layout as a high-tier item. - Tools count 4 -> 5; CLAUDE.md updated (tool list + security note). Verified: Deck mock for AC Revelations now headlines the Dankaishin keyboard-mapped Steam Input layout (Arrax's current fix), cites the thread at ?ctp=2, flags GE-Proton 10-21 as broken, and marks the old Proton 8.0-5 path stale. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Gate the fetch_page tool to a default set of trusted gaming domains (ProtonDB, PCGamingWiki, GitHub, Steam community/store, SteamDB); subdomains allowed. Other hosts return an error without a request, and the post-redirect host is re-checked so an allowed domain can't bounce the fetch off-list. Override via FUNFA_FETCH_ALLOWLIST (comma-separated) — this is a starting point, meant to be tuned. Scope note: this gates funfa's own fetch_page only. On the codex path, codex browses through its own sandbox and is NOT constrained by this list — funfa only runs codex with `-s read-only` + `--ignore-user-config`. Documented in CLAUDE.md. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The plan skill now tells the model to clean source URLs before citing them — drop locale/tracking query params (l=, lang=, utm_*, fbclid, gclid, snr=, curator_*) and #comment fragments, but KEEP params that change what the reader sees (forum page ?ctp=N, wiki anchors, GitHub line ranges). Fixes plans citing e.g. `?ctp=2&l=greek`. Enforced in the skill, not code: the CLI writes the plan markdown verbatim and never parses the body. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
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
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
Adds a zero-setup, zero-cost provider for users logged in to the OpenAI Codex CLI (ChatGPT subscription), plus research-quality improvements (page fetching, recency, Steam Deck awareness) that benefit both the codex and API-key paths.
1. Codex subprocess provider (
43ce7e2)When no API key is set (or
FUNFA_PROVIDER=codexforces it) and codex is installed + logged in, funfa delegates investigate→propose tocodex exec— codex runs the hardware probes via its own shell and researches via its own web search, returning the plan as--output-schemaJSON. funfa keeps the consent gate, on-disk write, and verbatim render client-side.resolveProviderIntent): forced codex → API key wins → codex fallback → missing-keys.ToolLoopAgentcan't drive codex with its own tools — codex owns the loop on this path. The API-key path is unchanged.codex-cli.ts:CodexCli,makeBunCodexCli(Bun.spawn, detect viaBun.which+auth.json, model fromconfig.toml/FUNFA_CODEX_MODEL),proposeViaCodex,buildCodexInstructions/Prompt,CodexUnavailableError.2.
fetch_pagetool + Steam Deck-aware research (a1f14af)fetch_pagetool (5th Vercel-path tool): reads a cited URL past search snippets — the accepted answer and the latest comments. The old "do not fetch" rule was self-defeating and is gone.?ctp=N), weight recent comments, a newer "no-longer-works" comment overrides the popular one.ID=steamos/VARIANT_ID=steamdeck/Van Gogh), read ProtonDB's Deck reports, treat controller/input as a first-class fix (Steam Input community layouts), allow a Steam Input layout as a high-tier item.3.
fetch_pagedomain allowlist (7d2fd6f)Limits
fetch_pageto trusted gaming domains (ProtonDB, PCGamingWiki, GitHub, Steam, SteamDB; subdomains ok), re-checks the post-redirect host, and is overridable viaFUNFA_FETCH_ALLOWLIST. Scope: gates funfa's tool only — the codex path browses through its own sandbox and isn't constrained by this list.4. Clean cited URLs (
72267dd)Plan skill strips locale/tracking params (
l=,utm_*, …) while keeping functional ones (?ctp=N). Enforced in the skill (the CLI never parses the plan body).Verification
check-types+ultracitegreen.funfa "Assassin's Creed Revelations"used codex and produced a PCGamingWiki-sourced plan.FUNFA_FAKE_HARDWARE): headlines the Dankaishin keyboard-mapped Steam Input layout (the current working fix), cites the thread at?ctp=2, flags GE-Proton 10-21 as broken, marks Proton 8.0-5 stale.New env vars
FUNFA_PROVIDER/FUNFA_USE_CODEX(force codex),FUNFA_CODEX_MODEL,FUNFA_CODEX_BINARY,CODEX_HOME,FUNFA_FETCH_ALLOWLIST,FUNFA_FAKE_HARDWARE(dev/test affordance for running off-Linux). All documented inCLAUDE.md.Security note
fetch_pageingests untrusted HTML; the no-fetch rule was lifted intentionally. Guardrails: domain allowlist (BYOK path), skill instructions to treat fetched text as data / ignore embedded instructions / never recommend a forum-only download, and codex runs-s read-only. The codex path's network is governed by codex itself, not funfa's allowlist.🤖 Generated with Claude Code