Skip to content

Conversation Decisions Audit

Date: 2026-05-23 Scope: ~16 of the largest Claude Code JSONL conversations from 2026-04-22 → 2026-05-23 (the full extent of locally persisted history; nothing earlier exists on disk). Method: Extracted user turns from each JSONL, scanned for decisions/deferrals/open questions, cross-referenced against ADR-001..065, CHANGELOG entries, and the eight files in docs/briefs/. Only items genuinely missing from the formal record are listed below.

Source-file shorthand (UUID prefix + rough date): - 5271c78c (May 2–6, agent-connector + personal scenarios) - 74391e42 (Apr 27 – May 9, teams + multiplayer + driver personas) - c7d99ac9 (May 5 – May 14, Cici implementation queue, synapses, packs) - 35c9eaf9 (Apr 22–23, Brink playtest fix sprint) - db0594ae (May 6–8, Hermes brief implementation L1–L5) - ffb273d9 (Apr 30 – May 1, consistency-vs-coherence philosophy, third-option problem) - 2d47f209 (Apr 30 – May 1, Dunia naming and approval) - 6d068328 (May 15–16, role-framing audit operational decisions) - 18ea62dc (Apr 29 – Apr 30, scenario voice + async multiplayer aside) - 82f4d7f9 (May 21, marketing-page rebrand + Pulse) - 129521f0 (May 15–17, VPN reachability + Supabase Pro cost call) - 48e0a429 (May 11–12, multiplayer-scenario quality parity) - cc20b56e (May 11, Fathom transcript automation) - 88e00f19 (May 6, ADR-010 re-examination) - b923559a (May 4–5, Sonnet 4.6 upgrade discussion) - 3e8aa6ed, 8abce74d, etc. (sampled, no additional standalone items)

1. Decisions made but not formally documented

Product / surface decisions

  1. Hide the public multiplayer lobby; multiplayer becomes invite-link-only. [74391e42, Apr 30] "if you're playing a multiplayer game, you're trying to create trust with other people, but if you don't know who those people are every single time you play, then you're not going to be able to make predictions about them. There's no point of having a multiplayer game." The browse-public-games surface was removed pending future re-introduction with privacy controls. CHANGELOG mentions teams but never logs the public-lobby removal as a deliberate design call.

  2. Team scoping rule: the round-starter's team determines team attribution; if user is on multiple teams, they pick. [74391e42, May 8] Captured in convo but not in ADR-033 (which defines the team primitive at a higher level). The "who owns this round" rule lives only in chat.

  3. Team context block on team profile (parallel to personal context). [74391e42, May 8] User explicitly designed: team owner can paste a team-context doc; multiplayer scenarios can then be generated against it, grey-locked behind "team profile filled in." Not in any ADR. Distinct from ADR-052 / ADR-064 which deal with multiplayer generation rules, not the team-context input.

  4. Owner-only edit rights on team profile / team context; ownership transfer deferred. [74391e42, Apr 30] "OK for owner that sounds fine for now, we can add more functionality later like transferring etc." Not logged.

  5. All multiplayer surfaces (Quick Round, Story, Campaign) get the same "Create / Join with code" pattern as Solo. [74391e42, May 8] Cross-mode consistency was decided as a UX principle in conversation. CHANGELOG records bug fixes; the underlying parity principle isn't captured.

  6. SyncDelta7d and per-driver 7-day trend display strategy: rely on the new profile_snapshots table written one row/user/day. [c7d99ac9, May 6] Discussed as B1 in Cici's implementation queue, agreed, and shipped — but no ADR for the snapshot table itself or the "delta-over-N-days" pattern.

  7. Driver-persona copy strategy: bespoke for all 56 driver pairs, not just the 8 single-driver entries. [c7d99ac9, May 8] "unless you see a major downside in implementing these, I think that we can go ahead and do that." The full 56-pair persona text exists in driverPersonas.ts per chat, but no ADR captures the deliberation (vs. composing on the fly, vs. only 8 singles) or the design rule ("frame as strengths; weakness framing held until mock-ups").

  8. Solo "personal" scenario category gating: only enabled once the personal-context field on profile is filled; otherwise greyed-out with link to fill it. [5271c78c, May 3] Codified in conversation as the UX rule. The design brief design-brief-personal-scenarios.md exists but does not enumerate the gate-state rule explicitly.

  9. Personal-context capture is decoupled from onboarding: a post-onboarding lightbox on hub, dismissible, re-appears on first personal-scenario play. [5271c78c, May 3] Same brief, but the specific UX pattern decision (lightbox vs. onboarding step) lives only in chat.

  10. "Stand/Change runner-up choice" UI surface should be added to multiplayer to match solo. [74391e42, May 8] "Yeah I say add it in so that we can see consistency between solo and multi." Quick lightweight decision, not in CHANGELOG or ADR-021.

  11. Demo /demo/credit-sketch page kept in repo unlinked as a "backup if a credit conversation comes back up" with Dunia. [2d47f209, May 1] CHANGELOG mentions the Dunia approval but not the explicit decision to retain the credit-sketch as latent infrastructure.

Agent / Hermes decisions

  1. Skill-install instructions intentionally avoid telling users which Hermes "category" to pick — let each user manage their own bucket. [db0594ae, May 8] "up to each individual to figure out their setup." This is now a documentation principle for SKILL.md / install docs that isn't in ADR-047 or any brief.

  2. Always-run-both predictor logic should be applied to multiplayer rounds the same way as solo. [88e00f19, May 6] "multiplayer is just an extension of solo… it's all collecting the same data to train my digital twin." ADR-044 covers always-run-both but doesn't explicitly cover the multiplayer parity rule.

  3. Sonnet 4.6 → Opus 4.7 NOT considered for the eval/prediction model — Sonnet held. [b923559a, May 4] Discussed; Jonathan asked about Opus cost vs. value and ultimately stayed on Sonnet. ADR-038 captures the Sonnet upgrade but not the "we considered Opus and decided no" deliberation.

  4. Loading-state UX after model upgrade: introduce a countdown / progress affordance for the ~20–30s generation window. [b923559a, May 5] Decided in conversation, no ADR, no CHANGELOG.

Content / scenario decisions

  1. No-batch audit policy: "go through one at a time, no batch, no bulking" for any scenario corpus review. [ffb273d9, May 1; 35c9eaf9, Apr 23] This is now an operational rule that produced two 152-scenario rewrites. ADR-029 covers static-arc audit pattern but not this no-batch rule as a standalone principle.

  2. All static campaigns and static rounds get a name-diversity pass (after the Brink playtest revealed name reuse). [35c9eaf9, Apr 23] Decided but I see no ADR or CHANGELOG; only project_static_arc_audit_pattern.md in memory references the broader audit.

  3. Preamble lives only in Story / Campaign surfaces — never in Quick Round case studies. [18ea62dc, Apr 30] "preambles only really exist in what are now called campaigns and worlds… The case studies themselves are very similar to how the quick rounds are now formatted." This boundary rule isn't in an ADR. The corollary check ("preamble bleed-through") drove a one-off fix but the architectural separation isn't documented.

  4. Multiplayer scenario quality bar = solo scenario quality bar. [48e0a429, May 11–12] "technically the scenario for multi should be as high calibre as a single player scenario regardless of theme." ADR-052 was superseded; ADR-064 covers difficulty but not the quality-parity rule directly.

Process / tooling decisions

  1. Fathom transcript ingestion approach: Python script + manual approval; raw paste only; saves to BeaconInfra/context-sharing, not the sync repo. [cc20b56e, May 11] Memory note reference_fathom_transcripts.md exists, but the design deliberation (MCP vs. API + script, filter rules, approval-UI shape) isn't anywhere formal.

2. Explicit deferrals without an ADR

  1. Async multiplayer is "an hour to build" but deferred until there's demand. [18ea62dc, Apr 29] "For now it's sync, but async is very… really easy to build in… say the word and I'll have it built and tested in like an hour." No ADR-style trigger condition captured anywhere.

  2. Voice equity / "no mechanic change beyond influence chain" deferred. [74391e42, Apr 29] Team metric considered, decided to rely on the influence-chain capture already shipped, but the explicit decision not to build a new voice-equity mechanic isn't logged.

  3. Cross-pair "no fault-line roll-up" team feature deferred until team-distinguished data is reliable. [74391e42, Apr 29] Discussed and parked.

  4. Public-figure-style sharing of weakness-framed driver personas is deferred until mockups exist. [c7d99ac9, May 8] User explicitly wants to see the strength/weakness framing in mocks before committing — held with no ADR.

  5. Score-gaming / governance-manipulation protections deferred to "later, when governance exists." [5271c78c, May 3] "score gaming is a possibility at some point, probably not soon." Tash also raised this in the May 4 all-hands per project_coordination_substrate_security.md. Still no ADR.

  6. Right-to-deletion as a plain editable text field on profile. [5271c78c, May 3] "Could we not have a text field that exists in their profile that they can always edit?" Treated as the working answer; not formalized as a privacy decision update to ADR-007.

  7. "Profile shaper" feature acknowledged as important but undefined — parked. [5271c78c, May 3] User flagged not understanding it; no follow-up documented.

  8. Feature-request capture / triage mechanism deferred — user count too low. [74391e42, May 9] "I'm not sure we should act on this yet since we have very few users, not sure where to capture this for now."

  9. Thumbs-up / thumbs-down feedback on driver personas — held off. [c7d99ac9, May 8] "let's hold off on that. I need to see the mock-ups first."

  10. Pack curator role / who curates the first packs. [c7d99ac9, May 6] ADR-040 stays Proposed; the who curates question Jonathan raised was never resolved.

3. Open questions never answered

  1. What are "themes" for packs — the five Sync dimensions, the existing scenario categories (boardroom / medieval / etc.), or a tag system? [c7d99ac9, May 6] Asked directly; conversation moved on without an answer. ADR-040 doesn't say.

  2. Whether packs should be connected to each other (sequenced, like campaigns) or standalone. [c7d99ac9, May 6] Open.

  3. How to handle a user whose authentic answer is consistently "none of the above / third option" — does that produce a meaningful sync score, or does the system characterise them as a third-option finder? [ffb273d9, May 1] Discussed but unresolved. Touches identity-vs-process but the edge case isn't covered by ADR-017 or ADR-036.

  4. How to test whether weakness-framed copy embarrasses users; is bespoke vs. composed copy actually distinguishable to readers? [c7d99ac9, May 8] "how could we tell that? Is this just through clicks on a page?" No measurement plan exists.

  5. How "register" tags (boardroom / improv / local-life / public-intellectual) map to scenario generation when a user has multiple registers active. [5271c78c, May 5] User raised improv as a distinct register; whether to split or merge into "Valencia local life" is open.

  6. Whether the personal-scenario generator should pick a different register per scenario or stay anchored to one per user. [5271c78c, May 5] "I do quite like the idea of each scenario looking at a different area of my life or a different register." Preference stated; no design commitment.

  7. What signal we expose to Hermes that captures "this is a high-conviction high-flag-rate user" without leaking raw post-reveal reactions. [c7d99ac9, May 14] Related to memory note project_pipeline_model_asymmetry.md and ADR-057, but the specific exposure question wasn't closed.

  8. Whether to expose synapses metric publicly or keep internal until the calculation is settled. [c7d99ac9, May 6] Jonathan loves the idea but explicitly asked "what does this actually look like and how is this actually calculated" — ADR-042 stays Proposed because of this.

  9. Whether the synapses quality groupings should mirror drivers or be a different taxonomy. [c7d99ac9, May 14] "is it the same as main drivers or another quality?" Open — and Jonathan noted drivers may not be how people naturally think.

  10. Whether the Hermes skill should access the user's LLM "memory" (work context / personal context / top-of-mind) at intake. [5271c78c, May 3] Discussed as potentially huge for accuracy; not resolved. The personal-scenarios brief touches the input field but doesn't address pulling from Claude/ChatGPT memory.

  11. Why a user's profile shows "still forming" at 60+ games (Cici × CoachJ exchange, May 4). [5271c78c, May 4] Jonathan flagged this as a product-trust issue — "what's the point if Hermes can't use my data after 60 games?" — and it kicked off the May 11 SKILL.md reframe, but the underlying root cause / threshold question isn't logged.

  12. The right balance between "ask clarifying questions" vs. "dive into recommendation" for Hermes responses. [5271c78c, May 4] "I don't know how to find the balance there of clarifying questions versus just diving into the recommendation." Tied to ADR-047 but not resolved inside it.

4. Feature ideas mentioned positively, not committed to

  1. Case-study writeups as marketing artefact (Ryan Sarver as first subject). [db0594ae, May 8; c7d99ac9, May 7] Discussed repeatedly as the "holy grail" — "my case study proves quite definitively that if you're using an agent without the judgment layer that is created through Sync, it will perform worse." No brief, no ADR, no scoped plan.

  2. Per-team "demo / training" arc that lets a team train its agent on a specific situation. [c7d99ac9, May 7] Came up in the PAX clarification; user said "we're gonna hold off on this for now." No deferral ADR.

  3. A "see all unresolved ADRs" surface inside the product (or at least a saved command). [c7d99ac9, May 8] "I want to be able to pull up all ADRs that are unresolved." The grep recipe exists as reference_unresolved_adrs.md but no in-product or scripted surface.

  4. Daily Hermes-driven sync routine (the "daily drop" idea Jonathan referenced). [c7d99ac9, May 8] Discussed in transcript he was importing; "the daily drop. I believe I shared the transcript with you." Never produced an ADR or brief.

  5. A monetization model: Jonathan flagged twice that monetization is unclear, and floated BYOM (memory note: project_monetization_roadmap.md). [129521f0, May 16; 5271c78c, May 3] Memory captures part of it; no ADR.

  6. Real-time decision capture (in-channel / agent-call) — Phase 2 / 3 of ADR-039. [c7d99ac9, May 5] "not sure how to design the real-time capture piece." ADR-039 punts to "deferred until Phase 1 validates"; the design problem itself is open.

  7. Auto-detect-and-import flow for Beacon-related Fathom meetings (vs. manual title-tag). [cc20b56e, May 11] User picked the manual path for now but expressed interest in the auto-detect path later; trigger condition not captured.

  8. Mobile-first responsive issue surfaced repeatedly (multiplayer demo "still in mobile mode" on wide screens). [74391e42, May 8] Recurring but no canonical fix decision; ADR-031 dashboard pattern was the closest formal pass.

5. Things discussed and ruled out

  1. Crediting Dunia.gg explicitly on the preamble surface — ruled out. [2d47f209, May 1] "I think it would be weird to credit Dunia" — and Dunia's creator's response ("all good with me") made it moot. CHANGELOG records the approval but not the deliberate decision not to add credit text.

  2. Showing the preamble on the chapter-summary page before play. [35c9eaf9, Apr 23] User explicitly ruled this out: "I was thinking more of a brief ai generated description, no? This seems almost like the preamble that I am about to expect." This drove ADR-029 (chapter synopsis dual-source) but the "not the preamble" framing is not captured there.

  3. Mass / bulk scenario audits — explicitly ruled out twice. [ffb273d9, May 1; 18ea62dc, Apr 30] "do not batch bulk anything, and do the rewrites here as needed." Operational rule, not ADR'd.

  4. Supabase Pro upgrade for custom-domain auth proxy — ruled out at the final cost ($45/mo, not $25). [129521f0, May 16] ADR-060 captures the technical deferral but the specific cost-trigger reversal (signed-up briefly, then re-deferred when the per-project compute charge surfaced) isn't in ADR-060's Discussion section in detail.

  5. Acting on the woman-Jonathan-met's "third-option finders" pushback as a corpus change. [ffb273d9, May 1] "I don't think it would get her coherence score that high, or would it?" Decided the existing none-of-the-above + character-shape signal is enough; ADR-017 covers the mechanic but the specific persona pushback resolution isn't logged.

  6. Adding new control surfaces to the personal-context input flow (e.g. structured fields). [5271c78c, May 3] "Agreed. We can always add more control later but sounds complicated to do that now." The "free text now, structured later" decision isn't in the personal-scenarios brief.

  7. Removing the "DTA Delegation: 61% to 95% target" line from the hub. [74391e42, Apr 27] "that math doesn't check out. I'm thinking we can just remove that entirely." Small UI removal; not in CHANGELOG.

  8. Renaming "Back to Campaign" → "Back to Menu" on results page — explicitly left as-is. [74391e42, Apr 27] Decided to keep current language because the World/Campaign rename (ADR-035) was already pending.

Summary

The biggest pattern: a lot of product-shape decisions get made in conversation and either land directly in CHANGELOG (as ship notes) or never get logged at all because they feel "obvious" once decided. The most under-documented categories are (a) multiplayer / team mechanics post-ADR-033 (items 1–5, 10, 13), (b) the personal-scenarios / register system (8–9, 35–36, 40), (c) Hermes interaction-design rules that ADR-047 explicitly deferred but that have continued accumulating (12, 37, 41–42), and (d) operational policies like "no bulk audits" and "scenario-quality parity across modes" that have driven real work without ADR backing (16, 19, 53). The open-questions list (31–42) is where the highest-leverage formalization work is — these are debates that ended in "we'll figure it out later" without a trigger condition for "later."

Second Pass — Additional Items From Smaller Conversations

Date: 2026-05-23 Scope: Remaining ~29 JSONL files not covered in the first pass (sizes ~30 KB to ~2 MB; all from Apr 27 → May 23). Method: Extracted user turns from every uncovered file, cross-referenced against ADRs 001–066, the full CHANGELOG, all eight briefs in docs/briefs/, and the items already listed above. The smaller files skew heavily toward execution (bug fixes, scenario backfills, scan-strip prompt runs, marketing-page tweaks) with little design content. Only items genuinely missing from the formal record and from the first-pass list are below.

Source-file shorthand for this pass: - e5feae1f (May 3, AI-generated scenario images exploration) - a5c335da, 126420e6 (May 21–22, per-category sync scores deep-dive) - c40ff464 (May 17, summary window cap raise) - 3237fad2 (May 16, multiplayer drop-off cleanup) - b43c54a0 (May 22, codebase security pass) - 955707c9 (May 22, MkDocs/Cloudflare/Vercel docs-site setup) - f0494366 (May 15, Obsidian vault question + ADR-059 implementation) - cefb3641 (May 23, SKILL.md tuning + adaptive depth open question) - b04295d2 (May 19, RaidGuild playtest debrief with Tay) - f9bf644f (May 22, Variant D criticism + master-constraints idea) - 2fa3639e (May 21, daily-sync triage) - eb62d4d0 (Apr 24, world/campaign progression-gating preview)

Additional decisions made but not formally documented

  1. Delete ~75% of historical multiplayer drop-off sessions before recording forward metrics. [3237fad2, May 16] "If we wanted to get rid of 75% of the drop-offs for multiplayer, I would have no problem with that. And then we can just start recording moving forward." The data-hygiene call (drop the test-seed multiplayer abandonments so go-forward funnel numbers are accurate) was made in chat and is the practical reason reference_metrics_filters.md exists, but the historical-purge decision itself is not in any ADR or CHANGELOG. Distinct from migration 104 which only added the filter columns.

  2. Raise the learning-summary window cap from 50 → 200 games. [c40ff464, May 17] "I think it's safe to take in more information… Let's raise it to 200." The cap had been set conservatively for cost/signal reasons; bumping to 200 is a deliberate tilt toward "more signal, cheap enough." Not in CHANGELOG, no ADR; quietly tunes the same learning pipeline ADR-057 modifies.

  3. Static-card share-twin is the "Path A" decision; Path B (live updating) explicitly ruled out as a feature of Strava-style sharing. [126420e6, May 21 reference; CHANGELOG covers ship] ADR-041 captures the rationale and CHANGELOG covers the ship, but the negative decision — "live updating would remove reasons to share" — was made in the May 6 Cici exchange and is the load-bearing reason Path B never came back. Worth noting because future "make the share live" requests should bounce off this.

  4. Operational rule for scan-strip backfill: write all patches at once, run inline (no claw.ai), self-contained prompts only. [79ab7ade, May 23] "Can you write down all of the patches at once instead of just one at a time? And also I don't want to run this in claw.ai. So my question is: are these prompts self-contained enough that an incognito window with no context could easily develop the output we're looking for?" This is the working operating procedure for v1.7.0 backfill (and arguably any future corpus-rewrite task), but lives nowhere outside the chat. Pairs with item 53's "no batch" rule — once prompts are self-contained, batching the prompt-writing is fine; what's not OK is batching the per-scenario quality review.

  5. Switched docs-site hosting from Cloudflare Pages → Vercel after Zero Trust dashboard friction. [955707c9, May 22] CHANGELOG mentions Vercel was "the simpler path" but doesn't capture the deliberate abandonment of Cloudflare Pages mid-setup (the Pages-domain dropdown silently refused to accept the right hostname; Zero Trust onboarding led to a recommendations screen with no usable application-setup link). Worth flagging because anyone re-evaluating hosting later should know this was tried and bailed on.

  6. Bring-your-own-MCP path is the preferred install for Hermes-into-Claude-Code, over project-local config or global config files. [126420e6, May 22] During the live install, Jonathan asked "Is there a static way at a level that affects ALL projects not just 1 project? Or do you think MCP is most effective?" and the working answer became: register via MCP at the user level. The MCP tool-name sanitization gotcha (no colons/periods/spaces) was discovered the hard way and resolved in-chat. None of this is in install docs or ADR-047.

  7. Codebase-security pass (migration 100) shipped as a bundled hardening sweep, no ADR. [b43c54a0, May 22] Triggered by "secure the codebase of this project." Migration 100 plus a handful of related changes shipped together. Jonathan paused mid-conversation to ask whether to interrupt his scenario-generator prompt work — answered "do this first so I can integrate." That sequencing call (security before in-flight scenario-prompt work) is implicit in commit history but not captured anywhere as a deliberate priority decision.

Additional explicit deferrals without an ADR

  1. Master-constraints document pattern (à la World mode) considered for fixed-theme scenarios and deferred on cost. [f9bf644f, May 22] "Is it worth exploring the master constraints doc type of thing we do in worlds? Or is that too expensive an API call at this point?" Memory note project_theme_master_doc_deferred.md captures the deferral, but the cost framing — that the deferral was specifically API-spend-driven, not technical — is not in that note or any ADR. Trigger condition: when monetization removes the cost objection.

  2. Auto-detect-vs-manual Fathom import trigger condition (item 49 in first pass) extended. [2fa3639e, May 21; b04295d2, May 19] Two further daily-sync transcripts processed manually since the original deferral confirmed the manual path is still working; no new pressure to flip to auto-detect yet. Worth logging that the trigger is still "not yet."

  3. Refresh-Stonemark-manually decision = ADR-059 Phase 2/3 (hub clearing mechanism) explicitly deferred because an existing self-heal already covers the case. [f0494366, May 15] "And just to make sure there is a hub clearing mechanism already, which is why we're not implementing phase 2 and 3 yeah?" ADR-059 Phase 1 shipped (host migration); the in-chat decision not to implement Phases 2 and 3 isn't captured in the ADR's own Status section.

Additional open questions never answered

  1. "Trust pick phase" counter UI question — should the multiplayer trust-pick phase show a live count of who has picked, and can we wire it to be consistent across players? [b04295d2, May 19] "Trust pick phase — got it. I guess we just have a count there at the last step? Unless we can wire something to show the players and make it consistent, which would be ideal." Raised, not resolved. Distinct from the ADR-059 presence work, which doesn't address the per-phase counter UI.

  2. Multiplayer difficulty calibration paradox — do we serve lowest-common-denominator scenarios when a 70-game host plays with a 1-game player, or do we always serve hard scenarios? [b04295d2, May 19] "Do we default to lowest common denominator when it comes to multiplayer or do we just play any scenarios? Hard scenarios that make me think… I don't know how to reconcile the multiplayer version." ADR-064 covers multiplayer difficulty/onboarding mechanics but the philosophical answer to this question — does mixed-experience multiplayer get hard or easy scenarios — was left open and is upstream of any concrete dial.

  3. Should the AI's "flip condition" in the why-expansion ever surface in the headline, or always live behind a follow-up? [cefb3641, May 23] "I'm actually interested. Do we need to even put that in the headline, or should we leave the headline as is and leave the onus on the user to ask for the flip condition or ask a question similar to, you know, when should I change my mind?" Resolved partially (ADR-055 says no situational flips; profile-derived flips landed in v3.3.0), but the headline-vs-tail placement question for profile-derived flips is still open.

  4. Adaptive response depth — should Hermes read whether a question is exploratory vs. decisive and adjust depth, or should the user set it explicitly? [cefb3641, May 23] Posed as five sub-questions including where a style preference would live (SKILL.md vs SOUL.md). Not answered. The current reco + why shape is treated as the default and the exploratory case is unhandled.

  5. What is the right way to "embed the CoachJ profile" into every Claude Code conversation so the agent works with Jonathan's judgment baseline by default — at the project level, repo level, MCP level, or all three? [126420e6, May 22] "I want this to be a regular part of things that as I continue to go, you continue to act more and more like my own thinking partner, like me." The MCP install (item 64) is a partial answer for new chats; whether mid-conversation chats can re-load and how settings.json hooks fit in is open.

  6. Per-category sync scores — are the five categories (Strategy, Money, People, Voice, +1) the right taxonomy, or should we research what users actually want their agents good at first? [a5c335da, 126420e6, May 21–22] "Let's do some research on what people are using OpenClaw and Hermes for, and what they want to be using them for, to better set these categories for us." The deep-research artefact lives at docs/research/briefs/openclaw-hermes-research.md (per memory/project_category_sync_scores.md) but the decision on categories has not been made.

  7. Does the per-category score get any actual computation, or is it purely visualization? [126420e6, May 22] Jonathan: "if it's just pure visualization then it's meaningless." Open — tied to whether the weighted average rolls up to the existing single score or runs in parallel.

  8. What does "Calls" / "Hard Calls" / "Stakes" actually mean as a per-category label? [126420e6, May 22] Jonathan didn't love "Hard" or "Stakes" as words; "ethics could work, but it could be too narrow and loses the irreversibility." Naming is unresolved and blocks any user-facing per-category surface.

  9. Should the per-category score sit in SKILL.md / agent-skill output, or only in the in-product UI, or both? [126420e6, May 22] "I don't know if we need to do any calculations on the actual homepage when you land and see those different numbers, or if it's just on the Hermes skill." Open.

  10. Embedding judgment-capture inline in every external decision moment ("anytime there's a judgment call made, find out why and record it"). [126420e6, May 22] Floated as the grand-vision version of ADR-039. Same domain but explicitly more ambitious than Phase 1 capture. Trigger condition: never set.

Additional feature ideas mentioned positively, not committed to

  1. "Lead product development with an AI that matches my profile" — set up a Hermes agent that uses the CoachJ profile to make product decisions and queue tasks autonomously. [126420e6, May 22] "I would love for this program and this design of this product to almost be led by AI… maybe even setting up a Hermes agent to do that and giving it tasks and it has my profile so it knows what decisions to make." No ADR, no brief. This is the natural extension of the trust-thesis line of work but it's never been scoped.

  2. Mine all historical Claude Code conversations (not just the ones flagged) for missing ADRs and undocumented decisions. [126420e6, May 22] Exactly the meta-task that produced this audit. Jonathan's framing: "I would even encourage you to go back through all the chats and look at decisions that were made, decisions that were not made, discussions that were had…" — so the recurring version of this audit (not just a one-shot) was the actual ask. No scheduled job exists.

  3. HTML render of ADRs that Jonathan can read while leaving markdown as source — exists via MkDocs site, but the broader idea of HTML-first reading + Obsidian-vault writing was the original ask. [955707c9, 126420e6, May 22] The shipped MkDocs site partly answers this; the Obsidian vault for cross-project memory angle (item 14 from chat, drives the cross-project memory ask in user_parallel_projects.md) is still unaddressed — no decision on whether docs/ should be opened in Obsidian directly or mirrored.

  4. Sycophancy mitigation as a measurable product claim — run a simulation that contradicts the profile to test whether Sync actually reduces sycophancy or just gives the user reason to act on pushback. [cefb3641, May 23] Memory note project_sycophancy_content_play.md captures the content angle. The proposed experiment — use the Supabase MCP to fetch the actual CoachJ profile, design questions that deliberately contradict it, then measure agent behavior — was a Jonathan-recommended next step, not committed to.

  5. Timestamping all SKILL.md tool calls and "why" requests for downstream measurement. [cefb3641, May 23] "Are we timestamping tool calls for skills and the why? Would be helpful for measurements." Partial shipping happened (migration 095 added why_requested_at); whether to also instrument every sync.get_profile / sync.get_recent_decisions call with timestamps was Jonathan-approved ("the more data we have, the better") but never scoped beyond the why-timestamp.

  6. Cinematic/visual preamble exploration — the brief exists (docs/briefs/visual-preamble-exploration-brief.md), but the original e5feae1f conversation also surfaced the idea that scenario bodies themselves could be image-augmented, not just preambles. [e5feae1f, May 3] The current brief scopes preambles only. The broader "every scenario gets a mood-setting image inserted in the right spot" idea is deferred without explicit trigger conditions; would compound the image-gen cost objection that already deferred world-mode preamble images.

Additional things discussed and ruled out

  1. Backfilling scan strips by chaining one-prompt-per-batch through a script — ruled out in favor of inline manual paste. [79ab7ade, May 23] "Yeah do it inline I have lots of tokens for this week left." Confirms the operational principle: when token budget is fine, prefer inline to a scripted batch run. Same family as items 53 and 62 but distinct: this one is about the executor, not the unit of review.

  2. API-key spend on internal eval simulations — ruled out by default, allowed only with explicit per-experiment permission. [f517b3fd, May 15; cefb3641, May 23] "Wait, are you using my API keys to do this? Can you not just do this here in this chat? I don't like that we're spending money on this. That's unnecessary." and later "if you think that there's anything you want to run with an API key just to make sure, I'll allow it for just this experiment." This is the operational rule behind feedback_no_api_spend_when_inline_works.md, but the carve-out — that one-off experimental API calls are OK with explicit permission — isn't in the memory note.

  3. Renaming sync.beaconinfra.xyz to a different post-Pulse host — discussed and ruled out for now; the marketing rebrand lands as Pulse but the runtime/auth host stays. [2f9a3ebf, May 23] The Pulse rebrand commit log shows the homepage swap; the deliberate choice to keep sync.beaconinfra.xyz running and not introduce a pulse.beaconinfra.xyz runtime hostname yet is implicit in the recent commit 3b24033 ("Switch all hardcoded URLs from sync.beaconinfra.xyz to pulse.beaconinfra.xyz"), which actually went the other way mid-week. The order of operations — marketing pages first, then runtime host swap — was a deliberate sequencing call that isn't captured.

  4. Surfacing flip conditions in the headline even when the flip is purely situational/common-sense — explicitly ruled out. [cefb3641, May 23] Three consecutive live test runs produced flips that were "all common sense things." Decision: do not surface situational flips at all; profile-derived flips only. Captured in ADR-055 and SKILL.md v3.1.0 — but the underlying test-driven nature of the decision (it was three concrete misses, not a priori reasoning) is worth preserving and isn't in the ADR.

Second-pass summary

The second pass added 30 items (59–88), heavily concentrated in three buckets: (a) operational/process rules for scenario backfill and audit (items 62, 85, 86), (b) the still-unresolved per-category-sync-scores design space (items 74–77), and (c) Hermes/SKILL.md interaction-design questions that ADR-055 partially resolved but that have continued accumulating (items 71–73, 78, 83). The smaller files were largely execution-heavy — bug fixes, scenario rewrites, marketing tweaks — but did surface several deferrals and open questions (especially around multiplayer difficulty philosophy, response depth adaptation, and the "AI-led product development" feature idea) that weren't in the first 58 items or any ADR.