Twitter Reply

Twitter Reply

Own ad-hoc reply work as an independent flat peer. Never activate or depend on twitter-automation; that peer may route here, but this skill owns preflight and the complete reply workflow.

Condense The Request First

Before preflight or any tw command, protect any exact target and final reply text. For every task-mode request, rewrite the remaining user wording as one concise, corrected sentence. Fix obvious typos from the full context and remove filler or repetition, while preserving every explicit count, topic, literal search phrase, account or relationship filter, engagement threshold, and other constraint. Do not add a topic or requirement the user did not state.

Use this condensed request as the task direction for task creation and any fresh-task restart. If the wording cannot be condensed without guessing the user's intent, ask one concise question and stop. Never rewrite an exact target/reply pair. For example, fin 10 let's connect tweets and draft replies means Find 10 tweets matching "let's connect" and draft replies, not a finance task.

Select One Mode

Select task mode when the user supplies a direction, topic, literal search anchor, account/surface criteria, target-discovery request, or a discovery count from 5 through 10.

Select exact-action mode only when the user supplies one exact target, one complete final reply, and explicit immediate-send intent.

Also select exact-action mode when the user explicitly retries one exact unchanged draft from the current task and its durable result conclusively proves that dispatch never started. A retry creates a fresh exact-action flow; it never reopens, reschedules, or executes the failed scheduled mutation.

• Default a missing discovery-task count to 5.
• Accept only an integer from 5 through 10 for target discovery.
• For a discovery count below 5 or above 10, ask the user to choose 5..10; never auto-chunk or reinterpret it as exact-action mode.
• Never infer an exact target from browser state, a screenshot, a profile URL, or “that tweet.”
• For several exact target/text pairs, ask whether the user wants one direction-based discovery task or separate exact actions.
• If exactness versus discovery/direction is unclear, ask one concise question before invoking tw.
• For an outcome or status question about earlier reply work, use Inspect Previous Work instead of either mode.

Task Template

When entering task mode — and whenever the user's wording is too vague to condense without guessing — offer one localized fill-in template instead of open-ended questions:

Reply task template:
Find <5–10> tweets containing "<keyword/phrase>" and draft replies
Optional: only/exclude specific accounts, minimum engagement, language
Example: Find 10 tweets containing "let's connect" and draft replies

回复任务模板：
找 <5–10> 条包含「关键词/短语」的推文并起草回复
可选：只看/排除某账号、最低互动量、语言
示例：找 10 条包含「let's connect」的推文并起草回复

Reuse the same template at a discovery shortfall such as no_valid_targets_found to suggest a rephrased direction.

Language

Respond in English or Simplified Chinese from explicit preference, latest message, conversation language, then English. Never translate or normalize an exact target/reply pair. Condense task directions as defined above without adding requirements.

Mandatory Preflight

After condensing the request and selecting the mode, activate threadwave-preflight by skill name at the start of each new reply task. Pass this skill name, selected mode, unchanged original request, and required capability families. It checks skill updates for this task, then uses regular CLI preflight, which reuses readiness until 12 hours of inactivity and runs the full check only when stale or invalidated. Do not rerun preflight for a review decision alone. Do not invoke tw without a current ready result.

Require CLI 1.0.35 plus task, draft, plan, scheduler, and action command families. Missing skill, CLI, or extension state routes to https://www.threadwave.xyz/cli/setup/agent.md while preserving the request.

CLI Result Authority

The skill owns UX: choose the documented command for the user's situation, preserve safe argument transport, present returned targets/content/choices, and collect approval. The CLI owns workflow execution, task/draft lineage, validation, recovery truth, and exact continuations.

• Require one complete parseable tw_cli_harness_v1 envelope. A process/session handle is transport state, not a result.
• For ok=true, treat data, refs, status fields, warnings, and next as authoritative. Never compare parent and child task refs, rebuild lineage, repeat CLI invariant checks, or reinterpret the accepted result as contract drift.
• Show every non-empty top-level warnings entry. When returned draft data contains generation_warnings, show every message beside the affected draft. These are CLI-sanitized customer-facing humanizer messages: preserve their meaning, never suppress them or replace them with a generic gate failure, and never add internal rule IDs, prompt text, or gate context.
• Execute returned read-only next commands to continue the workflow. Present review, restart, cancel, or X-mutation commands as choices and run only the exact command the user authorizes.
• For ok=false, report the returned error.code, error.message, and error.retryable; show the customer-facing message before any safe returned next choices, including when it describes a humanizer failure. Never reduce it to a generic blocked state or replace it with a model-invented diagnosis.
• Report a workflow failure stage only when the CLI returns failure_stage. Never infer a stage from source_status, draft_status, timing, or an error code; if the field is absent, report only the returned failure facts.
• If the command exits without a complete envelope, stop with task_dispatch_unconfirmed. This transport failure is the only host-side result check and never authorizes a duplicate task.

Approval Authority

Keep each approval inside its own gate:

Plan-free manual task creation uses the user's request as discovery and draft-generation authority. It creates no task-proposal or source-selection review. Deterministic source selection is not user approval and never authorizes an X mutation. No wording such as “approve all,” “continue,” or “retry” transfers authority between the remaining gates.

Task Mode

1. Create One Bounded Task

Pass direction as one safe argument; never concatenate user content into shell syntax. Run the semantic equivalent of:

On a packaged Windows install, use the structured downstream form retained from threadwave-preflight: add each CLI token and the complete condensed direction separately through ProcessStartInfo.ArgumentList. Never invoke tw.cmd with PowerShell & plus splatting, Start-Process -ArgumentList, ProcessStartInfo.Arguments, or a composed cmd.exe /c string. Stop with twitter_automation_cli_unconfirmed if the host cannot preserve one argument per value.

tw task create --surface reply --direction <condensed_request> --count <5..10> --json

Retain the complete command-execution result. If execution yields a session_id, poll that same session to terminal completion; keep state invoking until its complete CLI envelope arrives and never project only output while discarding the continuation handle.

2. Follow Automatic Discovery And Draft Generation

Follow the accepted envelope's exact next commands. When task creation or restart returns a repeated exact task-show continuation, consolidate that read-only continuation into one tw task wait <task_blueprint_ref> --timeout 900 --include-drafts --json command. It waits inside the CLI and returns the terminal task plus every exact draft, source, review ref, and review continuation together; do not model-poll task show or invoke separate draft-show commands for that batch. Never substitute a global list or latest record.

If task wait returns task_wait_timeout, show the current task state and its exact returned wait continuation as a user choice. Do not start another wait or create replacement work automatically.

When the CLI returns drafts and review choices, present the returned target/content pairs together and wait for per-item decisions. When it returns a shortfall, failure, stalled state, or fewer drafts than requested, render that state exactly and show only its returned choices. Never create replacement work automatically.

3. Review Drafts And Scheduled Mutations

• Use only commands from the accepted CLI envelope's next. Run read-only inspection continuations directly. Present approve, reject, skip, restart, cancel, and X-mutation commands as choices; after an explicit decision, invoke the matching exact command once and leave omitted reviews pending. If one decision fails, do not replay successful decisions; use the returned state to ask only for unresolved decisions.
• Capture every exact returned scheduled_task_ref. Scheduling is not completion: while any ref reports scheduled or running, keep the session active and poll that ref with tw scheduler show <scheduled_task_ref> --json at intervals no longer than 15 seconds. Do not invoke tw scheduler execute merely to accelerate the daemon-owned reservation.
• When a ref leaves scheduled or running, inspect it once with tw scheduler evidence <scheduled_task_ref> --json. Treat succeeded as sent only with its conclusive scheduler evidence. If a terminal reply still lacks a conclusive sent/not-sent result, run tw action snapshot verify --task <task_blueprint_ref> --json once for its exact logical task. Report item verified as sent, not_complete as not sent, and inconclusive as outcome unknown. Do not repeat the snapshot or reinterpret its result. A paused or waiting_review ref needs user action; show that blocker and keep every other ref under observation.
• After each polling round that changes any status, publish one compact update using the original numbered draft mapping and place every draft under exactly one heading: Sent, Not sent yet, or Needs attention. Use Sent only for succeeded with conclusive evidence and Not sent yet only for scheduled or running. Put blocked, failed, paused, recovery, and unknown outcomes under Needs attention; distinguish a conclusively undispatched result as not sent and an inconclusive result as outcome unknown. Never rewrite outcome unknown as nothing sent.
• If a polling round changes nothing, repeat the complete numbered split at least once every 60 seconds so the user can see that monitoring remains active. Continue until every returned ref is terminal, user-skipped, or waiting on a user decision, and never finish the reply flow merely because all drafts are scheduled.

When multiple content reviews are pending, present them together as one numbered set with each exact target, reply text, and review_ref. Require an explicit per-item decision map. Never collapse multiple replies into “approve all” or treat a bare decision as batch mutation authority. Each content approval authorizes one exact reply mutation only.

Never guess a target or ref.

4. Handle Changes And Failures

Every fresh task below runs the update check and regular preflight; a new task alone reuses fresh readiness instead of forcing the full check. Review, monitoring, and redraft reuse the current ready result unless readiness was invalidated.

• If the exact task reports execution_status=failed and restart_allowed=true, present its failure code and offer one explicit choice to restart or stop. On restart, invoke tw task restart <task_blueprint_ref> --json exactly once, verify ok=true, and resume the same automatic-discovery polling contract against the same logical task. Restart creates one fresh pre-mutation execution; it never creates a duplicate task or claims the old attempt was cancelled. After two failed restart rounds in the same session, stop and report the exact blocker.
• Whenever tw scheduler cancel returns logical_workflow_terminalization.status=pending, tell the user that linked workflow cleanup is pending, recheck the exact task within recheck_after_seconds (currently 60), and do not describe the immutable scheduler record itself as expiring.
• Once drafts exist and no reply mutation is scheduled, running, paused, waiting_review, or has unknown evidence, use tw draft redraft for a wording-only change to one unchanged source and require a new content review. For “start over,” “these candidates are wrong,” or a direction, source, angle, or target change, show every current content review and ask the user to confirm skipping the still-pending drafts. After confirmation, invoke tw plan review skip <review_ref> --json once per confirmed ref, verify data.review.status=skipped, then create one fresh task with the latest condensed direction and count. The historical task remains recorded; no authority transfers.
• If any reply mutation is scheduled, running, paused, waiting_review, or has unknown evidence, do not restart or create a replacement reply. Continue monitoring the exact scheduled ref until the user makes an explicit decision for it under the recovery decision below.
• For a recovery_required or otherwise inconclusive scheduled ref, present the exact scheduled_task_ref with its durable evidence classification and stop for one explicit user decision: keep monitoring, skip, cancel, or report. On explicit skip, invoke tw scheduler skip <scheduled_task_ref> --json once and require data.outcome_acknowledged=true; then end its monitoring and keep it listed under Needs attention as skipped. The durable outcome stays outcome unknown, the acknowledgment survives later sessions, and skip authorizes no retry, redraft, replacement, cancellation, or mutation. cancel only when the user explicitly names that exact ref: invoke tw scheduler cancel <scheduled_task_ref> --json once, verify the returned status, and follow the terminalization recheck rule above. A skipped ref no longer blocks completing the reply flow with the remaining refs.
• A rejected or skipped content review creates no mutation.
• Treat relay_unavailable with no dispatch ActionRecord, and action_preparation_failed with reply:blocked_before_dispatch, no dispatched_at, and confirmation_state=not_required, as conclusively not sent. An ActionRecord or EvidenceRecord count alone is audit evidence and does not prove dispatch.
• A retry accepts one exact unchanged draft only. Resolve its exact target/text pair from the current task's returned refs or from exact refs lineage-matched through Inspect Previous Work, run the update check and regular preflight for the fresh exact-action task without forcing a full check merely because the mode changed, and start one fresh exact-action dry-run. Show that pair and require fresh explicit approval before dispatching it once. For multiple retry requests, ask the user to choose one; each later pair requires a separate exact-action flow and approval. Never execute or reschedule the old scheduled-task ref or transfer its task/content approval.
• If the numbered draft mapping is stale, the target/text changed, or neither the current task nor a lineage-matched previous-work ref contains conclusive pre-dispatch evidence, ask for the exact current draft/ref or stop as unknown. Unknown, post-dispatch, unknown_confirmation, confirmed, or otherwise inconclusive outcomes never enter this retry flow; they require either a sanitized issue report that preserves outcome unknown or an explicit user skip decision under the recovery decision above.

Inspect Previous Work

Use this path when the user asks about the outcome of earlier reply work from a previous task or agent session. It is read-only evidence inspection plus honest reporting; it never reopens the old task's discovery, generation, or review flow.

1. Run mandatory preflight, then tw scheduler list --json. Match refs by the lineage fields shown in the list output (task_blueprint_ref, manual_request_ref, work_cycle_id) for the work the user means. Never guess a ref or infer one by timestamp. When lineage is ambiguous, show the candidate tasks and ask the user to identify the exact one.
2. Inspect each exact ref with tw scheduler show <scheduled_task_ref> --json and tw scheduler evidence <scheduled_task_ref> --json. For a terminal reply without a conclusive sent/not-sent result, run tw action snapshot verify --task <task_blueprint_ref> --json once for its exact logical task.
3. Report under the same truth rules as live monitoring: Sent for conclusive scheduler evidence or item verified, Not sent yet only for scheduled or running, and everything else under Needs attention; item not_complete is not sent and item inconclusive is outcome unknown.
4. A ref whose durable evidence is conclusively undispatched under the existing criteria may enter the retry flow; an inconclusive ref follows the keep-monitoring, skip, cancel, or report decision and never enters retry.

This path never re-approves, reschedules, or executes the old ref, and never claims an outcome the evidence does not prove.

Exact-Action Mode

Require exactly one target accepted by the CLI contract and one exact final reply. A profile URL is not a reply target.

1. Run tw action reply <exact_target> --text <exact_reply> --dry-run --json through safe argv/stdin handling.
2. Show the exact target and reply in full, the dry-run result, and the semantic operation; ask for explicit approval.
3. Treat any target or character change as a new payload requiring a new dry-run and approval.
4. After the user explicitly approves, reuse the current exact-action flow's preflight result and dispatch tw action reply <same_target> --text <same_reply> --json once.
5. Treat the accepted CLI outcome as authoritative. Report complete only when it returns complete; otherwise render its state, error, and safe choices. Never retry an unknown result or send a test reply.

Boundaries

• Task mode creates one materialized manual task with 5..10 possible reply drafts; it does not create task/source reviews or grant batch mutation approval.
• Exact-action mode controls one target/text pair only and remains outside task/plan lineage.
• Do not create original posts, quote, like, save, follow, or operate the browser UI directly.
• Do not expose reply text, targets, raw payloads, private refs, or local paths in diagnostics.

Issue Report

For explicit report requests, a returned CLI error, repeated unresolved setup, or unknown evidence, activate threadwave-preflight in issue-report-only mode with sanitized metadata. Exclude target, reply text, and task direction. Issue-report generation sends no reply; report each affected reply as sent, not sent, or outcome unknown exactly as its durable evidence supports.

Return Format

State: <discovering | generating | needs content approval | needs exact-action approval | monitoring | complete | blocked | unknown>
Mode: <task | exact-action>
Count: <requested / valid drafts / content-approved / scheduled, or n/a for exact-action>
Reviews: <current content reviews with exact review_ref; none before drafts or for exact-action>
CLI messages: <every returned warning or error message; none when empty>
Waiting for you: <matching per-item content decisions; approve this exact target/text pair; or none>
你可以 / You can: <two to four verbatim-sayable options valid at the current gate>
Next: <one returned ref/action or stop>
Outcome: <sent | not sent | not sent yet | outcome unknown, per reply>
Issue report: <copy/paste only; preserves the durable outcome>

The 你可以 / You can: line lists only options that are real at the current gate — exact numbered content decisions, the displayed exact-action pair, displayed refs, or the task template — worded so the user can reply verbatim, localized to the selected language; it never offers an action beyond the current gate's authority.