- Extend ProviderStatus with loadedModels[] populated from /running;
llamacpp panel now shows active model name instead of just "N models"
- Return DiscoverResult from discoverProvider (models + loadedModels)
- All llamacpp models get reasoning:true + compat.thinkingFormat:"qwen-chat-template"
so Ctrl+T / Shift+Tab work for both Qwen3 and Gemma4 (verified live)
- s3: guard resp.Body against undefined before calling transformToByteArray
- s3: throw instead of returning 0 when both pi-mtime metadata and
LastModified are absent (non-conformant S3 response)
- s3: parallelize HeadObjectCommand calls per list() page with Promise.all
instead of serial awaits, reducing N+1 round-trips to one per page
- sync: emit skip(local-only-in-pull) for local-only files in pull mode so
they appear in the report instead of being silently dropped
- manifest: convert readManifest/writeManifest to async fs.promises to
avoid blocking the Node.js event loop on slow/network filesystems
- panel: track lastError separately from lastReport so a sync exception
shows a persistent error line instead of leaving the previous report
- k8s/diag: set identity to "(auth whoami unavailable)" when the can-i
--list fallback succeeds, instead of leaving it as "(unknown identity)"
- k8s/index: warn when a configured helper model is not found in the
registry rather than silently falling back to the active model
- Remove session_start (pull) and session_shutdown (prune+push) handlers
from extensions/sync/index.ts; sync is now driven only by the /sync
panel and subcommands.
- /sync pull now downloads every file present in S3 regardless of local
mtime or manifest state. The first-pull manifest-absence gate is gone;
the download reason is renamed first-pull -> force-pull.
- Update sync.test.ts: the former LWW test now asserts pull overrides
local even when the manifest records the key.
- README: drop the "Lifecycle hooks" section and rewrite the conflict
model to reflect "remote always wins for pull".
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Summarize k8s_get/describe/logs/events through a helper model when
output exceeds a configurable threshold (default 4000 chars). Optional
`prompt` biases the summary; `raw=true` bypasses it. Config + helper
picker live in the /k8s panel. Naming mirrors websearch
(helperModel/helperTimeoutMs/helperMaxTokens/helperUsed).
- PI_K8S_INSECURE_SKIP_TLS now accepts 1/true/yes/on so SOPS can encrypt
the value as a string.
- classifyKubectlResult recognises ENOENT (empty stdout+stderr, code≠0)
as kind=missing-binary; diag short-circuits with a clear panel state
and session_start warns once.
- Sync: pull on a fresh machine no longer skips files where the local
default has mtime>remote — first-pull (no manifest entry) forces a
download. Fixes custom-providers.json reverting on every sandbox.
- websearch typecheck fix: replace DOM Document reference with `as never`.
New extensions/sync/ that bidirectionally syncs ~/.pi/agent/ configs +
memories with an S3 bucket so fresh OpenShell sandboxes (or new laptops)
boot with existing pi state.
- Allowlist: settings.json, k8s.json, websearch.json, custom-providers.json,
hosts.json, local-models.json, memory/**. Excludes auth.json
(provider API keys; injected per-sandbox).
- Per-file last-write-wins by real mtime carried in x-amz-meta-pi-mtime;
1s clock-jitter tolerance.
- Per-machine ~/.pi/agent/.sync-manifest.json records what was last seen
on S3 so auto-prune can't wipe a concurrent laptop's data.
- Manual /sync command opens a TUI panel; subcommands /sync pull, push,
prune, status work without the panel.
- session_start auto-pulls (5s soft timeout, non-blocking).
- session_shutdown auto-prunes then auto-pushes (10s hard timeout).
- Config via PI_AWS_* env vars only (chicken/egg: nothing to bootstrap).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The env-only constraint on KUBE_TOKEN only mattered for OpenShell's
placeholder rewriting; since the extension runs outside OpenShell, the
token is just a plain credential the user supplies. Mirror the cluster
URL flow: press `t` in /k8s to set the token, persisted to
~/.pi/agent/k8s.json (0600).
- config.ts: K8sConfig gains optional `token`
- resolveK8sEnv: env first, then config file, for both URL and token
- panel: `t` affordance + token-source status line; prompts never
prefill/echo the secret (ctx.ui.input has no masking)
- fix: promptApiServer now merges instead of overwriting (was dropping
a saved token); both prompts read-merge-write
- README documents the plaintext-on-disk tradeoff
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A new extensions/k8s/ extension giving the agent read-only Kubernetes
tools (k8s_get, k8s_describe, k8s_logs, k8s_events, k8s_top, k8s_explain,
k8s_context). Credentials: KUBE_TOKEN from the env, cluster URL from
KUBE_APISERVER env or ~/.pi/agent/k8s.json (editable via the /k8s panel).
All kubectl invocations go through one exec.ts chokepoint that prepends
--server/--token; the token never reaches any render path. A /k8s panel
runs reachability + can-i probes and lets the user set the cluster URL
(press e). A bundled "k8s" mode restricts the toolset to read + grep +
k8s_* for focused troubleshooting. README ships the SA + RBAC manifest.
Container changes (Dockerfile):
- install kubectl
- bake customizations into /app (OpenShell Landlock allowlists /app,
not /opt) and chmod world-readable so the remapped sandbox uid can
read them
- openshell-policy.yaml: baseline filesystem/landlock/process policy
Also deletes the unused status-line.ts demo extension.
Known limitation: does NOT work inside an OpenShell sandbox yet —
OpenShell hardcodes the k8s control-plane ports (6443 et al.) as
universally blocked in its SSRF engine with no override. Documented in
extensions/k8s/README.md ("Why not OpenShell yet"). The extension works
when pi runs directly on a host with cluster access.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
10-task plan for extensions/k8s/: exec chokepoint + 7 typed tools +
diag/panel + index wiring + bundled k8s mode + README. TDD per task
with vitest unit tests on all pure helpers.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
pi.exec returns one captured result; ssh-hosts uses its own streamer.
For the k8s extension, single-shot output with tail caps is fine and
simpler. Streaming can be added later if log latency proves to matter.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Brainstormed design for an extensions/k8s/ extension that gives the
agent read-only kubectl tools, credentialed via an NVIDIA OpenShell
provider injecting KUBE_APISERVER + KUBE_TOKEN.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A single todo_write tool (add/complete/update/remove/clear) writes
full-state snapshots to the session via appendEntry. The current list
is injected into the system prompt each turn and surfaced as a
done/total counter in the status line.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The pi ExtensionUI.input(title, placeholder?, opts?) second arg is a
placeholder hint, not an initial value — there is no API for an
initial value. So on edit, hitting Enter on an "empty" field was
actually submitting an empty string. The previous panel pattern
silently turned that into the empty string, making edit behave
identically to add. Fix:
- Show current value in the title: "X (current: Y, blank to keep)".
- Treat empty trimmed input as "keep current" for every field.
- Esc (input returns undefined) still cancels the whole flow.
- Removed the now-unused promptField helper.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two UX fixes after live testing:
- The select dialog has no native pre-fill, so on edit the current
kind was invisible. Reorder the options so the existing kind is
first (Enter accepts it) and append "(current: <kind>)" to the
prompt label.
- Asking ollama users for an "OpenAI-compatible base URL" is the
wrong mental model. Once kind is chosen, prompt for the bare
server URL ("Ollama server URL (e.g. http://localhost:11434)" or
"llama-swap server URL (e.g. http://localhost:9000)") and append
/v1 on save. Edit prefill strips /v1 + trailing slashes so users
see the bare host. Storage shape (baseUrl with /v1) is unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace the transitional kind placeholder with a proper ctx.ui.select
step in the add/edit flow. List row gains a [kind] column between
name and baseUrl. Required fields (name, kind, baseUrl) all cancel
with a consistent warning notify; apiKey and defaultCtx remain
optional.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cosmetic consistency fix surfaced in code review — both fetch URLs
now compose from the trimmed `baseUrl` local, making the trim/derive
pipeline read as a single chain. Behavior unchanged; the previous
regex /\/v1\/?\$/ matched both with and without trailing slash, but
the asymmetry was easy to misread.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add required kind: "ollama" | "llamacpp" to CustomProvider; remove
nativeUrl. Discovery now dispatches on kind: ollama probes /api/show
per model, llamacpp probes /running once. nativeUrl is derived inline
from baseUrl (strip trailing /v1). loadConfig validates kind and
all other field types; a missing or unrecognized kind makes the file
"corrupt" and the seed file is used in memory.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mechanical rename: directory, type (LocalProvider → CustomProvider),
slash command (/local-models → /custom-providers), config path
(local-models.json → custom-providers.json), panel title, and stale
comment references in subagent extensions. No behavioral change —
kind discriminator and nativeUrl removal land in the next commit.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The default header was a version line + keybinding crib + "press ctrl+o
for more" reminder, repeated every session start. Swap it for the block-
art π mascot example that ships in @earendil-works/pi-coding-agent's
examples/extensions/. Subtitle trimmed from the example's self-
deprecating tagline to a plain "pi vX.Y.Z".
Hits ctx.ui.setHeader() on session_start. Keybinding hints still
reachable via ctrl+o, and /builtin-header restores the old display if
needed.
The [Prompts] / [Extensions] block printed after the header is a
separate "loaded resources" view controlled by the quietStartup setting
in ~/.pi/agent/settings.json — toggle that to silence it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The /websearch panel had 11 rows after v1.1 wired up the new tunables —
too dense for the common case. Default view now shows only the 5
load-bearing fields (URL, max results, char limit, sandbox, helper);
press 'a' to reveal the advanced section (fetch/helper timeouts,
helper max tokens, user-agent, size cap multiplier, max redirects).
Advanced shortcuts (f/l/k/u/z/r) only trigger when the section is
visible — pressing them in basic view is a no-op rather than a silent
config edit.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
It auto-commits every session with "[pi] <first line of last assistant
message>" — those commits land in the middle of intentional work and
need to be soft-reset + reauthored constantly. Easier to remove than
to retrain around.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous renderResult sliced the first 200 chars of the markdown
output and put it on the result line. For a helper-summary fetch that
meant the first sentence of an LLM's answer; for raw fetches it was
opaque markdown ("# Configuration\n\nPaperless provides..."). Either
way the snippet repeated content the user can already see in the full
result panel and added no signal.
Replace with a length-on-success / kind-on-failure summary using the
structured details we already bubble up:
✓ 14300 chars
✓ 80012 chars (truncated)
✗ http 403
✗ cross-domain-redirect
✗ too-large
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A2 (commit 4978123) made it impossible for a successful fetch to end
on a different host than the request: followRedirects refuses any
cross-host hop, so the success path always returns a finalUrl on the
original host. The crossHostRedirect bookkeeping kept around after
that change was always false:
- classifyRedirect() / its prefix string / FetchSuccess.crossHostRedirect
/ successDetails.crossHostRedirect — all dead.
Removed across fetch.ts, index.ts, and the tests. ~25 lines net. If
we ever want to allow some cross-host redirects (trusted-host
allowlist), the field can come back then.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three public-surface cleanups now that the cache is gone.
fetchUrl(url, opts): charLimit is now part of FetchOpts instead of a
positional sibling. Everything tunable lives in one place. All call
sites updated.
SearchOutcome now carries `kind` and optional `status` on failure,
matching FetchOutcome's shape. SearchErrorKind = "not-configured" |
"network" | "http" | "body". The web_search tool surfaces both in
result.details so the calling agent can branch without parsing
strings (same pattern A5 introduced for fetch).
panel.ts exposes every config knob the v1.0 set added: fetch /
helper timeouts, helper max tokens, user agent, size cap multiplier,
max redirects. New single-letter shortcuts (f/l/k/u/z/r). The
prompt-for-positive-int pattern is collapsed into one helper.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The in-memory page cache was solving the wrong problem for pi's usage.
What it cost:
- ~120 lines across fetch.ts, config.ts, and the test files
- The "negative caching" footgun: a transient 503 or DNS blip poisoned
the cache for 15 minutes with no retry path
- Recurring maintenance to keep the cache key in sync with new opts
(we just did this round in df2b1cd)
What it gave back, in practice:
- A speed win on the rare case where an agent fetches the exact same
URL twice in a 15-min window. Typical agent loops fetch different
URLs per iteration, so this hit rate is near zero for interactive use.
Removed: CacheEntry / cache Map / cacheKey / cacheGet / cacheSet /
__clearCache, the cacheTtlMs config field, the clearFetchCache call in
session_start, and 6 cache-specific tests. fetchUrl is now a thin
wrapper around the renamed performFetch.
103 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two cache holes that mattered.
fetch.ts cache key was (url, charLimit) only. Other options that affect
the cached outcome — sizeCapMultiplier, maxRedirects, userAgent —
weren't keyed. Fetching the same URL with a smaller sizeCapMultiplier
after a successful larger-cap fetch returned the prior outcome and
silently skipped the new size check. Now the key includes every option
that changes the result content (timeoutMs and cacheTtlMs don't — same
outcome, just different timing).
loadConfig() was reading websearch.json from disk on every tool call —
two reads per fetch_url, one per web_search. Memoize per filePath, with
invalidation on save (saveConfig updates the cache) and on session_start
(picks up external edits). Tests get __invalidateConfigCache in
beforeEach/afterEach so they don't leak state across cases.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five internal cleanups that pay off in readability and remove a latent
ordering bug, none of which change observable output:
- dom-shim.d.ts: small local DOM type interface for linkedom (which
ships near-empty types). Replaces ~11 `any` / `as any` sites across
html-extract and site-extractors with typed parameters and accesses.
- html-extract: snapshot the body's outerHTML before Readability runs,
so the fallback path reparses a clean tree instead of mutating the
document Readability already operated on. stripChrome is now only
ever called on throwaway docs. Pure helpers bodyToMarkdown() and
runReadability() replace inline mutation.
- fetch: combine buildRedirectPrefix() + isCrossHost() into one
classifyRedirect() that parses the URL pair once and returns both
fields. Removes a duplicate try/catch.
- fetch: split fetchUrlNoCache from ~90 lines into named helpers
(isHtml, isTextual, truncateAtCharLimit, checkDeclaredSize,
readBodyAsMarkdown). Same control flow, the top-level function is
now scannable.
- index: rename wrapWarning → helperUnavailableResult with a real
SuccessDetails type for its parameter; the prior signature was an
any-typed assembly that grew accidentally during A5.
All 105 tests pass; live URLs (github PR, wikipedia, reddit) extract
the same byte counts as before.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two real bugs surfaced in actual use.
Size cap rejected normal pages. The streaming size guard was checking
total bytes against charLimit × multiplier, which for a user with the
old fetchCharLimit=10000 (the default before the v1 bump) gave a 100 KB
ceiling. Almost every modern web page is 200–700 KB of HTML, so every
fetch failed as too-large after we had already extracted the content.
The original intent — don't download multi-MB binary dumps — is now
served by:
- Streaming reader truncates instead of rejecting. We keep what we've
read up to the cap, mark truncated=true, and continue to extraction.
- Content-Length pre-check still refuses obvious oversize downloads.
- An absolute 2 MB floor on the byte cap, independent of charLimit, so
small charLimit values don't cause spurious rejections.
web_search count not clamped. The model could pass count=10 even when
the user had configured maxResults=8. Now Math.min(requested,
config.maxResults) — the config is the ceiling.
Live-tested all six URLs from the user's failing transcript at the
broken charLimit=10000 — all return content now.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
B3+B4: replace contextWindow * 0.5 * 4 magic with explicit math —
input_budget_tokens = contextWindow - responseMaxTokens
- SYSTEM_PROMPT_TOKEN_RESERVE,
clamped at MIN_INPUT_TOKEN_BUDGET, converted via CHARS_PER_TOKEN. Now
auto-scales when the helper switches between a 32k qwen and a 200k
Sonnet. Pi-ai exposes no per-model tokenizer; 3.5 chars/token errs short.
Also trim verbose comments added during the prod-ready pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Imperative extractGithubArticle becomes a data table in
site-extractors.ts. Adding a new site is now a data change (one entry,
host regex + selectors) instead of a code change. If a listed site
renames a selector, the extractor returns null and we fall through to
Readability + the density-based fallback — no broken output, just
slightly worse extraction until the table is updated.
The GitHub entry is a 1:1 port of the old code:
- titleSelectors: .gh-header-title, .js-issue-title, release-header
- bodySelectors: .markdown-body
- exclusiveBodySelectors: .release-main-section .markdown-body
(covers release pages where many .markdown-body blocks exist)
Live-tested on github PR / discussion / Wikipedia — identical char
counts to pre-refactor.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The hardcoded CHROME_PHRASE_MARKERS list ("provide feedback", "saved
searches", reddit sidebar copy …) was a maintenance treadmill — every
new site whose chrome happened to contain different stock phrases
needed a new entry. The detection signal was tied to whatever empty-
state copy GitHub or Reddit happened to ship that week.
Replace it with two principled signals:
- If Readability's output is shorter than MIN_READABILITY_CHARS (200),
it's almost certainly not a real article. Use the body-strip
fallback.
- If the body-strip fallback is at least FALLBACK_PREFERENCE_RATIO (3×)
larger than Readability's output, Readability probably picked a
chrome island. Use the fallback.
Both constants are signal-shaped, not site-shaped — they don't go stale
when a site changes its copy. The hardcoded phrase list is gone.
Net effect on the real-world URLs we've been testing:
- GitHub PR: 14,300 chars (site extractor still short-circuits before
density — unchanged)
- Wikipedia: 39,518 chars (Readability picks correctly — unchanged)
- Reddit comment thread: 1,846 chars → 35,282 chars (density rule now
grabs the full thread instead of just the OP — real upgrade)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Everything that was a magic number scattered through the module bodies
is now a named config field with a documented default:
- fetchTimeoutMs (was hardcoded 10_000)
- helperTimeoutMs (was hardcoded 30_000)
- helperMaxTokens (was hardcoded 2000)
- cacheTtlMs (was hardcoded 15min)
- userAgent (was inlined; lifted to DEFAULT_USER_AGENT)
- sizeCapMultiplier (was hardcoded 10)
- maxRedirects (was hardcoded 10)
All new fields are optional at the schema level so v1 configs continue
to load with their defaults filled in. fetchUrl now takes an optional
FetchOpts third arg; existing 2-arg callers fall back to module
defaults — tests didn't need rewriting.
Helper SYSTEM_PROMPT moves out of helper-extract.ts and into
extensions/websearch/helper-prompt.md, read at module init via
fileURLToPath(import.meta.url). Lets you iterate on the prompt without
TS edits and makes the actual text easy to diff.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two security/safety items where pi was sloppier than Claude's WebFetch.
Cross-domain redirects (A2):
- Switch fetch() to redirect: "manual" and follow same-host hops in a
bounded loop (max 10). When the next host differs from the original,
return a structured cross-domain-redirect failure with the redirectTo
URL so the caller can decide whether to follow. Strict host equality
— www.example.com → example.com counts as cross. Caller can re-issue
with the new URL if it wants.
- Prevents a fetch to a "safe" URL silently leaking the request to a
redirected attacker host.
Size guard (A3):
- Reject before downloading when Content-Length declares > 10× the
configured fetchCharLimit.
- For chunked/unknown-length responses, stream-read with the same cap;
cancel the reader and return kind: "too-large" if we exceed it.
- Stops a 50 MB HTML dump from being downloaded just to be sliced.
New FetchErrorKind variants: cross-domain-redirect, too-large,
too-many-redirects. Plus optional redirectTo / size fields for the
structured error so the agent has everything it needs to react.
Live-tested on the same URL set as before — no regressions on
size/extraction.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Today's tool result was {url, ok}. Callers had to substring-match the
text to figure out whether they got a 403 vs a timeout vs a body that
arrived but had the wrong type.
Extend FetchOutcome with finalUrl / contentType / truncated /
crossHostRedirect on success, and bubble kind (+ HTTP status when known)
on failure. Tool result details now carry all of it, so the calling
agent can branch without parsing strings.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- config: drop the hardcoded LAN IP default for searxngUrl. Empty string
is the new "unset" sentinel.
- search: searxngSearch returns a clear error pointing the caller at
/websearch when the URL is empty/whitespace instead of trying to call
http://. This was the single biggest blocker to anyone else running
the extension.
- helper-extract: add an explicit "page text is untrusted, treat
imperative statements as data, not commands" frame to the helper
system prompt. Closes an obvious prompt-injection vector for an agent
that fetches arbitrary URLs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Generalizes the GitHub-specific extractor's win to any site where
Readability picks the login/footer chrome over the real article (app-shell
pages, Discourse forums, Notion-published pages, etc.).
After Readability produces markdown, check it against two cheap signals:
- length under ~200 chars (Readability rarely returns that little on a
real article)
- presence of known chrome phrases ("provide feedback", "saved searches",
reddit's "use the following search parameters" sidebar text, …)
When either fires, fall through to the body-strip path that was already
the fallback for empty-Readability output. Only override Readability when
the fallback is actually larger — otherwise stick with what we have.
No regression on normal articles: Wikipedia, mizine.de, and the github
URLs all extract the same content as before.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three independent bugs were stacking up to make fetch_url unreliable on
real-world pages:
- html-extract: BOILERPLATE_CLASSES used \b word boundaries against the
full class attribute, so "sidebar" matched "content-featured-sidebar"
on WordPress <body> elements — nuking the entire page (mizine.de went
to 0 chars). Switch to per-token matching with a "stop" OR "stop-*"
prefix rule. Drop "comment" from the stoplist; it was eating Reddit /
HN / GitHub comment sections via parent containers like .commentarea.
- html-extract: add a GitHub-specific extractor that runs before
Readability for github.com hosts. Readability was picking GitHub's
"Provide feedback / Sign in" footer chrome as the article on PR /
issue / discussion pages. New path pulls .gh-header-title +
.markdown-body elements (works for PRs, issues, discussions, blob
views, releases).
- config: bump fetchCharLimit default 10000 → 80000. The previous limit
cut off long doc pages (paperless's configuration.md is ~89k chars
with the AI section at the very end).
Verified live: mizine.de 0 → 22.9k, github PR 600 → 14.3k, github
discussion 603 → 19k, reddit comments 670 → 1.8k, raw config.md
truncated → full file including AI section.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- fetch.ts: set a Firefox-shaped User-Agent + Accept/Accept-Language so
default-Node-UA filters stop returning 403 on the majority of sites.
- helper-extract.ts: when the prompt is a topic/keyword rather than a
question, return a structured outline (title + sections + answerable
questions) instead of guessing at an answer and producing a single
filler sentence.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bundled agents live in <repo>/agents/ — edit and delete now write/remove
those files directly, matching how a package author iterates on their own
shipped agents. The "c" duplicate shortcut is still there for users who
want a user-dir fork instead.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Dockerfile now runs `npm ci --omit=dev` inside /opt/pi-customizations so
v1.1's new runtime deps (linkedom + readability + turndown + turndown-plugin-gfm)
are present — they aren't part of pi's global install.
.dockerignore now excludes more host-only or test-only paths:
- .pi/, .claude/, .worktrees/ (local agent/IDE/git state)
- docs/ (planning docs, not needed in image)
- *.test.ts, **/fixtures/, vitest.config.ts (test code/data, 200 KB of fixtures)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Defines a researcher agent that surveys the open web, fetches relevant pages,
and returns a compressed brief. Uses the v1.1 websearch tools.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the regex-based HTML→Markdown pipeline with a DOM-based one and
extracts helper-orchestration into a pure function so index.ts::execute
becomes a thin coordinator.
- Declare engines.node >= 18 (AbortSignal.timeout requires it)
- Deps: linkedom, @mozilla/readability, turndown, turndown-plugin-gfm
- helper-extract.ts: new orchestrateHelper() owns truncate → extract → hybrid-error envelope
- html-extract.ts: new htmlToReadableMarkdown(html, baseUrl) — parseHTML → stripBoilerplateClasses → Readability → stripChrome → turndown(GFM). Old extractMain/htmlToMarkdown removed.
- index.ts::fetch_url execute delegates to orchestrateHelper; module-scope wrapWarning helper for the two "helper unavailable" branches
- New tests: 3 orchestrateHelper, 9 htmlToReadableMarkdown fixture, 1 real-world MDN smoke, 6 index.test.ts fetch_url execute scenarios. Old 38 regex-pipeline tests removed.
- Total: 193 tests across 20 files (was 220 across 20).
Resolves six known limitations:
- A: word-boundary regex on classes (main-nav-container-info false positive)
- B: AbortSignal.timeout needed Node >= 17.3 with no engines declaration
- C: leaf-only boilerplate stripping let wrapper elements with non-boilerplate children survive
- E: no integration tests for fetch_url::execute
- H: orchestration logic awkwardly inlined in index.ts
- I: brittle regex-based HTML→Markdown
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Eight tasks: deps + engines, orchestrateHelper, new html-extract pipeline
(additive), fetch.ts switch + old code removal, captured fixture smoke
test, index.ts refactor, six smoke tests, final verification. TDD-style
with hand-written HTML fixtures for each named regression (A, C) plus
one real-world capture.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- htmlToReadableMarkdown catches Readability exceptions and falls through
to the chrome-stripped body path; without this, parser-throw cases would
surface as opaque "Could not read response body" errors despite having
the page in hand.
- Test fixture strategy: hand-written minimal HTML for each named regression
plus one captured real article. Stops the implementer from picking
arbitrarily.
- Acceptance criterion 4 replaced: behavioral ("delegates to orchestrateHelper,
no inline extract logic") instead of cargo-cult line count.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces regex-based HTML extraction with linkedom + Mozilla Readability +
turndown, extracts orchestrateHelper from index.ts, declares engines.node
>= 18, and adds smoke tests for fetch_url execute. Resolves issues A, B,
C, E, H, I identified in the helper-model post-implementation review.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds an optional `prompt` argument to `fetch_url`. When supplied, the fetched
page is summarized through a helper model with the user's question, and only
the answer is returned. Without prompt, behavior is unchanged.
- helper-extract.ts: truncateForHelperContext, buildMessages, extractAnswer, extractWithHelper (with DI complete())
- SYSTEM_PROMPT for helper guidance
- config.helperModel field (with back-compat) — optional provider+id pair, falls back to active model
- /websearch panel: helper model picker
- index.ts: wire fetch_url to helper-extract when prompt is supplied; fall back to raw page on helper failure (hybrid-error envelope)
- imports consolidated at file tops
Includes design spec and implementation plan under docs/superpowers/.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When sandbox mode is on, fetch_url failures with kind="network"
(connection refused, DNS, timeout — the egress-denial signatures) get
a trailer pointing the model at the most likely cause: an unapproved
egress rule for the host. HTTP-level failures and body-read errors are
unaffected, since those indicate the egress actually worked.
Toggle lives in the /websearch panel (key: b). Config field is
optional in older files, defaulting to false, so existing
websearch.json keeps loading.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Collapses the agent's active tool schema to a focused subset for the task
at hand, and optionally appends a short system-prompt nudge. The win for
small local models: a 4-tool schema produces dramatically more reliable
tool calls than a 20-tool schema, and a one-paragraph nudge keeps the
model on-task.
Commands:
- /mode <name> switch (bundled or user mode)
- /mode reset restore the default tool set captured at session_start
- /mode open the management panel
Bundled, read-only modes (extensions/modes/defaults.ts):
- plan read-only investigation (read/ls/find/grep/web_search/memory_read)
- edit code editing without bash (read/write/edit/ls/grep/find)
- debug read + bash; investigate without writing
- chat no tools; answer from knowledge
User modes live as <name>.md under ~/.pi/agent/modes/ with YAML
frontmatter (name, description, tools) + body (the prompt nudge). The
/mode panel mirrors /subagents and /memory: list / view / new / edit /
delete / duplicate / switch. Bundled modes reject edit/delete; broken
files surface with a [broken] tag and only support delete. Tool-name
validation is lenient at save time and at switch time: unknown tools
trigger a warning notify but the save (or the switch) proceeds.
File layout:
- store.ts types, file IO, validate-name, serialize/parse
- defaults.ts bundled Mode[] (data, no IO)
- apply.ts pure: intersect tools, append prompt nudge
- panel.ts /mode TUI panel
- index.ts wires command, before_agent_start nudge, session_start capture
State is session-only: each new session re-captures the default tool set
and clears the active mode. Status line shows `mode: <name>` while active.
Tests: 20 new vitest cases — store IO (round-trip, dir resolution,
discover, save/delete), bundled defaults well-formedness, applyMode
intersection + dropped-tools tracking + prompt-append shape + mode-order
preservation.
Spec: docs/superpowers/specs/2026-05-16-modes-design.md
Plan: docs/superpowers/plans/2026-05-16-modes-extension.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A tool-mode switcher tailored to small local models: collapse the agent's
schema to a focused 3-7 tools per mode and append a task-specific prompt
nudge. Bundled defaults (plan/edit/debug/chat) are read-only; user modes
live as one markdown file per mode in ~/.pi/agent/modes/ and are managed
via the /mode panel (mirroring /subagents). /mode <name> switches; /mode
reset restores defaults; /mode opens the panel.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mirrors the /remember <text> pattern: with an arg, sends a templated
user message asking the agent to call web_search. Without args, opens
the config panel as before.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A new extension under extensions/memory/ that gives the agent a
cross-session memory, modeled on the Claude Code auto-memory system.
How it works:
- before_agent_start handler discovers memories from two roots
(~/.pi/agent/memory/ for global, <cwd>/.pi/memory/ for project),
builds a typed index grouped by user / feedback / project / reference,
and appends it to the system prompt every turn (zero tokens when both
stores are empty).
- memory_save tool: agent writes/updates an entry; same-name in same
scope under a different type triggers a move with old-file cleanup.
- memory_read tool: agent fetches one entry's body by name; both scopes
returned when ambiguous.
- /memory command: TUI panel for list/view/create/edit/delete/duplicate.
Mirrors /subagents bindings; broken entries surfaced with [broken] tag.
- /remember <text>: dispatches a templated user message so the agent
classifies + writes in one turn.
- /remember (no arg): opens the same form as /memory's "new" — direct
write, no model involvement.
File layout (mirrors existing extensions/subagent/ decomposition):
- store.ts: types, file IO, scope resolution, save/discover/delete
- inject.ts: buildMemoryIndex (system-prompt block builder)
- tools.ts: pure tool handlers + typebox schemas (no pi runtime)
- panel.ts: /memory + /remember form TUI
- index.ts: wires tools, commands, and the before_agent_start handler
Tests: 30 new vitest cases cover store IO end-to-end (round-trip,
scope merging, broken-entry surfacing, multi-stale cleanup, save
actions), the index-block builder (empty case, grouping, ordering),
and tool-handler happy paths + invalid-name rejection.
Spec: docs/superpowers/specs/2026-05-15-memory-design.md
Plan: docs/superpowers/plans/2026-05-15-memory-extension.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
9-task TDD plan covering store, inject, tools, panel, wiring, and
verification. Mirrors patterns from extensions/subagent and
extensions/local-models. Tests for store (round-trip, scope resolution,
discover, save/move), inject (block format, ordering, empty case), and
tool handlers (success + invalid name + multi-scope).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Persistent file-based memory across sessions. Index-only autoload via
before_agent_start; full entries fetched on demand. Both global and project
scope. Typed categories (user/feedback/project/reference). /remember
captures (LLM-assisted with arg, panel form without). /memory panel for
direct management. Mirrors the Claude Code auto-memory shape, adapted to
pi's extension API.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Origin-aware agent discovery (bundled/user/project, non-deduplicated) plus
serialize/write/delete helpers in agents.ts, a new panel.ts implementing the
/subagents ctx.ui.custom panel (list, detail view, create/edit/delete/
duplicate — bundled agents read-only), and the /subagents command in
index.ts. Frontmatter is collected via ctx.ui.input, the system prompt body
via the built-in ctx.ui.editor.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the embedded pi-tui Editor sub-panel approach with pi's
built-in multi-line editor dialog.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extract run execution from index.ts into runner.ts, add an in-memory
RunRegistry shared between the subagent tool and a new /subagent-runs
command, give each run its own AbortController for per-run cancellation,
and add the runs-panel.ts monitor with a live-refresh timer. Also adds
TypeScript type-checking (tsconfig.json + typecheck script) so the
refactor has an automated safety net.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
discovery.ts detects a llama-swap nativeUrl (responds to /running) and
reads each loaded model's real context window by parsing -c/--ctx-size
from its command line; not-loaded models keep defaultCtx. The llamacpp
seed entry gains nativeUrl.
Builds a node:22-slim image with pi installed globally and this package
registered via a local-path `pi install`. Meets OpenShell BYOC requirements:
sandbox user at uid/gid 1000660000, iproute2/iptables, writable /sandbox workdir.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
discovery.ts gains a llama-swap-aware probe: GET <nativeUrl>/running,
parse -c/--ctx-size from each running model's cmd, use the real context
for loaded models and defaultCtx for the rest. Detection by path
convention on nativeUrl; manual re-probe via the panel's existing 'r'.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Restructures the local-models extension into config.ts / discovery.ts /
panel.ts / index.ts. Providers live in ~/.pi/agent/local-models.json
(seeded on first run); a /local-models TUI panel adds/edits/removes/
refreshes them, re-registering with pi live (no /reload).
Make the local-models extension's providers editable via a /local-models
interactive panel, persisted to ~/.pi/agent/local-models.json, with live
register/unregister so edits take effect without /reload.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The hint line was a single long string that pi truncated to the panel
width, hiding "a add / e edit / d remove / i import / Esc close" on
narrower terminals. Now the hints pack into as many lines as needed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Selecting an entry from ~/.ssh/config now adds it to the registry
immediately using the parsed values, with no field prompts. The old
flow re-prompted every field because ctx.ui.input only takes a
placeholder (not an editable default), so the "pre-filled" import was
indistinguishable from a blank Add. Use 'e' to tweak a field afterward.
Also collapses the duplicate label-list derivation into one array.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EislM0203