Skip to content

research: X social discovery harvest run (candidate claims only, public URLs, date precision) - #1

Merged
bunnykong merged 15 commits into
mainfrom
research/hermes-x-insights
Jul 27, 2026
Merged

research: X social discovery harvest run (candidate claims only, public URLs, date precision)#1
bunnykong merged 15 commits into
mainfrom
research/hermes-x-insights

Conversation

@ruemic

@ruemic ruemic commented Jun 24, 2026

Copy link
Copy Markdown
Contributor

Public X/social discovery harvest only. This PR records candidate social signals for the watched Frontier projects without promoting any claim into findings, signals, digests, or profiles.

What changed:

  • Adds runs/2026-06-24-x-social-harvest-2026-06-24-frontier-v0/ with manifest.yml, harvest/*.md, verify/*.md, qa.md, research-journal.md, and social-cards/x-cards.yml.
  • Records 15 candidate harvest claims across 10 watchlist projects.
  • Adds 8 static social receipt cards for exact-dated X/social items. These render from repo data only, link to the source posts, and do not load Twitter/X embeds or scripts at runtime.
  • Adds a drama/code follow-up receipt map in verify/drama-code-followup.md so each GitHub release, PR, or official changelog receipt is tied to the exact bounded claim it supports.
  • Upgrades only narrow, source-backed social cards: Gemini Antigravity transition/migration support and Claude Code 2.1.187 release-note items. OpenHands ACP remains partial, Paperclip Maximizer remains maintainer intent, and OpenClaw/Hermes governance tension remains unverified social context.
  • Surfaces verify/*.md files in the run artifact list so the follow-up receipts are visible from the generated run page.
  • Adds docs/deep-social-research-loop.md and strengthens docs/x-social-harvest-workflow.md with the scout/critic/editor loop, receipt-to-claim mapping, reputational-claim firewall, and counterweight search rule.
  • Adds site/src/components/SocialReceiptCards.astro and run-page data plumbing so social cards appear on run pages as archived evidence cards.
  • Removes the loose unreceipted research/*-x-insights-2026-06-24.md narrative files from the original PR shape.

Boundaries:

  • This is discovery input only.
  • Product/version/capability claims still require source-contract cross-checks before promotion.
  • Drama/public-exchange cards are ecosystem context; they do not assign motives or verify governance claims.
  • Static cards receipt that a public post exists on a date. The claim inside the post remains unverified unless separately supported.
  • A secondary receipt upgrades only the exact bounded claim it proves; it cannot validate a different claim from the same project.
  • No local paths, session IDs, private prompts, raw private API payloads, or internal reviewer details are included.

Loop notes:

  • Hermes was used for X-native scout passes.
  • A skeptical critic pass rejected inferred-date, tracker-metric, self-referential, and overclaimed material.
  • A drama-specific follow-up found interesting social leads, but kept OpenClaw/Hermes governance tension, Gemini quota pain, and OpenClaw reliability complaints journal-only unless source receipts supported the exact claim.
  • Public code/release follow-up verified Gemini transition/migration affordances, Claude Code 2.1.187 release-note items, OpenHands ACP partial support, Paperclip OpenClaw/Hermes interoperability work, and OpenClaw reliability-fix counterweight.

Checks:

  • git diff --check
  • node site/scripts/check-integrity.mjs
  • npm --prefix site run check (0 errors; existing hint for an unused redirectChain local)
  • npm --prefix site run build
  • Rendered HTML sanity check for the run page showing upgraded cards and verify/*.md artifacts

ruemic added a commit that referenced this pull request Jun 24, 2026
- openhands, openclaw, pi-coding-agent, paperclip, agent-zero, flue, codex, claude-code, gemini-cli (hermes-agent already done in PR #1)

Each file contains maintainer signals, recent features, drama/community tension, and Bitter Frontier integration recommendations.

This completes the first full pass of the per-project X research loop.
@ruemic

ruemic commented Jun 24, 2026

Copy link
Copy Markdown
Contributor Author

Updated PR with structured run per review feedback:

  • Replaced loose research/*-x-insights files with proper run under runs/2026-06-24-x-social-harvest-2026-06-24-frontier-v0/
  • Added manifest.yml, qa.md, and structured harvest/ files
  • Every claim now includes primary_url, author, observed_at, evidence_kind, channel, and status (candidate or single-source-unconfirmed)
  • Removed all editorial/promotion language
  • Discovery input only — no findings promoted
  • No trailing whitespace issues in new files

This directly addresses the evidence model, receipt, and artifact chain concerns.

ruemic added 4 commits June 24, 2026 14:37
…with OpenClaw, maintainer signals

First per-project deep X detective output.

Covers the new /learn self-improving skills system, animated pets, smart cron, 406+ community repos, public drama between Teknium and OpenClaw founder Peter Steinberg, Hermes #1 OpenRouter ranking, and integration recommendations for Bitter Frontier.

Part of the per-project X research loop for sheetgenius/frontier.
- openhands, openclaw, pi-coding-agent, paperclip, agent-zero, flue, codex, claude-code, gemini-cli (hermes-agent already done in PR #1)

Each file contains maintainer signals, recent features, drama/community tension, and Bitter Frontier integration recommendations.

This completes the first full pass of the per-project X research loop.
- Created proper run directory following feedback
- manifest.yml with sources, window, and verification status
- Structured harvest/ files with claim_id, primary_url, author, observed_at, evidence_kind, channel, status
- qa.md documenting limitations and checks
- All claims marked as candidate or single-source-unconfirmed
- No editorial language
- Discovery input only (not findings)

This replaces the previous loose research/*-x-insights files.
…st run with date_precision, reproducibility notes, public-only claims; update qa for discovery-only scope
@ruemic ruemic changed the title research(hermes-agent): deep X intelligence digest — /learn, OpenClaw rivalry, maintainer signals research: X social discovery harvest run (candidate claims only, public URLs, date precision) Jun 24, 2026
@ruemic
ruemic force-pushed the research/hermes-x-insights branch from 191f972 to 383e031 Compare June 24, 2026 06:37
ruemic and others added 11 commits June 24, 2026 14:46
…rls; add verify/x-post-dates.md with per-cluster public resolution notes; update qa for crosscheck fields and discovery boundary; all claims now carry day-precision event_date or explicit reason
Revive the parked X-social-discovery branch on top of current main so it
can host the Hermes+Grok harvest workflow. Conflicts resolved additively:

- AGENTS.md: keep main's authorities list under its heading; reintroduce
  the humanizer-skill reference beside the house-style paragraph.
- runs/[runId].astro: union of main's robust source mapping + noindex Base
  and the branch's social-card editorial rendering (dropped the branch's
  non-ASCII middot title in favor of main's ASCII pipe).
- global.css: keep both the editorial-card styles and main's correction
  styles; close .editorial-evidence-list and the @media block explicitly.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add the operational half of the parked X-social workflow: drive the Hermes
agent (web/X search tools) on Grok via a SuperGrok / X Premium+ subscription
(Hermes xai-oauth provider), as an optional discovery lane of frontier-cycle.

- docs/hermes-grok-harvest-setup.md: setup runbook. Auth, config, per-cycle
  run, the secret boundary, and the two known limits of the subscription path
  (no Grok 4.5 on the OAuth catalog; xAI can 403 a valid subscriber).
- ops/hermes/config.template.yaml: credential-free config shape (real config
  and token live in ~/.hermes/, never the repo).
- ops/hermes/grok-harvest.sh: doctor / gateway / models / harvest. A 403 or
  unreachable gateway degrades the lane (exit 3) and records the gap instead
  of failing the cycle. Drives the agent via the OpenAI-compatible gateway.
- frontier-cycle SKILL.md: wire the lane in as discovery-only. Leads, never
  receipts; nothing promotes without a primary-source cross-check.

Verified against Hermes and xAI docs. Not run end to end here (Hermes is not
installed on this machine); doctor's preflight and degrade paths are exercised.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ootstrap

Correct the workflow against Hermes' real CLI and simplify the driving model:

- grok-harvest.sh: drive Hermes as a one-shot sub-agent with `hermes -z`
  (single prompt in, final text out) instead of standing up the HTTP gateway.
  Fix command names (`hermes auth add xai-oauth --type oauth`, `hermes doctor`,
  `hermes model`). Portable: no hard dependency on `timeout` (macOS lacks it).
  403/allowlist or empty output degrades the lane (exit 3), never fails the cycle.
- frontier-cycle SKILL.md: promote the lane from optional to a standard harvest
  step that runs every cycle. The receipt firewall is unchanged: leads only,
  never receipts; a down lane is noted in qa.md and the cycle proceeds.
- runbook + README: real bootstrap (install is a global curl|bash to ~/.hermes
  and ~/.local/bin, not repo-local; OAuth needs the operator's browser + sub).
  README carries the four-command bootstrap so a fresh clone can stand it up.
- config.template.yaml: select the model with `hermes model`, not a removed
  script subcommand.

Verified against Hermes install/CLI/OAuth docs. Not run end to end (Hermes is
not installed here); doctor's preflight and degrade paths are exercised.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
First live run of the lane on macOS surfaced three defects. All found by
running it, not by reading it.

- run_oneshot passed "${ONESHOT_ARGS[@]}" with the array empty by default.
  macOS ships bash 3.2.57, where expanding an empty array under `set -u` is an
  "unbound variable" error, so the probe died before reaching Hermes. Use the
  ${arr[@]+"${arr[@]}"} idiom and split the args inside the function.
- The probe classifier swallowed the failure text and reported a generic
  "check auth", which pointed at the wrong cause. Keep the raw output in
  PROBE_OUTPUT and print it on failure. Also check for a 403/allowlist refusal
  before the success case so a degraded lane is never misread as ok.
- The harvest prompt named the record fields but did not enumerate the allowed
  values, so the agent invented its own: 8 distinct evidence_kind values where
  the workflow defines 4, plus a release_channel of "official_social" that is
  not in the schema. Pin every enum in the prompt.

Also correct a wrong claim in the runbook and config template: Grok 4.5 is NOT
API-only. Hermes derives its catalog live from a models.dev cache, and as of
2026-07-22 (Hermes v0.19.0) grok-4.5 is offered by `hermes model` and served
normally over the SuperGrok / X Premium+ subscription. Guidance is now "run
`hermes model` and take what the catalog shows" rather than a pinned version.

Validated end to end: preflight green (hermes doctor OK, Grok probe OK) and a
real single-source harvest returned 21 claim records, every one carrying a
public post URL, a day-precision date, and needs_primary_crosscheck.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Pre-push audit of everything this branch adds over main. Secrets scan clean,
no local paths or private info, but 19 non-ASCII characters across 15 files
violated the house rule in AGENTS.md ("ASCII punctuation only, no em dashes").

Two of them rendered on the public site rather than sitting in an artifact:
`Static social cards -- <file>` and `Verification notes -- <file>` labels in
frontier.ts, and the date/precision separator in SocialReceiptCards.astro.
The rest were em dashes in run-artifact headings and smart quotes in a social
card title.

Punctuation only. No claim, date, receipt, or window judgment is altered.
Integrity check and build both clean (904 pages).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@bunnykong
bunnykong merged commit 1436b2c into main Jul 27, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants