FE-1431: Build the plugin authoring surface by lunelson · Pull Request #9337 · hashintel/hash

lunelson · 2026-08-26T08:30:42Z

🌟 What is the purpose of this PR?

Brunch interviews a domain expert and builds a formal model from what they say. ADR-0007 decided that the interviewer's craft — how to open, when to walk one case and when to sweep, how to probe, where not to dig, how to close — belongs to the harness, and that a target formalism contributes only data under a fixed set of harness-owned keys. This pull request makes that decision executable. It materialises the key catalogue as a schema in core, writes the harness's default teaching (the repertoire) against it, and co-authors both test-case plugins (SDCPN process models and Gherkin scenarios) against the same shape, so that a plugin is one validated YAML document rather than prose the harness interprets. It then puts the rendered layer under evaluation: baseline condition 4 tests the teaching as a prompt alone, and condition 5 runs the harness itself inside the baseline loop against the same expert simulator, recording the first turn-latency figures.

The branch also carries the steering pass that closed the condition-5 arc and selected the next frontier (S-011, the first human elicitation from the Petrinaut panel), together with the control-vocabulary and worktree protocols that let that work run as a parallel partition. This pull request was opened as a draft while cycle one was under review; cycle two (the freeze) and ADR-0008 (repertoire ownership) have since been built on branches stacked above it, so its scope is the cycle-one materialisation and the evaluation and control work that followed it.

🔗 Related links

FE-1431: Build the plugin authoring surface (internal; owning issue, Done)
FE-1406 (internal; ADR-0007, the decision this implements)
Decision: libs/@hashintel/brunch-agent/docs/adr/0007-harness-teaching-meets-plugin-content-at-fixed-keys.md
Contract: libs/@hashintel/brunch-agent/docs/specs/plugin-contract.md; schema changelog: libs/@hashintel/brunch-agent/packages/core/schema/CHANGELOG.md
Evidence: libs/@hashintel/brunch-agent/docs/evidence/evaluations/process-model-elicitation/baseline/readout.md (conditions 4 and 5); docs/evidence/proofs/design/plugin-keys-pressure-review-cycle-1.md
Strategy: libs/@hashintel/brunch-agent/docs/control/STRATEGY-LOG.md S-009 to S-011; frontier and partition: docs/control/STEERING.md

🚫 Blocked by

Nothing in code. Stacks on #9327 (ADR-0007). The branches above it — the panel run, timing, the cycle-two freeze, and the ADR-0008 move — depend on this one landing.

🔍 What does this change?

The plugin contract as code. PluginDefinitionSchema in packages/core/src/plugin-definition.ts defines the sixteen keys in four groups (contract data, guidance typed by mechanism, per-job runbooks, machinery). packages/core/schema/plugin.schema.json is derived from it and a test fails when they drift. Readers reject an unknown key, require the repertoire to fill every guidance and runbook key, and let a plugin leave cells blank but never add a key. The instruction renderer interleaves key → harness default → plugin cell, so cells add and never override.

The repertoire and both plugins. packages/repertoire/repertoire.yaml holds the harness's default filling of every key, admitted from the evidence ADR-0007 names (the catalogued elicitor failures, the v0 prompt's headings, the lifted rows). plugin-sdcpn/plugin.yaml and plugin-gherkin/plugin.yaml are written against the same schema; the SDCPN Markdown plugin file retires, its conversational triggers become techniques and sweep movements, and the precision ladder becomes harness vocabulary. The cycle-one pressure review (100 situations, two further plugins sketched) found every situation landing on an existing key and fixed one matching defect (patterns.items[*].on: [] never fired).

Evaluation. Condition 4 puts the rendered teaching into the baseline as a prompt only; condition 5 runs the harness inside the baseline loop, which required the Flue binding to accept the rendered instruction and the runner to observe per-turn timing. The retired condition-3 instrument is deleted. The condition-5 readout records the arc's headline: about 145 seconds per interviewer turn, which S-011 makes the first thing the next frontier measures.

Control and protocols. The strategy log gains S-009 to S-011; STEERING.md is rewritten to the dogsled shape (objective, frontier, governing concerns, beliefs, cuts, stop triggers) and records the parallel partition of the next arc into efforts. CONTEXT.md defines move, stream, sequence, effort, and partition. docs/agents/ gains the partition-worktrees protocol and the steering supplement's partition step; legibility.md, arc-close.md, issue-writing.md, and git-workflow.md are tightened from what the arc taught.

🏗️ Agent notes

Commits

63eeb6aff1 Materialise the plugin keys, the repertoire, and both plugin definitions
259255f5a9 / 1e58aa1677 Add and score baseline condition 4: the teaching layer as prompt only
f447cc2ba8 Record the cycle-1 pressure review and fix any-kind patterns
67d4ff4b15 / 0159d7ac84 Put the harness in the baseline loop as condition 5; record the first run
7501dcfa26 Close the condition-5 arc: steering, ledger, and the turn-latency assessment
312ddf095f Delete the retired condition-3 instrument and fix the condition-5 test stub
f2e52e298f … a0bf9c8ffc Reconcile steering with dogsled; select the black triangle as the frontier (S-011)
eaef0fdca9 … dcf5b37ebf Control vocabulary (moves, streams, sequences, partition) and the partition-worktrees protocol
323cdd09f8, 19f7e9247a Repair stale control docs; anchor the control-surface test on the governing-concerns section
f4ced883eb … 9543278418 Driver-side steering reconciliation during the partitioned arc

Verification

turbo run test:unit lint:eslint lint:tsc across @hashintel/brunch-agent, the repertoire, both plugins, binding-flue, transport-aisdk, and @apps/brunch-agent; the architecture suite (package direction, Flue entrypoint, hermetic runtime, control surfaces) passes. The schema-drift test ties plugin.schema.json to PluginDefinitionSchema. Condition-4 and condition-5 runs are recorded as immutable snapshots under docs/evidence/evaluations/process-model-elicitation/baseline/runs/.

Stack

main → #9322 (FE-1404) → #9325 (FE-1497) → #9327 (FE-1406) → this PR → #9345 (FE-1504) → FE-1506 → FE-1505 → FE-1431 cycle two → FE-1507. The packages/repertoire workspace this PR introduces is removed again by the FE-1507 branch under ADR-0008; it is kept here because the cycle-two freeze evidence was produced against it.

Pre-Merge Checklist 🚀

🚢 Has this modified a publishable library?

This PR:

does not modify any publishable blocks or libraries, or modifications do not need publishing

📜 Does this require a change to the docs?

The changes in this PR:

are internal and do not require a docs change

vercel · 2026-08-26T08:30:48Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
hash	Ready	Preview	Aug 27, 2026 5:56pm
petrinaut	Ready	Preview	Aug 27, 2026 5:56pm
petrinaut-docs	Ready	Preview	Aug 27, 2026 5:56pm

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
hashdotdesign-tokens	Ignored	Preview	Aug 27, 2026 5:56pm

lunelson · 2026-08-26T08:31:02Z

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

This stack of pull requests is managed by Graphite. Learn more about stacking.

Cycle one of ADR-0007's co-authoring convergence (S-009): the harness-owned key catalogue exists as code, both test-case plugins are written against the same schema, and the harness's own teaching ships as a package that the binding renders interleaved with each plugin's cells. Harness (`packages/core`): - `keys.ts` is the catalogue — jobs, the eight guidance keys, the two movements, the three runbook keys — with the one-paragraph definition the interviewer reads above every rendered key. - `plugin-definition.ts` reads `plugin.yaml` under a strict valibot schema (an unknown key anywhere fails to load) and cross-checks what a schema cannot say: rows name declared kinds, every kind has a row, the anchor is a counted `must_know` row, runbooks belong to declared jobs. The anchor is declared (`schema.anchor`), replacing `objective`-by-convention; precision words become harness vocabulary (`PRECISION_LADDER`). - `repertoire.ts` types the repertoire and requires every key filled and every entry sourced. - `instructions.ts` renders harness preamble → contract → each guidance key (definition, default, cell) → each declared job's runbook. The completion protocol sentences move from `cue.ts` into `HARNESS_PREAMBLE`. - `schema/plugin.schema.json` is derived from the valibot schema (`yarn schema:emit`) and a structural drift test keeps it honest; `schema/CHANGELOG.md` opens the cycle log with cycle one's shape and its open questions. - `plugin-file.ts` (the Markdown parser) and its test are removed; completion, fold, slot-assertion, cue, and `Plugin.definition` consume the definition. Fixture is now YAML. Repertoire (`packages/repertoire`, `@hashintel/brunch-agent-repertoire`): every guidance and runbook cell filled from the v0 prompt, the failure catalogue (FM-02/03/04/06/07/08/12/13/14/15), the condition-3 hints, GEN-Q02, the literature cards, and ADR-0007 decision 4; each entry names its source. Depends on core only; the boundaries gate now checks that only bindings depend on it, and the role regex and kernel §12.2 admit it. Plugins: `plugin-sdcpn/plugin.yaml` (objective anchor, ten kinds, 24 rows, eight patterns, formalism-specific cells only) replaces `plugin.md`; `plugin-gherkin/plugin.yaml` (feature anchor, four kinds, feature → rule → example → step) is written in the same cycle so the schema is pulled by two formalisms at once. Gherkin keeps its `statement-noted` proposal for now; the binding folds and cues only for definitions that declare `slot-asserted`. Binding: `renderInstructions(repertoire, definition)` replaces the plugin-file instructions and the separate completion fragments. Docs: `plugin-contract.md` amended to the key contract; INDEX rows, kernel §12.2, README, CONTEXT, and every `plugin.md` link repointed; STEERING's authoring lane names the simulated-interview re-run as each cycle's run stage. Gates: `test:unit`, `lint:tsc`, `lint:eslint`, `build` green across core, repertoire, binding-flue, plugin-gherkin, plugin-sdcpn, and apps/brunch-agent. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The first "run" stage of an ADR-0007 co-authoring cycle (S-009; STEERING's authoring lane): re-do the simulated interview with the interviewer taught by the rendered repertoire and SDCPN plugin definition, and nothing else. `run.ts 4` assembles the system prompt from `condition-4-prompt.md` — a framing that stands in for the harness preamble, telling a harness-less interviewer to keep the Must-know tally and the assumption ledger itself — plus `renderContract`, `renderGuidance`, and `renderRunbook(construct)` from the harness's built output over the shipped `repertoire.yaml` and `plugin-sdcpn/plugin.yaml`. The assembled prompt is written beside the transcript as `condition-4-system.md` so the run is reproducible against the exact text it saw. Mechanics are conditions 1–2's (legacy impatience probe, turn budget, delivery classifier), so the 2→4 delta isolates what the fixed keys and the repertoire buy over the seven-category prompt; a later harness-in-the-loop run isolates what the machinery buys over this text. Verified hermetically through the stub client (delivers on turn 1; the system prompt renders every key in catalogue order, ~5.7k words). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The first co-authoring cycle's validate step: a read-only desk review of the ADR-0007 key catalogue against 100 situations from the CPS process-modelling material, the literature, and the condition-2 run, with a discrete-event and a formal-verification plugin sketched against the keys. Placed as evidence under docs/evidence/proofs/design with a provenance header and indexed. Its verdict is folded into the schema changelog as the cycle-two input: no key changes; two shape changes proposed (an optional slot predicate on pattern rows; any-of precision on must-know rows), one applicability facet for repertoire entries, the content findings (motifs restating patterns, the posture nothing consumes, repertoire under-fill against the ADR's own rows, seven source contradictions resolved silently), and eight changes considered and left. One defect it found is fixed here: a pattern indexed on no kind (`on: []`, documented as "any node") never fired, because the sweep list tested `kinds.includes(node.kind)` on an empty list. sdcpn P08 (source-regime divergence) therefore never reached the interviewer as a harness fact. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

One run of the rendered ADR-0007 layer (condition-4 framing + repertoire + sdcpn definition, prompt only, no machinery) against the simulated master scheduler: 22 interviewer turns, delivered-after-forced-wrap. Transcripts, raw record, assembled system prompt, and delivered model added; the read-out appended to readout.md under its own heading with headline findings, the seven-category and pack-tier tables extended with a condition-4 column, the silent-assumption audit, stopping behaviour, the 2->4 delta, and a "Strains for cycle two" list keyed to plugin.yaml and repertoire.yaml entries. Headline: the best-disciplined questioning of any condition (objectives first, one to three questions a turn, quantiles on every changeover type, per-slot precision and an attributed assumption ledger in the deliverable) and the thinnest model (no run time, no arrival pattern, no QA spread; one tacit fact of nine fully surfaced). Stopping failed in a third form: the interviewer delivered on the expert's stop at turn 10 and then had no concept of ending. Must-know rows were graded rather than asked; the anchor's dependency slot was authored by the interviewer and drifted across three deliverables. Instrument: the haiku delivery classifier answered NO on the turn-10 deliverable (replayed 3/3) because it names its own gaps, so the stop reason overstates the failure; the raw record keeps classifier usage but not verdicts. Both noted for the rerun. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The baseline protocol's fourth condition measured the ADR-0007 teaching layer as prompt text; nothing yet measured the machinery around that text. Condition 5 drives the shipped SDCPN elicitor in the Flue runtime — the binding's ask suspension, settlement nudge, private sweep extraction into the capture store, fold, and computed completion — against the same simulated expert, probes, and turn budget as conditions 1, 2, and 4. The runner reads every harness fact from durable history and the store, takes interviewer tokens from Flue's observe() turn events, and needs no delivery classifier: the deliverable is the folded store, and the interviewer ends its own turn-taking by replying without a question. The runner lives with the protocol but is launched from the application package (`baseline:harness`), which owns the agent composition; the elicitor's model id gains a `BRUNCH_SDCPN_MODEL` override so the runner can use the same interviewer model as the other conditions. A hermetic test swaps both models for stand-ins — the existing Anthropic stub for the expert, a faux pi provider for the interviewer — and asserts the loop end to end: ask bound, sweep applied with completion, stall detected, artifacts written. Condition 3 is retired before its first model call (superseded by the harness it would have projected by hand; lock left sealed, amendment atop the preregistration) and conditions 1 and 2 are frozen as reference evidence, per the checkpoint review. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The shipped SDCPN elicitor (claude-opus-5) in the Flue runtime against the simulated expert, single-shot: twelve interviewer turns, 267 captures, 69 nodes, completion not reached (46 unsatisfied; 53 nodes outside every objective's slice). The runner's `stalled` rule fired on what the transcript shows to be a deliberate close: at turn 10, immediately after the impatience probe, the interviewer read the record back in the expert's words and declined to call it finished; the two turns after it are notes-to-self. Evidence only — the read-out and the cycle-two edits wait for review. A first attempt that reached turn 10 died in a network outage and is kept outside the tree. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…essment Arc-close for the first harness-in-the-loop run. Records S-010 (retire condition 3, freeze 1–2, conditions 4 and 5 as the live arms; harness facts replace the classifier), the Proof 1 first-run result and narrowed claim in STEERING, four belief rows, the candidate successor streams, and the per-turn latency as an immediate concern with its own evidence document: 29 min for 12 turns, ~97% of output tokens extraction on the critical path, actions R0–R5 and a deferred spike. Ledger rows §8.1, §8.2, §9.1, §9.5, §11.1 reflect what the run exercised. Notes the FE-1404 numbering gap (its "condition 3" is the protocol's condition 5). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…t stub Salvage assessed as none: the instrument's projection schema and semantic validators encoded the domain-keyed DemandTable that S-007 ruled the wrong level; the kind-level fold and evaluateCompletion in core now do that job on the production path. Removes condition-3-instrument.ts (1,317 lines), its lock, the operator, scoring, legibility, and pre-run-review documents, its unit test, and the condition-3 paths in run.ts (1,720 → 623 lines; conditions 1, 2, and 4 keep their reviewed resume/continue semantics). The preregistration and prompt stay as the record, with a dated amendment. Also: the condition-5 test passed replies inline while the shared expert stub has always read them from BASELINE_STUB_REPLIES_PATH, so the committed test could not pass; it now writes the replies file. S-010's Supersedes line is bare "none" so the control-surfaces gate accepts it. Protocol, STEERING, and S-010 record the deletion. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Latency targets adopted provisionally; the isolating spike as first task of the next arc marked provisional pending Lu's further next-steps input; the condition-5 read-out deferred until after the spike; and the black-triangle directive recorded as the frame the streams are worked inside: the full end-to-end flow through the Petrinaut assistant interface first, so the surface stabilises for the team and the voice-mode work while the implementation behind it keeps changing. S-011 is appended when Lu's input lands. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Reduce the local steering protocol to its Brunch-specific supplement and route strategic refreshes through ds-steer. Align Wayfinder commission, landing, issue-close, and concern vocabulary with the canonical skills while preserving local Linear approval and proof-execution rules. Refresh STEERING as current concern-first control without diary or mirrored tracker history.

Two jobs, one order: construct through the Petrinaut panel is Proof 0, review-and-revise sits on top. G0 (FE-1503) is the selected frontier with its cut, parallel streams, gap assessment, and tool inventory pinned as current truth against intent; G1 and G2 follow as a sequence. Records the plugin-API-in-flux and package-topology concerns, the voice-edge attach surface and provider gate, and defines legibility on the human in the legibility protocol. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The work inside G0 shares one proof and is a set of moves joined at Proof 0, not streams; "stream" is reserved for separately proven parallel work, and G0 -> G1 -> G2 is a sequence. Defines the three terms in CONTEXT.md and corrects STEERING, S-011 (addendum), and the legibility protocol. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

ds-steer selects one frontier and leaves other available work unplanned; Brunch adds a partition step: efforts with their own proof, projection, disjoint write set, and named join points, run in separate worktrees with control documents written only from the driver. Records today's partition (W1-W5 and the voice work) in STEERING and defines effort and join point in CONTEXT. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

S-011's link to the H-6763 plan pointed at a file that exists only on the voice branch; three older strategy entries linked STEERING headings that no longer exist; the E1 and E2 epicentre rows described gaps that landed on the stack. Adds the unmerged stack as an active gate, since G0's baseline is not on main. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

The steering control retired the `Governing strategic decisions:` line in 6903277; the architecture test still matched it and failed on every checkout. Match the `## Governing concerns` section instead, keeping the invariant that steering cites at least one strategy entry and every cited entry resolves and is unsuperseded. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The app's condition-5 test executes the context-owned runner, so package-isolated CI must copy that non-workspace path. Co-authored-by: Cursor <cursoragent@cursor.com>

The contract text and slice reader both assumed objective-shaped demand: a hardcoded "objective-relative" floor line, and a dependency reader that dropped a single kind:node string. Steering also linked ADR-0008 and cycle-one transcripts this branch does not yet carry. Co-authored-by: Cursor <cursoragent@cursor.com>

github-actions Bot added area/deps Relates to third-party dependencies (area) area/infra Relates to version control, CI, CD or IaC (area) labels Aug 26, 2026

lunelson mentioned this pull request Aug 26, 2026

FE-1406: Audit the harness-teaching lineage and propose the fixed-key plugin contract (ADR-0007) #9327

Open

3 tasks

github-actions Bot added area/libs Relates to first-party libraries/crates/packages (area) type/eng > frontend Owned by the @frontend team area/tests New or updated tests area/apps type/legal Owned by the @legal team labels Aug 26, 2026

vercel Bot deployed to Preview – petrinaut-docs August 26, 2026 08:32 View deployment

vercel Bot deployed to Preview – petrinaut August 26, 2026 08:34 View deployment

vercel Bot deployed to Preview – hash August 26, 2026 08:41 View deployment

lunelson changed the base branch from ln/fe-1406-harness-teaching-adr to graphite-base/9337 August 26, 2026 08:49

lunelson force-pushed the ln/fe-1431-plugin-authoring-cycle branch from 57631f3 to b3d5343 Compare August 26, 2026 10:42

lunelson force-pushed the graphite-base/9337 branch from 000d346 to 44fb37e Compare August 26, 2026 10:42

lunelson changed the base branch from graphite-base/9337 to ln/fe-1406-harness-teaching-adr August 26, 2026 10:42

vercel Bot deployed to Preview – petrinaut-docs August 26, 2026 10:44 View deployment

vercel Bot deployed to Preview – petrinaut August 26, 2026 10:46 View deployment

vercel Bot deployed to Preview – hash August 26, 2026 10:50 View deployment

lunelson changed the base branch from ln/fe-1406-harness-teaching-adr to graphite-base/9337 August 26, 2026 13:44

lunelson mentioned this pull request Aug 27, 2026

FE-1524: Prove idempotent capture sweep on the live Flue chat path #9384

Open

4 tasks

lunelson and others added 23 commits August 27, 2026 19:41

Define partition as a strategic-control term

227e48b

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Add a protocol for cutting worktrees from a recorded partition.

94c98dd

Co-authored-by: Cursor <cursoragent@cursor.com>

driver fixes for W1 overreach

157bd65

steering churn

69409e7

more steering churn

ca489f2

Record the completed pre-Proof-0 braid in steering

8e11c92

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Keep the Brunch evaluation runner in pruned app jobs

27ffd72

The app's condition-5 test executes the context-owned runner, so package-isolated CI must copy that non-workspace path. Co-authored-by: Cursor <cursoragent@cursor.com>

lunelson force-pushed the ln/fe-1406-harness-teaching-adr branch from a3724ed to bed9c08 Compare August 27, 2026 17:42

lunelson force-pushed the ln/fe-1431-plugin-authoring-cycle branch from 25e0280 to 3db1b81 Compare August 27, 2026 17:42

lunelson deployed to pull-request August 27, 2026 17:43 — with GitHub Actions Active

vercel Bot deployed to Preview – petrinaut-docs August 27, 2026 17:45 View deployment

vercel Bot deployed to Preview – petrinaut August 27, 2026 17:46 View deployment

Navigation Menu

Uh oh!

FE-1431: Build the plugin authoring surface - #9337

FE-1431: Build the plugin authoring surface#9337
lunelson wants to merge 23 commits into
ln/fe-1406-harness-teaching-adrfrom
ln/fe-1431-plugin-authoring-cycle

lunelson commented Aug 26, 2026 •

edited

Loading

Uh oh!

vercel Bot commented Aug 26, 2026 •

edited

Loading

Uh oh!

lunelson commented Aug 26, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

lunelson commented Aug 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🌟 What is the purpose of this PR?

🔗 Related links

🚫 Blocked by

🔍 What does this change?

Commits

Verification

Stack

Pre-Merge Checklist 🚀

🚢 Has this modified a publishable library?

📜 Does this require a change to the docs?

Uh oh!

vercel Bot commented Aug 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lunelson commented Aug 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

1 participant

lunelson commented Aug 26, 2026 •

edited

Loading

vercel Bot commented Aug 26, 2026 •

edited

Loading

lunelson commented Aug 26, 2026 •

edited

Loading