address-feedback-with-replies

# Address Feedback With Replies Use this workflow when the user wants feedback triaged, fixed, and answered in Slack in one pass. Reply only in the requested scope and keep replies as evidence-based as the code. Cluster identical symptoms under one Builder thread and follow `review-latest-feedback`'s single-owner reply rule. This workflow handles clear bugs and concrete design/UX feedback about existing repo-owned surfaces. In Slack sweeps, follow `review-latest-feedback` for eligibility: do not exclude Design feedback because it is visual or subjective; new capability requests still require the invoking identity's `:upvote:`. Praise, status updates, merge or review requests, bot forwards, duplicates, and other unrelated messages stay out of scope. Content remains with Alice unless the user explicitly assigns it. An `:upvote:` from the invoking identity promotes a new-capability request into scope - that reaction is the product decision, so build the smallest version rather than asking which variant is wanted. The upvote is the authorization — do not wait for a second sign-off. It does not transfer ownership: an upvoted Design or Content item still gets built, with Sid or Alice named in the recap row so the mapped owner is not surprised by a change in their area. Naming them is a courtesy, not a gate. Either workflow may complete an upvoted improvement and record the terminal **Shipped** disposition. Everything in this file about voice, evidence, and verification applies to an upvoted item unchanged. For each distinct defect, ask one direct question when reporter input changes reproduction or unblocks a safe fix. There is no per-run question cap; keep one pending question per thread. If this workflow eyed an out-of-scope item, keep the eye and post one concise **Skipped** reply if needed. Apply `review-latest-feedback`'s Reaction gate before `🎫`. Correct any mistaken reply to **Skipped**. New messages must pass the clear-bug gate before any external write. Follow `review-latest-feedback` for status reactions; never remove reactions. Follow `review-latest-feedback` for ownership and cluster status; a claimed report keeps its `👀`, and an unresolved eye cannot be taken over. ## Prerequisites - Read `address-feedback`, `concurrent-agents`, and `verifying-changes` first. - Read the linked Slack parent and every reply. If a message points to a Clips link, transcript, video, screenshot, file, or newer follow-up, inspect it too. - Inspect every linked artifact that is accessible, but track an artifact that is permission-gated, expired, or otherwise unreadable separately from evidence that is absent. Do not treat an inaccessible artifact as proof that its contents are missing. If that artifact is actually needed to identify or verify the fix, ask for access or a fresh/replacement link; if it is not needed, continue with the available evidence and record the limitation. - Before asking, build a known/missing-evidence ledger from the complete thread and linked artifacts: surface, repro, exact error, files, run IDs, and answers. Never request evidence already present or accessible. If a needed artifact is inaccessible, request access or a replacement link, not its contents again. - Scan for a substantive cause, repro, fix, or ownership signal first. Verify it or continue the handoff; never ask the reporter to repeat it. - If the report includes a run ID, use that ID first to inspect the persisted run, events, tool cards, and linked app state. Do not ask for the prompt or last tool card until the run ID and available observability paths have been exhausted; do not ask for evidence already present in Slack or the app. - Search recent Slack history, local Git history, and merged PRs for repeat reports and existing fixes before editing. - Re-read dirty files. In task worktrees, make needed branch changes without asking; preserve peer work and branches. Keep shared checkouts on their current branch. Explicit ship/PR requests authorize a managed task worktree from the base selected by `new-branch` when needed; carry only this task's changes there. Never reset, stash, or overwrite peer work. ## Slack identity Every Slack interaction in this workflow - channel history, reactions, thread replies and read-backs, permalinks, and Slack message/user/file metadata - uses the connected Slack identity of the user who invoked the workflow. Verify that identity before the first write and keep its stable `{ team_id, user_id }` tuple for all read-backs. Do not silently switch to a bot, another user's connector, or a different workspace. Do not load or use `SLACK_BOT_TOKEN`. Linked external artifacts are not Slack interactions. After extracting their reference from Slack, fetch them through the artifact's owning connector or public URL. - Read the current Slack profile and confirm its user ID and display name match the invoking user. If identity or workspace verification fails, record Slack as unavailable and do not write. - Use that same connected identity for channel history, thread replies, reactions, and Slack metadata. Use cursors until the requested history or thread is complete. A reply read-back must include author metadata matching the invoking `user_id`; missing author data is unverified and cannot satisfy the reply ledger. It must also match the target channel, timestamp, thread parent, and workspace. A reaction read-back must include the invoking user in the reaction's user list. - After every reaction or reply, re-read through the same connected identity and verify the reaction or reply exists before continuing. ## Decision Gate Apply the shared `address-feedback` **Choose the fix altitude** gate before reacting, editing code, or replying. It selects the smallest owning seam and prevents one subjective report from becoming a global instruction. For Slack sweeps, `review-latest-feedback` owns scope and reaction eligibility. Honor its inclusion of concrete Design/UX feedback about existing surfaces and the invoking user's assignment; do not override it with a generic UX exclusion. New-capability requests still need the invoking identity's `:upvote:`. Once an item is in scope, add `👀` before investigation or delegation. During triage, follow the single-owner rule. After merge, follow `review-latest-feedback`: reply once in every affected source thread, including clusters. A tracked clear bug or authorized upvoted improvement receives at most one disposition per run. **Verified locally**, **Built - live unverified**, **Deployed - live unverified**, **Not reproducible - attempted**, and **In progress** retain `👀` after the report has been claimed. **Asked**, **Clarification needed**, and **Blocked on reporter** retain `👀` after the report has been claimed. Terminal dispositions: **Fixed**, **Shipped**, **Live verified**, **Open - no question**, **Resolved elsewhere**, **Skipped**, **Clustered**, and **Abandoned - no answer in 4 days**, each with required evidence and `👀`. Follow `review-latest-feedback` for `✅`/`🎫`. An already-eyed skipped item gets one concise **Skipped** reply if absent. **Fixed** closes the issue after a verified source fix merges; publication, beta, and live verification follow separately. **In progress** is an open ownership state. Do not send an interim progress reply for work fixed in the same run. If work continues beyond this run, follow `review-latest-feedback` and post one concrete status before the next sweep. Continue or coordinate existing work instead of starting a duplicate. Never use a vague update or replace verification. The next run must revisit this state and resolve it to **Fixed**, **Clarification needed**, or evidence-backed **Open - no question** when no safe fix or reproduction remains. `Blocked`, `not fixed yet`, `still needs a fix`, and similar phrases are internal notes, never a complete Slack reply. **Open - no question** is terminal with our eye retained and no checkmark. A reply must state a verified resolution, concrete active work, a needed question, or one terminal disposition with its reason; never send a vague status alone. These are ledger states, not mandatory headings: keep the reporter-facing wording natural instead of opening with the robotic phrase “Clarification needed”. A substantive diagnosis, fix, or in-progress ownership statement from someone in the thread is not a reason to ask for clarification; verify it or continue the existing handoff first. **Clarification needed** is an open state, not a completed product fix. Asking the question creates a standing obligation to come back for the answer. It is the invoking identity's open disposition for the current cursor, not a terminal closure; keep the eye while waiting. The next `review-latest-feedback` run must re-read every thread it previously asked in before scanning newer messages; when this workflow runs on its own, do the same and act on the replies first. After four days, keep `👀` and post **Abandoned - no answer in 4 days** once. Apply `review-latest-feedback`'s Reaction gate; carry active bugs forward. **In progress** is also an open state. It records that the thread already has real ownership or an active fix, so the invoking identity must not ask the reporter to repeat the issue. Re-read it on the next run, verify the work, and replace the open state with **Fixed** when complete or **Clarification needed** only if a specific reporter or product input is still missing. Treat a clarification reply as new evidence, not a new report. Re-read the thread and try the fix before asking anything else. An answer or explicit resolution from any participant is sufficient when it supplies the requested detail. A partial reply leaves the one existing request pending. Ask again only for one specific, non-repeating detail that still blocks a clear-bug fix; never ask for a subjective product choice or evidence already present. For an inaccessible artifact, request access or a replacement link. ### One pending question per thread There is no per-run question cap. After inspecting supplied evidence and attempting reproduction, ask for each specific reporter detail or action that would change the next step. Do not leave a useful question unasked because other threads also need questions. There may be only one unanswered clarification request per thread. Before posting, re-read the complete thread for an earlier question from this workflow, the companion `review-latest-feedback` workflow, or `@agent-native`, and check whether its exact requested detail has been semantically answered or explicitly resolved anywhere in the thread. If it is still unresolved, including after a partial or unrelated reply, keep that request as the sole pending handoff and do not post another question. Once it is answered or resolved, attempt the fix from the new evidence first; ask at most one new, non-repeating question only if one specific required detail still blocks it. ## Workflow 1. Build a per-thread checklist with the symptom, expected behavior, evidence, owner, and disposition: bug, UX suggestion, unclear, policy, or out of scope. Use the shared `address-feedback` categorization and Fix-altitude gate when choosing the disposition and owning seam. When the same underlying issue appears in multiple threads, build one cluster checklist and drive one Builder thread for that cluster; do not create separate Builder threads unless the reports diverge in symptom, surface, or owner. 2. The reaction is the first external action after classification. Add `👀` to each clear bug and every actionable Design/UX item already classified in scope, one thread at a time as it enters scope. Do not batch reactions until after investigation, implementation, testing, or the final Slack pass. If the reaction fails, stop and retry or report the concrete Slack permission/API blocker before continuing the investigation. Do not react to out-of-scope product requests, policy or informational messages, bot forwards, status-only items, or non-repo-owned reports. For an authorized upvoted improvement, perform and read back that same eye reaction before investigation or delegation, then include it in the ledger. 3. Parallelize independent investigations and narrow fixes with disjoint write sets. For every actionable repo-owned bug, keep working toward a verified fix. If reporter or product input is missing, ask one concrete question for it; do not settle for a vague unresolved status. If only internal test, deployment, or tooling verification is unavailable, keep that blocker internal and do not turn it into a reporter question. 4. Verify each fix with the smallest relevant test, typecheck, action read-back, or browser path. Keep source-tested, built, installed, deployed, and live observations separate. 5. Prepare one short status for each unclustered clear bug and owning parent marked `👀`. Record clustered source links and disposition in the owner thread or linked work; do not send duplicate replies: - **Fixed** - follow `review-latest-feedback` for the post-merge reply: name the fix, say it merged, and give the correct beta or package timing. Never claim it is live without runtime proof. - **Shipped** - use for an authorized upvoted improvement after its requested behavior and verification check are complete. - **In progress** - only when work continues beyond this run; thank the reporter and name the active work. A same-run fix needs only its final verified status. This is an open handoff, not a fix. - **Clustered** - list source permalinks and state in owner-thread status or linked work. Reply once; answer a distinct question in its own thread. - **Clarification needed** - ask one concrete plain-language question only when missing reporter input or an inaccessible needed artifact blocks a safe clear-bug fix. Re-read immediately before posting and confirm the detail is absent and no one already owns or resolved the issue. Never post a vague or bare unresolved status, repeat a question, or expose internal test/deployment gaps. Keep replies shorter than the investigation and omit implementation details, IDs, tools, and internal ownership. 6. When the user explicitly asks to reply, post directly in each requested thread with `thread_ts` through the same connected Slack identity. Do not silently turn an authorized write into a draft. Re-read each thread afterward to confirm the reply landed under the intended parent. Use the exact parent timestamp as `thread_ts`; never reply to a search-result timestamp or an adjacent thread. Before ending the run, mechanically audit the reply ledger: for each claimed parent, record its reply timestamp or owner-thread link, disposition, and eye state. A terminal state in the thread or linked work needs no duplicate reply. Record **Owned elsewhere** only when the latest update confirms another active owner; preserve foreign eyes. Record non-owning **Clustered** rows in the owner thread or linked work, and send no duplicate reply unless a distinct question or update needs an answer. Record out-of-scope items with one **Skipped** status if missing. If any participant replies after the post, re-read the entire thread again before deciding whether to fix, close, or ask anything else. 7. If any participant supplies the requested detail or an explicit resolution, re-read the full thread and use that new evidence in the same follow-up pass. Attempt the fix now; do not repeat the question. If the reply is partial or unrelated, keep the existing clarification pending instead of asking a second question. Replace an open clarification with a **Fixed** reply once the fix is verified. 8. If the user says earlier replies were too technical, harsh, or incomplete, search for every reply authored in this sweep and edit the bad replies in place. Do not fix only the newest example or leave the other addressed threads with the old wording. ## Slack reply voice Write in the invoking user's voice and send from that same connected Slack identity: - Use lowercase, short conversational paragraphs, and clear, conversational wording. Keep the tone casual and collaborative - warm without being corny, and specific without sounding terse or demanding. - Every feedback reply starts by thanking the reporter. Use the natural short form `ty for the feedback -` (or `thanks for the feedback -`) before the status. Do not open with `agreed`, `valid request`, `ah`, or a diagnosis. - Every reply from this workflow also ends with `this was sent from a bot.` so future sweeps can rediscover it; historical replies may not contain the marker and must still be found by the companion clarification search. - An **In progress** reply starts with that thank-you and names the active work. Do not use it to ask for clarification the thread already answered. - Use lowercase and a short conversational paragraph. Natural phrases such as `ah`, `yeah`, and `good find` can follow the thank-you when they fit; do not force them into every reply. Prefer ` - ` over em dashes. - Thank the reporter by name when it is available, for example, “thanks Alexander -”. Ask for help rather than issuing a demand: prefer “if you can share ...” or “a deck URL or request ID would help us dig into this” over “send ...” or “provide ...”. Avoid canned enthusiasm, scolding, and robotic labels such as “Clarification needed” in the reporter-facing prose. - For a clarification reply, the order is mandatory: thank the reporter first, then ask the one essential question. A resolution or ownership statement already present in the thread is not a reason to ask that question. - The audience is product/design/feedback reporters, not developers. Never post technical explanations such as shared paths, transports, sessions, repro levels, payloads, schemas, CORS, auth domains, action names, or implementation details. Translate the result to: fixed, or one essential missing detail that is required to fix and verify it. - A clear, valid, repo-owned request is an instruction to fix it. Do not reply `valid request` and stop, and do not say `no ship timing yet` as a dead end. Implement the fix first. After its PR merges, follow `review-latest-feedback` to report the fix and give the correct beta or package timing. Do not promise beta timing before merge. - Never claim unverified fixes, deployments, live behavior, or ownership. For merged Slack fixes, follow `review-latest-feedback`; a merge is not live proof. - If it is not fixed, continue the work or post a concrete **In progress** status when it continues beyond this run. Ask one concrete reporter question only when a missing reporter-only fact changes the next action; name the fact and why it gates progress. Ask only after exhausting the Slack thread, linked files/transcript/video, app state, run ID, sessions, and history. Internal verification blockers do not justify a reporter question. When a needed artifact is inaccessible, ask for access or a fresh/replacement link, not its contents as though the evidence were absent. Never ask for a prompt, run ID, session, or file already present or available through an accessible source. Never send `Open` or `pending`, a symptom recap, or an evidence/source gap in place of the needed question. If no reporter detail gates progress, keep `👀` and continue internal work; when a progress reply is due, name the concrete next action and owner. **Open - no question** is a ledger-only disposition and must never be sent as a Slack reply. Apply `review-latest-feedback`'s Reaction gate before `🎫`. - When a request ID would help, make the path easy and optional: “at the end of the chat, hit the three dots and share the request ID if that option is available.” Pair it with the useful surface link when one exists, such as a deck URL; do not ask for a prompt or run ID when the source fix is already established. - Before finishing the sweep, search every reply authored in that sweep for vague unresolved wording and edit or remove it. Re-read the affected threads after each edit. Unclaimed subjective/product/policy items get no reply. If an item was claimed before being skipped, preserve its eye and ensure one concise **Skipped** status reply. Follow `review-latest-feedback` for `✅`/`🎫`. A useful reply shape is: ```text ty for the feedback - [short plain-language status]. [if the merged app fix's beta publisher is queued or running: say the fix should be on beta within about 24 hours; if it succeeded, say it is published; if it failed or is missing, state the next action without an ETA.] [if the report is mixed: name each unfixed item and why; state the human action or owner still needed, link an existing ticket if available, and ask one targeted question only if its answer would unblock the work.] [if in progress: we're already looking into this and will follow up once the fix is verified.] [if clarification is needed: if you can share the one missing detail, that would help us investigate, such as a deck URL and/or request ID.] ``` Keep it to one short paragraph whenever possible. Omit the release sentence only when the change is not complete; do not invent a ship date for an open item. For an unclear runtime report, inspect its run ID and linked app evidence first; ask for one missing detail only when those sources cannot adequately identify the failure. ## Release follow-up After the final ship, return to the same threads and post a brief follow-up saying the fix is live only after verifying the live path internally. If the ship has not happened yet or live verification is unavailable, do not imply that the fix is already live. Omit commit, release, and live-path details from the posted follow-up unless the reporter asks for them. ## Verification - Run focused checks for every changed surface and `git diff --check`. - Re-read the Slack threads after posting and confirm each requested reply is present under the intended parent. - Include the eye-to-reply ledger in the recap so no marked thread silently disappears from the handoff. - Report repeat reports and any similar-feedback cluster handled as one Builder thread, along with the fixed items, flagged items, clarification questions, verification gaps, and the exact release state. ## Related Skills - `address-feedback` - classification and fix boundaries. - `concurrent-agents` - shared-checkout coordination. - `verifying-changes` - proof before claiming done. - `writing-agent-instructions` - guidance quality and scope.