Release 1.5.0 - #45
Merged
Merged
Conversation
Recall trust: lexical recall could return a memory on one incidental
shared word ("storage format" recalled for "how should I format my
responses") and present it exactly like a real preference match. Each
recalled memory now carries a confidence label derived from
significant-token coverage:
- strong: query verbatim in title/TLDR, or all significant tokens in the
memory head (title/TLDR/tags)
- moderate: half the significant tokens anywhere, or a significant token
in the title
- weak: everything else above the recall score floor
When every match is weak, recall packets and CLI output now say so and
tell the agent to verify with the user instead of acting on it. Scoring
also gains light suffix stemming ("commits" finds "committing") with no
embeddings or external calls.
Demo: seed three additional realistic reviewed memories (style
preference, local-Markdown decision, viewer-port fact) alongside the
existing pending one, so the first recall, brief, and viewer walkthrough
show a believable memory system and the review loop at the same time.
Tests updated for the new demo corpus (16 pages, 4 memories) and new
coverage for confidence tiers, stemming, weak-only guidance, and the
seeded review states. Full suite: 753 passed.
Add recall confidence labels and a believable demo memory set
The designed landing page (live memory-graph hero, serif editorial layout) was replaced with a plain templated page during the docs simplification, which dropped the animations and the strongest visual proof on the site. Bring it back and refresh its content for the current product: - Canonical six-tool slim MCP surface (status, recall, remember, ingest, review, admin) with confidence-labeled recall in the tools section. - CLI story led by lnk proof / lnk try / lnk start / lnk onboard, and installed-command names corrected from `link` to `lnk`. - Tool count and demo commands updated; external font preconnects stay removed, so the page makes no third-party calls. test_docs_site: the landing page is exempt from the template-uniformity checks (it ships its own inline runtime) while every privacy check still applies to it; docs sub-pages gain a subtle reduced-motion-aware scroll reveal via the shared site.js/site.css. README: document recall confidence labels in the agent tool list, tighten the quick start and viewer copy, and move scale checks into the documentation table.
Restore the animated landing page and align it with the slim surface
Browsers cache favicons in a separate, long-lived store, so returning visitors kept seeing the old logo in the tab even though assets/logo.svg was updated to the new mark. Reference the icon as assets/favicon.svg (same new mark, fresh URL) across all doc pages and in the landing page's head-keeper, forcing every browser to refetch. assets/logo.svg stays for in-page brand images.
Serve the favicon from a new URL so stale browser caches refresh
Port the design handoff (ledger cards + T1 segment meter) into the server-rendered viewer: - Tokens: cream/ink/rust palette on the existing custom-property system, with the warm-dark theme (#191309 paper, #1e1810 surfaces, rust lifted to #cd7657) replacing the pure-black dark mode. Same markup drives both themes; the graph canvas stays dark in both. - Typography: system stacks only — sans body, serif headings, mono for labels, paths, and commands. - Components: mono tab-strip nav with accent underline, outlined type-scope badges, ledger memory cards with caution borders for pending review, serif stat row with accent alert counts, plain-surface health cards, rust copy buttons, dashed empty states. - Confidence: recalled memories render the segment meter (strong/moderate/weak); weak pending memories add a quiet ochre "verify before trusting" note, never red. The dark-theme test now asserts the warm-dark contract instead of the retired pure-black palette. Full suite: 753 passed.
Command/prompt blocks in light theme rendered near-invisible dark text on a dark well. Per the handoff, light-theme command wells are light insets with ink text; the dark terminal well remains for the dark theme.
Same filenames, new captures: the warm-dark memory dashboard (with stat row, review queue, and copy actions), the health page, and the graph view now show the shipped console design instead of the old UI.
…cally The staged command contains wiki\memories on Windows; normalize backslashes before asserting the wiki/memories path.
The UI tour now cycles real screenshots of the shipped console (home, memory dashboard, health, graph); the CLI and MCP tours use the real terminal and agent panels instead of the old synthetic mockups.
The viewer now honors a ?theme=light|dark URL parameter (persisted like the toggle), which also makes reproducible captures possible. README and docs screenshots plus all four tour GIFs are regenerated in light mode with uniform framing, and the README shows the animated console tour instead of two mismatched stills.
The seed wiki page listed file names and byte counts, so the first recall after 'lnk seed' returned a page about seeding rather than the project context itself. Include bounded (1200-char) excerpts of each secret-scanned file plus up to ten recent commit subjects in Key Context, with a regression test asserting the seeded content is present. Also tidy the README documentation table.
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.
Since 1.4.0, most of the work went into making the wiki pattern agent-native — less "here's a wiki your agent can browse," more "your agent reliably uses it without being told." Highlights:
Slim MCP surface. 1.4.0 exposed ~37 MCP tools; models facing six overlapping read tools would pick wrong or skip memory entirely. 1.5.0 defaults to six: status, recall, remember, ingest, review, admin. One obvious read path, one explicit write path. Also ships MCP prompts + resources, so /link workflows work in any MCP client without per-agent glue.
Honest recall. Lexical retrieval can match a memory on one incidental shared word and present it like a known preference — for an LLM consumer, a wrong memory is worse than no memory. Every recalled memory now carries a confidence label (strong/moderate/weak) derived from significant-token coverage, and when everything is weak the packet says so and tells the agent to verify with the user instead of acting. Still zero embeddings, zero network calls.
Day-one seeding. lnk seed . reads allowlisted context that already exists in your repo (README, AGENTS.md, CLAUDE.md, .cursorrules, recent git subjects), secret-scans it, and writes a source-backed page — so the first recall returns your actual project context instead of an empty wiki. Ingest still never silently becomes durable memory; that stays behind explicit review.
60-second proof. brew install gowtham0992/link/link && lnk proof creates a workspace, writes one reviewed memory, and recalls it through the same bounded path CLI/skills/MCP share — the cross-agent continuity claim, demonstrated locally in under a minute.
Trust plumbing. Append-only audit log now hash-chained (tampering is detectable), interrupted multi-file writes leave rollback snapshots, and lnk benchmark reports how much context the bounded packet avoided sending versus dumping the wiki.